WPS8561


Policy Research Working Paper                        8561




                    Creating New Positions?
             Direct and Indirect Effects of a Subsidized
                      Apprenticeship Program

                                    Bruno Crépon
                                   Patrick Premand




Social Protection and Jobs Global Practice
August 2018
Policy Research Working Paper 8561


  Abstract
 Evaluations of employment programs usually focus on                                workers with program participants. The share of youths in
 direct impacts on participants. Yet employment programs                            apprenticeships increased by 52.8 percentage points. This
 can have a range of indirect effects that are rarely quantified.                   estimate accounts for a significant windfall effect: 26 per-
 This paper analyzes the impact of a subsidized apprentice-                         cent of the formal apprentices who were placed substituted
 ship program offering dual on-the-job and theoretical                              out of traditional apprenticeships. The inflow of appren-
 training in Côte d’Ivoire. The experiment simultaneously                           tices into firms increased significantly, but also induced
 randomized whether apprenticeship positions opened by                              substitution effects, as firms hired 0.23 fewer traditional
 firms were filled by the program, and whether interested                           apprentices per formal apprentice placed. Overall, the net
 youths were assigned to a formal apprenticeship. This design                       number of apprenticeship positions created was between 51
 allows for estimating direct impacts on youths and indirect                        and 74 percent of the number of formal apprentices placed.
 impacts on firms selected to host apprentices. The analy-                          In the short term, impacts on earnings were not significant
 sis identifies whether individuals forgo other employment                          for youths, but firms benefited from an increase in the net
 or training opportunities, and whether firms replace other                         value of work provided by apprentices.




 This paper is a product of the Social Protection and Jobs Global Practice. It is part of a larger effort by the World Bank to
 provide open access to its research and make a contribution to development policy discussions around the world. Policy
 Research Working Papers are also posted on the Web at http://www.worldbank.org/research. The authors may be contacted
 at Bruno.Crepon@ensae.fr and ppremand@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
 Creating New Positions? Direct and Indirect Effects of
         a Subsidized Apprenticeship Program∗

             Bruno Crépon (Crest) and Patrick Premand (World Bank)




JEL classification : D22, J23, J24, O12, C93.
Keywords : Employment, Apprenticeship, Wage Subsidy, Training, Direct and Indirect Ef-
fects, Equilibrium Effects, Micro and Small Enterprises, Field experiment, Africa.


    ∗
      This paper is the result of a collaboration with the government of Côte d’Ivoire. We are particularly
thankful to Adama Bamba, Hermann Toualy, Ismahel Abdoul Barry and Cesar Gbeugré Toassa at BCPE;
Traoré Bamoudien and Ouattara Abdould Kadher at AGEFOP; as well as Hamoud Wedoud Abdel Kamil
at the World Bank. Funding from the Skills, DIME i2i and Jobs Umbrella Trust Funds at the World Bank
are gratefully acknowledged. We thank Sondo Eloi Somtinda for outstanding field coordination throughout
the study, Joël Farronato for excellent baseline survey coordination, as well as Nicolo Tomaselli, Henriette
Hanicotte and Dimanche Allo for leading follow-up data collection. Marina Tolchinsky, Matthew Olckers and
Fatine Guedira provided excellent contributions to data analysis. Johanne Buba, Deon Filmer, Leonardo
Iacovone and seminar participants in Cairo, ETH Zürich, Mannheim, Oxford and the World Bank provided
useful comments and suggestions. All errors and omissions are our own. The findings, interpretations, and
conclusions of the paper are those of the authors and do not necessarily represent the views of the World
Bank or the government of Côte d’Ivoire.
1       Introduction
Improving employment opportunities is one of the main policy challenges in developing coun-
tries. Governments are implementing a range of employment programs or skills development
initiatives, including public works, wage subsidies, trainings or apprenticeships. Policy mak-
ers often make strong claims about the number of jobs created, the number of people trained,
or the number of individuals placed through these programs. Employment programs have
been widely evaluated (see Card et al. (2018) for a meta analysis), but evaluations usu-
ally only focus on impacts on participants. At the same time, it has been emphasized that
employment programs can induce a range of indirect effects (Calmfors, 1994; Abbring and
Heckman, 2007). While the cost-effectiveness of employment programs largely depends on
the magnitude of indirect effects, they are rarely quantified.
    In this paper, we report the results of the randomized experiment of a subsidized ap-
prenticeship program offering dual on-the-job and theoretical training in Côte d’Ivoire. The
experiment simultaneously randomized whether apprenticeship positions opened by firms
were filled by the program, and whether interested youths were assigned to a formal ap-
prenticeship. This is a design tailored to estimate direct impacts and windfall effects among
youths, as well as indirect impacts and substitution effects on firms selected to host appren-
tices. On the one hand, participating individuals may forgo other human capital investments
or employment opportunities. Windfall effects arise when some program beneficiaries would
have taken a similar position absent any intervention. Such effects have been documented
in studies measuring direct impacts on program participants. On the other hand, when
programs seek to place individuals in firms, the total number of positions created is likely
smaller than the number of individuals placed. There can be substitution effects if firms
displace other workers with subsidized individuals. These substitution effects at the firm
level can rarely be estimated precisely. The innovation of this paper is to simultaneously
measure windfall effects among youths and substitution effects in firms. By considering both
sides of the labor market, namely labor supply and earnings from individuals, as well as
labor demand and profits from firms, we can estimate indirect effects and provide a more
comprehensive assessment of program performance than evaluations focusing on individuals
only.
    The program we study offered a subsidy of 30,000 FCFA per month (approximately USD
                                              1
54, or half the formal minimum wage), paid directly to apprentices. It included dual train-
ing, with practical on-the-job learning complemented by mentoring and theoretical courses.
The apprenticeships lasted 12 or 24 months, depending on occupations. We implemented a
randomization procedure simultaneously on both sides of the market, so that the study was
designed to measure both windfall and substitution effects, respectively among youths and
firms. Specifically, a sub-set of firms was randomly chosen to have pre-identified appren-
ticeship positions offered to program youths. This allows measuring how the inflow of new
workers is transmitted at the firm level. We can identify workforce response and substitution
effects on firms’ hiring of traditional apprentices in the private market. At the same time, we
randomly assigned a subset of youths interested in apprenticeship positions to the program.
We can then measure impacts on participation in apprenticeships, including windfall effects
driven by youths exiting traditional apprenticeships to enter formal apprenticeships. We can
also measure forgone earnings and employment for youths.
   Results show that the program induces a range of indirect effects. The share of youths
in apprenticeship increases by 52.8 percentage points. However, traditional apprenticeships
are common in Côte d’Ivoire, and there are substantial windfall effects: 26 percent of formal
apprentices placed by the program substitute out of traditional apprenticeships. The pro-
gram also induces a small substitution out of formal schooling and technical and vocational
training. On the side of firms, the program leads to an inflow of 1.4 formal apprentices into
firms, but only 1.1 new apprentices in total when also counting traditional apprentices. The
experiment thus documents substitution effects in firms: for each formal apprentice placed
in a firm, 0.23 traditional apprentices are displaced.
   Overall, we show that given the windfall effect among youths (0.26) and substitution effect
among firms (0.23), the net number of new apprenticeship positions created ranges from 1-
0.26-0.23=0.51 to 1-max(0.26,0.23)=0.74 for each youth placed in formal apprenticeship.
   The short-term cost-effectiveness of the apprenticeship scheme likewise depends on its
impact on earnings for both youths and firms. We find no average short-term impacts on
youths’ earnings. However, a large reorganization of activities is observed. Youths forgo
employment opportunities: the share of individuals engaged in wage employment or self-
employment decreases by 13.5 and 12.9 percentage points, respectively. As a result, labor
earnings decrease by 25.1 percent. Apprenticeship earnings paid by employers only com-

                                              2
pensate for approximately 25.3 percent of youths’ foregone labor earnings, highlighting the
large opportunity cost associated with participation in apprenticeship. However, the pro-
gram subsidy fully compensates youths for losses in average labor earnings. At the same
time, results suggest substantial indirect program impacts on firms. The additional inflow
of formal apprentices into firms leads to an increase in labor inputs. Formal apprentices are
found to be less assiduous and work fewer hours than traditional apprentices, but the aggre-
gated value of the work provided by apprentices in firms increases. Additionally, some of the
subsidies paid to youths are indirectly transferred to firms, as employers are found to reduce
their own payments to formal apprentices. Overall, results show a positive significant impact
on marginal profits at the firm level, as measured by the difference between the aggregated
value of the work provided by apprentices and the compensation they receive from firms.
      Our paper contributes to an active strand of the literature on the identification of in-
direct and equilibrium effects in program evaluation. In the case of employment programs,
this question was first raised by Calmfors (1994) and is further discussed by Abbring and
Heckman (2007). Some recent papers use large experiments or shocks to measure general
equilibrium effects. For instance, Breza and Kinnan (2018) use a micro-finance crisis in
India to identify the role of the aggregate demand channel, and Muralidharan et al. (2017)
use the phase-in of the modernization of the large-scale NREGA program to identify both
direct effects and equilibrium effects. Other papers use tailored experimental designs to iden-
tify specific types of indirect effects and mechanisms. Angelucci and De Giorgi (2009) and
Bandiera et al. (2017) examine the impact of cash transfers on the non eligible population
at the village level. Cunha et al. (forthcoming) compare impacts of in cash or in kind trans-
fers to identify price effects. Other designs use a double randomization approach proposed
by Moffitt (2001) to identify impacts on the eligible population that is not participating.
For example, Crépon et al. (2013) study displacement effects of a job search program and
Akram et al. (2017) identify a range of negative and positive spillover effects associated with
emigration. The identification of indirect effects requires powerful experiments and adapted
designs.1 While we do not measure general equilibrium effects, our paper adds to this liter-
ature by using a new design to identify specific types of indirect effects. Our randomization
procedure on the two sides of the market is well suited and powered to measure direct im-
  1
      See for example Baird et al. (forthcoming) in the case of saturation designs.


                                                       3
pacts on participants as well as a range of indirect effects on firms, including substitution
effects. In another recent paper, Alfonsi et al. (2017) analyze several active labor market
programs in which both youth and firms are randomly assigned to treatment and control, as
discussed further below.
       Our paper also contributes to the literature on wage subsidies. Most empirical studies fo-
cus on the impacts of wage subsidies on employment of individual workers. Positive impacts
have been documented in the short run, although results are mixed in the long run when
subsidies are phased out (for reviews, see Almeida et al. (2014); Card et al. (2018)).2 Berniell
and de la Mata (2017) is a recent example where a 12 months wage subsidy was found to have
large short and long run impacts on employment in Argentina. However, wage subsidies can
affect the labor market through a wide range of indirect effects including substitution effects
at the firm level and equilibrium wage effects (Layard et al., 1991; Calmfors, 1994). Lise
et al. (2004) use a general equilibrium model to study the impacts of a subsidized program,
and show that accounting for these effects can alter initial findings. There have been several
attempts to identify indirect effects. Blundell et al. (2004) compare the population of eligible
and non-eligible individuals across areas where the program has or has not been implemented
and find no equilibrium effect. Some other papers have attempted to identify more specif-
ically impacts on labor demand that could arise from substitution effects. Calmfors (1994)
show that, as long as subsidized and non-subsidized workers are perfect substitutes and firms
are on their labor demand curve, there is full eviction of non subsidized workers and zero
overall employment effect when marginal workers are non subsidized. Most results to date
come from non-experimental studies in high-income countries, with mixed results. While
some studies find no substitution effects (Kangasharju, 2007; Rotger and Arendt, 2011),
others find relatively large substitution effects (Dahlberg and Forslund (2005); Bishop and
Montgomery (1993)). The large variation in estimates is hard to interpret in the context
of non-experimental studies. Most of the existing literature also focuses on the impact of
providing wage subsidies in formal firms in developed countries, with less focus on micro and
small enterprises in the more informal contexts prevalent in developing countries. In an im-
portant contribution, De Mel et al. (2016) set-up a wage subsidy experiment to analyze firm
labor demand in Sri Lanka. They find that informal micro-enterprises increase employment
   2
    In their meta-analysis, Card et al. (2018) focus on subsidies paid to firms. In this paper, we consider a
case where subsidies are paid to employees.

                                                     4
while they receive the subsidies, an effect that does not last after the end of the subsidy.
Our paper adds to this literature by measuring simultaneously the impact on the demand
from firms and on program participants. This enables us to estimate more precisely the
net number of total positions created. Our double randomization design is powerful enough
to measure both windfall and substitution effects. We find large impacts on the number
of apprentices in firms but limited eviction effects, which suggests either that firms are off
their labor demand curve or that subsidized and non-subsidized apprentices are imperfect
substitutes.
       Our paper further contributes to the literature on training and apprenticeship programs.
Evidence on the effectiveness of training programs has been mixed overall (for reviews and
meta-analysis, see Blattman and Ralston (2015); Kluve et al. (2016); Bertrand et al. (2013);
McKenzie (2017)), though there have been recent studies with promising findings (Attanasio
et al., 2011, 2017; Alfonsi et al., 2017).3 Evidence on apprenticeship programs also remains
very thin despite apprenticeship being one of the most common sources of training in de-
veloping countries (Teal, 2016).4 On the side of youth, our paper focuses on the impact of
offering participation in the program on entry into apprenticeship. As such, it contributes
to the literature on the demand for training programs.5 Black et al. (2003) highlight that
low demand is a critical consideration in the analysis of training and employment programs.
The issue has been mostly analyzed in developed countries (Babcock et al., 2012). Here,
we complement the literature by considering the role of wage subsidies in the decision to
enter apprenticeship in a developing country context. Importantly, it has also been argued
that employers benefit from apprenticeships (e.g. Lerman (2014, 2017); Biggs et al. (1995)),
but empirical identification of benefits to firms has been challenging. Recent findings in
the literature tend to confirm this conjecture. Alfonsi et al. (2017) analyze vocational and
on-the-job training programs and show that firms absorb a significant part of the surplus.
Hardy and McCasland (2015) analyze whether a placement intervention improves the match-
   3
     Attanasio et al. (2017) show that a vocational training program combining on-the-job learning and
incentives for training centers to match trainees with firms had long-lasting effects in Colombia. Alfonsi
et al. (2017) show that vocational training delivered by high-quality training centers and on-the-job training
had positive long-lasting effects in Uganda, with vocational training outperforming on-the-job training.
   4
     A few studies estimate the returns to traditional apprenticeships (e.g. Frazer (2006); Monk et al. (2008)),
or short public apprenticeship schemes (Cho et al., 2013).
   5
     The objective of the paper is not to document long term impacts of apprenticeship on youths, which we
will study in future work based on additional data to be collected.


                                                       5
ing between potential apprentices and firms. By focusing on firms, they show that addressing
inefficiencies in the screening of apprentices leads to increases in employment and profits in
firms. Our analysis adds to these recent findings by also showing that the effects of a formal
apprenticeship program on firms’ hiring and revenues are substantial.
    Finally, our study relates to the theoretical framework developed by Acemoglu and Pis-
chke (1999), who show that a form of monopsony power is a condition for an apprenticeship
system to be viable. Firms need to be able to capture part of the surplus generated by
training in order for them to deliver training, as well as for workers to believe firms will
do so. This can lead to an under-provision of training, including through low participation.
Our paper presents several elements of evidence related to this framework. First, it illus-
trates the form the monopsony power can take. In traditional apprenticeships such as those
prevalent in Côte d’Ivoire, apprentices receive low remuneration during several years before
they are able to exit. Second, it documents large opportunity costs of training for potential
apprentices. Lastly, and importantly, our results show that a subsidized program increases
participation in apprenticeship by reducing the cost of training and providing a commitment
to a minimum level of training.
    The paper is structured as follows. Section 2 presents the intervention, experimental
design, data and estimation strategy. Section 3 documents windfall and substitution effects
on youths and firms. Section 4 documents impacts on youth earnings and profits in firms.
Section 5 discusses additional mechanisms. Section 6 concludes. Tables and figures are
presented in the annex.



2     Intervention, Experimental Design, Timeline and Data

2.1    Apprenticeships in Developing Economies

Traditional apprenticeships are one of the most prevalent types of training in the developing
world. In some African countries, recent data suggest that around 20 percent of youths
have been an apprentice, while less than 5 percent have attended technical and vocational
training (Filmer et al., 2014). The vast majority of apprentices are in traditional appren-
ticeships, which are one of the few sources of training accessible to the large number of
youths who exited the education system without completing primary or secondary school.

