WPS6832


Policy Research Working Paper                       6832




          The Impact of an Adolescent Girls
               Employment Program
                   The EPAG Project in Liberia

                               Franck Adoho
                            Shubha Chakravarty
                            Dala T. Korkoyah, Jr.
                             Mattias Lundberg
                               Afia Tasneem




The World Bank
Africa Region
Poverty Reduction and Economic Management Unit
  &
Human Development Network
Social Protection and Labor Unit
April 2014
Policy Research Working Paper 6832


  Abstract
  This paper presents findings from the impact evaluation                           percent. In addition, the impact evaluation documents
  of the Economic Empowerment of Adolescent Girls and                               positive effects on a variety of empowerment measures,
  Young Women (EPAG) project in Liberia. The EPAG                                   including access to money, self-confidence, and anxiety
  project was launched by the Liberian Ministry of Gender                           about circumstances and the future. The evaluation
  and Development in 2009 with the goal of increasing                               finds no net impact on fertility or sexual behavior. At
  the employment and income of 2,500 young Liberian                                 the household level, there is evidence of improved food
  women by providing livelihood and life skills training                            security and shifting attitudes toward gender norms.
  and facilitating their transition to productive work. The                         These results reinforce the highly positive feedback
  analysis in this paper is based on data collected during                          received from focus group discussions with program
  two rounds of quantitative surveys in 2010 and 2011,                              participants. Finally, preliminary cost-benefit analysis
  the second of which was conducted six months after the                            indicates that the budgetary cost of the EPAG business
  classroom-based phase of the training program ended.                              development training for young women is equivalent to
  Strong impacts are found on the employment and                                    the value of three years of the increase in income among
  earnings outcomes of program participants, relative to a                          program beneficiaries. These preliminary results provide
  control group of non-participants. The EPAG program                               strong evidence for further investment and research into
  increased employment by 47 percent and earnings by 80                             young women’s livelihood programs in Liberia.



  This paper is joint product of the Poverty Reduction and Economic Management Unit, Africa Region and the Social
  Protection and Labor Unit, Human Development Network. . 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://econ.worldbank.org. The authors may be contacted at schakravarty@
  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
         The Impact of an Adolescent Girls Employment Program: The EPAG Project in Liberia
    Franck Adoho, Shubha Chakravarty*, Dala T. Korkoyah, Jr., Mattias Lundberg, and Afia Tasneem




JEL Codes: J13, J24, O12
Keywords: Adolescents, Gender, Vocational Training, Youth Employment, Sub-Saharan Africa



* Corresponding author: schakravarty@worldbank.org
Acknowledgements: This research was made possible through generous financial support of the World
Bank’s Multi-donor trust fund for adolescent girls, the Umbrella Facility for Gender Equality, the Bank-
Netherlands Partnership Program, and the Strategic Impact Evaluation Fund. The evaluation benefitted
from close collaboration with the Ministry of Gender and Development of Liberia and the Bank’s project
management team. Special thanks to Eve Lotter, Peter Darvas, Hiska Reyes, Rui Benfica, Waafas Ofosu-
Amaah, Andrew Morrison, Markus Goldstein, Niklas Buehren, Ali Ahmed, Sarah Marie Murray, Sarah
Nedolast, David Newhouse, Marta Favara, Sarah Nedolast, and Sarah Haddock. The data were collected
by a consortium of consultants including Subah-Belleh Associates and GOPA Consultants. All errors are
our own.
1. Introduction
Sub-Saharan Africa is the youngest region in the world, and the world’s fastest-growing. Between 2010
and 2025, the number of people between 15 and 24 will grow to 250 million – a net increase of nearly
50 percent. Over the next decade, roughly one million young people will enter the labor market each
month in Sub-Saharan Africa. However, these young people often enter the labor market too early and
unprepared. Although access to education is growing, illiteracy remains high and schooling low: among
the 32 Sub-Saharan African countries in the Barro-Lee (2010) data set, nearly 40 percent of women aged
15 and above have received no education at all; and the most recent (2007-2011) statistics in the World
Bank’s Edstats data reveal that female literacy is less than 60 percent, on average. 1

This lack of preparedness contributes to a growing problem of youth unemployment. Quantifying the
level of unemployment is bedeviled by lack of data and measurement issues. Household and labor force
surveys throughout Africa usually record unemployment rates of less than 10 percent, 2 but those figures
belie the extent of underemployment and vulnerability. Those same surveys indicate that the vast
majority of working adults have insecure work in the informal sector, on the family farm, or in less-
productive or unremunerated labor. The most recent labor force survey from Liberia mirrors these
statistics: there are roughly 1.1 million people in the workforce, of whom 195,000 (about 18 percent)
are engaged in wage employment; the remaining 900,000-plus workers (82 percent) are considered in
vulnerable employment, working for themselves or working unpaid for their own households (LISGIS
2010).

Among young women (15-24) in Liberia, the unemployment rate is 8 percent, double the rate among
young men (LISGIS 2010). Most of these gaps can be explained by differences across individuals,
especially in educational attainment, skills training, and years of experience. But segregation, market
segmentation, and discrimination do play a role in determining these individual characteristics. Women
have fewer opportunities for education or training, less access to credit, a larger share of domestic
responsibilities, and less independence and control over their own lives. In Liberia, women comprise
half of the employed, but only about one-quarter of paid employment (LISGIS 2011).

Fourteen years of civil war in Liberia devastated the country’s infrastructure and institutions, and left a
generation of young people with very low levels of education and training. Girls were particularly
disadvantaged. In 2003, almost 60% of young girls and 40% of young boys had no formal schooling
(ILO/UNICEF 2005). Data from the Demographic and Health survey shows that more than 40 percent of
adult women have no education, compared to fewer than 20 percent of men, while 23 percent of
women and 44 percent of men have some secondary schooling (DHS 2007). Happily, access to
education is rising rapidly, especially for girls: according to the Liberian labor force survey from 2010, the
ratio of girls to boys enrolled in primary school has risen from 72 in 2000 to 90 in 2009. Enrollment
levels and sex ratios are lower among older children and youth, as they become increasingly engaged in


1
  Both are unweighted averages; Barro-Lee comprises 32 countries with data from 2010; Edstats comprises 43
countries with data from 2007-2011.
2
  Defined as without work, available for work, and actively looking for work (LISGIS 2010).


                                                                                                                 2
household and productive work. Just over one-third of young (15-24-year-old) Liberians are in the labor
force. Young women are more likely than young men to be out of the labor force because they are
engaged in household duties (30.4 percent v. 18.9 percent), and young men are more likely to be out of
the labor force because they are still in school (75.8 percent v. 63.0 percent) (LISGIS 2010).

Given their initial disadvantage relative to their male peers, and the sources of this disadvantage in
social norms, market failures, and poorly functioning institutions, adolescent girls may require targeted
policy and program efforts to achieve better outcomes. However, as in many other post-conflict
situations, emergency skills training and public works programs in Liberia have targeted male youth ex-
combatants, likely reinforcing rather than reducing adolescent girls’ disadvantage. The few skills
training programs for adolescent girls, run largely by NGOs, have focused on traditional female skills
(such as sewing, soap production, tie-dyeing) for which the market is already well-supplied.

Programs designed to enhance employment among young people can focus on the supply side of the
labor market, including skills training programs for both wage employment and entrepreneurship
training; or programs to augment demand, such as wage subsidies, public works, and community service
programs; or programs to help the labor market clear, such as job search assistance and placement
services. In addition, it may be that the constraints facing young people are not in the labor market
itself, but in other markets, such as for credit. The vast majority of jobs programs have focused on the
supply side: training programs make up about 79 percent of over 600 cases included in the World Bank’s
Youth Employment Inventory database. 3

Although rigorous and general evidence of success is limited, it seems that successful skills training
programs share a few key features: they are responsive to local market conditions, they provide more
than just technical skills in a specific area (including, for example, “life skills”), and they include ancillary
services that alleviate other constraints preventing successful labor market integration (e.g., access to
credit). Among the most celebrated are the Jovenes programs that provide demand-driven technical
training, plus social skills that help in the labor market, plus internships. These programs have had
positive impacts on wages and employment, especially for young women and those from more
disadvantaged backgrounds (Ibarrarán and Rosas 2009).

This paper evaluates the first round of the EPAG (“Economic Empowerment of Adolescent Girls and
Young Women”) skills training program implemented by the Government of Liberia from 2010 to 2011.
EPAG was designed to alleviate the barriers to entering the labor market faced by young women, while
avoiding the shortfalls of previous skills training programs offered in Liberia. The program combined six
months of classroom-based technical and life skills training, with a focus on skills with high market
demand, followed by six months of follow-up support to enter wage employment or start a business.
Roughly 1200 young women aged 16-27 participated in the first round of EPAG.
3
  See http://www.youth-employment-inventory.org/. Also, although focused on a narrower time frame and using
a different selection criterion, the joint ILO/World Bank Inventory of Policy responses to the Financial Crisis
(http://www.ilo.org/crisis-inventory) finds that “about 78 percent of reported policy measures focused on the
supply side”, essentially through training (ILO/WB 2012).


                                                                                                                    3
The outcomes of interest for this impact evaluation fall into three categories. First, we are interested in
the economic outcomes of EPAG participants in terms of employment, earnings, and savings and
investment behaviors. Second, because of the program’s focus on young women, a variety of non-
economic outcomes are explored. Early evidence suggests that empowerment in one realm (e.g.,
schooling) can have important impacts on other realms, such as early pregnancy, prevalence of STDs,
and risky behaviors such as transactional sex (Baird 2010). Related to these non-economic outcomes are
measures of social empowerment, including mobility, decision-making, and self-confidence, which are
thought to strengthen women’s agency, or capacity to exert choice over decisions involving herself, her
family, and her community. Finally, we investigate a third set of outcomes on spillover effects. Evidence
suggests that women tend to invest more of their income in their families, especially their children, than
do men (e.g. World Bank 2001, Hoddinott 1995, Pitt 1998, Borges 2007). Advocates frequently argue for
increased investment in adolescent girls by pointing to the potential for spillovers onto other family
members (including future children). Our evaluation explicitly investigated these spillovers by including
household-level indicators such as food security and attitudes of the household head toward gender
norms.

Our results show strong impacts on economic outcomes, including large and statistically significant
increases in employment and earnings of EPAG participants. We show mixed results on empowerment-
related outcomes, and very little evidence of spillovers on non-participants. Self-assessed measures of
self-confidence show huge gains, as does ownership and control over monetary resources such as
savings.

The remainder of the paper is organized as follows: Section 2 describes the EPAG project including some
of its innovative design features and implementation details. Section 3 reviews the methodology of the
evaluation and Section 4 presents results on the three groups of outcomes discussed above: economic,
empowerment, and spillovers. Section 5 includes a short discussion of cost-effectiveness. Section 6
presents a series of robustness checks and Section 7 concludes with a discussion of next steps and policy
implications.



2. The EPAG Project
The EPAG project is part of a larger Adolescent Girls Initiative (AGI) administered by the World Bank with
support from the Nike Foundation and the Governments of Australia, the United Kingdom, Norway,
Denmark, and Sweden. Launched in Washington DC in October 2008, the AGI was spearheaded by
President Ellen Johnson Sirleaf, who signed on to undertake the initiative’s first pilot project in Liberia.
The Liberian pilot was launched in March 2010 and has served as a role model to seven subsequent pilot
projects in Rwanda, South Sudan, Nepal, Afghanistan, Haiti, Jordan, and Lao PDR.

Under the global AGI, young women and adolescent girls are given a package of skills training and
complementary services in order to facilitate their successful transition to employment. In the case of
EPAG, the intervention consisted of a six month phase of classroom-based training, followed by a six


                                                                                                               4
month placement and support phase in which the trainees were supported in their transition to self or
wage employment. Upon recruitment, the participants are assigned to a "Job Skills (JS)" track or a
"Business Development Services (BDS)" track. When possible, the participant's track preference was
honored; however, the demand for the Job Skills track greatly exceeded the supply, so the remaining
trainees were placed into the BDS track. In the first round of training, the proportion of Job Skills track
places was limited to 35% of the total training places available given the expectation that few wage jobs
will be available in the Liberian job market. The Job Skills track provided training in six areas: 1)
hospitality, 2) professional cleaning / waste management, 3) office / computer skills, 4) professional
house / office painting, 5) security guard services, and 6) professional driving. These areas were
determined based on independent labor market assessments, a review of the available market data, and
input from EPAG’s private sector partners. All Job Skills trainees received training in entrepreneurship
skills as well. The BDS training taught young women how to identify micro-enterprise opportunities
based on an assessment of market needs, and how to grow and manage any existing businesses they
already had. The curriculum included entrepreneurship principles, market analysis, business
management, customer service, money management, and record-keeping.

The EPAG program was implemented by four NGOs who were selected by the Liberian Ministry of
Gender and Development through a competitive bidding process: the Community Empowerment
Program (CEP), Liberia Entrepreneurial and Economic Development (LEED), the International Rescue
Committee (IRC), and the American Refugee Committee (ARC). Two of these organizations (ARC and IRC)
further subcontracted to four Liberian NGOs. 4 The service providers were responsible for developing
training curricula, identifying training venues, 5 making arrangements for childcare services, assisting
with the mobilization of the nine target communities, and participating in the recruitment of training
participants.

The EPAG program differed from many training programs in a number of ways. First, performance
bonuses were awarded to training providers that successfully place their graduates in jobs or micro-
enterprises. The bonus was the last payment that the service providers received under their contracts.
These were paid about 12 months after the start of training, or around the same time as the midline
survey. Second, a variety of contests and competitions were also held among EPAG trainees (such as
attendance prizes, quizzing contests, business plan competitions, etc.). Third, the EPAG program was
designed around the girls' needs: service providers held both morning and afternoon sessions, to
accommodate the participants' busy schedules; trainings were held in the communities where the girls
reside; and every site offered free childcare. Fourth, frequent and unannounced monitoring visits by
MoGD staff ensured that the service providers created and maintained a high-quality learning

4
  There are: National Adult Education Association of Liberia (NAEAL), Community Empowerment Sustainable
Program (CESP), EduCare, and Children’s Assistance Program (CAP).
5
  A total of 19 training venues were used during the first round of training. They were chosen with the following
considerations in mind: 1. Girls’ safety, so that the buildings are not so isolated or otherwise dangerous, raising
security concerns for girls. 2. Conducive atmosphere for learning, spacious and sanitary with access to water and
latrine facilities. Reasonably outside community noise concentration. 3. Proximity to community center and to
security posts such as police depots. 4. Accessible to girls from various parts of the community.


