Policy Research Working Paper 11185 The Lasting Effects of Working while in School A Long-Term Follow-Up Mery Ferrando Noemi Katzkowicz Thomas Le Barbanchon Diego Ubfal Gender Group A verified reproducibility package for this paper is August 2025 available at http://reproducibility.worldbank.org, click here for direct access. Policy Research Working Paper 11185 Abstract This paper provides the first experimental evidence on the pivotal educational transitions and are larger for vulnerable long-term effects of work-study programs, leveraging a ran- youth and men, while remaining positive for women and domized lottery design from a national program in Uruguay. non-vulnerable youth. The program is highly cost-effec- Participation leads to a persistent 11 percent increase in tive, with average impacts exceeding those of job training formal labor earnings, observable seven years after the pro- programs and comparable to early childhood investments.. gram. Effects are stronger for youth who participate during This paper is a product of the Gender Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at dubfal@worldbank.org. A verified reproducibility package for this paper is available at http://reproducibility.worldbank.org, click here for direct access. RESEA CY LI R CH PO TRANSPARENT ANALYSIS S W R R E O KI P NG PA The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team The Lasting Effects of Working while in School: A Long-Term Follow-Up Mery Ferrando, Noemi Katzkowicz, Thomas Le Barbanchon, Diego Ubfal* JEL Codes: I21, I26, J13, J24, J31, O15. Keywords: Work-study Program, Youth Employment, School-to-Work Transi- tion, Long-term Effects. * Ferrando: Tilburg University. Email: m.ferrando@tilburguniversity.edu. Katzkowicz: Univer- ´ sidad de la Republica. Email: noma.katzkowicz@fcea.edu.uy. Le Barbanchon: Bocconi University, CEPR, IGIER, IZA, J-PAL. Email: thomas.lebarbanchon@unibocconi.it. Ubfal: World Bank, IZA. Email: dubfal@worldbank.org. This study uses confidential data from the Uruguayan Ministry of Labor and Social Security (MTSS) and Social Security Administration (BPS). These data were ac- cessed through a confidential agreement between Universidad de la Republica´ and MTSS. Diego Ubfal acknowledges support from the World Bank’s Gender Innovation Lab for Latin America and the Caribbean and the Umbrella Fund for Gender Equality. We thank David McKenzie for useful comments. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the World Bank and its affiliated orga- nizations, or those of the Executive Directors of the World Bank or the governments they represent. All errors are our own. Youth unemployment represents a critical issue across various global contexts, with rates consistently surpassing those of adult unemployment by a factor of three. This disparity underscores the unique challenges young individuals face in transi- tioning from education to employment. Exacerbating this issue is the prevalence of young people classified as NEET (Not in Education, Employment, or Training). In 2023, nearly 20 percent of the youth population in Latin America and the Caribbean were identified as NEET. One approach to increasing youth employment and fa- cilitating the school-to-work transition is through work-study programs. Working while studying is a relatively common practice among youth in some countries, though the overall levels remain low. In 2023, 17 percent of students aged 15- 19 in Latin America and the Caribbean were employed.1 Evidence on the long- term causal effects of in-school work remains limited to non-experimental designs (Ruhm, 1997; Hotz et al., 2002; Ashworth et al., 2021), and the experimental lit- erature on broader active labor market policies offers limited guidance, as most evaluations do not extend beyond three years post-program completion (McKen- zie, 2017; Card et al., 2018; Carranza and McKenzie, 2024).2 This paper fills this gap in the literature by providing causal evidence on the long- term effects of a work-study program. Leveraging a randomized lottery-based de- sign, we examine a program in Uruguay that offers youth aged 16 to 20 their first formal work experience within state-owned companies for up to one year. Le Bar- banchon et al. (2023) document significant short-term effects of this program on formal earnings. They also find that the program enhances initial educational at- tainment, alleviating concerns about potential crowding-out effects on educational investment.3 Despite these promising findings, whether short-run benefits of work- study programs lead to lasting earnings improvements remains unresolved. While the program initially provides participants with an advantage through early work experience, its impact may diminish over time as non-participants accumulate their own work experience, potentially narrowing labor market differences. How- ever, if labor market trajectories depend heavily on first jobs, program effects may 1 Incomparison, 14 percent of students in Europe and 22.5 percent of high school students in the United States (US) were employed in 2023, with higher employment rates observed among college students. 2 Notable exceptions include Ibarrar´ an et al. (2019), Attanasio et al. (2017), and Bandiera et al. (forthcoming), who provide experimental evidence on outcomes measured six, eight to fourteen, and six years after program completion, respectively. 3 The absence of crowding-out effects on academic performance is consistent with findings from a recent related study (Aucejo et al., 2024). 