                                             6
Traditional apprenticeships are also one of the main sources of skill acquisition for informal
operators. Despite its prevalence, traditional apprenticeship remains poorly understood and
documented. Traditional apprenticeships are private arrangements between youths (or their
families) and private sector firms. Although their form can vary, traditional apprenticeships
share a range of characteristics (Walther, 2008). They take place in micro and small firms,
many of which operate in the informal sector. With the help of their family, youths are often
placed with master craftsmen identified through connections. A fee (in-kind or in cash) is
paid for the placement. Arrangements are rarely formalized through a contract. Youth learn
the trade through practical, on-the-job training by working in enterprises under the mentor-
ing of a master craftsman, either an experienced worker or the enterprise owner. Over time,
youths start being paid. Traditional apprenticeships can last many years, and often do not
lead to certification, although master craftsmen typically need to grant departure to mark
the completion of an apprenticeship. After completing traditional apprenticeships, youth
transition either as an employee in the host firm, as a wage worker in another firm, or in
self-employment. Most youth remain in the informal sector given the lack of formal wage
jobs for workers with limited education.
   While traditional apprenticeships have developed over time through a private, market-
based system with little public intervention, the optimality of the model has been ques-
tioned. The improvement of apprenticeship systems has become an important objective in
many countries around the world (Fazio et al., 2016; OECD/ILO, 2017), particularly in West
Africa (Walther, 2008; UNESCO, 2015). One of the common rationale for reforms is that
policies can facilitate access to apprenticeships, while at the same time improving training
quality and returns for youths. The set-up of dual apprenticeship schemes combining theoret-
ical and practical training is often considered, modeled after institutions from high-income
countries such as Germany or Switzerland. Reforms also seek to integrate and formalize
apprenticeships as part of the postprimary vocational and education training system. Yet
some have argued that reforms should instead consider improving quality in traditional ap-
prenticeships that are prevalent in low-income and lower middle-income countries, rather
than setting-up parallel public systems (Filmer et al., 2014). Others have made the case
that resources for policies to support apprenticeships may be better spent on creating more
high-productivity wage jobs (Teal, 2016).

                                              7
       The effectiveness of public interventions in the market for apprenticeships, and ultimately
the rationale for reforms, depends on successfully addressing market failures faced by youths
or firms. The creation of a formal apprenticeship system may induce youths to forgo other
human capital investments, such as traditional apprenticeships, or employment opportuni-
ties. Beyond effects on youths, the cost-effectiveness of public apprenticeship schemes in part
also depends on indirect effects, such as whether they benefit firms. For instance, it hinges
on the absence of negative substitution effects on private traditional apprentices, meaning
that there is unmet demand and absorptive capacity for apprentices in firms. However,
impacts on firms, such as substitution effects on the hiring of other apprentices, are rarely
quantified. The literature on the benefits and costs of apprenticeships is extremely thin, and
mostly concentrated in a few high-income countries such as Germany or Switzerland (for a
review, see Lerman (2014, 2017)). The lack of evidence in developing countries is partic-
ularly striking given the prevalence of apprenticeship as one of the most common types of
training, its mostly private nature, and the importance it is given in skills development or
youth employment strategies. Most existing studies focus on access and returns for youths,
with little evidence on indirect effects on firms, and little distinction between traditional and
formal apprenticeships.


2.2       The Côte d’Ivoire Formal Apprenticeship Program

After steady economic development through the mid-1990s, Côte d’Ivoire entered a period
of conflict, punctuated by a post-electoral crisis in 2010-11. Stability returned after the
institution of a new government in 2011, and economic performance has improved since
then. A range of public investments and programs were launched in 2011. They included
an emergency youth employment and skills development project (PEJEDEC), which had an
objective to improve access to temporary employment and skills development opportunities
for young men and women in Côte d’Ivoire.6 Among other interventions, the project included
an apprenticeship component.7
       The PEJEDEC apprenticeship component is overseen by the office coordinating employ-
   6
    PEJEDEC: Projet Emploi Jeune et Développement des Compétences (www.pejedec.org).
   7
    See Bertrand et al. (2017) for evidence on the cost-effectiveness of a public works program supported by
the same project.



                                                    8
ment programs (BCP-Emploi) 8 , with the national training agency as implementing agency
(AGEFOP).9 The program puts in place a formal apprenticeship scheme lasting 12 or 24
months, depending on occupations. The program initially aimed to cover 3,000 youths, and
is in the process of being expanded to approximately 10,000 youths. Low-skilled youths be-
tween 18 and 24 years old are placed in firms, where they receive on-the-job training under
the supervision of a master craftsman, either the enterprise owner or an experienced em-
ployee. Youths sign a contract with the implementation agency (AGEFOP) and are paid a
monthly subsidy of 30,000 FCFA (approximately USD 54, or half the formal minimum wage),
which is aimed to cover meals and transport costs. They receive an insurance coverage and
work equipment. The apprenticeship is dual, since on-the-job practical training is comple-
mented by theoretical training (approximately 180 hours per year) tailored to the needs of
apprentices and delivered by local training institutions. Apprentices are also mentored by
AGEFOP apprenticeship counselors, who regularly visit master craftsmen and apprentices,
and have the authority to suspend subsidies in case there are issues with youths’ participation
or performance. Formal apprenticeships end with an assessment of youths’ skills, leading to
certification. Firms are not compensated for taking on apprentices, though they do receive
a small toolkit of material to facilitate practical learning. Moreover, employers commit not
to request the payment of tuition fees at the start of the apprenticeship, in contrast to the
traditional apprenticeship model in West Africa (Walther, 2008).
      The average program cost is estimated at FCFA 1,135,030 (approximately USD 2,045)
per youth for a 24 months apprenticeship. This includes FCFA 720,000 (or USD 1,297) for
subsidies for youths, FCFA 330,000 (or USD 595) of other direct costs (toolkit, theoreti-
cal training, equipment,...), and FCFA 85,030 (approximately USD 153) for indirect costs
(selection, counseling, and so forth).


2.3      Enrollment Process and Experimental Design

One of the main objectives to embed an experiment in the Côte d’Ivoire formal apprenticeship
program was to measure simultaneously windfall effects among youth and indirect effects in
firms. This requires a specific design that randomly assigns both youth and firms to treatment
  8
      BCP-Emploi: Bureau de Coordinnation des Programmes d’Emploi.
  9
      AGEFOP: Agence de la Formation Professionnelle.


                                                 9
and control groups. In this section, we present the experimental protocol and discuss how it
addresses potential interference.
       The program was implemented in 7 urban areas in the interior of the country.10 In each
locality, AGEFOP worked with private sector organizations (such as chambers of commerce
or trade associations) to identify firms interested in hosting formal apprentices. For each
firm, the number of available apprenticeship positions was collected. AGEFOP staff then
systematically visited all firms to explain the program, to check each firm’s ability to train
apprentices and confirm the number of apprenticeship positions they could offer. A baseline
firm survey was implemented right after the collection of apprenticeship positions. Once all
positions were identified in a given locality, they were grouped and advertised by trade.11
Youths between 18 and 24 years old were then invited to visit a central location in each
locality to apply for apprenticeship positions in available trades.12 They filled an applica-
tion form and indicated the trade they were interested in. The program targeted low-skilled
youths, but did require an ability to read and write to ensure youths could participate in
the theoretical training, de facto implying that youths had at least a few years of school-
ing. Youths who met basic eligibility criteria were invited to an interview with AGEFOP
apprenticeship counselors. The interview sought to confirm youth motivation for doing an
apprenticeship as well as their choice of trade. A baseline survey was implemented among
all youth who successfully passed the interview.13 This registration process led to the same
number of youths eligible and motivated than open positions, in each locality and in each
  10
     These included Man (35% of youth in the sample), Daoukro (15%), Gagnoa (14%), Divo (12%), Bouaké
(12%), Adzopé (7%) and Mankono (5%). It was planned that Abidjan would also participate in the ex-
periment, but demand for apprenticeship positions among youth was limited and there was not enough
oversubscription, so that Abidjan had to be dropped from the sample. The program was launched between
July 2014 (Adzopé), August 2014 (Daoukro, Gagnoa, Man and Mankono), September 2014 (Divo) and Oc-
tober 2014 (Bouaké). A target number of youth to include in the program was set for each of these localities
based on the estimated number of available apprenticeship positions and other considerations.
  11
     Throughout the paper, we make a distinction between "sectors" and "trades". Sectors refer to the
activity of the firm and "trades" refer to jobs taught to youth. The two concepts are often the same, but
in some cases firms in a given sector are active in several trades. A good example is the garage sector,
which includes apprenticeship positions in several trades: coach builder, car mechanic, car electrician, and
car painter.
  12
     The most popular trades included car or motor mechanic (21% of positions), metalworker, boilermaker,
welder (14%), bricklayer, painter, plumber (11%), carpenter (9%), car electrician (9%), electrician (8%),
coach-builder (8%), repairman for fridges and freezers (7%).
  13
     Despite the efforts made to advertise the program, it was not possible to find enough youths interested in
some trades in some localities. In such cases, a rationing occurred at the firm level: the number of positions
to be filled was reduced proportionally, while ensuring as much as possible that firms would keep at least
one open position. In a few cases, this was not possible and some firms had to be randomly excluded, even
though they had been initially registered and surveyed.

                                                     10
trade.
       In each locality, following the interviews, a double-sided randomization protocol was
implemented. The procedure was the following: firms were paired according to the number
of positions they opened per trade, and within each pair a firm was assigned to treatment
and another to control. The reason for implementing this pairing procedure, instead of
a theoretically more appealing stratification by trade, is that some firms opened positions
in different trades (see footnote 11). Once the firm randomization was implemented, the
number of open positions in treatment firms was counted by trade. This gave the exact
number of youths to select. The second step of the randomization was then implemented,
randomly assigning youth to treatment by trade. We assigned the exact same number of
youth to treatment as the number of open positions to fill. As a result, the probability of
youth assignment to treatment is trade-specific. On the side of youth, the experiment is
thus stratified by locality (since the randomization procedure was implemented separately
in each locality) and by trade. Since assignment probabilities are strata-specific, we include
strata-specific weights in the specification used to estimate impacts on youth, as discussed
further below. Figure A1 in appendix A1 shows the distribution of this ratio by stratum,
showing as expected a strong concentration around 0.5.14
       Once apprenticeship positions to be filled were selected, and youths were selected in each
locality, AGEFOP counsellors matched selected youths to selected firms with open positions
in the same trade. The matching took place based on criteria such as distance between the
firm and youth residence. Once assigned to firms, youth passed a medical visit and were
invited to sign a contract and start their apprenticeship.15
       Across the 7 localities covered by the study, 731 firms offered apprenticeship positions and
1,832 young applicants were eligible and passed the motivation interview. Approximately
half the firms (361), were randomly selected to host program apprentices. 911 eligible and
motivated youths were assigned to the program and 921 to the control group. Most firms
  14
      As can be seen from the figure, there are a few cases in which the assignment ratio is either 0 or 1. The
appendix explains how we deal with these cases.
   15
      This is a demanding experimental protocol requiring a lot of specific actions and close coordination
with the implementation agency in a short period of time. We had a team of three highly skilled research
associates based in the field, as well as a full data collection team implementing baseline surveys. Once the
experimental protocol was implemented in a given locality, a detailed summary report was written to list
all the specific implementation aspects. For example, the report registered the initial number of positions
offered in each trade, the number of firms involved, as well as any rationing that occurred and the number
and identity of any firm randomly excluded from the experimental protocol.

                                                     11
offered several positions, and on average treatment firms were assigned 2.52 apprentices.
   Importantly, substitution effects at the firm level and windfall effects at the youth level
imply that labor market flows are affected by the intervention. Some youth not registered
in the experiment (as well as youth assigned to the control group) could potentially be
displaced and have to find a different position than the one they would have obtained absent
the program. Similarly, in presence of windfall effects, some treated youth hired in treatment
firms may have left open positions in firms where they would have been working absent the
program. These effects are both of primary interest of the study, but can also lead to
interference in the measurement of program impacts.
   We summarize here how our experimental protocol helps address the challenges of mea-
suring windfall and substitution effects. Appendix A1 provides a more detailed discussion.
First, youth and firms have to be statistically identical in the treatment and control groups.
Proper implementation of the randomization protocol ensures this. In addition, our exper-
imental design is implemented at the level of micro-markets, defined as a given trade in a
given locality. In this context, a second requirement is that youth and firms in the control
group have the same opportunities in their micro-market as what they would have had ab-
sent the program (this is described as assumption A(1) in appendix A1). The experimental
design seeks to achieve this property by maintaining a similar share of control youth and
control positions in each micro-market. A third important dimension is that windfall and
substitution effects can potentially change the flows of non-registered youth in the appren-
ticeship market, as well as the flows of youth towards non-registered firms. An additional
requirement is thus that the corresponding match probabilities are not affected by the ex-
periment (this is described as assumption A(2) in appendix A1). For this property to hold,
the shares of treated youths and firms in the apprentice market need to remain small. Using
information from a recent national employment survey and population census, we can show
that this share is estimated to be 5% (see discussion in appendix A1 and Table A2). Overall,
the fact that these three requirements are met mitigates risks of interference.


2.4    Data and Estimation Strategy

The program was rolled out sequentially, locality after locality. As explained in the previous
section, baseline data was collected in each locality as part of the enrollment process. Specif-

                                              12
ically, after the apprenticeship positions offered by firms were validated by program staff,
a comprehensive baseline survey was implemented in each firm with confirmed positions.
Separately, baseline data were collected among youth deemed eligible after they successfully
passed the motivation interview. Baseline data collection among firms and youth took place
in each locality before the randomization was performed. The baseline and enrollment phase
took place between July 2014 and October 2014. The selection of firms and youth took place
shortly after, and placements were mostly completed by January 2015.
       The follow-up survey for the impact evaluation took place between March 2016 and June
2016. It was collected on average 20 months after the start of the program.16 Since most
apprenticeships last 24 months, results based on the follow-up survey should be interpreted
as providing short-term impacts while apprentices are still in the program.17
       Substantial efforts were made to minimize attrition during the follow-up survey.18 As a
result, 1,661 youth were surveyed, implying an attrition rate of 9.3% (or 171 youth, 84 in
Treatment and 87 in Control). Similarly 674 firms were surveyed, leading to 7.8% attrition
(or 57 firms, 26 in Treatment and 31 in Control).19
       Tables A3 and A4 present baseline characteristics and balance checks for youth, respec-
tively firms. Both tables have the same structure, the left panel is devoted to the analysis
of baseline data (including on the last row the share of youth or firms with available data).
The right panel presents baseline characteristics of follow-up respondents and related balance
checks.
       Table A3 shows that youth interested in formal apprenticeships are 20.7 years old on
average, and mostly men (87 percent). They have some (but limited) education, as 63
percent have completed primary school and 17 percent lower secondary school. 45 percent
of applicants aspire to a wage job, and 54 percent to become self-employed. There are few
significant differences between the treatment and control group of youth, who are largely
  16
     Figure A2 documents precisely the timing of surveys as a function of the randomization date.
  17
      754 of the 914 youths in the treatment group (or 82%) were in trades where the apprenticeship contract
lasted 24 months.
  18
     An unfortunate IT issue with the online server used for electronic data collection led to the loss of
baseline data for 26% of youths (475 youth, 250 in Treatment and 225 in Control) and 5% of firms (37
firms, 18 in Treatment and 19 in Control). The problem was concentrated in two localities. The loss of
some baseline data limited availability of contact information to track youth (and firms) at follow-up. This
contributed to a lower response rate among youths in localities where IT issues occurred.
  19
     It is worth to note that part of this attrition is due to firm closure. We designed a specific data collection
instrument for registered employers whose firm had closed by the time of the follow-up survey. 12 cases were
identified. "True" attrition is limited to 6.2% (45 firms, 23 in Treatment and 22 in Control).