                                                                                                                      5
environment. Any issues discovered during these monitoring visits were brought to the attention of
service providers and resolved swiftly in conjunction with the project coordination team at MoGD.

Eligibility: The EPAG program was targeted to young women who: i) were age 16 to 27, ii) possessed
basic literacy and numeracy skills, iii) were not enrolled in school within several months prior of the
program initiation, and iv) resided in one of nine target communities in and around Monrovia. 6 These
eligibility criteria stemmed from the project's objectives to reach young women at an early enough age
to significantly improve the trajectory of their working years, to focus on girls who already had the basic
literacy and numeracy skills needed to succeed in the labor market, and to avoid incentivizing applicants
to drop out of school. The literacy requirement in particular, although basic, made the EPAG program
out of reach for many of the most vulnerable young Liberian women; the requirement reflected a
deliberate choice on the part of program designers in the face of a tradeoff between serving the most
vulnerable and serving those who could most readily make use of this relatively short training program. 7

The team recognized that it was difficult, if not impossible, to require documentation from the
applicants to verify each of the eligibility criteria (especially age, since many Liberians do not have any
official form of identification). Hence the application process relied primarily on self-reported data. To
counter the likelihood that applicants would give false information in order to gain entry into the
program, the eligibility criteria were not made public; the mobilization and outreach campaigns did not
specify the precise age or education requirements for the program. During the recruitment events, each
applicant had to physically present herself, fill out an application form specifying her age, education
history, and residence. A simple literacy and numeracy assessment was also administered at the time of
application. Beyond these basic eligibility criteria, no further selection criteria were applied, and
program managers did not choose whom to train. A “randomized pipeline” design was adopted in which
eligible candidates were assigned by lottery to one of two rounds of the program, where those assigned
to the second round would serve as a control group for the evaluation. Every applicant who met the
eligibility criteria had an equal chance of participating based on the lottery results.

Retention and Attendance: After dividing EPAG trainees into two rounds, or cohorts, for the training, the
first round of training was held from March 2010 to February 2011. The classroom training was held
from March through August, during which the project achieved a 95% retention rate (far higher than
similar programs in Liberia and elsewhere), and an average attendance rate of nearly 90% during the
classroom training phase. EPAG trainees were given incentives to participate and to make the most of
their training: they signed "Trainee Commitment Forms" at the start of the training, they were paid
small stipends and a completion bonus contingent upon attendance, they were offered free childcare at
every training site, they were assisted to open a savings account at a local bank in which to save their



6
  Bassa Community, Battery Factory, Bentol, Doe Community, New Kru Town, Old Road, Red Light, and West Point
in Montserrado County and Kakata in Margibi County.
7
  To alleviate the burden of this requirement, the second round of the EPAG program included a preliminary basic
literacy program to help otherwise eligible young women to improve their literacy and numeracy skills in advance
of entering the program.


                                                                                                                   6
stipend money, and they were formed into small groups or "EPAG teams", each with a coach or mentor,
to foster support networks and boost attendance.


3. Methodology
     3.1. Research design
The impact evaluation of the EPAG project uses a randomized controlled trial, in which eligible
applicants to the program were randomly assigned to participate in one of two cohorts (or “rounds”) of
training. The treatment group is defined as those who were offered a space in the first round of training
and the control group comprises those assigned to the second round. Selection into the training rounds
was performed on a computer (using Excel) and was stratified by the track choice of the applicant (job
skills versus business development skills), community, and service provider. Data were collected using
three quantitative household surveys (baseline, midline, and endline) and two sets of qualitative focus
group discussions (one after each round of training). A timeline of the impact evaluation is depicted in
Figure 1. During both the baseline and midline surveys, the head of the household in which the EPAG
participant was residing was also interviewed, in order to examine potential spillover effects of the
program on non-treated household members. The household head interviews also provided an
opportunity to gather useful household-level information (including assets, housing characteristics, and
household head characteristics) to be used as control variables in our analysis. The baseline survey was
conducted before participants were informed of the outcome of the randomization, and the midline
survey was conducted one year after the baseline survey, one month after completion of the first round
of training and before the start of the second round.

The results in this paper are based on a panel data set that includes data from the baseline and midline
surveys as well as administrative and monitoring data from the program implementation team. Note
that although the midline data collection occurred just one month after the completion of the
intervention, the intervention itself includes 6 months of post-training follow-up; hence the midline
survey captures outcomes of the first cohort 7 months after they exited the classroom-based training.
This timing was necessary to prevent too long of a delay between cohorts; the second cohort of trainees
started immediately after the midline survey was completed. Future analysis using endline data will
examine the trends in outcomes of both groups after Round 1 completion, but will not involve
comparison against a pure control group of non-participants.

    3.2. Sample and attrition
Following the communications and outreach campaigns in each of the nine target communities, during
which 2,106 young women were originally recruited to be a part of the EPAG program, 8 a randomized




8
 The original recruitment effort fell short of the target of enrolling 2500 participants. Hence, before the second
round of training, another recruitment effort was launched and an additional 617 girls were enrolled. These girls
are not included in the impact evaluation since they were not subject to the initial random assignment.


                                                                                                                     7
assignment process was conducted to assign the trainees to the first or second round of training. 9 Of
those entered into the random selection, 1273 young women were assigned to the first round of
training, with the remaining 808 to serve as a control group (the control group would participate in the
second round of training starting in July 2011). Of the 1273 assigned to treatment, 118 women were not
found or chose not to participate after they were selected. 10 In order to fill at least some of these slots,
39 young women from the control group were randomly issued as replacements, resulting in a modified
control group of 769 individuals. In the end, 1191 young women entered the first round of training. 11
The assignment process and all post-randomization modifications are summarized in Figure 2.

Table 1 reports the baseline and midline survey response rates leading to the sample used for the
analysis in this paper. The target sample for both the baseline and midline survey consisted of the
original 2106 EPAG recruits, of which 1989 were successfully interviewed during the baseline survey. 12
At midline, 1736 were interviewed, including 56 who were not interviewed at baseline. For our analysis,
we drop individuals who were excluded from the randomization or who were manually re-assigned from
control to treatment as replacements after the randomization. We also drop individuals for whom we
lack household information at baseline, 13 resulting in a final panel of 1601 individuals. In section 6 on
“Robustness Checks” we investigate the potential biases introduced by sample attrition.

    3.3. Identification
Our identification strategy relies on a straightforward difference-in-difference OLS model, owing to the
randomized nature of selection into the program. The analysis will focus on estimating the intention-to-
treat (ITT) estimate, which considers all participants initially assigned to the first round of training as
part of the “treatment” group, regardless of whether they actually joined or completed the classroom
training phase of the program. In the results section we will show that these ITT estimates are identical
to the average outcomes among those who actually completed the program (the “ATT” estimates). 14
The main regression model is specified in equation (1) below.

         Yit= β1(Postt x Treati) + β2Postt + β3 Treati + β4 Xi0 + εit                                       (1)


9
  Of the 2106, twenty-five were pregnant and late into their pregnancies, so they were assigned to Round 2 directly
and not entered into the lottery at all. These 25 were included in the surveys but have been dropped from the
analysis because they were not assigned randomly.
10
   The reasons given for not entering the training included: 1) they were back in school, 2) they had moved to a
distant location, 3) they were seriously ill, 4) they had found full-time work, 5) they were not interested or able to
make such a big time commitment, or 6) they could not be located despite numerous efforts.
11
   This includes 1155 of the original 1273 assigned to treatment, plus 36 out of the 39 "replacements"- young
women from the control group who were offered a chance to be reassigned to round 1.
12
   Note that a previously released baseline report for this evaluation was based on 2008 observations. However,
after cleaning, 4 were found to be duplicate observations and 15 were not found in the program data and hence
were dropped from the midline analysis. This leaves 1989 baseline observations that are included in the midline
analysis.
13
   These are cases in which the adolescent girl was interviewed but the household head interview was not
conducted because the head was not available, could not be found, or declined to take part.
14
   This suggests that the loss of the 118 young women who were selected but declined to take part, and the 60
who started but did not complete the program, does not bias the results.


                                                                                                                         8
Yit is the outcome of interest for individual i and time t. Treati is an indicator which is equal to 1 for
treated individuals and 0 otherwise. Postt is an indicator equal to 1 for midline observations and 0 for
baseline. Xit is a vector of controls at baseline (t=0), including individual characteristics (education, age,
current pregnancy, marital, parental, and orphan status) and household characteristics (sex of
household head and household size). β1 is the coefficient of interest that defines the “impact” of the
program on individuals in the treatment group. The model also includes dummy variables for the
communities where the program was implemented (and where the trainees resided) as well as the
program track (business skills or job skills) to which the respondent was assigned. Finally, in order to
control for household wealth, we compute an index based on household asset ownership at baseline
using multiple component analysis, similar to the method described in Filmer and Pritchett (2001). After
constructing the index, which includes thirteen household assets and six indicators of housing
conditions, we control for the quintile of household’s overall asset position in all regressions. We
augment this basic specification with an individual fixed effects model and find that the results are
almost identical. For binary dependent variables, we estimate both linear probability models and probit
models analogous to the one given in equation (1), again finding that the results are nearly identical
across all outcomes. We cluster the standard errors by classroom in all models, and to account for the
interaction effect for variables such as (PosttxTreati) we additionally follow Ai and Norton (2003) to
correct the standard error for interaction terms in probit models.

     3.4. Baseline characteristics
Table 2 presents baseline balance tests for survey respondents from the treatment and control groups. 15
In addition to confirming the success of the randomization, as judged by the very few significant
differences between the two groups, the table provides a vivid profile of the average EPAG participant.
The study population has an average age of 23 years, with 55% falling between 20 and 24 years. The
majority have never been married, while 29% are cohabiting with a partner and only 5% are married.
The majority of the study population has started or completed high school, which is consistent with the
program’s target group of young women with basic literacy and numeracy, and with the program’s goal
not to encourage girls to drop out of school. Thirty-eight percent of the sample was already engaged in
at least one income-generating activity (IGA) at baseline. This is consistent with the national figures from
the 2007 DHS survey, which found labor force participation rates of 34% for women aged 15-19 and 49%
for women ages 20-24. It is also consistent with the Liberian 2010 Labor Force Survey, which found labor
force participation rates of 25% for women aged 15-19 and 47% for women aged 20-24.For the
purposes of this study, to be consistent with program objectives and the Liberian context, our definition
of income-generating activity encompasses the full range of activities through which people earn
money, including paid employment, either formal or informal, and self-employment in small business or
through petty trade. The most common types of IGAs reported at baseline were petty trade, including

15
   The balance tests are run on the same sample as will be used in the impact analysis in Section 4, that is, the
subset of individuals for whom we have a panel. Balance tests run on the full sample of baseline survey
respondents, regardless of whether they also participated in the midline survey, confirm the same findings. A
report summarizing the balance tests on the full sample, including comparisons to nationally representative data,
is available upon request from the authors.


                                                                                                                    9
street vending (48% of those with at least 1 IGA), food processing for sale, including baking, cooking, and
drying (16%), and home production of crops, livestock, and fish (11%).

It is important to note that the EPAG program was not targeted toward the most vulnerable segments of
Liberian society, but rather toward young women with enough education to be able to benefit from a
training program of this nature. Based on comparisons from the 2007 CWIQ (Core Welfare Indicators
Questionnaire) survey, the average EPAG participant is more educated, more literate, more likely to be
engaged in an income-generating activity, and owns more assets than the average Liberian women of
the same age group: mobile phone ownership was high (63%), as was the proportion reporting that they
had some money of their own (79%). Even compared to other similar residents of Monrovia, the EPAG
participants are better educated and have higher income.

A strong sense of female empowerment at baseline emerges from the sections of the survey instrument
having to do with self-confidence and agency. Survey respondents reported a high degree of freedom of
mobility, a high degree of control over money (both earned money and the money needed to pay for
items like food and medicine), and a high level of agreement with statements such as “If I had the
chance, I would like to become a leader in the community, as I would be a good one”. With such high
rates of self-confidence at baseline, there was limited scope for improvement on these measures. Their
self-confidence was a bit more tempered on issues related to running a business or searching for a job: a
composite score based on self-assessed ability to perform a series of tasks 16 yielded a mean of 2.4 out
of 4.0 (with standard deviation 0.86) at baseline, leaving room for improvement. Because of concerns
about the limited usefulness of the empowerment questions asked at baseline, some questions were
dropped and new measures were added for the midline survey only.

Panel 6 of Table 2A summarizes the family background characteristics of the study population, who tend
to come from large families and frequently have lost one or both parents. A large majority (78%) were
displaced during childhood or adolescence due to the long-running Liberian civil conflict (1989-1996;
1999-2003). Furthermore, about two-thirds of study respondents were already mothers before entering
the program, with small but significant differences between the treatment and control groups. Similarly,
7.5% of study respondents were pregnant at baseline, with a slightly higher rate in the control group.
Because of these significant differences in parental and pregnancy status, these variables are included as
controls in the subsequent analysis presented in Section 4. Panel 7 also investigates sexual behaviors of
study participants. Almost all have been sexually active, with a slightly higher age at sexual debut for the
control group. Sexual behaviors are very similar across treatment and control groups, with most girls
reporting one regular sexual partner, and one-third of respondents reporting ever having a casual sexual
partner. Seventy percent have used a condom, with about 55% reporting having used one in their last

16
  The specific questions asked the respondent to rate how well she would be able to perform the following
activities: find information about job opportunities in your area, run her own business, save in order to invest in
future business opportunities, manage business finances effectively, bargain with a supplier to obtain good prices
when purchasing, collect money from someone who owed her for purchases who are not repaying on time.
Numerical values were assigned to each response. The score is based on the average across all the activities on
which the respondent reported, conditional on having responded to at least half the activities.