1 persist.4 In the Uruguayan case, well-regarded positions in state-owned companies could serve as crucial stepping stones to future employment. Understanding the long-term effects of combining work and study is critical for several reasons. First, such evidence is essential for conducting cost-benefit anal- yses, as benefits realized over extended periods are more likely to outweigh the costs incurred during program implementation. Second, identifying long-term ef- fects can provide valuable insights into the extent of state dependency, which can inform broader policy decisions and program design. The Uruguayan work-study program offers a unique opportunity to examine the long-term causal effects of combining work and study, as program assignment is determined by randomized lotteries. Using administrative data from social security records, we analyze a sample of 90,423 teenagers who applied to the first three editions of the program, tracking their earnings for up to seven years after program completion. Participants were between 16 and 20 years old when they entered the program; thus, short-term effects reported in Le Barbanchon et al. (2023) are measured when they are 18 to 22 years old. We track them until they are 23-27 years old to assess long-term outcomes – a period during which most have completed their studies and entered the labor force. We find that participation in the work-study program increases yearly earnings by approximately 10 to 13 percent from the third year after program completion through the seventh year. This effect is comparable to, or larger than, the 8 percent increase observed two years post-program, as documented in Le Barbanchon et al. (2023). The long-term effect on yearly earnings, observed in Year 7, is driven by a 4 percent increase in the monthly probability of being employed (extensive margin) and a 6 percent increase in monthly wages (intensive margin). The earnings effects are not limited to the industries that provided the initial program jobs, and the work experience accumulated during the program appears to be valued in the long run by different employers across the economy. We provide suggestive evidence that the work experience channel and the edu- cation crowding-in channel increase long-term earnings by 2 percent each. The results on work experience indicate that its returns may not decline quickly, align- ing with the concept of state dependency, wherein early advantages in the labor market persist over time. Additionally, gaining work experience during the piv- 4 SeeSchwandt and von Wachter (2019) and Alves and Varvasino (2025), among others, for evi- dence on the long-term effects of entering the labor market during periods of high unemployment. 2 otal developmental stage of adolescence may shape participants’ long-term career trajectories more than if they had entered the labor force at a later stage.5 Our heterogeneity analysis suggests that the persistent effects of the program are most pronounced when teenagers participate at the conclusion of secondary school or the onset of tertiary education — a critical juncture for making pivotal educa- tional decisions. Furthermore, we find positive labor market effects for both vul- nerable and non-vulnerable individuals, suggesting that work-study programs like the one studied can be effective beyond the traditionally targeted disadvantaged groups. Although differences are not statistically significant, the estimates indicate potentially larger effects for vulnerable youth. The effects are also stronger for men than for women, though they are positive and statistically significant for both. The work-study program is highly cost-effective. Despite relatively high initial costs, the long-term earnings gains are projected to result in full fiscal recovery by age 34 (i.e., 15 years after program participation). This recovery is driven by increased tax revenues resulting from participants’ higher earnings. Compared to the main job training programs for youth studied in the literature, the work-study program performs substantially better in terms of long-run cost-effectiveness, as measured by the Marginal Value of Public Funds (MVPF; Hendren and Sprung- Keyser, 2020). Overall, our findings suggest that work-study programs can facilitate the transition from school to work, with benefits lasting at least seven years beyond program completion. Policies that integrate education with early labor market exposure may therefore serve as an effective strategy for improving long-term labor market outcomes. This study contributes to three strands of literature. First, it contributes to the literature on the effects of working while in school. Previous research, primarily based on non-experimental data from the US, has found mixed results on the long- term labor market returns of combining work and study (Ruhm, 1997; Hotz et al., 2002; Ashworth et al., 2021). By leveraging a randomized lottery-based assignment, our study provides the first causal evidence on the long-term impact of a work- study program, demonstrating that early work experience leads to significant and lasting earnings gains. 5 Neuroscience research highlights adolescence as a key period for neural development, marked by significant changes in brain regions associated with various cognitive functions (Sebastian et al., 2010; Blakemore and Robbins, 2012). 3 Second, this study contributes to the literature assessing the long-term effects of youth employment programs. Our results align with non-experimental evaluations of federal employment programs in the US, which find positive effects on long- term employment (Scott-Clayton and Minaya, 2016) and lifetime earnings (Aizer et al., 2024). We build on this research by providing experimental evidence from a middle-income country on a work-study program that explicitly requires partic- ipants not to drop out from school. This latter distinctive feature is important, as our findings suggest that combining work and study can yield lasting labor market benefits without crowding out educational attainment. Finally, we contribute to the broader literature on active labor market policies (ALMPs). Meta-analyses reveal that most experimental evaluations of ALMPs fo- cus on short-term outcomes, with impacts typically assessed within three years of program completion (McKenzie, 2017; Card et al., 2018; Agarwal and Mani, 2025; Carranza and McKenzie, 2024), with a few notable exceptions (see footnote 2). Moreover, many training programs show only modest or temporary impacts (Blattman and Ralston, 2015; McKenzie, 2017; Card et al., 2018). In contrast, our study extends the time horizon by examining earnings effects seven years after program participation. 