                                                       13
comparable and well-balanced at baseline. As can be seen from the table, the share of
available baseline data is not perfectly balanced (see footnote 18), but the response rate at
follow-up is well-balanced, which is what matters most since it is the basis for empirical
estimation.
   Table A4 highlights that most firms offering apprenticeship positions are informal micro
and small enterprises. 84 percent have no formal legal status and 68 percent do not keep
books. Firms have 6.3 permanent employees on average (counting the owner), of which 3.3
are apprentices. Traditional apprentices therefore constitute more than half the workforce
in these micro and small firms. Traditional apprentices are mostly hired through private
channels, 82 percent based on a request from their parents. About half of the apprentices
in firms at baseline pay fees to the master craftsmen. These traditional apprenticeships are
expected to last six years on average. Table A4 documents that the experiment led to good
baseline balance between the treatment and control firms: the few significant differences are
marginal and of small magnitude.
   The follow-up survey was collected by phone for youth in the treatment and control
groups. The most important variables for the analysis are youth activities at the moment of
the survey, as well as hours and earnings in those activities. The survey contains a detailed
employment module covering primary and secondary activities. We can thus describe the
portfolio of youth activities, distinguishing between occupations as apprentice (formal or
traditional), wage worker and self-employed. Other important variables relate to youth
human capital investments, including participation in apprenticeship (formal or traditional),
vocational training and schooling.
   The follow-up firm survey in the treatment and control groups took place in person. It
collected data on the characteristics of the firm, its workforce, as well as its sales and profits.
It also compiled a listing of all apprentices who entered or left the firm since the start of the
experiment (i.e. on or after the randomization date in each locality), and collected additional
information on each of these apprentices, both from enterprise owners and from apprentices
themselves. This employer-employee type of data enables us to accurately measure the flows
of apprentices in and out of firms, as well as their contribution to firm activity. For example,
we are able to compute the number of apprentices working in firms at the moment of the
survey, but also various interesting flows: the number of apprentices who entered firms since

                                               14
randomization, and among them those who left firms and those still in firms. We can measure
all these variables separately for formal and traditional apprentices. The survey also asks
about the number of days worked in the last seven days and the number of hours worked in
the last business day for each apprentice. In order to measure apprentices’ contribution to
firm activity, we asked employers about the amount he would have had to pay had he hired
a casual worker to perform the same tasks. This in turn allows us to compute a value for the
work performed by each apprentice. We also asked employers about the compensation paid
to each apprentice. These measures can be aggregated at the firm level.20 We also collected
several measures of sales and profits, addressing concerns about the measurement of these
variables raised by De Mel et al. (2009). Following the procedure recommended in De Mel
et al. (2009), we obtained direct measures of total profits and revenues. We then asked
the firm owner to recall all sales over the last month as well as related expenses. On that
basis, we collected another (updated) measure of total profits and revenues. Moreover, we
implemented near-systematic back-checks of key variables, including sales and profit. Thus,
for most firms in the sample, we have six measures of sales and profits.
       Given the double-sided randomization protocol, intent-to-treat (ITT) program impacts
on firms can be estimated by comparing outcomes between firms assigned to treatment
(i.e. where formal apprentices were assigned by the program to fill open positions), and
firms assigned to control (i.e. where open apprenticeship positions were not filled by the
program). The ITT analysis at the firm level is performed using OLS regressions with the
667 firm-level observations at follow-up:




(1)                          yi = a + bTi +       γv 1v +       δs 1s + ui
                                              v             s


       We compute robust standard errors. In this equation T is the assignment to treatment
variable, v stands for the locality and s for the sector. Sectors are a broader concept than
trade (see footnote 11).
       In parallel, intent-to-treat program impacts on youth can be estimated by comparing
  20
    We made a distinction between various forms of compensation. Employers usually provide meals and
cover expenses such as transportation and clothes. They also provide some money for the work done by
youth in order to "motivate" them. We measure each of these payments and aggregate them by youth and
by firm to get a total wage bill for apprentices.

                                                  15
outcomes between youth assigned to treatment (i.e. offered a formal apprenticeship position
in a treatment firm), and control youths. We have, however, to account for the fact that
youth were assigned to treatment and control with probabilities that were specific to each
trade in each locality, producing a set of corresponding strata St . We compute the empirical
assignment rate in each stratum and estimate inversely propensity weighted regressions.21 To
obtain accuracy gains from stratification, we run an inversely propensity weighted regression
with strata dummies on the 1,661 youth observations:




(2)                                  yi = a + bTi +         µSt 1St + ui
                                                       St


         We compute robust standard errors.22



3          Windfall and Substitution Effects for Youth and Firms

3.1         Youth Entry into Apprenticeship and Windfall Effects

Table 1 presents ITT estimates of human capital investments among youth, covering the
experiment period between randomization and the follow-up survey. We consider human
capital investments in the form of schooling and training. As part of training, we distin-
guish between technical and vocational training (TVET) and apprenticeship. Participation
in apprenticeship is further decomposed between traditional apprenticeships, i.e. private ar-
rangements that exist independently of the program, and formal apprenticeships of the type
promoted by the program.
         Table 1 clearly shows, as expected, that the program leads to a large increase in participa-
tion in formal apprenticeships: the share of youth in formal apprenticeship is 71.2 percentage
points larger in the treatment group than in the control group. A very low, but non-zero
participation in formal apprenticeships is observed in the control group. Overall, 75 per-
cent of youths in the treatment group participate in formal apprenticeships, which is highly
    21
    The empirical assignment rates are defined on the sample used in the regression.
    22
    There are 1,676 observations for which we were able to collect follow-up data. However, there are 15
youth observations for which the empirical assignment probability within their stratum is either 0 or 1, and
these are discarded from the youth regressions.


                                                    16
consistent with program take-up measures based on process evaluation or administrative
data.23
       Importantly, Table 1 documents substantial windfall effects: participation in formal ap-
prenticeship is in part explained by substitution out of other forms of human capital invest-
ments, such as youth in the treatment group substituting out of traditional apprenticeships.
22.5 percent of youth in the control group participate in traditional apprenticeships, a pro-
portion that is reduced by 18.5 percentage points in the treatment group. The windfall effect
is of substantial magnitude. It can also be expressed in relative terms, as the share of youths
exiting traditional apprenticeships (18.5 percentage points) relatively to the share of youths
entering formal apprenticeships (71.2 percentage points), or 26 percent. In other words,
for each formal apprentice placed, 0.26 youths substituted out of traditional apprenticeship.
Overall, the net program impact on the share of youth in any form of apprenticeship over
the course of the experiment is 52.8 percentage points.
       The program also induced a small substitution out of other forms of human capital
investments such as schooling or technical and vocational training (TVET). The participation
in TVET in the control group is quite low (7.2 percent), but the program has a relatively
large impact in reducing this proportion by 5.7 percentage points.24 Last, the program also
has a significant and negative impact on youths attending school. 20.5 percent of youth in
the control group report going to school, and this proportion is reduced by 5.7 percentage
points in the treatment group.
       Despite these indirect effects, a large proportion of youth makes no human capital invest-
ment of any form in the control group (51.4 percent). Offering participation in the program
sharply reduces this proportion by 36.3 percentage points, approximately half the program
take-up. In the end, only 12.3 percent of youth in the treatment group do not make any
form of human capital investment.
       So far, we have focused on impacts on entry into apprenticeship between the baseline and
  23
    Appendix A2 and Table A5 provide additional information on take-up measures from other data sources.
  24
    To build human capital investment variables, we use a question in the follow-up survey about partic-
ipation in apprenticeship and TVET, as well as a separate question about youth participation in public
apprenticeship programs. We classify as participation in formal apprenticeship those who are either in
apprenticeship or TVET and say they were registered in a public apprenticeship program. Although not
perfect, and subject to some recall bias, this is the classification that is most consistent with information
from the process evaluation survey, for the subsample of treated youth that can be matched in both surveys.
See Appendix A2 for a more detailed discussion.


                                                    17
follow-up surveys, which captures inflow into apprenticeships. We can also analyze impacts
on participation in apprenticeship at the time of the follow-up survey. In Table 1, the right
part of the lower panel presents ITT estimates of impacts on youth participation in appren-
ticeship at the time of the follow-up survey (i.e. approximately 20 months after the start
of the program). The data is obtained from a survey module on youth occupations, which
includes apprenticeship as an option. The information covers all types of apprenticeships,
including traditional and formal apprenticeships. The left part of the lower panel of Table 1
presents ITT estimates for the share of youth who started apprenticeship but dropped-out
by the time of the follow-up survey. Those estimates (labeled as d in the table) are simply
obtained as the difference between estimates of entry in apprenticeship since the start of the
experiment (labeled as e ) and estimates of youth still in apprenticeship at the moment of
the survey (labeled as c ).
       Drop-out is a common issue in many employment or training programs. It has been shown
to be important in apprenticeship as well (e.g. Cho et al. (2013)). We can compute drop-out
rates from Table 1. Drop-out is estimated as 0.222/0.712=31.2% in formal apprenticeships
and 0.060/0.185=32.5% in traditional apprenticeships.25 These results show that drop-out is
important, but also that it is not an issue specific to formal apprenticeships as the drop-out
rate in traditional apprenticeship is close.26 Although it is reassuring that drop-out is not
higher in formal apprenticeship, these results show that the program did not reduce the
prevalence of drop-out compared to traditional apprenticeships.
       Results on participation in apprenticeship at the time of the follow-up survey confirm
previous findings on increased participation but also significant windfall effects. The share
of youth in formal apprenticeship increases by 49 percentage points. However, the share of
youth in any type of apprenticeship increases by only 36.5 percentage points. The difference
between the two estimates, 12.5 percentage points, provides another estimate of the windfall
effect, namely the share of youth who substituted out of traditional apprenticeships to be in
formal apprenticeships at the time of the follow-up survey. Expressed relatively to the share
of formal apprentices placed, the windfall effect is again 26 percent (12.5/49), which is the
  25
      We could also obtain these estimates by running a regression of dropout status on entry into formal or
traditional apprenticeship using the treatment assignment variable as an instrument.
   26
      These numbers overestimate drop-out in the treatment group because a small share of formal appren-
ticeships only lasted 12 months and these contracts were over at the moment of the follow-up survey. See
footnote 17.

                                                    18
same as the estimate over the period between baseline and follow-up.
       These results show that some youth decided to forgo undertaking traditional apprentice-
ship to participate in formal apprenticeship, again pointing to substantial windfall effects.
Overall, 54.4 percent of youth in the treatment group are in apprenticeship at follow-up,
compared to 17.9 percent of youth in the control group. Most youth in the control group
are in traditional apprenticeship (16.1 percent out of 17.9 percent), while most youth in the
treatment group are in formal apprenticeship (50.8 percent out of 54.4 percent). The fact
that youth in the control group were able to find traditional apprenticeship positions by
themselves through the private market, and that some formal apprentices substituted out of
traditional positions, thus creates a windfall effect that mitigates the net program impacts
on the share of youth in apprenticeship.27


3.2       Intake of Apprentices and Substitution Effects in Firms

We now turn to the other core research question of whether the program induced indirect
employment effects among firms. The analysis is based on a survey module asking employers
to list all apprentices who have worked in the firm over the course of the experiment, including
those who left the firm since the randomization. We can thus measure flows of apprentices
in and out of firms between the start of the experiment and the follow-up survey. We
can also distinguish between formal and traditional apprentices. Table 2 (upper panel)
documents the impact of the program on the flow of apprentices into firms since the date of
the randomization. The program led to an increase in the total number of apprentices that
entered by 1.08 new apprentice per firm over the course of the experiment. Yet the total
number of youth who entered formal apprenticeship in these firms increased by 1.398. As
such, the program induces a significant substitution effect: fewer traditional apprentices were
hired as additional formal apprentices entered firms. Substitution effects were of moderate
magnitude, however, with (1.398-1.08)/1.398 = 0.227 informal apprentice substituted per
formal apprentice placed.
       The net increase in formal apprentices in treatment firms can be contrasted with the
number of apprenticeship positions offered by these firms. On average, firms offered 2.51
  27
    Additional results are presented in Table A7, which documents the time evolution of participation in
apprenticeship using information from a retrospective calendar of occupations covering the four quarters
before the follow-up survey.

                                                  19
apprenticeship positions.28 A substantial share of these positions were not filled, and the
placement ratio is relatively weak: only 1.398/2.51 = 55.5% of positions offered by firms were
effectively filled. These patterns are consistent with substantial drop-out among selected
youths. This also suggests that firms face challenges in attracting youths, even despite the
program subsidy.
       Figure 1 provides additional information about the impact of the program on the inflow of
apprentices into firms. The figure shows monthly inflows of apprentices in treated and control
firms by date (with zero being the randomization date). The figure makes a distinction
between inflows of formal apprentices and traditional apprentices in treated firms. The
figure clearly shows a spike of entry of formal apprentices in treated firms shortly after
randomization. It also shows that flows of traditional apprentices were weakly affected by
the program.
       Table 2 (intermediate panel) documents the impact of the program on exit of apprentices
out of firms. Exits from apprenticeship are substantial. Out of 1.08 apprentices who entered
firm due to the intervention, 0.467 apprentices had exited by the time of the follow-up survey.
This amounts to a 43.2% drop-out rate. As discussed in the previous section, drop-out is a
pervasive aspect of apprenticeship. While it is large, it is not higher in the treatment group
than in the control group.29
       Table 2 (lower panel) documents net program impacts on the number of apprentices
who entered since randomization and are still in firms at the time of the follow-up survey.
Results can simply be deduced as the difference between the top two panels. They show
that, 20 months after the launch of the program, there are 0.613 total apprentices per firm
that entered since the date of randomization, and 0.787 additional formal apprentices per
firm. Substitution effects can also be measured at the time of the follow-up survey. They
are again positive and significant, but remain of relatively moderate magnitude, with (0.787-
0.613)/0.787 = 0.221 traditional apprentice substituted per formal apprentice placed. This
estimate is very close to the one obtained looking at inflows between the baseline and follow-
up surveys.30
  28
     There are 911 youth assigned to 361 firms in the full treated sample (see section 2.3), resulting in an
average of 2.52 per firm. In the regression sample, 864 youths were assigned to 334 firms resulting in an
average of 2.51 per firm.
  29
     If we consider youth who entered firms within 6 months after randomization, we find drop-out rates of
43.3% in the treatment group and 47.1% in the control group.
  30
     As can be seen in the top two panels of Table 2, there is also some non-compliance with the experimental
                                                     20
    The follow-up firm survey also provides information about the total workforce in firms
(including apprentices and other types of employees). Results are presented in appendix
Table A8. While there is a significant impact on the inflow of youth that entered firms since
the beginning of the experiment, there is no significant impact on the overall number of
apprentices in firms at follow-up. The estimated impact is 0.464 with a standard error of
0.362. The impact on flows are not large enough to affect stocks significantly, which may
be due in part to large standard errors in the stock variables.31 Consistent with the lack of
significant impact on the total number of apprentices in firms, the program does not have
an impact on the overall workforce in firms. No significant impact is found on the number
of interns or occasional workers either.
    Table A8 shows that the program reduced the share of firms without any apprentice.
Some firms that opened apprenticeship positions had no apprentice. In the control group,
20.3% of firms have no apprentice at the time of the follow-up survey. This proportion is
significantly lower by 5.4 percentage points in the treatment group, but there are still 14.9
percent of firms without apprentice despite these firms having offered positions that were
selected to be filled by the program.
    In terms of measurement, the study shows the importance of carefully defining outcome
measures. If we had implemented a simple survey only asking about the number of employees
and apprentices in firms, we would have concluded that there was no significant impact on
employment. However, this stock measure would have missed important impacts of the
program. Indeed, the impact on stock is the combination of impact on inflows and outflows
(see footnote 31). In addition, the flows between the start of the experiment and follow-up
combine both entry and dropouts. With an employer-employee survey, data on the dates of
entry and exit for each apprentice enable us to build a more precise set of measures of both
entry and exit, providing a richer understanding of indirect impacts on firms.
protocol. The average number of formal apprentices who entered since the start of the experiment per control
group firm is 0.188. As can also be seen from the table, this imperfect compliance did not last long: most
(0.13) formal apprentices in control firms had left by the time of the follow-up survey. The non-compliance
was highly concentrated in two out of the seven localities in which the program was implemented.
  31
     For completeness, the impact of 0.464 on the stock of apprentice combines the previous impact of 0.613
on entry by follow-up with the impact on total number of apprentices who were in the firm before the
randomization and are still in the firm at the moment of the survey, -0.154 (with a standard error of 0.239).