                                                                                                                      10
sexual encounter. Finally, about 10% of study respondents have experienced a forced sexual encounter
in their lifetimes, which is consistent with the national figure of 13% for this age group (DHS 2007). 17

Table 2B presents baseline balance tests for the household characteristics of respondents in the study
sample. More than 40% of households are headed by females, and the average household has slightly
fewer than five members. The mothers of EPAG respondents tended to have very low education levels
(almost 60% had never been to formal school), while the fathers had more variance in their education
(about a quarter had no schooling, but over 60% had at least some secondary education). In both
treatment and control households, a high proportion of school-aged children are in fact enrolled in
school. Less than half of young people aged 13-30 in study households have any employment. Housing
conditions are also similar across experimental groups. Tables 2A and 2B include a representative, but
not exhaustive, list of indicators that were tested for balance by the authors. We conclude that the
presence of very few significant differences between the individual or household characteristics
between the two experimental groups indicates a high degree of internal validity for the study.

     3.5. Limitations of the approach
As is always the case for these studies, there are limits to the generalizability (external validity) of these
results. First, the roll-out design may not ensure a perfect comparison of “treatment” with “no
treatment.” Both groups may alter their behavior in response to the presence of the program as well as
to their selection into the various experimental groups. The impact of this behavioral change is most
likely insignificant. In the case of the roll-out design, the members of the control group know that they
will receive the intervention in twelve months or so. This may lead them to exert less effort to find
alternative employment outside the program. This will increase the apparent results of the program, if
the control group outcomes are depressed in the short run by the expectation of treatment. 18 Ideally,
one would have access to several months of data prior to selection into the program and assignment to
treatment, in order to see the impact of this expectations-related behavior change. Such data were not
available for this evaluation, so it is not possible to ensure that the control group did not modify its
behavior in anticipation of the program. However, the program itself was designed to fit within the
young women’s lives, specifically with regard to their employment, education, and childcare duties.
Training schedules were flexible, to allow participants to continue with their pre-existing educational
and income-generating activities (and many participants did report continuing with these activities) and
free childcare was provided. Hence, at least for these three dimensions of life, there would not have
been much incentive to change one’s behavior prior to starting the program. While these explanations
do not erase concerns about anticipatory behavior, they at least mitigate them.

The generalizability of these results is also limited by the differences between the EPAG target group
and the population of young women in Liberia. First, a high proportion of adolescent girls and young
women in Liberia are illiterate or have very low literacy, while the participants recruited for the EPAG


17
   Gender-based violence questions were administered in line with international ethical protocols, with additional
informed consent procedures and referral mechanisms as needed.
18
   See Ashenfelter (1978).


                                                                                                                     11
project had to meet a basic minimum literacy level in order to qualify for admission to the program;
whereas only half of the 15-29-year-old female respondents to the nationally-representative Core
Welfare Indicators Questionnaire (CWIQ) survey report that they can read and write (LISGIS 2007).
Fewer than two percent of the girls in the EPAG study responded that they had no education, which is
much lower than in the CWIQ survey. Second, the majority of adolescent girls and young women in
Liberia reside in rural areas, whereas the survey participants were residing in urban and peri-urban
areas, where access to basic social services may be much more improved. Consequently, the results are
not representative of adolescent girls and young women in Liberia overall. The results are neither
indicative of the average Liberian girl and young woman; nor are they indicative of the average Liberian
girl or young woman in the project communities. They are only indicative of the average girl and young
woman who are part of the EPAG project.

Finally, many of the variables that we examine in this study are measures of self-assessed levels of
satisfaction or belief. These are entirely subjective variables, and are subject to significant
measurement error. There is considerable evidence that the wording of these questions can affect the
answers given, as can the order in which the questions are asked. Social desirability bias also affects the
answers: respondents can interpret from the wording of the questions or the attitudes of the
interviewers what sort of answers are more or less acceptable or expected. Behavior that is private or
about which there are social strictures is more likely to be misreported. In addition, we are all bad at
understanding our own motivations or beliefs, and we are all tempted to report beliefs that justify or at
least are consistent with our behavior: if we do not go to the cinema, we can say that it is because we
did not want to go, and we discount the reasons or constraints that might have influenced our decision
not to go (Bertrand and Mullainathan 2001). That said, these questions do contain real information
about attitudes, beliefs, and behavior, and there is no reason a priori to expect that the errors differ
systematically across treatment and control groups. In fact, ex-ante tests of balance show that these
two groups are identical in these observed characteristics.



4. Results
This section provides an overview of the impact of participating in the EPAG program on a variety of
outcomes. As described above, "impact" is defined as the change in outcomes observed for the
treatment group between the baseline and midline surveys, relative to a control group of respondents
who did not participate in the first round of training. We present the results for the primary outcomes
of interest – income and employment, and on a range of other outcomes, including savings behavior,
various measures of empowerment (including self-efficacy and self-confidence), fertility and sexual
behaviors, and, finally, household level outcomes.

    4.1. Economic Activities
Table 3 presents the main results of the EPAG program using two sets of estimates. The first are the
“Intention to Treat” estimates: these compare the outcomes of everyone who was originally assigned to
receive the treatment with those that were assigned to the control group, regardless of whether the
participants actually received the training. The second set presents instrumental variables results, in


                                                                                                              12
which receipt of training is instrumented by the initial assignment to treatment. All results are
estimated as OLS (or IV); nonlinear estimations (probits) for binary variables are consistent with the OLS
results and are available on request.

Recall that at baseline, nearly four in ten (38 percent) young women in the study population – treatment
and control – have at least one income-generating activity. At mid-line, the number of participants in
the treatment group engaged in activities to earn income increased by 18 percentage points, relative to
the control group. Economic activity also increased among members of the control group between the
baseline and midline surveys, but the increase was much greater among the treatment group. The
estimate of 18 percentage points translates into a 47% increase in employment, from a baseline
employment rate of 38%. This is a significantly greater improvement than has been seen in many other
similar programs around the world. Other interventions have occasionally found large impacts (e.g. the
Chile and Peru Jovenes programs), but most programs have yielded relatively modest impacts.
Moreover, as one review noted, the successful Jovenes programs rely heavily on a positive
macroeconomic context – particularly in terms of job creation (Ibarrarán and Rosas 2009). In the
absence of robust private employment growth, job skills development programs have generally
performed poorly.

Table 3a disaggregates the outcome into those EPAG participants who received job skills training and
those who received training and support in enterprise development. This table shows first that the
training worked, in the sense that those who received job skills (JS) training were more likely to find
wage employment than those in the control group, and more likely to find wage employment than to
open their own business. Similarly, those who received business development (BDS) training were more
likely to establish their own businesses than those in the control group, and also more likely to open
their own business than to find wage employment. This table also shows that own-account employment
is more common than wage employment, and that the EPAG program was more successful in enhancing
self-employment than wage employment. At baseline, about five percent of the EPAG JS participants
were engaged in wage labor; at midline, this increased by 15 percentage points among the JS
beneficiaries. At baseline, about 37 percent of the EPAG BDS participants were engaged in own
businesses; at midline, this increased by 20 percentage points among the BDS participants. On average,
the probability that participants in the BDS track found employment increased roughly twice as much as
among those in the JS track. This does not necessarily say anything about the relative quality of the
training offered – it most likely reflects the structure of the labor market in urban Liberia. The vast
majority of employment is in the informal sector, and in own-account activities. According to the 2010
Labor Force Survey, 68 percent of all employed Liberians, and 75 percent of employed women, work in
the informal sector (LISGIS 2011), and nearly 70 percent of employed women are self-employed. Even
among participants in the JS track, those with some income-generating activity at midline were equally
likely to be self-employed as employed in wage labor. The JS training, although primarily focused on
wage employment, did include a short module on self-employment basics in order to equip participants
for the possibility that many would not find wage employment.




                                                                                                             13
Tables 3b and 3c disaggregate the results by location and by individual characteristics. Of the nine
communities in which the EPAG program was delivered, the training was effective in six. Only in the
neighborhoods of Battery Factory and New Kru Town were the impacts not statistically significant. The
estimate of the impact of training in the West Point community is negative, suggesting that participants
are ten percent less likely to be employed than non-participants. This was an exceptionally difficult
community to work in, and had unusually high attrition, especially among the control group.19 Table 3c
presents the impact of the EPAG program across education and wealth class. Although significant and
positive impacts are seen across the entire wealth distribution, the program was significantly more
effective among participants in the middle wealth classes than among the very poorest and very
wealthiest. Among the poorest and wealthiest quintiles of our study population, the program had no
statistically significant impact, while the impact on the middle wealth quintile is a whopping 34
percentage points. In addition, that the program had a greater impact among those with more
education than less, and there is no discernible impact among those with no education (the estimate is
large, but statistically insignificant, which may be due to the relatively small number of participants with
no education); the greatest impact is found among those who have completed high school. On the one
hand, these findings are not surprising, since EPAG was not designed to reach the poorest or most
vulnerable; rather, the first round was designed to show a demonstration effect among high-potential
girls. Bearing in mind that our study population is better off than the Liberian population as a whole, the
relative ineffectiveness of the EPAG program for its poorest and least educated participants in the first
round has deep implications for the success of the program if it were to be expanded. Subsequent
rounds of EPAG have already begun to address this issue, by adding a preliminary basic literacy training
component before the second round of training to better prepare girls with low literacy before starting
the program. These results indicate that in order to effectively serve the even poorer and less educated
girls outside the capital as EPAG scales up, further adaptation may be needed. Finally, this table shows
that the impact among younger (15-19) participants seems to be greater than that among older (20-24
and 25-27) participants.

Beyond employment rates, the EPAG program also increased the intensity of employment among those
in the treatment group relative to those in the control group. One can disaggregate the total impact on
employment, and on earnings, as the sum of two separate processes: first, it can bring people into
employment who were previously not engaged in the labor market (the extensive margin); second, it
can increase the productivity or the level of activity among those who are already working (the intensive
margin). As table 4 shows, the EPAG did increase the average number of days employed among the
entire group of participants, but among those who were employed, it increased the number of days
worked in a typical week only from 5.3 to 5.5. Similarly, while EPAG did increase the number of hours
worked, on average, it did not increase the number of hours worked among those who had
employment. In other words, the program succeeded in enhancing employment primarily on the
extensive margin, i.e., by bringing more people into the labor force rather than by enhancing the
productivity of those who are already working.


19
     These are the OLS results; the probit results for West Point are not statistically significant.


                                                                                                               14
Panel 2 of Table 4 shows that the program was also highly successful in increasing the incomes of EPAG
participants. Participants in the program saw their incomes rise by 80%, relative to those in the control
group, from a mean of LD 700 (about US$10) per week to LD 1260 per week (about US$18). This result
was obtained using a simple OLS model that compares the income of the treatment and control groups
without regard to current employment status: individuals without employment, and hence have zero
income, are also included in the comparison. The same regression using logged earnings as the
dependent variable yields an even higher estimate of a 130% increase in earnings.

This table also shows the impact of the program on earnings for those individuals who are actually
working. Earnings are determined in a two-step process: first the individual makes a decision of whether
or not to engage in employment: conditional on that choice, the final outcome of earnings is realized.
We use a Tobit model to examine the impact of EPAG on earnings conditional on having some income.
In the first stage, the Tobit model estimates a 16 percentage point impact on the likelihood of having
earnings greater than zero, compared to the 18 percentage point coefficient in Table 3 on the likelihood
of employment. Conditional on having non-zero earnings, the Tobit model predicts an increase of 418 LD
in weekly earnings, which is only a bit less than the estimate of 563 LD obtained in the unconditional
OLS regression. We interpret these results to indicate that EPAG improved earnings through two
channels: by increasing the number of people with non-zero earnings by bringing more people into
employment, and by increasing the productivity of those who were already engaged in income-
generating activities. These effects are significant, both statistically and economically, and much larger
in magnitude compared to other vocational training programs that have been rigorously evaluated.

As was discussed earlier, self-employment is much more common than wage employment: graduates of
the BDS program are more likely than graduates of the JS program to have found employment, and
about half of those who went through the job skills training and found subsequent employment are self-
employed. These results are confirmed by a disaggregation of the earnings results, shown in Table 4A.
Although those in the JS track have higher earnings at baseline, there appears to have been no
significant impact on earnings (unconditional) among JS beneficiaries, whereas the BDS graduates see
their earnings more than double. Happily, even though the impact of the program on employment
seems to vary across the wealth spectrum, the impact of the program on earnings seems relatively
equitably distributed. The increase in earnings appears to be progressively distributed across education
classes, apart from those with no education, for whom the program does not appear to have had any
significant benefit. Those with primary education only saw their incomes increase by 700 LD, double the
increase experienced by beneficiaries with a high-school education. Finally, similar to the results on
employment, younger (16-19) EPAG participants experienced a larger increase in earnings than older
(20-27) participants.

   4.2. Savings and Loans
Much has been written about the importance of monetary and other assets, especially for young
women, as a way to smooth consumption, improve bargaining power, invest in productive activities, and




                                                                                                             15
cope with the emergencies that arise all too often.20 To facilitate successful coping and provide a safe
place to save, the EPAG program assisted each participant to set up a savings account at a local bank if
she did not already have one. Table 5 presents midline survey results that indicate that the treatment
group were nearly 50 percentage points more likely to have savings than the control group, and were
saving on average LD 2500 (nearly US$35) more than the control group. EPAG graduates were also
twice as likely as the control group to have outstanding loans (six percent v. three percent), and have
loans from formal lenders (five percent v. two percent), 21 although the overall rate of obtaining credit
remains extremely low.

    4.3. Empowerment
The Adolescent Girls’ Initiative is based on the hypothesis that livelihood and life skills training for young
women will improve their lives in more than just narrowly-defined economic dimensions. In addition,
evidence is increasing that these soft skills are also essential for success in employment (Borghans et al.
2008). Surveys of employers consistently find that more than hard, technical skills, employers value
these harder-to-quantify skills of honesty and integrity, problem-solving ability, work ethic,
communication skills, the ability to work productively with others, responsibility and dependability
(Blom and Hobbs 2007). The AGI program has focused on the development and measurement of these
softer attributes that matter for employment as well as those that matter more to the individual, such
as self-confidence and empowerment. Despite the challenges of measuring such subjective outcomes,
the survey instruments included panels of questions designed to elucidate a nuanced picture of the
personality and psychosocial characteristics that are most relevant for labor market success.