1 Institutions, Data, and Empirical Design In this section, we describe the Uruguayan work-study program, the data and our empirical design. We follow closely the related sections in Le Barbanchon et al. (2023). 1.1 YET Program Since 2012, the work-study program ”Yo Estudio y Trabajo” (YET) provides youth aged 16 to 20 living in Uruguay with a first formal work experience in state-owned companies for up to one year (see Online Appendix Table A1 and Section B for more institutional details). All youth aged 16 to 20 in Uruguay are eligible to apply for YET if they meet two key conditions: 1) they are enrolled in an educational institution, and 2) they have not worked formally for more than 90 consecutive days at the time of application. Assignment to the program is done by lottery at the locality level. The number of 4 participants in each locality depends on the number of jobs offered by state-owned firms that partner with the program in that locality. Lottery candidates are ran- domly ranked within locality, and program offers are made sequentially until local slots are filled. Starting with the third edition of the program in 2014, quotas were introduced in the largest localities to guarantee participation of minority youth: 8 percent of African origin, 4 percent with disabilities, and 2 percent transgender youth. Importantly, firms cannot choose the youth they want to hire, and candidates can- not select the firm in which they want to work. The program administration per- forms the matching of participants to available job positions, taking into account commuting distance and school schedules (but not skills). The program offers part-time jobs of 20 to 30 hours per week, with no overtime allowed. Participants are expected to work during the firm’s normal operating hours, ensuring their school attendance is not hindered. Contracts are temporary, lasting 9 to 12 months, and are non-renewable. In 2016, the fixed remuneration was $446 per month for a 30-hour-per-week job (around $3.7 per hour).6 The program wage is higher than the national minimum wage of $372 per month for full-time work. All program firms are in the public sector and pay wages from their own budget. Most are large state-owned firms, with only a few positions offered in the pub- lic administration. For example, the top five employers in the first three yearly editions were the following state-owned companies: the electricity company (hir- ing 22 percent of participants), the water company (21 percent), the oil and gas company (16 percent), the commercial bank of Uruguay (10 percent), and the tele- phone company (9 percent). The program stipulates that work activities must be in administration or operations, primarily focusing on support tasks. 1.2 Data We use two main data sources: YET program administrative records and social security records. All data can be matched at the youth level. First, data from the online application form, which youth must complete to par- ticipate in YET lotteries, include basic demographics (age, gender, locality) and 6 Throughout the paper, Uruguayan pesos are adjusted to January 2016 values using the Con- sumer Price Index (CPI), and these values are then converted to US dollars at the January 2016 exchange rate of 0.033 dollar per peso. 5 educational level. From YET administrative records, we also have information on lottery draws, subsequent offers, and program participation. Second, social se- curity records include monthly labor earnings from formal jobs for all applicants from 2011 to 2022. To balance sample size and long-term analysis, we focus on the first three editions of YET (2012, 2013, and 2014), the same sample as in Le Barbanchon et al. (2023). This ensures reasonable statistical power and allows us to observe earnings for seven years post-program for these applicants. We describe our sample of appli- cants and verify balance between groups that received offers and those that did not in Online Appendix Table A2. 1.3 Empirical Design In our primary analysis, we focus on the Treatment-on-the-Treated (ToT) effect of the program. We define treatment as working at least one month in a program job and define the variable Offered as ever receiving a program job offer. To obtain the causal treatment effect, we leverage the lottery design and use the Offered variable as an instrument for the treatment dummy. Under this definition of treatment, the local average treatment effect is equal to the ToT, as no youth can work in a program job without an offer (i.e., there are no always-takers). This effect is identified based on the exclusion restriction that the only reason the outcomes of youth offered a program job change is due to their participation in the program. We analyze data at the applicant level and handle applicants who apply multiple times as follows. We randomly select one application for each youth in the control group (who are never offered a program job). To maximize statistical power, we select the application that generated an offer for each applicant who receives at least one offer.7 We consider the following specification at the applicant level i in edition e: Yi,t = α1 + γt Treatedi + Locality × EditionFE + QuotaFE + # Appi + ρt Xi,0 + ϵi,t (1) Treatedi = α2 + δO f f eredi + Locality × EditionFE + QuotaFE + # Appi + β Xi,0 + υi (2) 7 Asshown in Le Barbanchon et al. (2023), short-term effects are robust when selecting a random application in the ever-offered group, or when analyzing the data at the application level. 