                                                    21
3.3      Net Impact on Apprenticeship Positions Created

So far, we have discussed ITT impacts of offering youth to enter formal apprenticeships, and
on assigning formal apprentices to firms with open positions. Results show that there are
windfall effects on youth as well as substitution effects on firms. These effects imply that
the net number of positions created by the program is smaller than the number of formal
apprentices placed. In this section, we discuss more precisely what is the overall impact of
the intervention by discussing how both effects combine, and by quantifying the net impact
on the number of apprenticeship positions created.
      We first turn ITT estimates into LATE estimates that represent impacts per youth en-
tering formal apprenticeships. On the youth side, we estimate:


(3)                           ai = α + (1 − ω )fi +          µSt 1St + ui
                                                        St


where ai stands for having started an apprenticeship since the beginning of the experiment
and fi for starting a formal apprenticeship. We estimate this equation using the treatment
assignment variable as an instrument. In this equation, (1 − ω ) represents the proportion of
formal apprentices who entered apprenticeship due to the program. Thus, ω measures the
proportion of formal apprentices who would have started an apprenticeship anyway absent
the program. Said differently: for each youth starting a formal apprenticeship, there are ω
fewer youth starting a traditional apprenticeship.
      We also analyze the impact of entries into formal apprenticeships on the total number of
entries of apprentices in firms. We estimate:


(4)                       ei = a + (1 − σ )ef
                                            i +        γv 1v +       δs 1s + ui
                                                   v             s


where ei is the total number of youth entering apprenticeship in firms and ef
                                                                           i is the total

number of formal apprentices hired by firms. In this equation, (1 − σ ) measures the number
of youth entering the firm per formal apprentice placed. Thus, for each formal apprentice
entering the firm, there is σ less youth entering as a traditional apprentice. This equation is
estimated using the treatment assignment variable as an instrument.
      Table 3 presents the results. The fist two columns present the reduced form, which are

                                                  22
the ITT estimates presented above. The third column presents IV estimates for (1-ω ) and
(1-σ ), which are simply the ratio of the first two columns. The last column provides the
estimated substitution and windfall parameters σ and ω . As can be seen from the table, the
estimated value of the windfall parameter among youths is 0.259, with a standard error of
0.022. On the firm side, there are 0.773 youths entering firms per formal apprentice placed,
thus leading to an estimated substitution parameter of 0.227 with a standard error of 0.128.
   One important question is how to combine these estimates to derive the net number
of apprenticeship positions created by the program per apprentice placed. The windfall
effect ω can be interpreted as the proportion of youth who would have entered traditional
apprenticeships absent the program. Consistent with the fact that the size of the experiment
is small compared to the size of local apprenticeship markets (see Section 2.3 and Appendix
A1), traditional apprenticeship positions which youth did not enter are located in the group of
firms not registered in the experiment. Thus, we consider that, for each program youth, there
are ω additional positions to fill in the set of non-registered firms. Similarly, the substitution
effect σ is interpreted as the number of positions that firms would have filled with traditional
apprentices, per program youth placed, had the program not been implemented. Again
since the size of the experiment is small, those traditional apprentices would have been hired
from the set of non-registered youth. For each entry of one additional program youth in a
treated firm, there are thus σ additional non-registered youth in search for a position. Our
understanding of the flows created by the implementation of the experiment is thus quite
simple: there are σ additional positions to fill in the group of non-registered firms, and there
is ω additional youth in search of apprenticeship positions. Figure 2 summarizes those flows.
Part of the additional youths searching for traditional apprenticeships may match with firms
with open positions (me ). The total number of lost traditional apprenticeship position le due
to the implementation of the experiment is thus le = σ + ω − me for one additional formal
apprentice placed. Because we do not follow non-registered youth and firms, we are unable to
measure the number of matches me . However, it can easily be bounded: 0 < me < min(σ, ω ).
As a result, the number of lost apprenticeship positions can be bounded by


               σ + ω − min(σ, ω ) = max(σ, ω ) = 0.259 < le < σ + ω = 0.486


We directly derive the bounds for the net number of apprenticeship positions created per
                                              23
additional apprentice placed ne as


                     1 − (σ + ω ) = 0.514 < ne < 1 − max(σ, ω ) = 0.741


Overall, the net number of apprenticeship positions created by the program was between 51
and 74 percent of the number of formal apprentices placed.



4     Earnings for Youths and Firms
So far, we have focused on results on youth participation in apprenticeships, and indirect
effects related to the number of new apprenticeship positions in firms. The discussion has
shed light on the presence and magnitude of windfall and substitution effects, respectively
among youth and firms. We now turn to analyzing the short-term impacts of the appren-
ticeship program on earnings for both youth and firms. This provides additional information
on opportunity costs from participation in apprenticeship among youth, as well as potential
indirect benefits to firms.


4.1    Youth Employment and Earnings

We first analyze short-term impacts of the program on youth employment, activities and
earnings. Results show that there are substantial opportunity costs for youth to participate
in apprenticeships. Table 4 documents ITT estimates (equation 2) for employment, hours
worked and earnings by type of employment.
    The upper panel of the table presents results on youth activities. It shows that youth in
the control group are mostly active, as 91 percent are engaged in some economic activity.
Moreover, the average number of activities in the control group is larger than one, indicating
that some youths have several activities. In this section, the program only induces a small
increase in participation in economic activity (by 3.4 percentage point), and a small increase
in the average number of activities (by 0.05). However, the program induces youth to
reorganize their portfolio of activities and forgo some employment opportunities. Specifically,
individuals in the treatment group are less likely to hold wage jobs (by 13.5 percentage point)
or to be self-employed (by 12.9 percentage point), and more likely to become apprentices


                                              24
(by 36.5 percentage points).
   The intermediate panel of the table presents results about hours worked and shows similar
effects. Total hours of work only marginally increase (by 3.7 hours per week). The increase in
hours worked as apprentices (+18.2 hours per week) is offset by a decrease in hours worked in
wage employment (-6.5 hours per week) and in self-employment (-7.7 hours per week). These
results are broadly consistent with the overall employment situation in Côte d’Ivoire, where
unemployment is relatively low, and most youth are engaged in some type of employment,
often in agriculture, non-agricultural self-employment or informal wage jobs (Christiaensen
and Premand, 2017). In this context, Bertrand et al. (2017) also find that the impacts of a
public works program on employment mostly take the form of a reorganization of economic
activities, as opposed to an increase in overall employment or activity rates.
   The third panel of the table presents estimates of program impact on total earnings,
earnings by source of employment, and non-labor earnings. Overall, the program has no
short-term impacts on average earnings among youth. Results show that labor earnings
decrease by FCFA 10 494 (or 25 percent), while non-labor earnings increase by 10 213 FCFA
(or 135 percent). The decrease in employment earnings is driven by a decrease in earnings in
wage employment (- FCFA 6 414) and self-employment (-FCFA 6 381), which is only partly
offset by an increase in apprenticeship earnings paid by employers (+ FCFA 3 238). The
program subsidy, which is paid by the implementing agency (and not the firm), is included
in non-labor income. Non labor earnings increase by FCFA 10 213 in the treatment group,
driven by the subsidy. As such, it is only after accounting for the program subsidy that
forgone labor earnings are fully compensated. Overall, although the total number of hours
worked increases, employment earnings decrease, and total earnings remain stable.
   The bottom panel of the table presents average hourly earnings in the different occupa-
tions across the treatment and control groups. Those average hourly earnings are simply
obtained by dividing earnings in a given occupation by the number of hours in that occupa-
tion. Comparisons of hourly earnings between activities and across groups are informative,
although they should not be interpreted as causal, since there is selection into different occu-
pations. The table shows that, in the control group, youths involved in apprenticeship earn
on average FCFA 628 per hour from their employers, far lower than hourly earnings in wage
employment or self-employment (respectively FCFA 1030 and FCFA 1082). This suggests

                                              25
large opportunity costs of apprenticeships. Interestingly, average hourly labor earnings of
apprentices in the treatment group (FCFA 310) is far lower than average hourly earnings of
apprentices in the control group (FCFA 628). However, accounting for the program subsidy,
the average hourly earnings of apprentices in the treatment group (FCFA 706) is larger,
although it remains lower than earnings in wage employment or self-employment. This illus-
trates how the subsidy changes the structure of payments made by employers to apprentices:
the provision of the subsidy leads to a behavioral response from employers, who in turn pay
apprentices less.32
       Overall, results show that the opportunity costs of participating in apprenticeship are
quite large. Individuals are foregoing earnings in wage jobs and in self-employment, and the
program subsidy contributes to balancing the financial costs of undertaking apprenticeships.
       The estimated average treatment effect of offering participation in formal apprenticeship
on earnings is zero. However, we expect that heterogeneity in the employment situation of
participants generates some heterogeneity in impacts on earnings. For some youth with lim-
ited outside opportunities, participation in formal apprenticeship might lead to an increase
in earnings, for example because of the subsidy. For other youth with better opportunities,
the impact on earnings might be smaller, and even possibly negative. Figure 3 illustrates im-
pact heterogeneity. We first consider the strong assumption of homogeneous zero treatment
effect. The intermediate panel provides the results of the corresponding Mann-Whitney test.
The test is implemented using a large number of permutation, which enables to obtain an
exact p-value (Imbens and Rubin (2015)). The test has the advantage to being robust to
outliers, and clearly rejects homogeneity. The upper panel of the figure displays estimates of
the cumulative distributions of potential outcomes in the groups assigned to treatment (blue
doted line) and to control (red doted line), as well as the confidence interval of the difference
between the two. As can be seen from the figure, the cumulative distribution in the treat-
ment group is first below and then above the cumulative distribution in the control group,
meaning that there is no stochastic dominance of one distribution over the other. However,
only 71% of youth assigned to the treatment group entered formal apprenticeship. Even if
the impact on participation in formal apprenticeship is constant, imperfect take-up can lead
  32
    While assignment to treatment has no impact on youth earnings in the short-term, some small positive
effects are found on the share of youth able to save. This may be due to the fact that the subsidy is paid
regularly into bank accounts. Positive effects are also found on youth self-esteem in the short-term.


                                                   26
to the observed patterns in the distribution of potential outcomes. The lower panel of figure
3 presents quantile treatment effects on the population of youth complying with assignment
to enter formal apprenticeship.33 The figure clearly shows a positive and increasing effect for
low quantiles, but then a declining pattern and a change in sign. Quantile treatment effects
at large quantiles are negative and significant. Although these quantile treatment effects
cannot be estimated as effects at quantile (unless assuming rank preservation), the observed
patterns are consistent with heterogeneity in treatment effects (see Heckman et al. (1997);
Djebbari and Smith (2008)).


4.2       Value of Work from Apprentices and Profits in Firms

We now analyze how the increase in inflow of apprentices indirectly affects firms, including
through changes in labor input (time worked), the value of work provided by apprentices, as
well as firm profits. During the follow-up survey, firm owners were asked questions about each
apprentice who entered the firm since randomization. We measure how these apprentices
contribute to firm activities, their hours worked, and whether they are involved in productive
tasks. We aggregate apprentice-level measures in each firm across all apprentices who joined
since randomization.
       Table 5 documents impact on total labor input from apprentices who entered firms since
randomization. The total time worked by apprentices at the time of the follow-up survey
increase slightly. Firms see a small increase in labor input by apprentices of 6.9 days or 55.9
hours per month. This effect represents a 23 percent increase in days worked by apprentices.
These results are significant, but given the size of the increase in the number of apprentices
at follow-up (0.613, as can be seen from Table 2), they are actually relatively small. For
example, the number of days worked per new apprentice entering firms is 6.9/0.613 ≈ 11.
One simple explanation for this limited impact is that there are more youth working, but
they work less hours. In section 5, we use disaggregated apprentice-level data to show that
absenteeism (a pervasive phenomenon for apprenticeship) is in fact higher among formal
apprentices.
       One critical question pertains to the overall value of work provided by apprentices. In the
  33
    We estimate unconditional instrumental variable treatment effects as developed by Frölich and Melly
(2013), in which entry into formal apprenticeship is instrumented by the assignment variable.


                                                 27
follow-up survey, we ask enterprise owners to recall the work performed by each apprentice
during their last working day, and to estimate how much they would have had to pay an
occasional worker to accomplish the same tasks. We can then estimate the value of work
performed by each apprentice by multiplying this estimated value of work by the number of
days worked in the last month. This apprentice-level measure is then aggregated at the firm
level across all apprentices who started since randomization. The third column of Table 5
shows that the program led to a strongly positive and significant increase of the value of work
by apprentices in treatment firms. The estimated value of work by apprentices increases by
25 543 FCFA per month, a significant 62 percent increase.
       Separately, we can estimate the payments made by firms to apprentices (wage bill).
The survey describes precisely the types of compensation received by apprentice for meals,
transportation, clothing and “motivation”, including both in cash and in-kind payments. We
sum all these components at the apprentice level and aggregate again at the firm level. We
also compute a net value of work at the firm level by taking the difference between the value
of work and wage bill for apprentices. Columns 4 and 5 of Table 5 show that although
the number of apprentices substantially and significantly increased, the total wage bill for
apprentices did not increase significantly in treatment firms. A small increase in the wage bill
is observed as employers provide some payments to program apprentices, but the increase
is not significant. As such, the impact on the net value of work (value of work minus wage
bill) remains large. This increase amounts to 21 380 FCFA per month, more than doubling
the net value of work by apprentices in control firms.
       The increase in the net value of work provided by apprentices in firms can be interpreted
as marginal profits being positive. It is possible that the net value of work may provide
an overestimation of marginal profits, since it accounts for the value of work performed by
apprentices and their direct labor costs, but does not take into account indirect training costs
incurred by firms. Still, these indirect training costs are unlikely to be large.34 It thus likely
that the marginal value of an additional apprentices is positive, even net of indirect training
costs in firms. A positive marginal profit would be consistent with a form of compensation
for the training provided by firms to apprentices.
  34
    For instance, the number of hours that apprentices spend working under direct supervision of master
trainers is rather limited, suggesting youth learn largely by working alongside master craftsmen, inducing
limited investments in time spent by master craftsmen solely teaching apprentices.


                                                   28
      Does the increase in the net value of work by apprentices, or marginal profits, affect
average firm profits? Measures of firm sales and profits are notoriously noisy. In the context
of the study, we asked firm owners to directly report sales and profits (as recommended by
De Mel et al. (2009)). We also asked firm owners to list all their sales from the previous
months to obtain a second measure. Based on this listing, we then asked them to report again
sales and profits, which provides a third (and our preferred) measure. Moreover, experienced
supervisors conducted near-systematic back-check of firm surveys to obtain repeat measures
(for 598 out of 677 firms). In total, this means we have six measure yim , m ∈ {1, . . . , 6} of
sales and profits for most firms in the sample. We estimate the following regression pooling
all the observations together, and including dummies specific to each type of measure:




(5)                                yi,m = a + bTi + δm + vi,m


      Results are presented in Table A9. As for results on firm workforce, we do not detect
any significant effect on average sales or profits, either using the variables themselves or an
inverse hyperbolic sine transformation. These results suggest that, while marginal profits
increase, no significant impacts are found on average firm profits. This could be due in part
to more limited statistical power given the dispersion of the profit variables.
      Figure A3 provides additional information on impacts on the distribution of revenues and
profits. It shows that the point estimates for program impacts on firm profits and revenues are
positive for most of the distribution. However, while these effects are significantly positive
in a few parts of the distribution, they are not significant overall. They also tend to be
negative at the top of the distribution, driven by a few firms in the control group. The
observed patterns are consistent with large variability in the profit and revenue variables.
To put results in perspective, the estimated increase in the net value-of-work provided by
apprentices (21 380 FCFA per month) amounts to 13 percent of average firm profits in the
control group. Overall, the results are suggestive that the increase in the value of work by
apprentices (or marginal profits), while strongly significant and positive, are not sufficient
to increase average revenues and profits in firms.




                                               29
4.3       Summing up Impacts on Earnings

As in section 3.3, we now re-express ITT estimates as LATE estimates for impacts on earnings
per formal apprentice placed by the program. Again, we consider direct impacts on youths
and indirect impacts on firms.
       We first estimate impacts on youth who were in formal apprenticeships at some point
since randomization. We simply estimate parameter bI of the following equation:


(6)                           Incomei = aI + bI × fi +              µSt 1St + ui
                                                              St


where fi captures participation in formal apprenticeship (since the start of the experiment),
which we again instrument by the youth treatment assignment variable.
       On the side of firms, we measure the impact of the entry of one formal apprentice on the
“net value of work” in firms as parameter bS of the following equation:


(7)                Net value of worki = aS + bS × ef
                                                   i +               γv 1v +       δs 1s + ui
                                                                v              s


where ef
       i stands for the number of formal apprentices entering firms. Again, we use firm

treatment assignment as an instrument.
       However, other indirect effects also need to be considered when accounting for program
impacts on youth and firms. As seen in section 3.3, the entry of program apprentices in
treatment firms crowds out some traditional apprentices and leaves some other firms with
unfilled positions. These flows should also be taken into account in a comprehensive cost-
benefit analysis. They are, however, complicated to value.35
       Table 6 presents the results. The upper panel presents the reduced-form estimates dis-
cussed earlier. The lower panel presents instrumental variable estimates, which are simply
the ratio of the first two rows in the upper panel. The first column of the table presents
results for youth and the second column results for firms. Results show that, unsurprisingly,
there is no significant impact on youth, with a non-significant reduction of FCFA -1 977
  35
     Assuming py is the probability of a youth finding an apprenticeship position, and wa and w0 his/her
earnings when finding or not, the contribution of crowded-out youth is σ (1 − py )(w0 − wa ). Similarly, for
firms with unfilled positions, contribution to the net value of work would be −ω (1 − pv )(ya − wa ), where pv
is the probability for an apprenticeship position to be filled, and ya − wa the net value of the work of such a
position.

                                                     30
of total earnings per youth entering formal apprenticeship. On the other hand, at the firm
level, the increase in the net value of work is positive, large and significant, with a value of
FCFA 27,165 per formal apprentice. The sum of the two effects is FCFA 25 188, which is
not significantly different from the subsidy of 30,000 FCFA paid to program apprentices.
    Thus, at first glance, even if program impacts on average firm profits are not significant,
the estimated net value-of-work provided by apprentices or marginal profits are nearly equal
to the subsidy paid by the program. This suggests that indirect effects on firm benefits are
nearly sufficient to make the program cost-effective in the short-term. However, there are
several considerations to keep in mind. First, as previously explained, there are potential
(and likely negative) additional indirect effects on both youths and firms outside the ex-
periment. Recall that for each apprentice hired there is σ = 0.227 traditional apprentice
crowded out and a windfall effect of ω = 0.259 that we interpret as an increase in unfilled
apprenticeship positions. We are unable to value the associated losses, but they are likely to
reduce the net surplus documented above. Second, we do not account for the training costs
incurred by firms, although we do not expect them to be high, as already mentioned.