Table 6A presents results on empowerment and decision-making. The first series of questions have to
do with control over resources, spending decisions and earnings. Respondents were asked how much
control they had over how to spend their own earnings; also, whether they had money of their own for
basic uses that they alone could decide how to use, without having to ask for permission. The EPAG
baseline survey found that respondents reported a high degree of control over resources even before
the program started. Because of this, the program impacts on control over resources were small, albeit
statistically significant: 80 percent of respondents at baseline said they controlled their own resources,
with a seven percentage point increase for the treatment group (relative to control). Similarly, 80
percent of those engaged in income-generating activities said that they controlled the money they
earned. The midline results indicated that among the treatment group, this had increased by roughly
eight percentage points. As shown in the second panel, EPAG graduates report that they worry less
than those in the control group. They are less likely to worry about their jobs or incomes or that they
won’t be able to pay for basic necessities, and those with partners are less worried about their
relationships breaking up. The impact on subjective well-being, as measured by a series of questions
about the respondent’s satisfaction with various dimensions on her life, indicate that EPAG was most



20
  See for example Ashraf et. al. (2010); Morcos and Sebstad (2010); and Austrian and Ghati (2010).
21
  “Formal” loans are those from banks, credit groups, susu, or money lenders; “informal loans” are those from
parents, friends, relatives, or business partners.


                                                                                                                 16
successful in improving the degree of satisfaction with one’s job or business, in line with the program’s
primary objective.

Measures such as self-confidence, self-control, and assessments of own capabilities were already
remarkably high among the young women in the study population at baseline, so we didn’t expect to
see a great improvement among the treatment group at midline. Nonetheless, two measures of non-
cognitive skills were administered at baseline and midline, and results are presented in the fourth panel
of Table 6A. The first indicator is a measure self-regulation: 22 whether the young woman can formulate a
goal, make a plan, stay on course despite setbacks, modulate intense emotions, and so on. Self-
regulation and self-control have been shown to be strong correlates of labor market success.
Respondents were asked whether they agreed with a set of eleven statements about themselves; these
responses were then aggregated into a single score reflecting the proportion of responses indicating an
ability to self-regulate. The impact of the program on this aggregate measure of self-regulation is not
statistically significant. This may reflect the relative stability of these characteristics over time. In
contrast, the self-assessed entrepreneurial score is based on questions asking how well the respondent
believes she could perform a series of six tasks related to starting or running a business, and can be
considered a task-oriented measure of self-efficacy. The aggregate measure of entrepreneurial ability
increased by roughly nine percentage points among EPAG beneficiaries relative to those in the control
group, equivalent to a quarter of a standard deviation. Enhancing participants’ self-confidence to
perform these tasks was one of the main immediate objectives of the BDS training program.

Table 6B summarizes the results on a series of questions on attitudes and self-confidence that were
added during the midline survey only (hence no panel analysis is possible). EPAG graduates report a
more positive attitude: they feel more in control and more comfortable, and they have greater
confidence in their own business abilities as well as in their personal and social lives. They are also more
confident than the control group in their personal relationships with spouses and partners, consistent
with the findings in Table 6A. Both treatment and control group respondents report equally high
confidence in their ability to return to school “should [she] decide to do so.”

These findings from the quantitative impact evaluation complement the results from a set of qualitative
focus group discussions that were held with participants at the end of the 6 months of classroom
training. Twenty-five percent of the trainees from Round 1 participated in a total of 34 focus group
discussions that covered a variety of topics including their satisfaction with the program and their
empowerment in both social and economic realms. The trainees overwhelmingly voiced a high degree of
satisfaction with the training, and trainers commented on how the motivation or “seriousness” of the
participants grew over the 6 month period. The trainees credited the transport allowance and free
childcare in particular as features that facilitated their full participation; as one trainee commented,



22
  Questions adapted from the Adolescent Self-Regulation Inventory, developed and validated for youth in the
United States by Moilanen, 2006. The questions were revised and translated into an 11-item for the Liberian
context. In the future, we plan to conduct basic testing on this scale on internal consistency and reliability.


                                                                                                                  17
“The [childcare] service made it difficult to have any excuse for missing class.” 23 In terms of
empowerment, the most striking finding concerned the participants’ aspirations and visions for the
future that seemed to spring from their participation in EPAG. Far from being content with the vocations
in which they were being trained, the participants expressed repeatedly their goal to use these
vocations as a “stepping stone” to higher goals, citing professions such as medicine, business, nursing,
and politics. The trainees appeared to understand fully that these professions would require formal
education; many girls stated their primary goal in pursuing employment was to save money to be able to
return to formal education. Although educational enrolment was not detected as a significant impact in
the quantitative survey (at midline, the treatment group was not more likely to be enrolled in school,
including night school, than the control group), the high aspirations of both the treatment and control
groups, consistent with the qualitative results, is apparent in their positive responses on the question
regarding their perceived ability to return to school (Table 6B). Other qualitative findings related to
empowerment, particularly with regard to self-confidence and improved social skills, were extremely
positive, consistent with the findings of the quantitative work.

    4.4. Fertility and Personal Behaviors
Fertility in Liberia is among the highest in the world, although it has declined from a total fertility rate of
more than six children per woman in 1990 to 5.2 in 2010. Observed fertility has many underlying
determinants including biological and behavioral factors (e.g., frequency of intercourse, contraceptive
use) as well as economic and social factors (e.g., norms, opportunity costs, and access to family
planning). Although reducing fertility was not a primary objective of EPAG, the program did include a
strong focus on the empowerment and independence of young women to exert control over their lives,
as well as lessons on reproductive health and family planning methods. The interrelated nature of young
women’s reproductive health decisions and the economic conditions in which they live is well
established in the public health literature, with poverty being a significant risk factor for risky sexual
behavior (Dunbar 2010, Ssewamala 2010). Table 7 explores outcomes related to fertility and sexual
behaviors. The first outcome concerns the desired number of children, which was roughly 3.5 at
baseline among both the treatment and control groups, with no significant change between baseline
and midline.

In spite of the lack of impact on desired fertility, the EPAG program does appear to have an impact on
realized fertility. We use three outcomes to explore fertility: whether or not the respondent currently
has at least one living child, the number of children she has, and whether or not she is currently
pregnant at the time of her interview. For the first outcome, the basic specification yields a significant
decline of four percentage points. However, this basic specification may include a purely mechanical
result because the control group was more likely to be pregnant at baseline (ten percent compared to
six percent among the treatment group, as shown in Table 2). The second specification interacts
treatment with an indicator for not being pregnant at baseline. This regression yields a smaller two
percentage point reduction in the likelihood of being a mother. The impact on number of children
further elucidates this finding. Among those women who have at least one child already, there are no

23
     Exit Poll Report from Round 1. SBA and GOPA consultants. December 2010.


                                                                                                                  18
impacts on the number of children the woman has. However, among all women, the unconditional
measure of number of children does fall among the treated relative to the control group. The results
also show a large and significant increase in pregnancies among the treatment group. Once again, this
may be a purely mechanical result of the control group’s higher pregnancy rate at baseline (making it
less likely that they would be pregnant again just one year later). For this reason, we once again interact
treatment with an indicator of not being pregnant at baseline. This interacted regression yields a smaller
but still large and significant increase of two percentage points in the likelihood of being pregnant.
Taking the three outcomes together (a weak reduction in the likelihood of having any children, a
significant reduction in the number of children, and a significant increase in pregnancy among the
treated), it appears that EPAG did not have any net impact on actual fertility. A simple cross-sectional
comparison of the likelihood of being a mother and the likelihood of being pregnant, excluding those
who were pregnant at baseline, confirms that there are no significant differences in either indicator at
midline. This may not represent a true reduction in lifetime fertility: the treatment group may simply be
delaying pregnancy for the duration of their participation in the program (although this may
demonstrate empowerment over the timing of pregnancy).

Aside from fertility-related outcomes, the EPAG program does not appear to have affected sexual
behaviors among its participants. Although control over income is often emphasized as a determinant of
women’s sexual empowerment and control, there appears to be no impact, at least in this short run, on
the number of sexual partners or the decision to use a condom. On average, these young women have
one regular partner, and use condoms about half the time. There is no distinction by treatment group.
Reported condom use, in particular, is much higher in this survey than among similar age groups in
nationally representative surveys such as DHS, and it is not clear whether the non-result is driven by
measurement error or simply the already high baseline rate of condom usage.

On the one hand, the lack of impacts on fertility and sexual behaviors are not surprising, since one might
expect that behavior changes of this sort more specific and longer-term interventions. Reducing fertility
was not, after all, an objective of the EPAG program. On the other hand, recent studies have shown a
causal link between cash transfers conditioned on school attendance and reduced fertility (Baird 2010)
and economic empowerment and reduced fertility (Bandiera 2012, Ibarrarán 2012), even in the short
run. The lack of impact in this study points to a need for further research on the conditions under which
economic empowerment may lead to reduced fertility, the mechanisms of change, and the length of
time required to affect change.

    4.5. Effects on the household
A central question of this impact evaluation is to what extent any improvements in the employment of
young women spill over onto the household in which she resides. Numerous studies have found that
increased monetary resources in the hands of women lead to improved well-being of other household
members, especially children, 24 leading researchers to hypothesize that women are more likely to spend

24
  See for example Pitt 1998, Hoddinott 1995, and King 2001. Borges 2007, in particular, has a frequently cited
claim that women reinvest 90% of their income in their household compared to 30-40% for men.


                                                                                                                 19
their earnings on household expenses than men. Most of these studies have focused on adult women,
specifically married women with children. It is not known whether the same holds true for young
women, who may have other spending priorities, have less experience in managing households, and
have younger children. Given the large increases in employment and earnings documented above, the
EPAG program serves as a good setting to examine these types of spillovers.

The evaluation included detailed interviews with the heads of the household in which EPAG participants
were residing. The purpose of the household questionnaires was precisely to examine the hypothesis
that investing in young girls would benefit her household. A secondary hypothesis was that EPAG
participation may change gender-related attitudes in the participants’ households. Household data was
collected for 1601 out of the 1622 individuals who were interviewed at both baseline and midline; this
same sample of 1601 individuals serves as the basis for both the individual and household level analysis
in this paper.

The estimated impact of the program on a broad range of household outcomes is summarized in Tables
8 and 9. Panel A of Table 8 examines the household size. The household and family dynamics in Liberia,
as in other African settings, can be complex, with families often sending their children to live with
relatives who are more able to provide for their schooling and basic needs. Young parents, in particular,
often leave home to migrate for work, leaving their children back home with relatives until they are able
to establish themselves and send for their children. If the economic success of the EPAG participant
allowed her to bring non-resident family members into her household (including but not limited to her
own children), then overall household size may have been expected to increase as a result of the
program. This does not appear to have happened, at least in the short term. The results in Table 8 show
that overall household size was not affected by the program. It is possible that the increase in earnings
due to EPAG was too small, or too short-lived, to have induced the kinds of migrations described above.

Other measures of household well-being, including food security and asset ownership, reflect shorter-
term investments that might be influenced by the economic success of EPAG participants. Panel B of
Table 8 shows the impact of EPAG on a broad range of food security measures. Household food security
was measured using two methods: dietary frequency of high-value protein-rich foods, and two
subjective questions on food shortages adapted from USAID’s FANTA questions (Coates 2007). On two
of the four of the dietary frequency questions and on both of the food shortage questions, the
treatment groups’ dietary situation improved as a result of the EPAG program. Weekly consumption of
fish and meat rose significantly by four percentage points in treatment households (from a high baseline
value of 84% for meat/chicken and 90% for fish), and weekly consumption of dairy and eggs did not
change significantly. Household heads report worrying less about insufficiency of household food
supplies, and the reported incidence of household members going to bed hungry also decreased in
treatment households relative to control. Combined, the impacts across these indicators portray a
situation of improved food security and dietary composition, consistent with the hypothesis that the
increased earnings of the EPAG participants were spent in part on food.




                                                                                                            20
Apart from questions of resource allocation, a secondary hypothesis of this evaluation was that
participation in EPAG might shift gender norms and attitudes among the participants’ household
members. This could theoretically occur if the participant actively shares the knowledge from the life
skills trainings that she receives in the program, or if the household head, upon observing the realized
productivity of the EPAG participant, revises his/her expectations about young women in general. To
investigate this, a series of questions regarding various gender norms were posed to the household
head, including the suitable age of marriage, the division of labor of various household tasks, and
appropriate types of work and socializing for young unmarried women. The results, presented in Table
9, show no impact of the program on opinions of the suitable age of marriage. This is not surprising
given the high prevalence of cohabitation and the absence of strict norms regarding age of marriage in
Liberia. The results do show some changes in attitudes toward gender equality in household duties,
which fall predominantly to women in Liberia. Respondents were asked who should be responsible for a
series of five household tasks, with response choices of “men”, “women”, or “both”. Responses were
coded as “1” for “both” (meaning that the task ought to be shared), and “0” otherwise. From a baseline
share of 12%, there was a 6 percentage point impact in the likelihood of respondents reporting that
both men and women should be responsible for washing, cooking, and cleaning. Similarly, from a
baseline share of 16%, there was a 4 percentage point increase in responses of “both” for the task of
feeding and bathing children. There were no significant impacts in the responses for other categories,
including earning money, fetching water, and helping children with studies. By summing the responses
across categories (again with a “1” indicating a gender-neutral response and a “0” indicating either men
or women alone), a “gender equality score” of 0 to 5 was computed. The impact of the program on this
score was an increase of 0.2 standard deviations, meaning that household heads of EPAG participants
experienced a small but significant shift toward more egalitarian attitudes related to the division of
household labor. Finally, respondents were asked to rate (hypothetically) whether they would approve
of the young unmarried female household members participating in a variety of activities. The first set
of activities had to do with socializing with friends and dating, while a second set dealt with various
types of employment, education, and training. Other than a small reduction in the approval rate for
socializing with friends, there were no impacts of the program on any of the indicators, probably owing
mostly to the high rate of approval at baseline for the majority of these activities.

Overall, the results in Table 9 indicate that the EPAG program did not have a significant impact on the
attitudes of household heads toward the economic or social activities of young women living in their
households, but did have a small significant impact on their attitudes toward gender norms regarding
domestic duties. This might be driven by the distinction between asking about the respondent’s
household members directly versus asking about men and women generally. When asked about specific
activities in relation to the young women in their household, the household heads’ responses did not
change over time, but when asked about sharing of household duties in theory (not in relation to their
particular household), there was a significant difference by treatment group. Given that the EPAG
participants themselves, whether or not they are engaged in income-generating activities, do not report




                                                                                                           21
any change in the amount of time spent on domestic work, 25 we have no evidence that these shifting
norms have affected the division of labor in practice. Further research would be needed to disentangle
the issues of changing gender norms in theory versus practice, and attitudes toward norms in general
versus those at play in one’s own household.