6 where Yi,t is the outcome of individual i, t periods after the application date in edi- tion e. Treatedi indicates whether individual i worked in a program job offered in edition e, while O f f eredi indicates whether individual i received a program job of- fer. To control for the lottery design, we include Locality × Edition fixed effects and quota fixed effects. This accounts for variation in the probability of receiving a job offer across lotteries, depending on the local number of program jobs offered and potential quotas. To further control for individual variation in the offer probability (and thus in the treatment probability), we include the number of applications of individual i in different localities during edition e (# Appi ). To increase precision, we include a vector of covariates Xi,0 measured at the application date. It comprises gender, age, household vulnerability status, earnings, and level of education in the year before applying to the program. Our parameter of interest is γt , which we estimate using two-stage least squares, as explained above, and which captures the ToT effect t periods after application. ToT effects are compared to the control com- plier mean (i.e., the mean for youth who would have participated in the program if they had won the program lottery). In practice, the first stage is strong: 77 percent of youth receiving an offer work in a program job, and it is homogeneous across program editions (see Online Appendix Table A3). The F-statistic from the first stage is well above the threshold value of 104.71 suggested by Lee et al. (2022) for a strong instrument when using a 5 percent critical value in the second stage. In Online Appendix A.1, we show that our results are robust to alternative specifi- cations: omitting controls, clustering standard errors at the locality level, not win- sorizing the earning variables, computing intention-to-treat (ITT) estimates, and defining treatment alternatively as working in any firm while being enrolled in school during the program year. 2 Long-Term Effects on Labor Market Outcomes In this section, we present the long-term effects of the program on labor market outcomes, measured up to seven years after participation. 2.1 Earnings Effects In Figure 1, we present a graphical visualization of the evolution of average quar- terly labor earnings for both treated and control compliers, as well as the treat- 7 ment effects on quarterly labor earnings. The dashed line represents the average quarterly earnings of treated youth, while the solid line represents the average for control compliers. As required by the program eligibility criteria, earnings are near zero in the year before applying. Quarterly average earnings of control compliers increase to $500 during the program year and continue to grow, reaching approx- imately $1,470 by the end of Year 7 (see values on the y-axis on the right-hand side). In contrast, the average earnings of treated youth peak at around $1,000 per quarter in the second quarter of the program year. Immediately after the end of the program, both trends converge, as program participants cannot remain employed by program firms. From then onward, average earnings of treated youth grow at a faster rate than those of control compliers, reaching approximately $1,650 per quarter seven years after the program. Figure 1 also plots treatment effects on earnings and their 95 percent confidence in- tervals. During the program year, quarterly treatment effects reach approximately $700 from the second quarter. Immediately after the program, treatment effects naturally decline, while they steadily increase thereafter, becoming statistically sig- nificant in Year 2. They reach approximately $180 toward the end of the period (y-axis on the left-hand side of the figure). Table 1 presents our main treatment effects for the program year and for each year after program participation. We report treatment effects on yearly earnings (Column 1), employment (Columns 2 and 3), wages (Column 4), and contract type (Column 5). During the program year, treated youth earn almost three times as much as control compliers. After the program, treatment effects on yearly earnings are positive in all years and statistically significant at the 10 percent level in Year 2, the 5 percent level in Year 3, and the 1 percent level from Year 4 onward. These effects correspond to an 8 percent increase in Year 2, and between 10 and 13 percent in Years 3 to 7 (see Column 1). We find that the short-term effects identified in Le Barbanchon et al. (2023) not only persist in the long-term but also increase over time, reaching 11 percent seven years after the program. Our administrative data capture earnings exclusively in the formal sector. To as- sess the potential role of sectoral shifts in total earnings, we use data on informality from the 2022 Household Survey in Uruguay (ECH, Instituto Nacional de Estadis- tica Uruguay, 2022), restricting the sample to individuals aged 24 to 28, the age group that treated youth reach seven years after program participation. We find 8 that 14 percent of young workers in this age bracket work in the informal sector in the country, earning an average of approximately $3,500 annually. To explain the observed $650 increase in formal earnings in Year 7 solely through a realloca- tion from informal to formal employment (assuming no change in total earnings), a shift of 19 percent of treated individuals from the informal to the formal sector would be required. However, since only 14 percent of individuals in this age group are employed informally in the country, it is unlikely that the increase in formal earnings can be attributed solely to sectoral reallocation. This suggests the rise in formal earnings reflects higher total earnings. 