5     Additional Mechanisms
We now turn to analyzing a range of additional mechanisms that shed further light on the
results. First, as we have shown, the intervention leads youths to enter apprenticeships,
which some of them would not have done absent the program. There is a new population of
apprentices entering firms, which partly substitutes for a population of traditional appren-
tices. These populations can be compared to document patterns of selection into (formal)
apprenticeships. Second, we analyze the performance of formal apprentices in firms, to bet-
ter understand their higher productivity and absenteeism. Third, we show how the program
(and the wage subsidy) affected the contractual relationship between apprentices and firms.


5.1    Youths’ Selection into Apprenticeship

We first explore the characteristics of youths who entered any form of apprenticeship (a = 1)
thanks to the intervention. Are those youths similar to those who would have entered
apprenticeship anyway? This can be analyzed by comparing youths who participate in ap-

                                              31
prenticeship in the control group (called "Always-takers"), youths in the treatment group
who enter apprenticeship due to the intervention (called "Compliers"), and youths in the
treatment group who do not enter apprenticeship (called "Never-takers"). We are particu-
larly interested in the comparison of Compliers and Always-takers.36 Results are presented
in Table 7. We use the same baseline characteristics as when checking balance between
the treatment and control groups. The first column shows the mean of characteristics for
Always-takers, and the second column the estimated mean for Compliers. For completeness,
the third column presents the mean for Never-takers. The last column contains the p-value
for the test of equality in means between Compliers and Always-takers. The variables are
organized by domains: demographics, skills, employment and earnings, aspirations and jobs
search, socio-economic background and financial constraints.
       The program can affect the decision to enter apprenticeship through two main channels.
The first channel relates to expected long-term gains from apprenticeships. There is a distri-
bution of potential gains in the population. In a classical framework, an individual decides
to participate if expected returns are larger than a given threshold. The provision of a sub-
sidy should then push some youths with lower expected returns to participate. Assuming
expected returns are negatively correlated with employment opportunities, Compliers would
be expected to have better opportunities than Always-takers. The table does not provide
strong support for this. The only noticeable difference is that the share of Compliers aspiring
to become self-employed is lower. It is not clear, however, that this indicates more limited
economic opportunities.
       A second channel through which the program can affect youths’ decision to enter appren-
ticeship is by relaxing financial constraints. This is usually one of the main motivations to
offer subsidies. In this case, Compliers would be expected to be more financially vulnerable
than Always-takers. Results do not provide evidence that compliers face more financial con-
straints or have lower socio-economic background. However, this is only based on measures
of financial constraints from baseline. One of the key findings from the previous section is
  36
    Following Abadie (2003), the average characteristic x of "Compliers" is obtained through the regression
of ax on a using T as an instrumental variable. The "Always-taker" population is directly observable as those
for whom a = 1 and T = 0. Testing the equality of means between Compliers and Always-takers is simply
obtained as the test of αaT in the regression E (x|a, T ) = αaT at + αT (1−a) T (1 − a)+ α(1−T )(1−a) (1 − T )(1 − a),
in which the excluded category is a(1 − T ) = 1, the "Always-taker" population and aT = 1 identifies the
population of Always-takers and Compliers.



                                                         32
that entering apprenticeship entails large opportunity costs in terms of forgone employment
opportunities. Since there are few financial instruments to smooth consumption over the
long duration of an apprenticeship, these opportunity costs are likely the main constraint in
access to apprenticeship that the program addresses.
       We can also document changes in the population of youths entering firms as apprentices.
The follow-up survey asks each apprentice in firms a set of questions about their background
characteristics. Three main populations of youths can be compared: formal apprentices f ,
traditional apprentices in treatment firm tT and traditional apprentices in control firms. We
use the following regression to describe heterogeneity in the apprenticeship population:


(8)                         x = a + bf f + btT tT +            γs 1s +       δt 1t + u
                                                           s             t


       Traditional apprentices entering control firms constitute the excluded category. The two
important coefficients in this regression are bf and btT . The first compares the population
of formal apprentices with the population of traditional apprentices in control firms, which
helps document the selection effect. The second parameter compares the population of
traditional apprentices between treatment and control firms. This allows to document the
characteristics of youth crowded out of apprenticeships by the entry of formal apprentices.
       The top panel of Table 8 presents the results. It shows that formal apprentices are older
(by 2.10 years), and more likely to be women (by 11 percentage points) than apprentices in
control firms. Formal apprentices also have a higher education level: compared to traditional
apprentices in control firms, formal apprentices are 48 percentage points less likely to have
no education, 35 percentage points more likely to have completed primary school, and 13
percentage points more likely to have completed lower secondary school. The program
therefore places youths that are more educated than the workforce traditionally hired by
firms. There is no indication that the program helped insert youths with more limited
networks, or more disadvantaged socio-economic backgrounds. They are also no more or
less likely to know a master craftsmen, or have a master craftsman as a relative or family
acquaintance. Formal apprentices do not report having been more affected by the crisis
period in the country either.37
  37
     The table also shows that there are very few differences in characteristics between traditional apprentices
in treatment and control firms (see row labeled "Traditional" in top panel). This suggests that the crowding-

                                                      33
5.2     Apprentices’ Performance

Section 4.2 showed that, despite an increase in the total number of apprentices entering
firms, there is only a weak increase in the total number of days worked, but a large increase
in the value of work by apprentices in firms. In this section, we further explain these results
by looking at disaggregated data at the apprentice level. The upper panel of Table 8 showed
results estimating equation 8 with dependent variables for apprentice attributes. The purpose
was to document selection into apprenticeship. The middle panel of Table 8 presents the
results of the estimation of equation 8 on a set variable obtained from the follow-up firm
survey. This allows documenting differences by type of apprentices at the moment of the
follow-up survey. These differences cannot be given a causal interpretation. They can be
explained by selection effects, by some of the effects of program participation, but also by the
difference in experience across apprentices.38 . Still, these results are useful to shed further
light on the aggregate impacts observed in firms.
    We first analyze apprentices’ participation in firm activities and their productivity. We
find two main results. First, absenteeism is more important among formal apprentices than
other apprentices: traditional apprentices in control firms worked on average 20.14 days, but
the average for formal apprentices is smaller by 7.09 days. Second, formal apprentices are
more productive than traditional apprentices in control firms. The value of tasks performed
the last day of work is on average FCFA 1,296 for traditional apprentices in control firms.
It is higher by FCFA 839 for formal apprentices.39
    To analyze differences in performance, we build several indices of skills: a technical
skill index, a behavioral skill index and a learning skill index.40 Results show that formal
out taking place was not associated with stronger selection of a particular profile of apprentice.
  38
     The inflow of formal apprentices took place within six month after the start of the experiment, while
other apprentices entered over the whole 20 month period (see figure 1).
  39
     The table shows that traditional apprentices who entered treatment firms are also more productive
than traditional apprentices in control firms. There might be various explanations for the differences in
performance. First, employers might have selected traditional apprentices more carefully in treatment firms
(particularly since substitution effects are observed within these firms, as we already documented). However,
results in the upper panel of Table 8 do not suggest strong selection effects. Very few significant differences
are observed between traditional apprentices in treatment and control firms. Second, formal apprentices are
found to be more productive, and this may spill over to traditional apprentices.
  40
     Skills are measured using a set of questions asked to the employer about each apprentice. The technical
skills measure includes two general questions about how well apprentices master techniques, tools and safety
procedures. It also includes questions specific to each trade: for each trade, we worked with the national
training agency (AGEFOP) to identify a list of 2 to 7 technical tasks and asked the employer how well each
apprentice performed these tasks (on a scale from 0 to 10). The apprentice-level technical skill index is the
average of the scores obtained across the trade-specific questions and the two general questions. The learning

                                                     34
apprentices in treatment firms have higher technical skills than traditional apprentices, which
can contribute to explain their higher productivity.
    The results can also explain explain the higher absenteeism observed among formal ap-
prentices. One striking result is that employers rate formal apprentices as having much lower
behavioral skills than traditional apprentices. Looking into the components of the index, the
items that drive the results are related to absenteeism and punctuality. There are various
possible explanations for a higher absenteeism of formal apprentices. First the program
might attract youths who are less interested in apprenticeships in the first place. A second
explanation is that the program might not meet youth expectations. The bottom panel of
the table shows levels of satisfaction reported by apprentices in treatment and control firms.
Formal apprentices are more dissatisfied in general than traditional apprentices, and this is
largely driven by a dissatisfaction with the level of earnings received from firms. Interest-
ingly, they are less satisfied with their labor earnings from apprenticeship than with their
income in general. These results are consistent with the impacts on earnings documented
above.


5.3      Contractual Relationships between Employers and Apprentices

The program affects the bargaining relationship between youths and firms. Indeed, youths
receive lower payments from firms than traditional apprentices. We now document how
the program modified the contractual arrangements and payments between apprentices and
firms. The middle panel of Table 8 details the financial arrangements between employers
and youths or their families. Traditional apprenticeship arrangements involve the apprentices
(or their family) paying a fee to the master craftsmen. Over time, firms start compensating
apprentices, with payments divided between a regular payment (said to be for “soap”) to
cover transport costs, room and board, and a “bonus” payment to motivate apprentices. The
payment for transport, room and board is often made weekly, and can be partly in kind (e.g.
meals). The “bonus” payment is typically paid monthly.
    Results show that formal apprentices pay significantly lower fees to firms compared to
skills and behavioral skills indices each average several general questions. For the learning skills index, these
include: ability to learn, quantity of work, quality of work, speed at work. For the behavioral skills index,
these include attitude at work such as: absenteeism, punctuality, respect of clients and boss, seriousness and
motivation.


                                                       35
traditional apprentices in treatment and control firms. The fact that master craftsmen do
not charge a fee is consistent with the program intent in subsidizing access. Employers were
requested not to charge fees. In parallel, firms make lower payments to formal apprentices,
and these lower payments are mostly driven by a decrease in payments for transport, room
and board. The firms internalize that the program subsidy covers such costs, so they strongly
reduce their contribution. This behavioral response from firms implies that part of the
program subsidy is in fact transmitted to firms. Formal apprentices still receive similar
“bonus” payments than traditional apprentices in treatment firms. This shows that firms
complement the subsidy offered by the program and attempt to directly motivate apprentices.
The decrease in overall payment by firms is consistent with dissatisfaction with labor earnings
among apprentices, as reported above.
    The table also provides some information on the indirect costs incurred by firms to
train apprentices. These costs are key to understand firms’ training decision (Acemoglu and
Pischke (1999)). The interpretation of the increase in the net value of work provided by
apprentices in firms as marginal profits depends on the magnitude of indirect training costs,
as mentioned above. The follow-up survey asked apprentices about the number of hours they
spent working independently, under direct supervision of master trainers or watching their
master. The lower panel of Table 8 presents the results. In the control group, youth spent
on average 2.5, 2.6 and 1.7 hours during their last day of work, working independently, under
the supervision of their master, respectively watching him/her. These figures suggest that
youth learn mostly by working alongside master craftsmen, inducing limited investments in
time spent solely teaching apprentices from master craftsmen. There are also few differences
between the various types of apprentices.



6    Conclusion
Evaluations of employment programs usually focus on direct impacts on participants, but
these programs can have a range of indirect effects that are rarely taken into account. This
paper quantifies the net number of new positions created by a subsidized apprenticeship
program in Côte d’Ivoire. We report results from a double-sided randomized control trial
specifically designed to measure direct effects among youths and indirect effects in firms.


                                             36
   Results show that the apprenticeship program leads to an increase in youth participating
in apprenticeship by 52.8 percentage points. This increase accounts for a significant windfall
effect: 26% of youth placed in formal apprenticeships actually substituted out of traditional
apprenticeships. On the side of firms, the program leads to an increase in the entry of formal
apprentices. Substitution effects are also observed, however, as firms hire 0.23 less traditional
apprentice per formal apprentice placed. Overall, the net number of apprenticeship positions
created by the program is between 51 and 74 percent of individuals placed. We interpret
these results as showing that indirect effects are meaningful. Still, the magnitude of the
substitution effects is moderate, and it is relatively far from full substitution. Following
Calmfors (1994), this suggests that firms are off their labor demand curve or that subsidized
and non-subsidized apprentices are imperfect substitutes.
   Results point to high opportunity costs as an important constraint for youth to partic-
ipate in apprenticeships. Youth reorganize their activities to enter formal apprenticeships,
and the net average impact on youth earnings is zero in the short-term. Participating in
apprenticeship has large opportunity costs in terms of forgone earnings from wage of self-
employment. By providing wage subsidies, the program increases the flow of young people
who are able to afford these opportunity costs. This finding highlights another failure of the
traditional apprenticeship system that the public program addresses.
   Results also show that there is an increase in the number of apprentices entering firms and
that marginal profits increase. This suggests that firms are off-equilibrium and constrained
in the recruitment of new apprentices, with an excess demand at the wage they offer. A
natural question is why firms do not increase the wage they offer to apprentices. Our exper-
iment cannot identify the specific underlying market friction, and this topic would deserve
additional research. However, findings are consistent with the mechanisms highlighted by
Hardy and McCasland (2015), who show that entry fees can serve as a self-selection device.
A low wage might also work as a device to select more motivated youths. Offering the
subsidy eliminates this selection mechanism, and potentially implies that the program leads
to an inflow of formal apprentices who are less motivated than traditional apprentices. We
find consistent evidence that formal apprentices are less assiduous and are less satisfied than
traditional apprentices.
   An important contribution of the paper is to document how formal apprenticeships have

                                              37
indirect benefits for firms that host apprentices. The net value of work provided by appren-
tices, which can be interpreted as marginal profits, strongly increases. This is consistent with
the framework by Acemoglu and Pischke (1999) showing that firms provide training if they
can capture some of its benefits. The increase in marginal profits is of substantial economic
magnitude. In fact, this indirect effect may be sufficient to make the program cost-effective
in the short-term. Direct impacts on youth earnings are not significant. The finding that
a "supply-side" apprenticeship program has impacts on firms on the "demand-side" of the
labor-market is important. It shows that indirect effects need to be quantified to provide a
robust assessment of program performance. It also shows that human capital interventions
targeting youths can have broader benefits for the economy.




                                              38
References
Abadie, A. (2003): “Semiparametric instrumental variable estimation of treatment response
  models,” Journal of econometrics, 113, 231–263.

Abbring, J. H. and J. J. Heckman (2007): “Econometric evaluation of social programs,
  part III: Distributional treatment effects, dynamic treatment effects, dynamic discrete
  choice, and general equilibrium policy evaluation,” Handbook of Econometrics, 6, 5145–
  5303.

Acemoglu, D. and J.-S. Pischke (1999): “The structure of wages and investment in
  general training,” Journal of Political Economy, 107, 539–572.

Akram, A. A., S. Chowdhury, and A. M. Mobarak (2017): “Effects of Emigration on
  Rural Labor Markets,” Working Paper 23929, National Bureau of Economic Research.

Alfonsi, L., O. Bandiera, V. Bassi, R. Burgess, I. Rasul, M. Sulaiman, A. Vi-
  tali, et al. (2017): “Tackling Youth Unemployment: Evidence from a Labor Market
  Experiment in Uganda,” Working Paper eopp64, Suntory and Toyota International Cen-
  tres for Economics and Related Disciplines, LSE.

Almeida, R., L. Orr, and D. Robalino (2014): “Wage subsidies in developing countries
  as a tool to build human capital: Design and implementation issues,” IZA Journal of Labor
  Policy, 3, 12.

Angelucci, M. and G. De Giorgi (2009): “Indirect effects of an aid program: how do
  cash transfers affect ineligibles’ consumption?” American Economic Review, 99, 486–508.

Attanasio, O., A. Guarín, C. Medina, and C. Meghir (2017): “Vocational Training
  for Disadvantaged Youth in Colombia: A Long-Term Follow-Up,” American Economic
  Journal: Applied Economics, 9, 131–43.

Attanasio, O., A. Kugler, and C. Meghir (2011): “Subsidizing vocational training for
  disadvantaged youth in Colombia: Evidence from a randomized trial,” American Economic
  Journal: Applied Economics, 3, 188–220.



                                            39
Babcock, L., W. J. Congdon, L. F. Katz, and S. Mullainathan (2012): “Notes on
  behavioral economics and labor market policy,” IZA Journal of Labor Policy, 1, 2.

Baird, S., J. A. Bohren, C. McIntosh, and B. Ozler (forthcoming): “Optimal Design
  of Experiments in the Presence of Interference,” The Review of Economics and Statistics.