5. Cost Effectiveness
Although the first task of any evaluation is to demonstrate the effectiveness of an intervention – that is,
whether or not the program actually has a measurable and attributable impact – this is not enough to
recommend the program to policymakers. This requires also that the program can show that it is worth
spending scarce public resources to do it. Ideally, a program worth doing will be both effective and cost-
effective. One can measure cost-effectiveness in terms of the number of physical outputs produced or
outcomes achieved, e.g. the number of people employed per dollar spent, or one can measure
achievements in terms of the value of the benefits acquired relative to the amount of money spent. In
the case of the EPAG, the unit cost of training in Round 1 was roughly $1200 for the Business Skills track
and $1650 for the Job Skills track. The EPAG Job Skills track consisted of more hours of training, and
required purchasing practical equipment for each trade area. Some Job Skills trainers, especially those
with more specialized skills and experience, were also more costly. The estimated unit costs cover
trainer salaries (EPAG has mostly college-educated trainers), training materials (not including curriculum
development), training venue rental, administrative and overhead costs of the training provider,
childcare costs, event costs (job fairs, etc.), stipends to mentors, trainee transport allowances and
completion bonuses. The costs also cover the withheld incentive payment to the training provider,
based on how many trainees find employment.

Although high relative to most developing-country budgets, these costs are well within the range of the
Jovenes youth training programs implemented in Latin America (cf. Ibarrarán and Rosas 2009). The
Jovenes programs were estimated to cost between $700 and $2000 per participant, depending on the
country (Betcherman 2007). Not only were the costs of the EPAG program within international norms,
the program itself is also cost-effective. Given the large earnings gains demonstrated in this study (44
USD per month for Business Skills, and 11 USD per month for Job Skills, as shown in Table 4A), and
assuming the gains persist at the same level, the cost of the Business Skills training is recouped within
three years. 26 For the Job Skills training, given the higher costs and lower benefits, it would take
approximately 12 years to recoup the training costs, making it less cost-effective. The project has
revamped the Job Skills training substantially in light of these findings, condensing the training timetable
and seeking formal agreements in advance with employers to hire EPAG graduates.



25
   A limited number of time use questions were included in the survey of EPAG participants, but the results are not
reported in this paper.
26
   In addition to the assumption of persistence of earnings gains, other assumptions include income accruing from
month 7, no discounting, and an exchange rate of 70 LD= 1USD. With an annual discount rate of 5%, the Business
Skills training costs are still recouped within 3 years while the Job Skills training costs would take 17 to 20 years to
recover.


                                                                                                                           22
6. Robustness Checks

     6.1. Sample attrition
Many challenges were encountered during survey implementation, due to the context – the densely
populated and impoverished communities in and around Monrovia are difficult settings in which to find
and track respondents—as well as the transience of a young study population. Despite vigorous efforts
to track and interview each individual in the sample, a certain amount of survey attrition was expected.
As the survey response rates in Table 1 show, 1622 (or 80%) of the individuals in the study sample were
successfully interviewed in both the baseline and midline surveys. Another 305 respondents were
interviewed at baseline but not at midline and hence are not in the panel used for the analysis in this
paper. 27 This survey attrition, while not much higher than other program evaluations in Africa, may
cause concern that the results of this evaluation are biased, especially if the loss to follow up is
correlated with individual characteristics that might affect the outcomes. To address this concern, Table
10 presents regressions on the likelihood of panel inclusion, that is, the likelihood of being interviewed
at both baseline and midline. The first column indicates that treated individuals are significantly more
likely than control to have been interviewed twice. This result persists even after controlling for
individual characteristics and community dummies. Columns 2, 3, and 4 show that age, school
attendance, and employment status at baseline are all correlated with survey attrition. To further
investigate whether these characteristics lead to differential attrition between the treatment and
control groups, we interact treatment with particular characteristics at baseline. Columns 5 and 6 show
that, conditional on being in the control group, the most likely predictor of attrition is having a child. By
itself, being a mother does not predict attrition, but when interacted with treatment, we see that
control group mothers are more likely to have dropped out of the panel than their treated counterparts.
Employment at baseline, while positively correlated with attrition, does not differentially affect treated
and untreated individuals.

Because the differential attrition between treatment and control groups may bias our results, we use
Inverse Probability Weighting (IPW) as outlined in Wooldridge (2002) to adjust the estimates of our key
outcomes, using the inverse probability of inclusion in the panel as a probability weight. As a first step,
we use the probit model in Column (6) of Table 10 to regress the likelihood of being observed twice on
baseline individual characteristics, including those likely to affect attrition, such as employment and
parental status. In the second step, we use the inverse of the predicted values from that probit model as
probability weights to redo the difference-in-difference regressions for our key outcomes of interest.
This method gives more weight to the individuals with the highest chance of attrition, giving them more
influence on the estimate of the impact than those with a low probability of attrition. The results are
reported in Table 11. The results show a high degree of similarity between the original (unadjusted) and
the adjusted estimates. Across all outcomes, the point estimates and standard errors vary only slightly.


27
  Of these 305 cases, nine are dropped from the attrition analysis because the household head was not
interviewed at baseline, hence the household level control variables are not available.


                                                                                                                23
Because the adjusted estimates so closely mirror our original findings, we conclude that selective survey
attrition is unlikely to be biasing our results.

7. Conclusion
Liberia’s history of civil conflict compounds the challenges that face many African countries: large
populations of low-skilled youth and limited labor market opportunities. A growing policy dialogue
around youth employment in Africa has tended to focus on engaging young men in productive work to
reduce the seeming threat of conflict, crime, and civil unrest. We argue that promoting young women’s
engagement in the labor market is just as pressing: getting women on the right track early in their
working lives has the potential to affect not just their employment trajectories but, through the channel
of empowerment, can also transform their roles in their families and communities. By focusing on both
young men and women, African policy makers and development partners can ensure that countries like
Liberia are able to reap the demographic dividend that comes only once in a country’s development as
fertility rates slow to a point where the ratio of working-age adults to dependents falls to a minimum.

As this paper demonstrates, skills training programs for young women can be wonderfully effective in
increasing employment and incomes, at least in the short term. The EPAG program, which delivered six
months of classroom-based skills training followed by six months of job placement support for either
self or wage employment, led to a 47% increase in employment and 80% growth in earnings, relative to
a randomly selected control group of non-participants. The program’s Business Skills track had markedly
higher impacts on employment and earnings than the Job Skills track, which focused on wage
employment. These impacts vary somewhat but remain consistently positive and significant across
almost all communities, educational backgrounds, and wealth levels. The highest impacts were obtained
for those in the middle of the wealth distribution, and for girls with higher educational levels, which is
consistent with the program’s initial screening for young women with basic literacy who would be able
to make use of a classroom-based skills course.

These strong impacts on employment and earnings translated into positive impacts in other realms of
the participants’ lives. Our results show striking improvements in various empowerment measures,
including access to and control over monetary resources, including savings, where the program led to a
sizeable difference of 35 USD in savings between treated and control individuals. The study also
documents significant improvements in a wide range of subjective outcomes including measures of
worry, life satisfaction, self-regulation, self-confidence, and self-perceptions of social abilities. In the
area of fertility and sexual behaviors, the results paint a somewhat more nuanced picture. The EPAG
program had no discernible effect on the desired number of children or on the actual number of
children, conditional on having any children. There was a weak reduction in the likelihood of having any
children, and a stronger increase in the likelihood of being pregnant, even after excluding those who
were pregnant at baseline. On net, these impacts appear to cancel each other out, consistent with a
hypothesis that treated individuals waited until the end of the EPAG program to become pregnant.

The third main area of outcomes looks at household-level measures. Consistent with the wide body of
literature on the benefits to the household of women’s increased resources, the results show a


                                                                                                               24
significant improvement in household food security, judged by increased consumption of high-value
animal proteins and lowered incidence of food shortages. Beyond food security, we find no
generalizable impact on household asset holdings, plausibly due to the short term nature of this
analysis, but we do find significant improvements in subjective measures of attitudes toward gender
responsibilities among the heads of EPAG households. These outcomes provide some evidence that
norms are shifting at the household level in ways that may benefit other female household members.
Taken together with the results on food security, this suggests that the program had positive, though
small, impacts on the households of treated individuals.

We end with a mention of the limitations of the study. First, as mentioned above, is the issue of
selective attrition from the sample. Second, given the short time horizon of this study, in which follow
up data was collected just one month after the conclusion of the first round of the program (which was
seven months after the conclusion of the classroom-based training), questions remain about the
persistence of the effects found and the potential for further impacts to emerge. The research design of
this evaluation, in which the control group took part in the second round of the EPAG program,
precludes a rigorous experimental comparison of treated and untreated individuals after the midline
survey. However, an endline survey was conducted after the second round of the EPAG program as per
the timeline depicted in Figure 1. Examination of this endline survey data will permit a descriptive
analysis of the outcomes of the first group of trainees 12 months after they completed the EPAG
program, as well as examination of the outcomes of the second batch of trainees. The second round
included not only the control group from this impact evaluation but also newly recruited participants
who were offered brief basic literacy and numeracy training program prior to program entry. Work is
already underway to design and implement the third round of EPAG, with a substantial redesign of the
Job Skills track, an emphasis on reaching younger girls with lower literacy, and expansion to
communities outside of Monrovia. If the high success rates found in this study are replicated for these
future cohorts, the EPAG program should serve as a model for policy makers in Africa and the world
seeking to improve lives and livelihoods of all youth, male and female.




                                                                                                           25
8. References
1) Ai, Chunrong, and Edward C. Norton (2003). Interaction terms in logit and probit models. Economics
    Letters 80, p. 123-129.
2) Ashenfelter, Orley (1978). Estimating the effect of training programs on earnings, Review of
    Economics and Statistics, vol. 60, pp. 47–57.
3) Ashraf, Nava, Dean Karlan, and Wesley Yin (2010). Female empowerment: Impact of a commitment
    savings product in the Philippines. World development, 38(3), 333-344.
4) Attanasio, Orazio, Adriana Kugler, and Costas Meghir (2008). Training Disadvantaged Youth in Latin
    America: Evidence from a Randomized Trial. NBER Working Paper 13931.
5) Austrian, Karen and Dennitah Ghati. 2010. Girl-Centered Program Design: A Toolkit to Develop,
    Strengthen & Expand Adolescent Girls Programs. Nairobi, Kenya: The Population Council. Available
    at: http://www.ungei.org/files/2010PGY_AdolGirlToolkitComplete.pdf.
6) Baird, Sarah, Ephraim Chirwa, Craig McIntosh, and Berk Özler. The short‐term impacts of a schooling
    conditional cash transfer program on the sexual behavior of young women. Health economics 19,
    no. S1 (2010): 55-68.
7) Bandiera, Oriana, Niklas Buehren, Robin Burgess, Markus Goldstein, Selim Gulesci, Imran Rasul, and
    Munshi Sulaiman (2012). Empowering Adolescent Girls: Evidence from a Randomized Control Trial in
    Uganda. Working Paper.
8) Bertrand, Marianne, and Sendhil Mullainathan. 2001. "Do People Mean What They Say?
    Implications for Subjective Survey Data." American Economic Review, 91(2): 67-72.
9) Betcherman, Gordon, Martin Godfrey, Susana Puerto, Friederike Rother, and Antoneta Stavreska
    (2007). A Review of Intervention to Support Young Workers: Findings of the Youth Employment
    Inventory. World Bank Social Protection Discussion Paper No. 0715.
10) Blom, Andreas and Cynthia Hobbs (2007). School and Work : Does the Eastern Caribbean Education
    System Adequately Prepare Youth for the Global Economy?. World Bank, Washington, DC.
11) Borges, Phil (2007). Women Empowered: Inspiring Change in the Emerging World. New York.
12) Borghans, Lex., Angela L. Duckworth, James J. Heckman, and Bas Ter Weel (2008). The economics
    and psychology of personality traits. Journal of Human Resources, 43(4), 972-1059.
13) Card, David, Pablo Ibarrarán, Ferdinando Regalia, David Rosas, and Yuri Soares (2007). The Labor
    Market Impacts of Youth Training in the Dominican Republic: Evidence from a Randomized
    Evaluation. NBER Working paper 12883.
14) Coates, Jennifer, Anne Swindale, Paula Bilinsky (2007). Household Food Insecurity Access Scale
    (HFIAS) for Measurement of Food Access: Indicator Guide. Version 3. Food and Nutrition Technical
    Assistance (FANTA) project. USAID.
15) Dunbar, Megan S., M. Catherine Maternowska , Mi-Suk J. Kang , Susan M. Laver , Imelda
    Mudekunye-Mahaka and Nancy S. Padian (2010). Findings from SHAZ!: A Feasibility Study of a
    Microcredit and Life-Skills HIV Prevention Intervention to Reduce Risk Among Adolescent Female
    Orphans in Zimbabwe. Journal of Prevention & Intervention in the Community, 38:2, 147-161.
16) Empower. 2010. The Girls Entrepreneurship Toolkit, “It’s her Business: A Handbook for Preparing
    Young, At-risk Women to Become Entrepreneurs.” Available at: www.itsherbusiness.org




                                                                                                         26
17) Filmer, Deon and Lant Pritchett (2001). Estimating Wealth Effects without Expenditure Data—or
    Tears: An Application to Educational Enrollments in States of India. Demography, Volume 38-
    Number 1, February 2001: 115–132.
18) Hicks, Joan Hamory, Michael Kremer, Isaac Mbiti, and Edward Miguel (2011). “Vocational Education
    Voucher Delivery and Labor Market Returns: A Randomized Evaluation among Kenyan Youth”.
    Working Paper.
19) Hoddinott, John, and Lawrence Haddad (1995). Does Female Income Share Influence Household
    Expenditures? Evidence from Cote d’Ivoire. Oxford Bulletin of Economics and Statistics, 57, 1.
20) Ibarrarán, Pablo, Laura Ripani, Bibiana Taboada, Juan Miguel Villa, and Brigida Garcia (2012). Life
    skills, employability, and training for disadvantaged youth: Evidence from a randomized evaluation
    design. Institute for the Study of Labor (IZA) Discussion Paper No. 6617.
21) Ibarrarán, Pablo and David Rosas (2009). Evaluating the Impact of Job Training Programs in Latin
    America: Evidence from IDB funded operation. Journal of Development Effectiveness.
22) International Labour Organisation / World Bank (ILO/WB) (2010). Joint Synthesis Report: Inventory of
    Policy Responses to the Financial and Economic Crisis. Geneva and Washington, DC.
23) King, Elizabeth M., and Andrew D. Mason (2001). Engendering development: Through gender
    equality in rights, resources, and voice. Policy Research Report, Washington, DC: World Bank
    Publications.
24) Liberia Institute of Statistics and Geo-Information Services (LISGIS) (2007). Report on the Core
    Welfare Indicators Questionnaire. Monrovia: LISGIS.
25) Liberia Institute of Statistics and Geo‐Information Services (LISGIS) (2010). Report on the Liberia
    Labour Force Survey 2010. Monrovia: LISGIS.
26) Measure DHS. Final report of the 2007 Liberia Demographic and Health Survey, June 2008. Available
    at: http://www.measuredhs.com/publications/publication-FR201-DHS-Final-Reports.cfm
27) Mollanen, Kristin L. (2006). The Adolescent Self-Regulation Inventory: The Development and
    Validation of a Questionnaire of Short-term and Long-term Self-Regulation. Journal of Youth and
    Adolescence 36:835-848.
28) Morcos, Carmen and Jennifer Sebstad (2010). Financial Education for Adolescent Girls, Nike
    Foundation.
29) Pitt, Mark M. and Shahidur R. Khandker (1998). The Impact of Group-based Credit on Poor
    Households in Bangladesh: Does the Gender of Participants Matter? Journal of Political Economy,
    vol. 106, no. 5.
30) Ssewamala, F. M., Ismayilova, L., McKay, M., Sperber, E., Bannon, W., & Alicea, S. (2010). Gender
    and the effects of an economic empowerment program on attitudes toward sexual risk-taking
    among AIDS-orphaned adolescent youth in Uganda. Journal of Adolescent Health, 46(4), 372-378.
31) Wooldridge, Jeffrey M. Econometric Analysis of Cross Section and Panel Data, MIT Press:
    Cambridge MA. 2002.