2.2 Employment Effects The positive treatment effects on earnings are partially explained by positive em- ployment effects at the extensive margin. Column 2 presents treatment effects on the number of months with positive formal earnings within a year. During the pro- gram year, treated youth work in a formal job for almost seven additional months, compared to fewer than three months for control compliers. Employment effects at the extensive margin grow steadily after the program, becoming statistically significant and stabilizing by Year 4. From Year 4 onward, treated youth work approximately one-third of a month more in formal jobs than control compliers, equivalent to roughly a 5 percent increase. Treatment effects are somewhat weaker when employment is measured at the extensive margin using an indicator for positive earnings in any month over 12 months (Column 3). Still, the long-run employment effects observed in Year 7 are positive and statistically significant, reflecting a 2 percentage point increase in the probability of being employed — equivalent to 3 percent of the control complier mean. 2.3 Wage Effects Column 4 shows treatment effects on monthly wages, conditional on employment, indicating that the intensive margin also contributes to the long-term earnings gains. Monthly wages are defined as total earnings divided by the number of months with positive earnings, restricting the sample to individuals with at least one month of positive earnings within a year. During the program year, we find negative treatment effects, consistent with the program offering part-time jobs. Af- 9 ter the program, the effects are consistently positive and statistically significant from Year 2 onward. The short-term effects persist in the long run. By Year 7, treated youth earn $44 more than control compliers, whose average monthly wages are $709, representing a 6 percent increase. To address potential selection into employment, we follow Lee (2009) to estimate bounds on the treatment effects on monthly wages. Table A13 in the Online Ap- pendix shows that, for most years, the treatment effects on wages remain statisti- cally significant after accounting for selection. 2.4 Effects on Labor Contract Types Next, we explore whether participation in the program has long-term effects on the type of contracts that workers hold. Different contractual arrangements can in- fluence job stability and earnings potential, which may help explain the observed treatment effects on earnings. In Uruguay, dependent employees can be hired as regular workers, who are paid a fixed monthly salary, or as daily/hourly workers, who are paid per day or hour of work. Column 5 reports treatment effects on the number of months working under a regular contract within a year. As expected, given that the program jobs offer regular contracts, treatment effects during the program year are large: treated individuals hold a regular job for over seven ad- ditional months, compared to fewer than two months on average among control compliers. After the program, treatment effects on the number of months with a regular job are positive and statistically significant at all horizons. By Year 7, treated youth work in a regular job for 0.6 additional month, relative to an average of six months among control compliers — a 10 percent increase. Because regular jobs typically yield higher earnings, increased regular job tenure may help explain the treatment effects on earnings. In Online Appendix Table A4, we estimate the returns to regular employment by regressing yearly earnings on the number of months worked under a regular contract, controlling for covariates. To avoid confounding treatment effects, we restrict the sample to control group workers. Each additional month under a regular contract is associated with a positive and statistically significant increase in average yearly earnings. By Year 7, each additional month with a regular contract is associated with nearly $700 higher yearly earnings, a 35 percent increase relative to the average earnings of workers with non-regular contracts. These findings suggest that part of the earnings gains for treated individuals stems from greater access to regular, better-paying jobs. 10 2.5 Comparison with Existing Evidence Our findings are next compared with non-experimental studies examining the long-term effects of in-school work and youth employment programs. First, we focus on a few non-experimental studies that specifically investigate the long-term effects of working while studying in the US. Ruhm (1997) examines the returns to working while in high school up to nine years after graduation, controlling for observable differences between employed and non-employed students, while Hotz et al. (2002) study returns up to ten years after graduation. Ashworth et al. (2021), in turn, examine wage returns at age 29 from early work experience in both high school and college. The latter two studies control for dynamic selection into em- ployment. Our estimates are smaller than those of Ruhm (1997), who finds a 22 percent increase in earnings and a 9 percent increase in monthly wages following a 20-hour student job, but they are larger than the non-statistically significant returns reported by Hotz et al. (2002). Our effects on wages are close to those reported by Ashworth et al. (2021) for work experience during college. Hence, our estimates fall within the range of findings for the