Bandiera, O., R. Burgess, N. Das, S. Gulesci, I. Rasul, and M. Sulaiman (2017):
  “Labor markets and poverty in village economies,” The Quarterly Journal of Economics,
  132, 811–870.

Berniell, L. and D. de la Mata (2017): “Starting on the right track? The effects of first
  job experience on short and long term labor market outcomes,” Working Paper 2017/26,
  CAF.

Bertrand, M., B. Crépon, A. Chuan, R. Haget, M. Mahoney, D. Murphy, and
  K. Takavarasha (2013): “J-PAL Youth Initiative review paper,” Abdul Latif Jameel
  Poverty Action Lab, Cambridge, MA.

Bertrand, M., B. Crépon, A. Marguerie, and P. Premand (2017): “Contempo-
  raneous and Post-Program Impacts of a Public Works Program,” Working paper, World
  Bank, Washington, DC.

Biggs, T., M. Shah, and P. Srivastava (1995): “Training and Productivity in African
  Manufacturing Enterprises,” Working Paper 15101, World Bank, Washington DC.

Bishop, J. H. and M. Montgomery (1993): “Does the targeted jobs tax credit create
  jobs at subsidized firms?” Industrial Relations: A Journal of Economy and Society, 32,
  289–306.

Black, D. A., J. A. Smith, M. C. Berger, and B. J. Noel (2003): “Is the threat of
  reemployment services more effective than the services themselves? Evidence from random
  assignment in the UI system,” American Economic Review, 93, 1313–1327.

Blattman, C. and L. Ralston (2015): “Generating employment in poor and fragile
  states: Evidence from labor market and entrepreneurship programs,” Working paper.



                                           40
Blundell, R., M. C. Dias, C. Meghir, and J. Van Reenen (2004): “Evaluating
  the employment impact of a mandatory job search program,” Journal of the European
  Economic Association, 2, 569–606.

Breza, E. and C. Kinnan (2018): “Measuring the Equilibrium Impacts of Credit: Ev-
  idence from the Indian Microfinance Crisis,” Working Paper 24329, National Bureau of
  Economic Research.

Calmfors, L. (1994): “Active labour market policy and unemployment: A framework for
  the analysis of crucial design features,” OECD Economic Studies, 22, 7–47.

Card, D., J. Kluve, and A. Weber (2018): “What works? A meta analysis of recent
  active labor market program evaluations,” Journal of the European Economic Association,
  16, 894–931.

Cho, Y., D. Kalomba, A. M. Mobarak, and V. Orozco (2013): “Gender differences
  in the effects of vocational training: Constraints on women and drop-out behavior,” Policy
  Research Working Paper 6545, World Bank, Washington DC.

Christiaensen, L. and P. Premand (2017): Cote d’Ivoire Jobs Diagnostic, World Bank,
  Washington DC.

Crépon, B., E. Duflo, M. Gurgand, R. Rathelot, and P. Zamora (2013): “Do
  labor market policies have displacement effects? Evidence from a clustered randomized
  experiment,” The Quarterly Journal of Economics, 128, 531–580.

Cunha, J. M., G. De Giorgi, and S. Jayachandran (forthcoming): “The Price Effects
  of Cash Versus In-Kind Transfers,” Review of Economic Studies.

Dahlberg, M. and A. Forslund (2005): “Direct displacement effects of labour market
  programmes,” The Scandinavian Journal of Economics, 107, 475–494.

De Mel, S., D. McKenzie, and C. Woodruff (2016): “Labor Drops: Experimental
  Evidence on the Return to Additional Labor in Microenterprises,” Working Paper 23005,
  National Bureau of Economic Research.



                                            41
De Mel, S., D. J. McKenzie, and C. Woodruff (2009): “Measuring microenterprise
  profits: Must we ask how the sausage is made?” Journal of Development Economics, 88,
  19–31.

Djebbari, H. and J. Smith (2008): “Heterogeneous impacts in PROGRESA,” Journal of
  Econometrics, 145, 64–80.

Fazio, M. V., R. Fernández-Coto, and L. Ripani (2016): Apprenticeships for the XXI
  Century: A Model for Latin America and the Caribbean?, Inter-American Development
  Bank, Washington DC.

Filmer, D., L. Fox, K. Brooks, A. Goya, T. Mengistae, P. Premand, D. Ringold,
  S. Sharma, and S. Zorya (2014): Youth employment in sub-Saharan Africa, World
  Bank, Africa Development Series.

Frazer, G. (2006): “Learning the master’s trade: apprenticeship and human capital in
  Ghana,” Journal of Development Economics, 81, 259–298.

Frölich, M. and B. Melly (2013): “Unconditional quantile treatment effects under en-
  dogeneity,” Journal of Business & Economic Statistics, 31, 346–357.

Hardy, M. and J. McCasland (2015): “Are small firms labor constrained? experimental
  evidence from Ghana,” Working paper.

Heckman, J. J., J. Smith, and N. Clements (1997): “Making the most out of pro-
  gramme evaluations and social experiments: Accounting for heterogeneity in programme
  impacts,” The Review of Economic Studies, 64, 487–535.

Imbens, G. W. and D. B. Rubin (2015): Causal inference in statistics, social, and
  biomedical sciences, Cambridge University Press.

Kangasharju, A. (2007): “Do wage subsidies increase employment in subsidized firms?”
  Economica, 74, 51–67.

Kluve, J., S. Puerto, D. Robalino, J. M. Romero, F. Rother, J. Stöterau,
  F. Weidenkaff, and M. Witte (2016): “Do Youth Employment Programs Improve
  Labor Market Outcomes? A Systematic Review,” IZA Working Paper 10263.
                                           42
Layard, R., S. Nickell, and R. Jackman (1991): Unemployment: Macroeconomic
  Performance and the Labour Market, Oxford University Press.

Lerman, R. (2014): “Do firms benefit from apprenticeship investments?” IZA World of
  Labor, 55.

Lerman, R. I. (2017): “Skill Development in Middle-Level Occupations,” The Oxford
  Handbook of Skills and Training, 180.

Lise, J., S. Seitz, and J. Smith (2004): “Equilibrium Policy Experiments and the Evalu-
  ation of Social Programs,” Working Paper 10283, National Bureau of Economic Research.

McKenzie, D. (2017): “How effective are active labor market policies in developing coun-
  tries? a critical review of recent evidence,” The World Bank Research Observer, 32,
  127–154.

Moffitt, R. A. (2001): “Policy interventions, low-level equilibria, and social interactions,”
  Social dynamics, 4, 6–17.

Monk, C., J. Sandefur, and F. Teal (2008): “Does doing an apprenticeship pay off?:
  evidence from Ghana,” Working paper, Centre for the Study of African Economies, Uni-
  versity of Oxford.

Muralidharan, K., P. Niehaus, and S. Sukhtankar (2017): “General Equilibrium
  Effects of (Improving) Public Employment Programs: Experimental Evidence from India,”
  Working Paper 23838, National Bureau of Economic Research.

OECD/ILO (2017): Engaging Employers in Apprenticeship Opportunities, OECD Publish-
  ing.

Rotger, G. P. and J. N. Arendt (2011): “The Effect of a Wage Subsidy on Employment
  in the Subsidised Firm,” in Sixteenth Annual Meetings of the Society of Labor Economists
  (SOLE), 29–30.

Teal, F. (2016): “Are apprenticeships beneficial in sub-saharan africa?” IZA World of
  Labor, 268.


                                             43
UNESCO (2015): Delivering TVET through Quality Apprenticeships, UNESCO Publica-
  tion.

Walther, R. (2008): Towards a renewal of apprenticeship in West Africa, Agence Française
  de Développement (AFD), Paris.




                                          44
Tables and figures

Figure 1: Entry of traditional and formal apprentices in treatment and control firms, by
month




        Source: Firm follow-up survey (674 observations)
        Notes: Number of youth entering treated and control firms as formal or
       traditional apprentices, by month. (0 is the randomization date.)




                                          45
   Figure 2: Summarizing net flows of traditional apprentices


          Registered                   Non-registered


 Firm     σ less hires                 ω position to fill
                              ?        O
                                         potential matches
                                       ranging from 0 to M in(σ, ω )
 Youth   ω less entry                  σ youth not hired


 Notes: For each youth hired as a formal apprentice in a firm, there
is σ less youth hired as traditional apprentices, and thus σ more
youth searching for an apprenticeship position outside the exper-
iment. Similarly, for each youth entering a formal apprenticeship,
there is ω less youth starting traditional apprenticeships and thus
ω open traditional apprenticeship positions.




                                      46
Figure 3: Distributions of potential outcomes and Unconditional quantile treatment effects
on Compliers for Hours worked and Income

                Weekly hours                                   Monthly income

                               Distribution of potential outcomes




       Mann Whitney test: p values obtained from 10,000 permutations within strataa

                 p = 0/10000                                     p = 53/10000

                Unconditional instrumental variable quantile treatment effects




 Source: Youth follow-up survey (1661 observations)
 Notes: The figures in the upper panel show the results of the estimation of the cumulative
distribution of potential outcomes in the two assigned groups(they are based on the estimation
of equation 2, with variables defined as 1(y < t) for t varying over the support of y ). The
intermediate panel presents the result of the Mann-Whitney rank test implemented using
10,000 permutations within randomization strata. The figures in the lower panel presents
the results of the estimation of unconditional instrumental variables quantile treamtent effect
Frölich and Melly (2013). The doted blue line 47provides, for a given q , the estimated parameter
and the shaded area its confidence interval.
 a - The p-value is computed as the ratio of the number of times the statistics from a permuted
assignment variable was found larger than the statistic obtained with the true assignment
variable to the total number of permutations.
                                  Table 1: Human Capital Investments

                          Apprentice                      TVET         Any         School
            Formal     Traditional         Total                     Training                    None
              e1           e2               e3
           Started since randomization
 Treated    0.712***    -0.185***        0.528***        -0.057***   0.471***     -0.057***    -0.363***
             (0.017)     (0.016)          (0.021)         (0.010)     (0.021)      (0.019)      (0.021)

 Mean        0.038        0.225            0.263           0.072       0.335        0.205        0.514

           Started apprenticeship and dropped out                         Apprentice at follow-up
            Formal Traditional          Total                         Formal Traditional       Total
              d1          d2             d3                             c1          c2           c3
 Treated    0.222***    -0.060***        0.163***         Treated    0.490***     -0.125***     0.365***
             (0.016)     (0.016)          (0.023)                     (0.018)      (0.014)       (0.022)

 Mean        0.020        0.064            0.084           Mean        0.018        0.161        0.179

 Source: Youth follow-up survey (1661 observations)
 Notes: Estimation of equation 2.
 Upper panel (e) uses information from the human capital module of the follow-up survey, covering the
duration of the experiment. (See footnote 24 and section A2 for definition of variables). Column “None”
means neither school nor any training.
 The right part of the lower panel (c) uses information from the employment module of the follow-up survey.
It measures occupation in apprenticeship at that time.
 The left part of the lower panel measures impacts on dropouts (d). (For each category of youth :
(e)=(d)+(c)).




                                               48
  Table 2: Inflow of apprentices into firms

             Formal     Traditional      Total
 Inflow since randomization (e)
 Treated    1.398***        -0.318*    1.080***
             (0.096)        (0.178)     (0.208)
 Mean         0.188          1.942       2.130
 Exits since randomization (x)
 Treated    0.611***     -0.144**      0.467***
             (0.071)      (0.068)       (0.093)
 Mean         0.130        0.430         0.561
 In the firm at follow-up (s)
 Treated    0.787***         -0.174    0.613***
             (0.065)        (0.149)     (0.172)
 Mean         0.058           1.512      1.570
 Source: Firm follow-up survey (674 observa-
tions).
 Notes: Estimation of equation 1. The upper
panel gives the total number of new apprentices
since randomization (e). The intermediate panel
gives the number of apprentices who left the firm
since randomization (x), and the lower panel the
number of apprentices still in the firm at the time
of the follow-up survey (s=e-x).




                       49
        Table 3: Overall impact on number of apprentices

                    Formal       Total         Per formal apprentice
 On Youth side
 Treated youth     0.712***    0.528***       0.741*** ω       0.259***
                    (0.016)     (0.021)        (0.022)          (0.022)
 Mean                0.038       0.263
 On firm side
 Treated firm       1.398*** 1.080***          0.773***    σ      0.227
                    (0.096)  (0.208)           (0.128)          (0.128)
 Mean                0.188    2.130
 Sources: Firm and youth follow-up survey (respectively 674 and 1661
observations).
 Notes: The first two columns present ITT estimates of equations 2 (up-
per panel) and 1 (lower panel). The third column presents IV estimates
of equations 3 (upper panel) and 4. The last column expresses estima-
tion results of column (3) in terms of parameters σ and ω . The outcome
variables are: entry into formal apprenticeship and entry into any appren-
ticeship since randomization (upper panel), and total number of formal
apprentices and of apprentices of any type since randomization (lower
panel).




                                   50
                           Table 4: Youth Activities, Hours and Earnings


 Activities
                                                       Other       Total #         At least
            Apprentice    Wage empl.     Self-empl.
                                                      activities   activities        one

 Treated      0.365***     -0.135***     -0.129***      -0.014      0.053*        0.034***
               (0.022)      (0.022)       (0.024)      (0.011)      (0.031)        (0.013)

 Mean          0.179         0.356          0.471       0.056        1.191          0.910
 Hours
              As an          As              As       In other
                                                                     Total
            apprentice    wage empl.     self-empl.   activities

 Treated    18.200***      -6.462***     -7.692***      -0.418      3.687**
             (1.170)        (1.235)       (1.302)      (0.401)      (1.492)

 Mean          7.558         14.954        17.637       1.748        41.880
 Earnings
                                                      In other       Total
            Apprentice    Wage empl.     Self-empl.                              Non-labora      Total
                                                      activities     Labor



 Treated      3,238***     -6,414***     -6,381***      -167.6     -10,494***    10,213***       -1,408
               (749.3)      (1,407)       (2,157)      (221.7)       (2,654)      (870.3)       (3,295)

 Mean          4746          15398         19089        799.9        41776          7540        51484
 Average Hourly Earnings
                                                                                       Adjustedb
                                                      In other       Total
            Apprentice    Wage empl.     Self-empl.                             Apprentices      Total
                                                      activities     Labor
 Control        628           1030          1082         458          998
 Treated        310           1058          1278         475          687            706           1099
 Source: Youth follow-up survey (1661 observations).
 Notes: The first three panels present ITT estimates of equation 2 for outcome variables related to occu-
pation, hours worked and earnings. (See footnote 24 and section A2 for definitions of variables).
 The lower panel presents estimates of hourly earnings across different occupations, obtained as the ratio
of average earnings to the average number of hours worked.
 a - Includes the program stipend
 b - Adding non-labor earnings (including program stipend) to apprenticeship earnings and total labor
earnings, respectively.

                                                51
             Table 5: Apprentices’ participation in firms’ activities


            # days of                                          Wage      Net value
                         # hours of work      Value of work
              work                                             Bill       of work

 Treated     6.945**          55.91**          25,543***       4,162 21,380***
             (3.253)          (26.71)           (7,023)       (3,094) (5,932)



 Mean         30.38            252.6             41080         21854       19226
 Source: Follow-up firm survey (674 observations).
 ITT estimates from equation 1 for outcome variables related to apprentice participa-
tion in firm activity. The variables are first defined at the apprentice level and then
aggregated at the firm level across all apprentices who started apprenticeship after
the randomization date.




                                         52
           Table 6: Overall impacts on earnings and net value of work in firms

 Reduced form
                        Youth Earnings                                Net Value of work

 Treated youth               -1,408                Treated firm             21,380***
                            (3,295)                                         (5,932)

                  In Formal Apprenticeship                         # of Formal Apprentices

 Treated youth             0.712***                Treated firm             0.787***
                            (0.016)                                         (0.065)

 Control Mean                0.038                                           0.058

 Second stage

 bI                          -1,977                bS                      27,165***
                            (4,473)                                         (7,315)
 Control Mean               51,484                                          19,226
 Source: Firm and youth follow-up survey (respectively 674 and 1661 observations).
 Notes: The first column presents ITT estimates of equation 2 on youth total earnings (up-
per panel) and youth participation in formal apprenticeship (intermediate panel), and then IV
estimates of equation 3 using assignment to treatment as an instrument (lower panel).
 The second column presents ITT estimates of equation 1 for net value of work from apprentices
in firms (upper panel) and the number of formal apprentices in the firm at the moment of the
follow-up survey (intermediate panel). IV estimates of equation 4 using treatment assignment as
an instrument are presented in the lower panel.