                                                                                                           27
Figure 1. Impact Evaluation Timeline




Figure 2. Summary of Assignment Process and Modifications




                                                            28
Table 1. Response Rates for Baseline and Midline Surveys
                Neither Baseline                                                Baseline and
                                      Baseline Only        Midline Only                                  Total
                  nor Midline                                                     Midline

Treatment              33                  169                   45                 1026                 1273
Control                26                  136                   11                  596                 769
Total                     59                  305                    56                   1622                  2042
Note: A total of 2106 young women were originally recruited for this study and the EPAG program. This table excludes: 39
individuals from the control group who were issued as replacements and 25 exempt from randomization due to pregnancy.
Please note that of the 1622 who were interviewed twice, 21 are dropped from the analysis because household information at
baseline is missing. This leaves a sample of 1601 individuals for the main analysis in this paper. Of the 305 interviewed at
baseline only, nine are missing baseline household information, so they are dropped from the attrition analysis in the
“Robustness checks” section.




                                                                                                                               29
Table 2A. Baseline balance tests of individual characteristics
                                                     Panel 1: Demographics
                                            Control          Treatment        Difference      P Value    Observations
 Average age                                22.849            22.780           -0.069          0.623        1601
 Age 16-19                                   0.131             0.163             0.032         0.086*       1601
 Age 20-24                                   0.582             0.529           -0.053          0.038**      1601
 Age 25-27                                   0.287             0.309             0.022         0.365        1601
 Married                                     0.048             0.063             0.015         0.225        1601
 Cohabiting                                  0.284             0.302             0.018         0.445        1601
 Separated/Widowed/Divorced                  0.010             0.009            -0.001         0.792        1601
 Never married                               0.658             0.627            -0.031         0.209        1601
                                          Panel 2: Education- Highest level completed
                                            Control          Treatment        Difference      P Value    Observations
 None/Informal only                          0.020             0.031             0.011         0.187        1601
 Primary (grades 1-6)                        0.196             0.189            -0.007         0.730        1601
 Junior High (grades 7-9)                    0.224             0.243             0.019         0.380        1601
 High School (grades 10-12)                  0.269             0.257            -0.012         0.597        1601
 Completed High School or above              0.291             0.280            -0.011         0.624        1601
                                          Panel 3: Income-Generating Activities (IGA)
                                            Control          Treatment        Difference      P Value    Observations
 Any IGA                                      0.364              0.391          0.027         0.284         1601
 Self-employment                              0.316              0.333          0.017         0.476         1601
 Wage-employment                              0.048              0.058          0.010         0.410         1601
 Monthly earnings in LD                  2,377.181          3038.042          660.861         0.084*        1597
 (unconditional on IGA)
 Monthly earnings in LD (conditional     6,570.887          7875.539        1,304.652          0.145         602
 on IGA)
 Any savings                                 0.333            0.342               0.009       0.708         1601
 Any loan outstanding                        0.041            0.03               -0.011       0.232         1601
                                     Panel 4: Ownership of Assets (either Jointly or Alone)
                                            Control        Treatment          Difference      P Value    Observations
 Owns cell phone                             0.694            0.632              -0.062        0.012**      1601
 Owns electronics                            0.255            0.284               0.029        0.213        1599
 Owns jewelry                                0.403            0.394              -0.009        0.729        1600
 Owns furniture                              0.429            0.391              -0.038        0.139        1601
 Owns poultry                                0.186            0.19                0.004        0.850        1600
 Owns vehicle                                0.017            0.016              -0.001        0.853        1601
 Owns dwelling                               0.187            0.175              -0.012        0.541        1600
 Owns land                                   0.152            0.147              -0.005        0.806        1600
                                      Panel 5: Empowerment and Psychosocial Measures
                                             Control       Treatment          Difference      P Value    Observations
                            28
 Any money of your own?                       0.798           0.790              -0.008        0.708        1601
 Entrepreneurial Ability score (0-1)          0.509           0.501              -0.008        0.698        1070
 Self-regulation score (0-1)                  0.655           0.650              -0.005        0.582        1598
 Satisfaction w/ life overall (0-1)           0.429           0.421              -0.008        0.766        1600
 Satisfaction w/ job/ business, if
 applicable (0-1)                             0.276           0.326               0.050       0.075**       1161


28
  The wording of this yes/no question was “Do you most of the time have money of your own for basic uses that
you alone can decide how to use, without having to ask for permission?”


                                                                                                                        30
 Table 2A. Baseline balance tests of individual characteristics (cont.)
                                                             Panel 6: Family
                                               Control          Treatment           Difference          P Value         Observations
 Number of brothers and sisters                 6.182              6.017             -0.165              0.330              1598
 Is father alive                                0.714              0.675             -0.039              0.104              1601
 Is mother alive                                0.872              0.863             -0.009              0.622              1601
 Ever displaced during the war                  0.781              0.773             -0.008              0.723              1601
 Any living children                            0.665              0.715              0.050              0.037**            1601
 If yes, how many?                              1.570              1.537             -0.033              0.510              1115
 Are you pregnant?                              0.097              0.063             -0.034              0.014**            1601
 Desired number of children                     3.501              3.470             -0.031              0.618              1588
                                                        Panel 7: Sexual Behaviors
                                                Control          Treatment          Difference          P-Value         Observations
 Have you ever had sex?                          0.976              0.976             0.000              0.988              1601
 If yes, age at sexual initiation
 (conditional on ever having sex)               17.198             16.981             -0.217             0.022**            1530
 Have you ever had a regular partner?            0.939              0.947             0.008              0.485              1563

 If yes, number of regular partners              0.993              1.009             0.016              0.418              1561
 Have you ever had a casual partner?             0.331              0.336             0.005              0.850              1563
 If yes, number of casual partners               0.539              0.542             0.003              0.961              1560
 Have you ever used a condom?                    0.701              0.716             0.015              0.524              1601
 If yes, did you use a condom last time
 you had sex with your regular
 partner?                                        0.547              0.556             0.010              0.762              1087

 Were you ever forced to have sex?               0.092              0.114             0.022              0.175              1563

Notes: The Self-Regulation Score is based on a set of eleven questions about self-reported behavioral patterns, adapted from
the Adolescent Self-Regulatory Inventory (Mollanen 2006). Responses were coded as follows: 1=strongly disagree, 2=somewhat
disagree, 3=somewhat agree, 4=strongly agree. The score is equal to the proportion of responses that reflect positive behaviors
or attitudes, of the items that the respondent reported on, conditional on having responded to at least 6 items. The
Entrepreneurial Ability Score is based on a set of questions that asked the respondent to rate how well she had been able to
perform the following activities in the past year: (i) find information about job opportunities in your area, run her own business,
(ii) save in order to invest in future business opportunities, (iii) manage business finances effectively, (iv) bargain with a supplier
to obtain good prices when purchasing, (v) collect money from someone who owed her for purchases who are not repaying on
time. The response categories were: very well, somewhat well, not very well, and not at all well. Numerical values were
assigned to each response. The score is based on the proportion of items to which the respondent answered “very well” and
“somewhat well” as opposed to “not very well” and “not at all well,” conditional on having responded to at least three of the
activities.




                                                                                                                                          31
Table 2B. Baseline balance tests of household characteristics
                                             Panel 1: Household Characteristics
                                     Control          Treatment        Difference   P Value   Observations
 Household head is male               0.597             0.582             -0.015     0.544       1601
 Household size                       4.736             4.738             0.002      0.988       1601
 Father’s education
 None or informal                     0.247              0.218           -0.029     0.176        1601
 Primary                              0.051              0.034           -0.017     0.086*       1601
 Junior School                        0.061              0.062           0.001      0.938        1601
 High School                          0.235              0.243           0.008      0.713        1601
 Beyond High School                   0.391              0.389           -0.002     0.930        1601
 Mother’s education
 None or informal                     0.583              0.563           -0.020     0.444        1601
 Primary                              0.085              0.097           0.012      0.436        1601
 Junior School                        0.082              0.078           -0.004     0.795        1601
 High School                          0.117              0.121           0.004      0.809        1601
 Beyond High School                   0.117              0.117           0.000      0.994        1601
 Number of hh members
 aged 5-11                            0.587              0.631           0.044      0.344        1587
 Proportion of children per
 household aged 5-11 who
 are attending school,
 conditional on the
 household having any
 children aged 5-11                  85.693             82.739           -2.954     0.296         646
 Number of hh members
 aged 5-14                            0.842              0.894           0.052      0.376        1587
 Proportion of children per
 household aged 5-14 who
 are attending school,
 conditional on the
 household having any
 children aged 5-14                  87.850             81.862           -5.988     0.016**       779
 Number hh members aged
 13-30                                2.313              2.279           -0.034     0.652        1601
 Proportion of individuals per
 household aged 13-30 with
 IGA conditional on the
 household having individuals
 aged 13-30                          48.351              44.819          -3.532     0.128        1559




                                                                                                             32
Table 2B. Baseline balance tests of household characteristics (cont.)
                                                Panel 2: Housing Characteristics
                                    Control            Treatment          Difference   P Value    Observations
 Primary Source of Household Drinking Water
 Private connection to
 pipeline                            0.131                0.119            -0.012      0.501         1601
 Bottled / bagged water              0.010                0.014            0.004       0.530         1601
 Borehole / rainwater
 catchments                          0.015                0.020            0.005       0.441         1601
 Vendor / tanker                     0.104                0.163            0.059       0.001***      1601
 Public tap                          0.247                0.220            -0.027      0.208         1601
 Protected spring                    0.102                0.100            -0.002      0.881         1601
 Unprotected well / spring or
 open source                         0.012                0.021            0.009       0.194         1601
 Hand pump                           0.376                0.342            -0.034      0.167         1601

 Sanitation
 Own flush toilet                     0.308               0.276            -0.032      0.167         1601
 Public flush toilet                  0.214               0.185            -0.029      0.163         1601
 Neighbour’s flush toilet             0.019               0.017            -0.002      0.777         1601
 Own improved pit latrine             0.184               0.180            -0.004      0.841         1601
 Own unimproved pit latrine           0.043               0.047             0.004      0.718         1601
 Neighbour's improved pit
 latrine                              0.068               0.060            -0.008      0.536         1601
 Neighbour's unimproved pit
 latrine                              0.019               0.027             0.008      0.314         1601
 Open source (bush, river,
 beach, stream)                       0.145               0.210             0.065      0.001***      1601
 Household owns their house           0.378               0.394             0.016      0.518         1601




                                                                                                                 33
Table 3. Impact of EPAG on Participation in Income Generating Activities
                                                      ITT Estimates                             ATT Estimates
                        Baseline
                         Mean             OLS             OLS              OLS                  Using IV
Any IGA                  0.381           0.181***        0.181***         0.181***               0.190***
                        [0.486]         (0.026)         (0.026)          (0.037)                (0.026)

Observations                 1601             3200           3200            3200                 3200
Controls                                       No             Yes             No                   Yes
Using individual fixed
effects                                        No             No              Yes                  No
For all tables: Standard deviation in brackets. Standard error in parentheses, clustered by classroom.
ITT: Intent to treat estimator. ATT: Average treatment effect on the treated. 46 individuals were offered
a space but declined to join the first round of training. We use treatment as an instrumental variable for
participation in the first round of training.
Notes: Two observations are missing for participation in income generating activities at midline. Controls
used are type of training, household size, sex of household head, asset based quintiles, community,
age of adolescent, educational status of adolescent, marital status of respondent, parental status of
respondent (indicator variables for adolescent having a child and being pregnant), orphan status of
respondent (one parent or both parents being deceased), and month of interview at midline; all controls
are from baseline data except month of midline interview.


Table 3a. Impact of EPAG on Participation in income generating activities, by training track
                                         Job Skills Trainees                                      BDS Trainees
                                                       +                                                     +
                                             (n= 608)                                               (n=993 )
                            Baseline-          OLS                                Baseline-            OLS
                              Mean        (w controls)       Observations           Mean           (w controls)     Observations
Any IGA                       0.309           0.101**           3200                0.425             0.226***           3200
                             [0.463]         (0.041)                               [0.495]           (0.031)
Wage-employment               0.051           0.145***          3200                0.055             0.017              3200
                             [0.220]         (0.039)                               [0.229]           (0.018)
Self-employment               0.258          -0.053             3200                0.370             0.204***           3200
                             [0.438]         (0.042)                               [0.483]           (0.033)
+ Table reports the sum of coefficients and joint p-values of the interaction terms “post x treat” and “post x treat x type” where
type is “job skills” or “BDS”. Controls used are household size, sex of household head, asset based quintiles, community, age of
adolescent, educational status of adolescent, marital status of respondent, parental status of respondent (indicator variables
for adolescent having a child and being pregnant), orphan status of respondent (one parent or both parents being deceased),
and month of interview at midline; all controls are from baseline data except for month of midline interview.