US labor market. Second, we compare our results to non-experimental evaluations of youth employ- ment programs in the US. Consistent with our findings on long-term employment effects, Scott-Clayton and Minaya (2016) find that the US Federal Work-Study Pro- gram increases the youth employment rate by 2.4 percentage points six years after college entry. Similarly, Aizer et al. (2024) find that participating in the largest youth employment and training program in the US leads to a 5.2 percent increase in lifetime earnings. One key distinction between this study and the existing literature on student em- ployment lies in the nature of the jobs we examine. Unlike much of the research focused on the US, our analysis centers on well-paid positions in state-owned en- terprises that involve complex tasks, such as computer use and report writing, offering greater opportunities for learning and human capital accumulation. Another important feature of the YET program is its requirement that participants continue attending school (a condition also present in Scott-Clayton and Minaya, 2016). This limits educational crowding-out and supports formal human capital accumulation, which we explore next. 11 3 Potential Mechanisms We study two potential channels driving the persistent earnings effects of the work- study program. The first main channel relates to accumulated work experience. The second channel relates to formal education. Starting with the education channel, formal education raises workers’ human capi- tal and has lasting effects on labor market earnings. The program effects on educa- tion are a priori ambiguous: the classical crowding-out effect of work on study may be offset by a crowding-in effect due to the program enrollment condition. Le Bar- banchon et al. (2023) find evidence of stronger crowding-in in the short run, as the program increases the number of years of high-school education by 0.17 year, with no effect on college enrollment. We confirm these results using additional data: in Online Appendix Table A5, we find no effect on university completion seven years after the program. We thus conclude that the program increases educational attainment by 0.17 year. Based on a Mincerian wage regression using 2022 house- hold survey data for workers aged 25 to 50, this would correspond to a 2 percent increase in earnings. In turn, work experience can raise earnings through various channels, including on-the-job training, learning from coworkers, and improved job matching.8 The extent to which these effects contribute to long-term earnings depends on how quickly returns to experience decline. If returns diminish rapidly, as both par- ticipants and non-participants accumulate work experience, we would expect the earnings of program participants to eventually converge with those of the control group. Using 2022 household survey data for workers aged 25 to 30, we estimate — based on a Mincerian wage regression — that an additional 0.7 year of cumulative experience is associated with a 2 percent increase in earnings.9 Overall, these back-of-the-envelope computations suggest that the two potential channels contribute equally to the 10 percent earnings effect. This evidence is only suggestive, as the experience-based estimate does not fully account for the education channel, and vice versa. Importantly, the extent to which skills learned through education or work expe- 8 See Adhvaryu et al. (2023) for the role of on-the-job training, Demir et al. (2024) for learning from coworkers, and Cahuc et al. (2021) for signaling. 9 We compute the additional cumulative experience from Column 2 of Table 1. This is calculated by summing the treatment effects on months with positive earnings from years one to seven after the program. 12 rience are transferable across sectors is key to explaining the earnings effects. If long-term effects were primarily concentrated in the sectors where participants worked during the program, this would suggest that the program experience pro- vides sector-specific skills. Conversely, if significant effects appeared in other sec- tors, it would indicate that the skills acquired are broadly applicable and valued across different labor market contexts. The employment register dataset classifies employer industries according to the International Standard Industry Classification (ISIC Rev. 4). To analyze earnings effects across sectors, we group firms into four broad categories: Agriculture, Man- ufacturing and Energy Production, Market Services, and Non-Market Services. In Table 2, we show that during the program year, the largest treatment effects are observed in the Manufacturing and Energy Production sector, which includes ma- jor program employers such as the state-owned electricity and water companies. The second-largest effects are found in the Market Services sector, which includes the national commercial bank of Uruguay, representing 10 percent of program jobs. This sector also employs the majority of control group youth during the program year, as it includes retail trade. In contrast, treatment effects on earnings in the Non-Market Services sector are smaller and primarily driven by the limited num- ber of public administration positions offered through the program. Seven years after program completion, treatment effects on earnings remain con- centrated in the Manufacturing and Energy Production sector, while no significant effects persist in Non-Market Services. The long-term effects in the Market Ser- vices sector, however, are larger in absolute value than those in Manufacturing and Energy Production. This shift suggests that skills acquired in Manufacturing and Energy Production, and in Non-Market Services are transferable and increasingly valued in Market Services. 