                                              53
                         Table 7: Youth selection into apprenticeship

            Variable Names                  Always-takers   Compliers   Never-takers   p-val
Male                                            0.900         0.812        0.860       0.038
Age                                            20.743        20.907       20.650       0.536
Married                                         0.025         0.015        0.015       0.622
No diploma                                      0.261         0.156        0.205       0.062
Primary education                               0.605         0.679        0.575       0.256
Lower secondary education or above              0.133         0.166        0.220       0.481
Skill Index (All)                                1.766        1.713        1.696       0.337
Learning Skill Index                             0.831        0.776        0.765       0.208
Behavioral Skill Index                           0.935        0.937        0.932       0.954
Has activity                                      0.855       0.863        0.917       0.862
Total nb of activities                           1.474        1.291       1.376        0.182
Total income (KCFA)                              64.893       75.960      87.298       0.450
Total income (KCFA) (hyperbolic sin)             3.088        3.079       3.610        0.979
Searching for a job                              0.415        0.489        0.401       0.262
Aspires to wage job                              0.366        0.509        0.467       0.026
Aspires to self-employment                       0.634        0.479        0.527       0.016
Nb of hhd members in wage jobs                   0.671        0.674        0.587       0.980
Has relatives in wage jobs                       0.466        0.548        0.558       0.213
Has friends in wage jobs                         0.540        0.572        0.529       0.626
Nb of hhd members with IGA                       1.756        1.754        1.728       0.989
Has relatives with IGA                           0.745        0.682        0.722       0.277
Has friends with IGA                             0.756        0.675        0.706       0.163
Parent were present when 15                      0.779        0.717        0.789       0.271
Household subject to crisis                      0.112        0.112        0.158       0.985
Family subject to crisis                         0.189        0.147        0.135       0.405
Lost employment during crisis                    0.022        0.038        0.009       0.435
Nb financial constraints                          2.843        2.823        2.700       0.950
Saved during last 3 months                       0.463        0.495        0.473       0.623
Has saving account                               0.040        0.057        0.089       0.523
Forced to use savings to face emergencies        0.846        0.813        0.841       0.500
Has debt                                         0.319        0.309        0.342       0.876
Has problem paying back debt                     0.190        0.149        0.176       0.418
Is credit constrained                            0.503        0.470        0.503       0.611
                                                                  Continued on next page...




                                            54
... table 7 continued
             Variable Names                 Always Taker    Complier     Never Taker   p-val
Source: Youth baseline survey
Notes: The first column presents average baseline characteristics for "Always-
Takers", i.e. youth assigned to the control group who entered traditional appren-
ticeship.
The second column presents average baseline characteristics of "Compliers", as in
Abadie (2003).
The third column presents average baseline characteristics of "Never-Takers", i.e.
youth assigned to the treatment group who did not start an apprenticeship.
The last column gives the p-value for the test of equality of means between "Always-
takers" and "Compliers" (see footnote 36).




                                           55
                                          Table 8: Apprentice characteristics, performance and satisfaction

                                  Apprentice characteristics (Follow-up firm survey, apprentice module)
                                            Education             Affected                Mastercraftman                                       Wealth
                  Age        Male     No       Prim.   Second       by    Knows       from        from.                        family         index
                                                                   crisis  one       family      neighb.                        acqu.
      Formal     2.10***   -0.11***    -0.48***     0.35***    0.13***        0.04      -0.03       -0.05          0.03         -0.06           0.12
                  (0.29)    (0.02)      (0.04)       (0.04)     (0.03)      (0.03)     (0.04)      (0.03)         (0.04)       (0.04)          (0.08)
      Tradi     -1.01***   0.02***      0.08**      -0.08**      -0.00       -0.02      -0.06       -0.02        -0.08**        -0.03            0.05
      tional      (0.37)    (0.01)      (0.04)       (0.03)     (0.02)      (0.02)     (0.04)      (0.03)         (0.04)       (0.04)          (0.07)
      Ref.        20.67      0.95        0.68         0.25        0.07        0.11       0.57       0.21           0.26         0.38            -0.04
                                      Apprentice performance (Follow-up firm survey, employer module)
                       Work
                 Days      Value               Skills Index                              Fees                            Compensation
                  Last      Last         Tech.   Learning Behav.            Entry       Last        Exit       Meals and  Bonus or             Total
                 month      day                                                        month                   Transport Motivation




56
      Formal    -7.09***    839***      1.01***       0.27     -0.49**    -5,303***     -703      -9,416***     -3,697***        -356        -4,053**
                 (0.87)      (272)       (0.21)      (0.20)     (0.21)      (879)      (451)       (2,301)       (1,293)        (716)         (1,744)
      Tradi        0.04     338**         0.08        0.16       -0.01       320        274         -3,210        -1,189          646           -543
      tional     (0.61)      (163)       (0.18)      (0.17)     (0.19)     (1,023)     (568)       (2,232)       (1,080)        (647)         (1,448)
      Ref.        20.14      1296        4.152       5.187      6.795        4481      992.2         7991          9507         4287           13794
                                     Apprentice satisfaction (Follow-up firm survey, apprentice module)
                                  Satisfaction with                            Hours of work                                         Aspiration
                                                                    Total  Auton- Superv. by Observing                        Salaried         Self
                 Tasks       Hours    Earnings Income      Work            -omous     Master    Master                                    Employment
      Formal     -0.10*       -0.06    -0.31***       0.06     -0.12**       -0.09     0.55**       -0.02         -0.28        0.23***       -0.23***
                 (0.06)      (0.06)     (0.09)       (0.09)     (0.06)      (0.21)     (0.25)      (0.27)        (0.19)         (0.04)        (0.04)
      Traditi     0.00        0.01       -0.05        0.05       0.06        0.27       -0.09       -0.01         0.25           0.00           0.00
      tional     (0.06)      (0.05)     (0.08)       (0.08)     (0.06)      (0.19)     (0.22)      (0.25)        (0.19)         (0.03)        (0.03)
      Ref.        3.645      3.587       3.036       2.969      3.645       8.170      2.506       2.628         1.726          0.182          0.791
      Follow-up firm survey, apprentice module, for upper and lower panels (948 observations);
     Follow-up survey, employer module, for intermediate panel (1260 observations).
      Estimation of equation 8. Rows “Formal” present the difference in means between formal apprentices in treated firms and traditional apprentices
     in control firms who entered within 6 month of the randomization date. Rows “Traditional” compares average characteristics traditional apprentices
     who entered within 6 months of randomization in treatment and control firms. Other coefficients of equation 8 are not reported.
     Traditional apprentices who entered control firms within 6 months of randomization are the reference category.
A1      Experimental design and implementation
In this appendix, we present the framework used for the analysis. The experiment uses a
design in which a strata is a micro market defined as a trade in a given locality. We identify a
set of firms that are interested to host apprentices and the number of apprenticeship positions
they open. This gives a number of positions to fill in a given micro market. The next step is
to register youth into the experiment. Youth are registered by micro market and there is the
same number of youth registered in a micro-market as the number of opened positions. We
then randomly assign firms to treatment and control, in order to have an equal number of
positions in treatment and control groups in each micro market. One practical complication
is that firms do not offer the same number of positions in each micro market and that firms
can open positions in several (closely linked) micro-markets. To address this, we paired
firms according to the structure of their open positions in the set of micro-markets, and then
performed randomized assignment within each pair. The outcome of this procedure can be
described by the ratio of the number of positions assigned to the control group over the total
number of positions in each micro-market.
   The next step of the experimental protocol is to randomly assign exactly the same num-
ber of youth to treatment in each micro-market as the number of treated positions. The
consequence of such a randomization is that the tightness of the treated and control set of
youth and positions is kept identical:

                        # Treated Youthm       # Control Youthm
                                            =
                       # Treated Positionsm   # Control Positionsm

Figure A1 presents the distribution of the ratio. For most micro-markets, we expect that the
ratio will be close to 0.5, which is indeed the case. However, as a result of firms being some-
times in several micro-markets, this is not always the case. Still, the distribution is closely
concentrated around 0.5. The pairing procedure was not perfect. When a small number of
positions were offered for some trades, and when those positions were offered together with
positions in other trades, the firm randomization process can lead to all the positions in a
given trade assigned to treatment or to control. In such a case, the youth assignation proba-
bility is either 0 or 1. We kept the firms in the dataset, but the corresponding youth were not
included in youth regressions (this is done automatically by using weighted regressions (see

                                              57
equation 2)). The case arises for 10 youth. As a result, although 1,842 youth were registered
in the experiment, only 1,832 can potentially be used in the regressions. The same logic
applies for observations used in regressions on follow-up outcome variables. We were able to
survey 1,676 youth out of 1,842. However 15 strata had just one observation (the 10 listed
before and 5 additional due to individuals that we were unable to survey at follow-up). In
the end our final regressions use 1,661 observations: individuals who we were able to survey
at endline and who were not alone in their stratum.
   Addressing potential issues related to spillovers is a key aspect of our experiment. For
each firm, we define the number of entries as e and we make a distinction between entries
of youths registered in the experiment ery and youth not registered in the experiment eny .
Similarly, among youths, we define being hired as an apprentice as a and make a similar
distinction between being hired in a firm registered in the experiment arf and a firm not
registered in the experiment anf . We also define potential outcomes in the following way: for
a variable y , we define y (1, 1) as the potential outcome when treated, y (0, 1) the potential
outcome when not treated but when some competitors in the labor market are. This is
defined for youth either registered in the experiment or not. Last, we define the potential
outcome when the program is not implemented as y (0, 0). Table A1 summarizes the flows
of registered and non-registered youth into registered and non-registered firms.
   Randomized assignment to treatment ensures that, conditional on being registered in the
experiment, potential outcomes are independent from treatment.


                    e(1, 1)x , e(0, 1)x , e(0, 0)x ⊥Tf |Rf = 1 for x ∈ {ry , ny }


and
                   a(1, 1)x , a(0, 1)x , a(0, 0)x ⊥Ty |Ry = 1 for x ∈ {rf , nf }

   However, the relations between e(0, 1)x , e(0, 0)x , a(0, 1)x and a(0, 0)x are of primary
importance.
   A first set of assumptions is that the chances for control firms to hire from the pool of
registered youth has to be the same as what it would have been absent the experiment.
Similarly the chances of control youth to be hired by a registered firm has to be the same as



                                                 58
what it would have been absent the experiment:
                        
                         l(e(0, 1)ry |R = 1, T ) = l(e(0, 0)ry |R = 1, T )
                                        f      f                  f      f
                 A(1) :
                         l(a(0, 1)rf |R = 1, T ) = l(a(0, 0)rf |R = 1, T )
                                        y      y                  y      y


   The experimental design maintains a same share of treated youth and treated firms,
which ensures that assumption A(1) is satisfied.
   More challenging, however, is the case of non-registered youth: h(0, 1)ny and a(0, 1)nf .
The complication arises from equilibrium flows of each category of youth in the micro-market.
We rely on the assumption that the size of the experiment is small compared to the rest of
the apprenticeship market.
                        
                         # Non registered firms       # Registered firms
                 A(2) :
                         # Non registered youth       # Registered youth

   This is a standard way to deal with spillovers in experiments. Table A2 shows that this is
a reasonable assumption in our context. The table aims to determine the share of youths in
the treatment group as a percentage of the total number of youths in apprenticeship in the
study localities. We start by using data from the 2013 national employment survey, collected
in February 2014. The data are representative at the district level for urban and rural areas
(12 districts with urban and rural areas, plus Abidjan, for a total of 25 strata). The 7 study
localities (column 2 in table A2) are located in 6 districts (column 1). We estimate the
share of youths aged 15-24 that are apprentices in urban areas of these districts. Column
3 provides the share of “employed” youths who are apprentices, and column 4 the share of
all youths who are apprentices in urban areas of these districts. We then use data from
the 2014 national census to obtain the total population of the locality. We estimate the
total population of youths aged 15-24. 45.9% of the national population is aged between
15 and 34. Since disaggregated data are not readily available, we estimate that youths
aged 15-24 constitute half this share (a lower bound), or 23% of the total population. We
then estimate the total number of youths aged 15-24 in apprenticeship in these localities
(column 7). Column (8) provides the total number of youths in the experiment (including
both treatment and control groups), and column (9) the ratio of youths in the experiment
over the total number of youths in apprenticeships in the localities. This proportion varies

                                             59
substantially. In Bouake, a large city, treated youth only represent 1.8% of the population of
youths in apprenticeship positions. In contrast, in small localities like Mankono or Daoukro,
the share is 15.7%, respectively 16.5%. On average, treated youth only represent 4.6% of
the population of youths in apprenticeship in the study localities.
   We now elaborate on how assumptions (A1) and (A2) combine in our setting. For exam-
ple, for non-registered youths, we get:


                          e(1, 1)ny +                 e(0, 1)ny +           e(0, 1)ny =           a(0, 1)
            Tf =1,Rf =1                 Tf =0,Rf =1                 Rf =0                 Ry =0


Any shock (for example negative) on the number of non-registered youth entering treated
firms affects the entry of these youth into apprenticeship, and is partially absorbed by changes
in the number of youths entering control firms and non-registered firms. We thus cannot
consider that these flows are the same as what they would have been absent the experiment.
However, assuming the number of non-registered firms in the experiment is large compared
to the number of control firms (Assumption A(2)), we can consider that the flow of non-
registered youth into control firms is not affected by the experiment, which gives:


                           l(e(0, 1)ny |Rf = 1, Tf ) ≈ l(e(0, 0)ny |Rf = 1, Tf )

   Similarly, for non-registered youth, we obtain the accounting equation:


                          a(1, 1)nf +                 a(0, 1)nf +           a(0, 1)nf =           h(0, 1)
            Ty =1,Ry =1                 Ty =0,Ry =1                 Ry =0                 Rf =0


As before, any shock (again negative) on the number of treated youth hired in non-registered
firms will affect the total inflow into non-registered firms, but will be partially absorbed by
changes in the number youth hired from the group of control youth and the group of non-
registered youth. However, assuming the number of non-registered youths is large compared
to the number of registered youth (assumption A(2)), we can consider that the flow of control
youth into non-registered firms is unaffected by the experiment:


                           l(a(0, 1)nf |Rf = 1, Tf ) ≈ l(a(0, 0)nf |Rf = 1, Tf )



                                                         60
       Note that another possible design would have been to randomly assign micro-markets
to treatment and control. In such a case, all the registered firms in a micro-market would
have been assigned to treatment or control. This design would have been valid under a
larger set of assumptions, especially as assumption A(2) would not have been needed and
assumption A(1) would have been satisfied. The implementation of such a design was not
possible for practical reasons related to program implementation. Moreover, there are 111
micro-markets, which might not provide enough randomization units. There is also some
heterogeneity between markets, for instance related to the size of the localities.



A2         Definition of apprenticeship and training variables
The formal apprenticeship program we study is part of the PEJEDEC project, but was
implemented by AGEFOP, the national training agency. AGEFOP also runs a smaller,
similar but independent program in some localities.
       Not all youth assigned to treatment started an apprenticeship. Table A5 provides in-
formation on take-up of formal apprenticeship for youths assigned to the treatment group.
The results are based on a short process evaluation survey collected to assess quality of
program implementation among treated youths and firms (column 1) and from the program
administrative data (column 2).
       The process evaluation took place in September 2015, between the baseline and follow-up
survey, and on average 12 months into the program. The process evaluation survey asked
youth several questions to understand take-up and the timing of potential drop-outs. Youth
dropped out at various points. 83.4 percent of the overall sample of treated youths signed
a contract, and 74.7 percent report that they started an apprenticeship.41 Table A5 also
documents dropouts within 12 months of the start of the apprenticeships, by the time of the
process evaluation survey, showing that drop-out was substantial. 61 percent of youths in
the treatment group were still in formal apprenticeships. This implies a dropout rate of 18.3
percent among youths who started apprenticeships.
       As discussed in section 3.1, the dropout rate measured from the follow-up survey is large,
  41
    There are several reasons for imperfect take-up. 11.5 percent of selected youths could not be re-contacted
by the implementing agency. An additional 5.1 percent of youths were contacted but did not sign the contract.
This can be considered as early dropout and might be due to imperfections in the process of matching youths
to firms. Finally, 8.7 percent of youth report having signed a contract but did not start the apprenticeship.