                                                                                                                                     34
Table 3b. Heterogeneous Impacts of EPAG on Participation in income generating activities—by community
                                                  OLS
 Communities                       Any IGA           Observations         Controls
                                    0.181***
 All
                                   (0.026)               3200                Yes
                                   -0.109**
 West Point
                                   (0.041)                252                Yes
                                    0.058
 Battery Factory
                                   (0.065)                314                Yes
                                   0.141
 New Kru Town
                                   (0.117)                338                Yes
                                    0.245*
 Doe Community
                                   (0.112)                358                Yes
                                   0.347***
 Bassa Community
                                   (0.075)                321                Yes
                                   0.121*
 Red Light
                                   (0.063)                558                Yes
                                    0.239***
 Old Road
                                   (0.032)                429                Yes
                                   0.354**
 Bentol
                                     (0.087)               184                Yes
                                      0.184***
  Kakata
                                     (0.056)               446                Yes
Controls used are household size, sex of household head, asset based quintiles, age of adolescent, educational status of
adolescent, marital status of respondent, parental status of respondent (indicator variables for adolescent having a child and
being pregnant), orphan status of respondent (one parent or both parents being deceased), and month of interview at midline;
all controls are from baseline data except for month of midline interview.




                                                                                                                                 35
Table 3c. Heterogeneous Impacts of EPAG on Participation in Income Generating Activities—by Wealth, Education and Age
                                         Panel 1: By Wealth Quintile
                                                                            OLS
                                                          Quintiles of asset-based welfare index
                       All                 Q1              Q2               Q3              Q4            Q5

 Any IGA             0.181***               0.006          0.178***          0.337***        0.256***     0.074
                     (0.026)               (0.049)         (0.055)          (0.051)         (0.048)       (0.046)
 Controls             Yes                   Yes             Yes              Yes             Yes           Yes
 Observations         3200                  606            646               659             634           655
                                               Panel 2: By Education
                                                                                 OLS
                                                                      Education level of the Girl
                                            None/                                                         High school
                                            Informal                         Junior high                  graduate
                      All                   only            Primary          school         High school   and above


 Any IGA             0.181***               0.219           0.192***         0.171***        0.140***     0.222***
                    (0.026)                 (0.149)         (0.047)          (0.050)        (0.044)       (0.036)
 Controls            Yes                      Yes            Yes              Yes             Yes          Yes
 Observations        3200                     88             612              758             835          907
                                                Panel 3: By Age Group
                                                                                OLS
                                                                        Age group of the girl

                      All                                   16-19            20-24          25-27

 Any IGA             0.181***                                0.227***         0.187***       0.151***
                    (0.026)                                 (0.049)          (0.032)        (0.038)

 Controls             Yes                                    Yes              Yes             Yes
 Observations          3200                                  484             1753          963
Controls used are household size, sex of household head, asset based quintiles (excluded for “by wealth” regressions),
community, age of adolescent (excluded for “by age group” regressions), educational status of adolescent (excluded for “by
education” regressions), marital status of respondent, parental status of respondent (indicator variables for adolescent having a
child and being pregnant), orphan status of respondent (one parent or both parents being deceased), and month of interview at
midline; all controls are from baseline data except month of midline interview.




                                                                                                                                    36
Table 4. Impact of EPAG on Hours and Earnings
                                              Panel 1: Intensive Employment
                                           Baseline
                                            Mean               OLS                               Controls       Observations
Number of days worked past week             2.000             1.107***                             Yes             3155
(unconditional)                            [2.758]           (0.153)

Number of days worked past week              5.321             0.251***                            Yes              1494
(conditional on IGA)                        [1.604]           (0.091)

Number of hours worked on a typical          2.712             1.282***                            Yes              3134
day (unconditional)                         [4.033]           (0.216)

Number of hours worked on a typical          7.247            -0.240                               Yes              1473
day (conditional on IGA)                    [3.253]           (0.193)

                                                      Panel 2: Earnings
                                           Baseline
                                            Mean                OLS              Tobit           Controls       Observations
Total earnings past week (LD),              698.9             563.216***       417.799***
unconditional*                             [1843.0]          (124.839)         (77.518)            Yes              3178

Probability that total earnings past                                             0.157***
week (LD) > 0                                                                   (0.024)            Yes              3178

Total earnings past week (log),
                                             2.592            1.316***           0.982***
unconditional*
                                            [3.443]           (0.197)           (0.151)            Yes              3178

Probability that total earnings past                                             0.155***
week (log) > 0                                                                  (0.024)            Yes              3178

Total earnings past week (LD),              1854.0            437.573**
conditional on IGA                         [2622.0]          (186.749)                             Yes              1517

Total earnings past week (log),                6.877              0.162
conditional on IGA                            [1.403]            (0.128)                               Yes              1517
Standard deviation in brackets. Robust standard errors in parentheses, clustered by classroom.
* For the unconditional earnings regression, a zero was imputed for respondents who report no income-generating activities.
Controls used are household size, sex of household head, asset based quintiles, community, age of adolescent, educational
status of adolescent, marital status of respondent, parental status of respondent (indicator variables for adolescent having a
child and being pregnant), orphan status of respondent (one parent or both parents being deceased), and month of interview at
midline; all controls are from baseline data except month of midline interview. Three outliers (LD 105,000, 133,625 and
218,750) were top-coded to 28,800 and a control indicator variable (1 for the ones top-coded and 0 otherwise) was added to
the regressions. Days per week over 7 days were recoded to 7 days and a covariate indicating recoded observations was added.
                                       th                                                                    th
When hours per day exceeded the 99 percentile (which is 15 hours per day), they were recoded to the 99 percentile and a
covariate indicating recoded observations was added.




                                                                                                                                 37
Table 4a. Impact of EPAG on Participation on Earnings, by Training Track
                                         Job Skills Trainees                                        BDS Trainees
                                                       +                                                      +
                                              (n= 607)                                                (n=990)
                            Baseline-          OLS                                Baseline-            OLS
                             Mean         (w controls)       Observations           Mean           (w controls)      Observations
Total earnings last           729.5          200.1              3178                 680.1            768.0***            3178
week (LD)                   [2322.7]        (185.6)                               [1474.8]           (169.6)
unconditional
+ Table reports the sum of coefficients and joint p-values of the interaction terms “post x treat” and “post x treat x type” where
type is “job skills” or “BDS”. Controls used are household size, sex of household head, asset based quintiles, community, age of
adolescent, educational status of adolescent, marital status of respondent, parental status of respondent (indicator variables
for adolescent having a child and being pregnant), orphan status of respondent (one parent or both parents being deceased),
and month of interview at midline; all controls are from baseline data except for month of midline interview.

Table 4b. Heterogeneous Impacts of EPAG on Participation on Earnings—by Wealth, Education and Age
                                         Panel 1: By Wealth Quintile
                                                                           OLS
                                                         Quintiles of asset-based welfare index
                              All            Q1              Q2               Q3              Q4            Q5

 Total earnings             563.216***      444.259**       521.720**        436.218**      805.215***     477.683*
 (LD) unconditional        (124.839)        (219.093)      (228.871)        (210.061)       (216.182)      (277.570)
 Controls                     Yes             Yes            Yes              Yes             Yes           Yes
 Observations                 3178            601           641               653             631           652
                                                Panel 2: By Education
                                                                                 OLS
                                                                      Education level of the Girl
                                             None/                                                         High school
                                             Informal                        Junior high                   graduate
                              All            only           Primary          school          High school   and above


 Total earnings            563.216***         809.749       724.562***        646.673***      341.710*     559.497***
                          (124.839)          (772.746)      (217.737)        (219.883)       (196.143)     (200.076)

 Controls                    Yes               Yes            Yes             Yes               Yes         Yes
 Observations                3178              87            604              753               833         901
                                                Panel 3: By Age Group
                                                                                OLS
                                                                        Age group of the girl

                              All                             16-19            20-24            25-27


 Total earnings            563.216***                        730.499***      578.846***      461.749***
                           (124.839)                        (162.166)        (178.138)      (158.549)
 Controls                    Yes                               Yes              Yes             Yes
  Observations              3178                              478            1742            958
Controls used are household size, sex of household head, asset based quintiles (excluded for “by wealth” regressions),
community, age of adolescent, educational status of adolescent (excluded for “by education” regressions), marital status of
respondent, parental status of respondent (indicator variables for adolescent having a child and being pregnant), orphan status
of respondent (one parent or both parents being deceased), and month of interview at midline; all controls are from baseline
data except month of midline interview.




                                                                                                                                     38
Table 5. Impact of EPAG on Savings, Loans, and Transfers

                                          Baseline
                                           Mean                  OLS                 Controls            Observations
                                            0.339                0.471***
Any savings                                [0.474]              (0.033)                 Yes                  3202

Total amount of savings (LD),              1420.4               2490.1***
unconditional on any saving               [4613.2]              (314.3)                 Yes                  3202

                                            0.034                0.031***
Any loans to be repaid                                                                  Yes                  3184
                                           [0.181]              (0.008)


                                            0.016                0.029***
Any formal loans                                                                        Yes                  3184
                                           [0.126]              (0.008)


                                            0.097               -0.002
Any loans to others                        [0.296]              (0.015)                 Yes                  3181

Note: “Formal” loans are those from banks, credit groups, susu, or money lenders. These are as opposed to “informal loans”
from parents, friends, relatives, or business partners. Controls used are household size, sex of household head, asset based
quintiles, community, age of adolescent, educational status of adolescent, marital status of respondent, parental status of
respondent (indicator variables for adolescent having a child and being pregnant), orphan status of respondent (one parent or
both parents being deceased), and month of interview at midline; all controls are from baseline data except month of midline
                               th                                                th
interview. Savings over the 99 percentile (40,000 LD) were top-coded to the 99 percentile and a control indicator variable (1
for the ones top-coded and 0 otherwise) was added to the regressions.




                                                                                                                                39
Table 6a. Impact of EPAG on Empowerment and Well-being
                                                                Baseline-                             Controls      Observations
                                                                                      OLS
 VARIABLES                                                       Mean
 Do you most of the time have money of your own for the
                                                                  0.793            0.068***
 basic uses that you alone can decide how to use, without                                               Yes             3201
                                                                 [0.406]           (0.019)
 having to ask for permission?

 (Among those with IGA) Do you control the money you               0.804            0.081***
                                                                                                        Yes             1509
 earn from this IGA? (Share reporting “Much” or                   [0.397]           (0.023)
 “completely”)
 Worries: Proportion reporting “Most/all of the time in the past year” to the following statements:
                                                                   0.381            -0.054**
 Worry about job/ income                                                                                Yes             3178
                                                                  [0.486]            (0.023)
 Worry that household will not have enough money for               0.305             -0.150***
                                                                                                        Yes             3142
 basics                                                           [0.460]            (0.025)
 (If married/cohabiting): Worry that you and your                  0.152             -0.056**
                                                                                                        Yes             1012
 husband/partner might split                                      [0.359]            (0.023)
 (If never married): Worry that you will not find a suitable       0.267             -0.041
                                                                                                        Yes             1776
 husband                                                          [0.443]            (0.028)
 Satisfaction
                                                                  0.558              0.026**
 Satisfaction score (0-1)                                                                               Yes             3194
                                                                 [0.235]             (0.012)
                                                                  0.307               0.140***
 Satisfied with job/business, if applicable                                                             Yes
                                                                 [0.462]             (0.024)                            2474
                                                                  0.347               0.045*
 Satisfied with income, if applicable                                                                   Yes
                                                                 [0.476]             (0.023)                            2945
                                                                  0.424               0.028
 Satisfied with life overall                                                                            Yes
                                                                 [0.494]             (0.028)                            3186
                                                                  0.221             -0.018
 Satisfied with education level                                                                         Yes
                                                                 [0.415]             (0.019)                            3190
                                                                  0.904              -0.001
 Satisfied with relationship with family                                                                Yes
                                                                 [0.295]             (0.012)                            3189
 Satisfied with relationship with husband, partner, or            0.793               0.034
 boyfriend, if applicable                                        [0.405]             (0.021)            Yes             3003
                                                                  0.640               0.017
 Satisfied with house                                                                                   Yes
                                                                 [0.480]             (0.019)                            3186
                                                                  0.736               0.012
 Satisfied with the community                                                                           Yes
                                                                 [0.441]             (0.017)                            3189
 Non-cognitive skills
                                                                  0.652             -0.008
 Self-regulation score (0-1)                                                                            Yes             3195
                                                                 [0.171]            (0.009)
                                                                    0.504             0.086***
 Entrepreneurial ability score (0-1)                                                                     Yes              2194
                                                                   [0.324]           (0.024)
For all regressions, controls include household size, sex of household head, asset-based quintiles, community, age of
respondent, educational status of respondent, marital status of respondent, parental status of respondent (indicator variables
for having a child and being pregnant), orphan status of respondent (one parent or both parents being deceased), and month of
interview at midline; all controls are from baseline data except month of midline interview.
Note: For the Worries regressions, the dependent variable is equal to 1 if the respondent is “Most/all of the time” and 0 for “at
least monthly,” “a few times” or “never.” For the Satisfaction regressions, the dependent variable is equal to 1 if the
respondent is “very satisfied” or “somewhat satisfied” as opposed to “not very satisfied” and “not at all satisfied”. The
Satisfaction score is equal to the proportion of items the respondents reported being “very satisfied” or “somewhat satisfied”
on, conditional on having responded to at least 4 items.
Note: The Self-Regulation Score is based on an 11-item scale adapted from an adolescent self-regulation scale developed in the
US. The responses were first converted to indicator variables; strongly agree and somewhat agree are grouped together as are
strongly disagree and somewhat disagree. Negative statements are reversed-scored so that higher scores indicate higher levels



                                                                                                                                    40
of self-regulation. The score is equal to the proportion of responses that reflect positive behaviors or attitudes, of the items
that the respondent reported on, conditional on having responded to at least 6 items. Alternate methods of aggregating
responses into a single score yielded similar results.