4 Heterogeneous Effects Next, we examine heterogeneous effects across youth characteristics: gender, eco- nomic vulnerability status, age and education (all measured at application). Since treatment effects on earnings stabilize around Year 3 (see Table 1), to increase statis- tical power, we pool data from Years 3 to 7 to study treatment effect heterogeneity. We first investigate whether treatment effects differ by gender. Panel (a) of Figure 2 shows that, while both men and women experience statistically significant long- 13 term treatment effects on earnings, the effects are more than double for men, and this difference is statistically significant. Online Appendix Table A6 confirms this heterogeneity in relative terms: the average treatment effect on earnings over Years 3–7 amounts to 8 percent for women but reaches 17 percent for men. While most of the literature on working while studying focuses on men (e.g., Hotz et al., 2002; Ashworth et al., 2021), this finding contrasts with the limited existing evidence (e.g., Ruhm, 1997) and the broader literature on ALMPs, which generally finds stronger effects for women (Card et al., 2018). Second, we explore heterogeneous treatment effects on earnings by vulnerability status (see Panel (b) in Figure 2). Households receiving benefits from the condi- tional cash transfer program in Uruguay (AFAM-PE) are categorized as vulnerable, as eligibility is determined by a poverty score. Our findings highlight two main points. First, we find statistically significant positive effects for non-vulnerable treated individuals, while many ALMPs are primarily targeted at disadvantaged or low-skilled workers. Hence, our findings suggest that broadening access beyond disadvantaged groups could still yield meaningful labor market gains. Second, al- though the difference is not statistically significant, our point estimates suggest that treatment effects may be larger for vulnerable youth. As shown in Table A7 in the Online Appendix, treatment effects on earnings for non-vulnerable youth amount to 10 percent in Years 3–7, while those of vulnerable youth amount to 19 percent. This latter finding aligns with prior evidence on youth employment programs (Scott-Clayton and Minaya, 2016) and, more broadly, with ALMPs (Es- cudero, 2018), which indicate that such policies tend to be particularly beneficial for low-income or low-skilled individuals. Finally, Panels (c) to (e) of Figure 2 present treatment effects on earnings by age and education at baseline. While we are underpowered to detect statistically significant differences, our results suggest that individuals aged 19 in either academic high school or technical school, and those aged 18 in university, exhibit larger treatment effects. This suggestive finding is consistent with effects being more pronounced for individuals at the margin of transitioning between educational levels. 5 Cost-Benefit Analysis To assess the cost-effectiveness of the work-study program, we calculate the Marginal Value of Public Funds (MVPF), following the approach of Hendren and Sprung- 14 Keyser (2020). The MVPF is defined as the ratio of the benefits to recipients, mea- sured by their willingness to pay (WTP) for the program, to the net cost to the government. We estimate both a long-run MVPF based on observed data up to seven years after the program and a projected MVPF over the life cycle. The net costs include two main components: the direct costs during the program year and changes in tax revenues driven by changes in earnings. The direct costs consist of the average net salary paid to participants ($2,223). During the program year, the government incurs an additional loss of $258 in tax revenues, as some treated youth would have worked in the absence of the program.10 Increased par- ticipants’ earnings after the program result in higher tax revenues, $943 in present value over seven years, reducing the total net cost to $1,538.11 The WTP of beneficiaries is measured by changes in net earnings. This approach assumes that any increase in earnings is not driven by higher effort. If increased effort were a factor, incorporating it into the estimation would lead to a lower WTP. Conversely, if individuals derived non-monetary benefits from their increased work, accounting for this would lead to a higher WTP.12 The discounted increase in nominal earnings during the first seven years following the program amounts to $4,841. By applying annual tax and transfer rates, we calculate a WTP of $3,193. When combining the net cost with the WTP, we obtain an MVPF of 2.1 over the seven-year horizon. For the life-cycle analysis, we project lifetime earnings effects using data from the 2022 Uruguay household survey (ECH). This analysis relies on two key assump- tions: (i) the ratio of average earnings in the control group at age 26 — seven years post-program — to average earnings for the same cohort in the ECH remains con- stant throughout the life cycle, and (ii) the percentage increase in earnings for the treatment group remains constant over the life cycle (Hendren and Sprung-Keyser, 2020). We first observe that the ratio specified in assumption (i) is approximately 1.02, indicating that average earnings in the control group are only 2 percent higher than those of the same cohort in the ECH survey. 10 On the other hand, program participants likely generated value through their contributions during the program, which is not accounted for in our estimates due to the lack of information on the services they provided. 11 Earnings are first deflated to January 2016 using the CPI, then converted to US dollars, and finally discounted using a 3 percent real interest rate. 12 Survey evidence from the 5th edition of the work-study program showed an increase in job satisfaction among participants (Le Barbanchon et al., 2023), which could potentially translate into higher well-being even after program completion. 