                                                     61
indicating that around 40 percent of youths from the treatment and the control groups
dropped out in the first year of the program (see table 1). These figures remain consistent
with most drop-out taking place early in the program.
   The administrative dataset provides information consistent with the process evaluation
survey. It shows that 72 percent of youths in the treatment group signed a contract and
started their apprenticeship. The administrative data also contains additional information on
youths having completed the program. It shows that 53.2 percent of youth in the treatment
group (or 73.5 percent of those who started apprenticeships) completed the full program,
while 19.1 percent of the treatment group (or 26.5 percent of those who started apprentice-
ships) dropped-out before the end of their contract.
   We now turn to the measurement of participation in apprenticeship and other human
capital investments. The follow-up survey asks youth whether they were involved in PE-
JEDEC or AGEFOP apprenticeship programs. Youth sometimes confused the two. The
survey also asked youth whether they had been involved in apprenticeship or TVET. The
two answers are mutually exclusive. Youth involved in apprenticeship programs with dual
practical and theoretical training also at times confused whether it was an apprenticeship or
TVET program. We define a "formal apprentice" as a youth who reported being involved in
the PEJEDEC or AGEFOP program, and reported being in either apprenticeship or TVET.
   Table A6 presents some results supporting the choice of the definition of a formal ap-
prentice. In the first column (Take-up), we consider the answer to the question about
participation in a public program such as PEJEDEC (in the first panel) or AGEFOP in
the second panel, and any of the two (in the third panel). The second column considers a
boolean variable for youth answering they have been involved in apprenticeship and each
government program. The third column does the same using the TVET variable instead
of the apprenticeship variable. Last, the third column shows results when considering the
apprenticeship and TVET variables together.
   Results show that some youths confused the AGEFOP and PEJEDEC programs: when
we define take-up as participation in the AGEFOP program, the treatment effect on the take-
up variable is large. We thus consider both programs together (third panel). The second
result is that many youths considered dual apprenticeships as TVET. The treatment effect
for the second and third columns are of a similar order of magnitude. Based on this, we thus

                                             62
define participation in formal apprenticeship as youth answering they are enrolled in any of
the governmental programs (AGEFOP or PEJEDEC), and reporting they participated in
either TVET or apprenticeship training.




                                            63
A3     Additional Tables and Figures

 Figure A1: Ratio of treated positions to total number of positions, by micro-market

             50
             40  30
         Frequency
        20   10
             0




                      0    .2              .4              .6            .8    1
                                # Treated vacancies/Total # of vacancies



      Source: Administrative dataset used for randomization.
      Notes: 111 micro-markets, defined as locality × trade.
      Total number of positions in treated firms in a micro-market divided by total
     number of positions in registered firms in the micro-market. By construction,
     this ratio is the same as the ratio of the number of treated youth to the total
     number of youth in a micro-market.




                                               64
                  Figure A2: Timing of firm and youth follow-up surveys

                   Youth survey                             Firm survey




                              Figure A3: Impact on Sales and Profit

                   Revenues                                            Profit




      Mann Whitney test, with p values obtained from 10,000 permutations within strataa

                 p=2158/10000                                     p=4484/10000
 Source: Firm follow-up survey (674 observations)
 Notes: Estimation of equation 1, with variables defined as 1(y < t) for t varying over the support
of y . The doted red line provides, for a given t, the average in the control group. The solid blue
line provides the sum of the average in the control group and the estimated coefficient.The shaded
area represents the confidence interval of the estimated coefficient. The Mann Whitney test is
implemented using 10,000 permutations within randomization strata.
 a - see figure 3

                                             65
       Table A1: Flows of registered and non-registered youth into registered and non-registered firms

                                              Registered firms R = 1                                    Non-registered
                                   Treated T = 1                 Control T = 0                             firms
     Under the experiment   h(1, 1) = h(1, 1)ry + h(1, 1)ny   h(0, 1) = h(0, 1)ry + h(0, 1)ny   h(0, 1) = h(0, 1)ry + h(0, 1)ny

     Absent the program                                       h(0, 0) = h(0, 0)ry + h(0, 0)ny


                                             Registered youth R = 1                                    Non-registered
                                   Treated T = 1                Control T = 0                             youth
     Under the experiment   a(1, 1) = a(1, 1)rf + a(1, 1)nf   a(0, 1) = a(0, 1)rf + a(0, 1)nf   a(0, 1) = a(0, 1)rf + a(0, 1)nf

     Absent the program                                       a(0, 0) = a(0, 0)rf + a(0, 0)nf




66
                                            Table A2: Experiment Size Ratio in Study Localities

                     Locality included    Share of urban youths                                                                  Experiment
      District                                                                Population in the locality        Treated youths
                      (among others)     (15-24) in apprenticeship                                                                size ratio
                                         employed         all         total               Youth 15-24
                                                                                       all        apprentices
                            (2)             (3)          (4)           (5)       (6)=(5)*0.2295 (7)=(4)*(6)          (8)         (9)=(8)/(7)
      Lagune             Adzope           15.45%       4.82%          58722          13477             650            61            9.4%
      Bandama            Bouake           14.41%       5.17%         536719          123177           6368           113            1.8%
      Lacs               Daoukro          28.23%       8.05%          44342          10177             819           135           16.5%
      Goh-Djiboua         Divo            25.17%       9.50%         105397          24189            2298           114             5%
                         Gagnoa           25.17%       9.50%         160465          36827            3499           125            3.6%




67
      Montagnes           Man             26.40%       13.13%        149041          34205            4491           318            7.1%
      Woroba             Mankono          19.75%       8.20%          15118           3470             281           44            15.7%
      Total                               21.97%        8.15%        1069804         245520           20010          910            4.6%
      Sources: Column (3) and (4) are based on the 2013 national employment survey (collected in February 2014), which is representative at
     both urban and rural levels in each district. They provide the share of apprentices in the population of youths aged 15-24.
      Column (5) comes from the 2014 national census. The average share of youth 15-24 in Côte d’Ivoire is used to compute the number of
     youth in each localities (column (6)).
      Notes: Information from the census and the employment survey are combined in column (7) to provide an estimate of the number of youth
     in apprenticeship in each locality.
      Column (8) provides the number of youth assigned to treatment in each locality. The last column (9) provides the experiment size ratio
     defined as the number of youth assigned to treatment divided by the estimated number of apprentices in each locality.
                             Table A3: Balance for Youth

                                                         Baseline                Follow-up
                Variables                     Cont        Coef p-val     Cont     Coef     p-val
Demographics
Male                                           0.87       -0.02   0.35    0.87 -0.02       0.26
Age                                           20.74        0.09   0.43   20.75 0.08        0.51
Married                                        0.02        0.01   0.47    0.02 0.00        0.60
No diploma                                     0.20       -0.00   0.81    0.20 -0.00       0.84
Primary education                              0.63       0.00    0.88   0.64 0.00         0.98
Lower secondary education or above             0.17       0.00    0.96   0.16 0.00         0.86
Has received training                          0.22        0.03   0.13    0.22 0.04        0.08
Skills
Skill Index (All)                                 1.72    0.01    0.65   1.72      0.01    0.75
Learning Skill Index                              0.79    0.00    0.90   0.79     -0.00    0.92
Behavioral Skill Index                            0.93    0.01    0.43   0.93      0.01    0.36
Economic Activity
Has activity                                   0.87 0.00          0.81   0.87      0.01    0.56
Nb of agricultural activities                  0.20 -0.02         0.32    0.21    -0.03    0.22
Total nb of activities                        1.36 -0.01          0.86   1.38     -0.02    0.75
Nb of non agricultural activities              1.17 0.01          0.69   1.17      0.01    0.72
Total income (KCFA)                           70.53 4.57          0.51   70.86     5.21    0.47
Total income (KCFA) (hyperbolic sin)          3.25 -0.07          0.57   3.27     -0.07    0.61
Employment aspirations
Searching for a job                               0.44    0.01    0.66   0.44      0.01    0.75
Aspires to wage job                               0.46    0.01    0.72   0.46      0.01    0.84
Aspires to self-employment                        0.54    -0.01   0.69   0.54     -0.01    0.80
Nb of hhd members in wage jobs                    0.70    -0.04   0.38   0.70     -0.04    0.44
Has relatives in wage jobs                        0.50    0.02    0.57   0.51      0.02    0.52
Has friends in wage jobs                          0.52    0.03    0.22   0.53      0.03    0.34
Nb of hhd members with IGA                        1.78    -0.04   0.64   1.78     -0.02    0.82
Has relatives with IGA                            0.71    -0.00   0.88   0.71     -0.01    0.84
Has friends with IGA                              0.78    -0.07   0.00   0.78     -0.07    0.00
Exposure to crisis
Parents were present when 15                      0.76    -0.01   0.72   0.76     -0.01    0.74
Household subject to crisis                       0.12    -0.01   0.62   0.13     -0.01    0.77
Family subject to crisis                          0.19    -0.04   0.07   0.19     -0.04    0.05
Lost employment during crisis                     0.03     0.00   0.88   0.02      0.00    0.77
Financial constraints
Nb financial constraints                           2.78    -0.02   0.89   2.83     -0.05    0.73
Saved during last 3 months                        0.49    -0.01   0.77   0.49     -0.01    0.84
Has saving account                                0.05     0.01   0.59   0.05      0.01    0.49
Forced to use savings to faced emergencies        0.85    -0.02   0.38   0.85     -0.02    0.36
                                                                    Continued on next page...

                                             68
... table A3 continued
                 Variables                    Cont      Coef    p-val    Cont       Coef     p-val
Has debt                                         0.31    0.01   0.76      0.31      0.01     0.68
Has problem paying back debt                     0.16   0.00    0.87      0.16      0.00     0.82
Is credit constrained                            0.52   -0.03   0.24      0.52     -0.03     0.37
                                                            Respondent to survey
1832 youth registered                           0.76    -0.04 0.03 0.91 0.00          0.92
Sources: Youth baseline and follow-up surveys
Notes: Each row in the table considers a specific baseline characteristic and presents
the result of the estimation of equation (2) on the whole sample for which the baseline is
available (left panel - 1357 youths), or the sample with both baseline and follow-up survey
respondent (right panel - 1299 youths). In each panel, the first column gives the number
of observations used in the regression. The second column gives the estimated coefficient
and the third column the p-value. The last row provides the survey response rate. (For
the baseline survey, the response rate captures the share of available data following an IT
issue with the online server.)




                              Table A4: Balance for Firms

                                              Baseline                  Follow-up
                Variables                 Cont Coef p-val           Cont Coef     p-val

  Nb of open apprenticeship positions     2.51      0.02    0.87       2.45      0.09      0.47
  Firm Status
  No legal status                         0.84      -0.00   0.95       0.86      -0.02     0.55
  No accounting                           0.68      -0.04   0.26       0.69      -0.04     0.26
  No salary slip                          0.97       0.01   0.66       0.98      0.00      0.89
  Workforce
  Permanent workers                       6.32      0.39    0.39       6.19       0.16     0.73
  Autonomous workers                      3.33      0.12    0.63       3.23      -0.03     0.88
  Supervisors                             2.37      0.12    0.38       2.32      0.14      0.33
  Apprentices                             3.38      0.13    0.67       3.38      0.04      0.89
  Channels to recruit apprentices
  Spontaneous application                 0.10      -0.03   0.20       0.10      -0.03     0.10
  Parents asked                           0.82       0.03   0.25       0.82       0.03     0.23
  Referral                                0.04      -0.02   0.18       0.03      -0.01     0.34
  National agency                         0.02      0.01    0.58       0.02       0.00     0.80
  Other recruitment channel               0.03       0.01   0.55       0.02       0.01     0.31
  Reasons to hire apprentices
                                                                Continued on next page...
                                            69
... table A4 continued
              Variables                 Cont    Coef    p-val   Cont    Coef    p-val
To get workers                          0.09    -0.01   0.71    0.08    0.00    0.87
To transmit knowledge                   0.45     0.02   0.60    0.48    -0.01   0.89
To help youth                           0.41    -0.01   0.70    0.40    -0.00   0.92
Because it pays                         0.01    -0.00   0.67    0.01    0.00    0.69
Other reasons                           0.04    0.01    0.71    0.04     0.00   0.86
First criterion to select apprentices
Skills                                  0.03    0.01    0.55    0.03    0.01    0.57
Motivation                              0.29    -0.02   0.56    0.29    -0.03   0.44
Respect                                 0.60    0.02    0.51    0.60    0.03    0.44
Tuition requested
At start                                 0.51   0.01    0.75     0.53   0.01    0.75
Amount                                  37944   -920    0.75    38059   -570    0.85
During training                          0.26   0.02    0.47     0.26   0.02    0.48
Apprenticeship dropouts between 2012-2014 and reasons for       dropping out
Total                               2.00    0.35 0.09            1.96  0.37     0.09
Unable                              0.07    0.05 0.09            0.06  0.05     0.09
Not interested                      0.21    0.02 0.66            0.20  0.02     0.63
Financial reasons                   0.13 -0.07 0.01              0.14 -0.08     0.01
No work perspec                     0.00    0.00 0.61            0.00 -0.00     0.93
Found a job                         0.08    0.00 0.99            0.09 -0.00     0.93
Disciplinary reason                 0.17 -0.03 0.41              0.17 -0.03     0.44
Apprenticeship finishers between 2012-2014
Number                               1.21        0.02   0.92    1.20    -0.05   0.77
Hired in firm                         0.24       0.03    0.69    0.20     0.02   0.72
Hired outside                        0.25       -0.04   0.49    0.25    -0.05   0.39
Starting business                    0.67       0.07    0.58    0.69     0.03   0.82
                                                    Respondent to survey
731 registered firms                       0.95   0.00 0.86     0.91    0.01     0.48
Notes: Each row in the table considers a specific baseline characteristic and presents
the result of the estimation of equation 1 on the whole sample for which the baseline
is available (left panel - 694 firms), or the sample with both baseline and follow-up
survey respondents (right panel - 643 firms). In each panel, the first column gives
the number of observations used in the regression. Some variables (for example
in the dropout section) are only defined conditionally on another variable in the
table (e.g. among those who had at least one dropout). The second column gives
the estimated coefficient and the third column the p-value. The first row contains
the number of apprenticeship positions offered firms before randomization (from
administrative data sources). The last row provides the survey response rate. (For
the baseline survey, the response rate captures the share of available data following
an IT issue with the online server.)



                                         70
Table A5: Program Take-up and Dropout (process evaluation and administrative data)


             Process Evaluation data                           Administrative data

                                  Count     As %                             Count     As%
   Enrolled                         914    100.00      Enrolled               914     100.00
   Contacted by implementer         809    88.51       Did not sign           253      27.68
   Signed a contract                762     83.37      Signed a contract      661      72.32
   Started apprenticeship           683    74.73       Dropped out            175      19.14
   Still in apprenticeship          558    61.05       Finished               486     53.17
Sources: Process evaluation survey (left panel) and administrative dataset (right panel).




                                             71
                 Table A6: Definitions of formal apprenticeship participation

                                 Take-up    Apprentice   TVET       Any training
             PEJEDEC program

             Treated youth     0.643***      0.472***    0.135***    0.607***
                                (0.017)       (0.017)     (0.012)     (0.017)
             Control Mean        0.028         0.015       0.004       0.019
             AGEFOP program

             Treated youth     0.708***      0.504***    0.168***    0.672***
                                (0.017)       (0.018)     (0.013)     (0.017)
             Control Mean        0.063         0.024       0.012       0.036
             Any of the 2 governmental programs

             Treated youth     0.747***      0.528***    0.184***    0.712***
                                (0.016)       (0.018)     (0.014)     (0.016)
             Control Mean        0.069         0.025       0.013       0.038
           Source: Youth follow-up survey (1661 observations)
           Notes: The table documents options to build human capital vari-
           ables (see discussion in Appendix A2).




  Table A7: Participation in apprenticeship (Status at follow-up and in retrospective calendar)

                                    19-21      16-18      13-15      10-12         10-21
                   Apprentice                                                                  Since
                                       # months after randomization             (last year)
                  at follow–up                                                                 start

 Treated youth      0.365***      0.377***    0.414***   0.462***   0.465***    0.424***      0.528***
                     (0.022)       (0.021)     (0.021)    (0.021)    (0.021)     (0.022)       (0.021)

 Control Mean         0.179         0.169      0.167      0.148      0.157         0.242       0.263
 Source: Youth follow-up survey (1661 observations).
 Note: Estimation of equation 2 on various apprenticeship variables using either occupation
at the time of the follow-up survey (first column), or information from a
retrospective calendar of occupations (column 2-6). The last column reports participation in ap-
prenticeship over the duration of the experiment.



                                               72
               Table A8: Stock of apprentices and other employees in firms

               Total #   Full-time                               Occasional No apprentices
                                   Apprentices      Interns
            of employees workers                                  workers    in the firm

 Treated       0.495         0.030       0.464       0.001          0.165              -0.054**
              (0.554)       (0.247)     (0.362)     (0.016)        (0.133)              (0.027)


 Mean          7.128        3.379        3.719         0.030           0.970             0.203
Source: Firm follow-up survey (674 observations)
Notes: Estimation of equation 1 on various workforce variables obtained from the employer
module of the firm follow-up survey.




                            Table A9: Sales and Profit of firms



                                      Sales                              Profit
                             Level    Inv Hyper sine           Level     Inv Hyper sine

         Treated            -63,474        0.122           -22,682             0.0284
                           (46,934)       (0.173)         (15,857)             (0.211)

         Control Mean       469338        12.47            169330              11.11
         Source: Firm follow-up survey, 3810 observations (3 measures for all 674
        firms, and 3 measures for 596 back-checked firms).
         Estimation of equation 5 for sales (left panel) and profit (right panel). In
        each panel, the first column considers the variable as measured. The second
        column √considers an inverse hyperbolic sine transformation: arsinh(x) =
        log (x + x2 + 1).




                                            73