Note: The Entrepreneurial Ability score is based on a set of questions that asked the respondent to rate how well she had been
able to perform the following activities in the past year: (i) find information about job opportunities in your area, run her own
business, (ii) save in order to invest in future business opportunities, (iii) manage business finances effectively, (iv) bargain with
a supplier to obtain good prices when purchasing, (v) collect money from someone who owed her for purchases who are not
repaying on time. The response categories were: very well, somewhat well, not very well, and not at all well. Please note that
“don’t know” and “no response” were also allowable responses, but were dropped from the analysis, leading to a low sample
size for this measure. The score is based on the proportion of items to which the respondent answered “very well” and
“somewhat well” as opposed to “not very well” and “not at all well,” conditional on having responded to at least three of the
activities. Alternate methods of aggregating responses into a single score yielded similar results.

Table 6b. Alternate Measures of Self-Confidence
                                                                              Ordered
                                                              Mean            Probit            Controls            Observations
                                                              3.368            0.419***
 Attitude score (1-4)                                                                              Yes                   1595
                                                             [0.615]          (0.047)
 I feel more able to work well with people now                3.566            0.344***
                                                                                                   Yes                   1589
 than a year ago                                             [0.794]           (0.056)
 I feel more comfortable with who I am now                    3.578            0.469***            Yes                   1593
 than a year ago                                             [0.761]           (0.064)
 I feel more in control of my life now than a year            3.525            0.347***            Yes                   1593
 ago                                                         [0.820]           (0.054)
 I feel more able to call upon my friends for                 3.089            0.236***
                                                                                                   Yes                   1590
 support than I was a year ago                               [1.062]           (0.038)
 I am more able to help my friends now than I                 3.283            0.386***
                                                                                                   Yes                   1592
 was a year ago                                              [0.950]           (0.050)
 I am more comfortable in situations now with                 3.446            0.356***
                                                                                                   Yes                   1588
 people I do not know than I was a year ago                  [0.824]           (0.044)
                                                              3.098            0.270***
 I am more outgoing now than I was a year ago                                                      Yes                   1573
                                                             [1.097]           (0.039)

                                                             5.610             0.308***
 Confidence score (1-6)                                                                            Yes                   1599
                                                            [0.669]            (0.048)
                                                             5.502             0.379***
 Confidence in business/job abilities                                                              Yes                   1420
                                                            [1.187]            (0.050)
 Confidence in respondent’s ability to present               5.605             0.256***            Yes                   1578
 yourselves positively in front of others                   [0 .839]           (0.053)
  [If single/unmarried] Ability to rely on self for         5.467            0.351***          Yes                  870
  money rather than rely on boyfriends                     [1.125]           (0.086)
  [If not currently enrolled in school] Ability to
                                                            5.655             0.013
  return to school should respondent decide to                                                 Yes                 1088
                                                           [0 .845]           (0.058)
  do so
  Anticipation that respondent will be in a                 5.665             0.130
                                                                                               Yes                 1572
  rewarding job if not now in the near future              [0 .792]           (0.051)
  Anticipation that respondent will be in a                 5.735              0.300***
                                                                                               Yes                 1441
  position to be able to provide for own children          [0 .633]           (0.056)
Note: These alternate measures of self-confidence were introduced at midline. The first batch of questions on “attitude”
measures changes in attitudes over the past year and responses are coded as follows: 1=strongly disagree, 2=somewhat
disagree, 3=somewhat agree, 4=strongly agree The second batch of questions on “confidence” measures the confidence level
of the respondent in comparison to a year before the survey and is coded as: 1=much lower, 2=somewhat lower, 3=not much
lower, 4=not much higher, 5=somewhat higher, 6=much higher. The table reports the results of regressions using the midline
cross-sectional data. The overall scores are the average response across all categories on which the respondent reported.



                                                                                                                                         41
Controls include household size, sex of household head, asset-based quintiles, community, age of respondent, educational
status of respondent, marital status of respondent, parental status of respondent (indicator variables for having a child and
being pregnant), orphan status of respondent (one parent or both parents being deceased), and month of interview at midline;
all controls are from baseline data except month of midline interview.

Table 7. Impact of EPAG on Fertility and Sexual Behaviors
                                        Panel 1: Fertility
                             Baseline mean              OLS              Controls      Observations
                                    0.696             -0.045***
Any living children                [0.460]            (0.012)               Yes            3196

Any living children x Not                           -0.023**
pregnant at baseline                                 (0.011)                Yes            3196
Number of children                 1.549             -0.024
(conditional)                     [0.798]            (0.026)                Yes            2353
Number of children                 1.079            -0.071**
(unconditional)                   [0.975]            (0.028)                Yes            3188

Are you pregnant right             0.076              0.057***
now?                              [0.264]            (0.012)                Yes            3169

Pregnant x Not                                       0.019*
pregnant at baseline                                 (0.010)                Yes            3169

Desired number of                  3.481            -0.038
children                          [1.203]            (0.050)                Yes            3182

                                     Panel 2: Sexual Behavior
                               Baseline mean           OLS              Controls       Observations

Number of regular                  1.003              0.016
partners                           [0.378]           (0.023)               Yes             3106

Number of casual                   0.541              0.065
partners                          [1.080]            (0.050)               Yes             3094

Ever used condom in                0.710             -0.016
your life?                        [0.454]            (0.019)               Yes             3150

If yes, did you use a
condom last time you
had sex with your                  0.553              0.024
regular partner?                  [0.497]            (0.026)               Yes             2150

Notes: Controls used are household size, sex of household head, asset based quintiles, community,
age of adolescent, educational status of adolescent, marital status of respondent, orphan status of
respondent (one parent or both parents being deceased), and month of interview at midline;
all controls are from baseline data except month of midline interview.




                                                                                                                                42
Table 8. Impact of EPAG on Household Characteristics
                                                   Panel A: Household Size
                                  Baseline mean               OLS                     Controls                Observations
Number of household                    4.738                 -0.049
members                               [2.603]               (0.088)                      Yes                       3182
                                               Panel B: Household Food Security
Egg included at least once in          0.314                 0.034
                                                                                         Yes                       3154
household meal each week              [0.464]               (0.020)
Fish included at least once in         0.900                 0.038**
                                                                                         Yes                       3165
household meal each week              [0.301]               (0.015)
Meat/chicken included at
                                       0.840                0.039***
least once in household meal                                                             Yes                       3161
                                      [0.367]               (0.013)
each week
Dairy products included at
least once in household meal           0.546                 0.021                       Yes                       3141
each week                             [0.498]               (0.026)
Worry that household
member would not have                  0.619                -0.126***                    Yes                       3171
enough food                           [0.486]               (0.021)
A household member went to             0.306               -0.103***
                                                                                         Yes                       3170
bed hungry in past 2 weeks            [0.461]               (0.021)
Notes: Controls include household size, sex of household head, asset-based quintiles all at baseline, and month of interview at
midline. Standard errors clustered by classroom.




                                                                                                                                  43
Table 9. Gender Norms and Attitudes of Household Head

                                                     Baseline mean           OLS            Controls        Observations
                                                        28.243             -0.076
 Suitable minimum age of marriage for men
                                                        (4.886)            (0.218)            Yes               3143
 Suitable minimum age of marriage for                   23.475             -0.161
 women                                                  (4.508)            (0.184)            Yes               3168
 Gender equality score (scale of 0-5, 5                  1.661              0.207***
 meaning highest degree of gender equality)             (1.247)            (0.070)            Yes               3202
                                                         0.645              0.027
 Thinks both should earn money for family
                                                        (0.479)            (0.022)            Yes               3176
 Thinks both should be responsible for                   0.121              0.062***
 washing, cooking, cleaning                             (0.326)            (0.021)            Yes               3171
 Thinks both should be responsible for                   0.216              0.031
 fetching water                                         (0.412)            (0.027)            Yes               3168
 Thinks both should be responsible for feeding           0.164              0.043**
 and bathing children                                   (0.370)            (0.018)            Yes               3172
                                                         0.523              0.030
 Thinks both should help children with studies
                                                        (0.500)            (0.019)            Yes               3141
 Parent or caregiver approves of an unmarried
 woman engaging in the following:
                                                            0.313          -0.034
 Spending time with friends at a friend's house
                                                           (0.464)         (0.025)            Yes                 2531
  Spending time with friends at a video club,               0.229          -0.040*
  community center, night club or bar                      (0.420)         (0.020)            Yes                 2517
                                                            0.825           0.011
  Continuing or re-entering formal schooling
                                                           (0.380)         (0.027)            Yes                 2515
                                                            0.847           0.024
  Vocational training course
                                                           (0.360)         (0.022)            Yes                 2524
  Unpaid internship (job practice to get                    0.783           0.042
  experience, but without pay)                             (0.412)         (0.028)            Yes                 2521
                                                            0.823           0.004
  Self-employment
                                                           (0.382)         (0.017)            Yes                 2520
  Dating a man of roughly the same age, that is,            0.604          -0.022
  no more than five years older                            (0.489)         (0.033)            Yes                 2430
                                                            0.306          -0.042*
  Dating a man at least 10 years older
                                                           (0.461)         (0.024)            Yes                 2409
                                                            0.763           0.013
  Wage employment
                                                           (0.426)         (0.026)            Yes                 2455
Notes: Lower panel on parent or caregiver approval is conditional on having at least one unmarried woman over age 15
in the household. Controls include household size, sex of household head, asset-based quintiles all at baseline, and month
of interview at midline. Standard errors clustered by classroom.



                                                                                                                             44
Table 10. Correlates of Survey Attrition
Dependent Variable: In Panel (1=yes, 0= no)
                              (1)              (2)              (3)        (4)         (5)         (6)
                              OLS              OLS              OLS       Probit       OLS        Probit
Treated                      0.047*            0.084**         0.081**     0.087**    -0.039      -0.035
                            (0.025)           (0.011)         (0.013)     (0.014)     (0.067)     (0.066)

Aged 16 to 19                                 -0.082***       -0.069***   -0.074***   -0.082***   -0.082***
                                              (0.023)         (0.023)     (0.026)     (0.011)     (0.015)

Aged 20 to 24                                 -0.018          -0.017      -0.018      -0.017      -0.018
                                              (0.012)         (0.013)     (0.015)     (0.014)     (0.016)

In School                                      0.058***        0.054***    0.054***    0.070***    0.062***
                                              (0.017)         (0.018)     (0.015)     (0.002)     (0.004)

Completed at least                            -0.019          -0.011      -0.013      -0.057***   -0.049***
some primary school                           (0.048)         (0.050)     (0.048)     (0.007)     (0.007)

Married or                                     0.006           0.010       0.008      -0.007***   -0.007***
cohabitating                                  (0.012)         (0.011)     (0.012)     (0.002)     (0.002)

Has Child                                     -0.017          -0.018      -0.015      -0.051***   -0.045***
                                              (0.028)         (0.027)     (0.025)     (0.002)     (0.002)

Is Pregnant                                    0.033           0.028       0.024       0.016**     0.013**
                                              (0.023)         (0.022)     (0.021)     (0.006)     (0.005)

Any IGA                                       -0.046***       -0.041**    -0.039**    -0.052***   -0.046***
                                              (0.017)         (0.018)     (0.017)     (0.003)     (0.004)

Program Dropout                               -0.203***       -0.198***   -0.219***   -0.197***   -0.215***
                                              (0.070)         (0.068)     (0.074)     (0.069)     (0.076)

Treated x Age 16-19                                                                    0.021       0.012
                                                                                      (0.031)     (0.027)

Treated x In School                                                                   -0.028      -0.019
                                                                                      (0.023)     (0.029)

Treated x Completed                                                                    0.068       0.068
at least some primary                                                                 (0.060)     (0.065)
school
Treated x Married or                                                                   0.028       0.026
Cohabitating                                                                          (0.019)     (0.020)

Treated x Has Child                                                                   0.055*      0.054*
                                                                                      (0.030)     (0.029)

Treated x Pregnant                                                                     0.018       0.018
                                                                                      (0.050)     (0.050)

Treated x Any IGA                                                                      0.018       0.012
                                                                                      (0.023)     (0.024)
Observations                 1895              1895            1895        1895        1895        1895
Community Dummies             No                No              Yes         Yes         Yes         Yes

Note: Probit models in columns (4) and (6) report marginal effects.




                                                               45
Table 11. Estimates of key outcome variables, adjusted for survey attrition
                                                  Weighted ITT           ITT Estimates
                                               Estimates Adjusted       Unadjusted for
                                                   for Attrition            Attrition     Observations
 Any IGA                                             0.183***               0.181***         3200
                                                     (0.026)                (0.026)
 Weekly income (LD, unconditional)                  551.401***            563.216***         3178
                                                   (126.302)             (124.839)
 Any savings                                          0.460***               0.471***        3202
                                                     (0.036)                 (0.033)
 Total amount of savings (LD)                    2,411.490***            2,490.057***        3202
                                                  (317.827)               (314.278)
 Any loans                                            0.033***                0.031***       3184
                                                     (0.008)                 (0.008)
 Any money of your own                               0.070***                 0.068***       3201
                                                     (0.020)                  (0.019)
 Entrepreneurial Ability score                        0.089***                0.086***       2194
                                                     (0.024)                  (0.024)
 Satisfaction score                                   0.023*                   0.026**       3194
                                                     (0.012)                  (0.012)
 Satisfaction with job / business                     0.131***                0.140***       2474
                                                      (0.024)                 (0.024)
 Any living children                                  -0.033**                -0.045***      3196
                                                     (0.013)                  (0.012)
 Are you pregnant right now?                          0.056***                 0.057***      3169
                                                     (0.012)                  (0.012)
 Eggs included at least once                           0.032                   0.034         3154
  in household meal each week                         (0.020)                 (0.020)
 Fish included at least once                          0.037**                  0.038**       3165
  in household meal each week                        (0.015)                  (0.015)
 Meat / chicken included at least once                0.038***                 0.039***      3161
  in household meal each week                        (0.013)                  (0.013)
 Dairy products included at least once                 0.024                   0.021         3141
  in household meal each week                        (0.026)                  (0.026)
 Worry household will not have                       -0.128***                -0.126***      3171
  enough food                                        (0.021)                  (0.021)
 A household member went to bed hungry               -0.107***               -0.103***       3170
  in past two weeks                                  (0.022)                  (0.021)




                                                              46