15 Based on these assumptions, the present value of additional earnings from ages 27 (eight years after the program) to 65 is $18,316. This results in an estimated fiscal gain of $7,137 in tax revenues. After accounting for all program costs, the total net fiscal impact is -$5,599, indicating full recovery of the initial investment. This im- plies an infinite MVPF, as the willingness to pay is strictly positive.13 Furthermore, under these assumptions, the program is projected to fully repay its initial costs (i.e., to have zero net cost to the government) 15 years after participation, at age 34. We compare our results with the findings provided by Hendren and Sprung-Keyser (2020). Online Appendix Figure A2 displays the MVPF for several US-based job training programs, alongside estimates for YET, the program we study, up to 8 and 21 years after completion. The Uruguayan work-study program performs substan- tially better than most job training programs targeting beneficiaries of similar age (18 to 21). Furthermore, the projected infinite MVPF of YET is comparable to those of US public policies targeting children, such as investments in early childhood education, child health insurance, and college access. 6 Conclusions In this study, we provide the first experimental evaluation of the long-term effects of a work-study program that offers young students formal work experience in state-owned companies. Leveraging a randomized lottery design, we find that program participants experience a sustained increase in formal earnings of between 10 and 13 percent, beginning in the third year and persisting through seven years after completion. This effect is driven by both employment and monthly wages gains. The benefits are particularly pronounced for individuals who participate in the program at crucial educational stages, as well as for men and vulnerable youth. Our findings suggest that well-designed work-study programs can be a cost-effective tool for facilitating school-to-work transitions, particularly in contexts with high youth unemployment. By integrating education with early exposure to the labor market, such programs can shape career trajectories and generate long-term ben- efits. Future research could explore whether similar effects hold in private sector work-study arrangements. 13 See the MVPF by projected age in Online Appendix Figure A1. 16 References Adhvaryu, A., N. Kala, and A. Nyshadham (2023): “Returns to On-the-Job Soft Skills Training,” Journal of Political Economy, 131, 2165–2208. Agarwal, N. and S. Mani (2025): “New Evidence on Vocational and Appren- ticeship Training Programs in Developing Countries,” in The Handbook of Experi- mental Development Economics, edited by Utteeyo Dasgupta and Pushkar Maitra. Cheltenham: Edward Elgar Publishing Ltd. Aizer, A., N. Early, S. Eli, G. Imbens, K. Lee, A. Lleras-Muney, and A. Strand (2024): “The Lifetime Impacts of the New Deal’s Youth Employment Program,” The Quarterly Journal of Economics, 139, 2579–2635. Alves, G. and J. Varvasino (2025): “Lasting Scars: Unemployment, Inequality, and Structural Transformation in Latin America,” SSRN Electronic Journal 5258234. Ashworth, J., V. J. Hotz, A. Maurel, and T. Ransom (2021): “Changes across Cohorts in Wage Returns to Schooling and Early Work Experiences,” Journal of Labor Economics, 39, 931–964. Attanasio, O., A. Guari ´n, C. Medina, and C. Meghir (2017): “Vocational Train- ing for Disadvantaged Youth in Colombia: A Long-Term Follow-Up,” American Economic Journal: Applied Economics, 9, 131–143. Aucejo, E. M., A. S. Perry, and B. Zafar (2024): “Assessing the Costs of Balancing College and Work Activities: The Gig Economy Meets Online Education,” NBER Working Paper 32357. Bandiera, O., V. Bassi, R. Burgess, I. Rasul, M. Sulaiman, and A. Vitali (forth- coming): “The Search for Good Jobs: Evidence from a Six-Year Field Experiment in Uganda,” Journal of Labor Economics. Blakemore, S.-J. and T. W. Robbins (2012): “Decision-Making in the Adolescent Brain,” Nature Neuroscience, 15, 1184–1191. Blattman, C. and L. Ralston (2015): “Generating Employment in Poor and Frag- ile States: Evidence from Labor Market and Entrepreneurship Programs,” SSRN Electronic Journal 2622220. Cahuc, P., S. Carcillo, and A. Minea (2021): “The Difficult School-to-Work Tran- sition of High School Dropouts: Evidence from a Field Experiment,” Journal of Human Resources, 56, 159–183. Card, D., J. Kluve, and A. Weber (2018): “What Works? A Meta Analysis of Re- cent Active Labor Market Program Evaluations,” Journal of the European Economic Association, 16, 894–931. 17 Carranza, E. and D. McKenzie (2024): “Job Training and Job Search Assistance Policies in Developing Countries,” Journal of Economic Perspectives, 38, 221–244. Demir, G., F. Hertweck, M. Sandner, and I. Yukselen (2024): “Coworker Net- works from Student Jobs: A Flying Start at Labor Market Entry?” SSRN Elec- tronic Journal 5065244. Escudero, V. (2018): “Are Active Labour Market Policies Effective in Activating and Integrating Low-Skilled Individuals? An International Comparison,” IZA Journal of Labor Policy, 7, 4. Hendren, N. and B. Sprung-Keyser (2020): “A Unified Welfare Analysis of Gov- ernment Policies,” The Quarterly Journal of Economics, 135, 1209–1318. Hotz, V. J., L. C. Xu, M. Tienda, and A. Ahituv (2002): “Are There Returns to the Wages of Young Men from Working While in School?” The Review of Economics and Statistics, 84, 221–236. ´ , P., J. Kluve, L. Ripani, and D. Rosas Shady (2019): “Experimental Ibarraran Evidence on the Long-Term Effects of a Youth Training Program,” ILR Review, 72, 185–222. Instituto Nacional de Estadistica Uruguay (2022): “Encuesta Continua de Hogares 2022,” Database retrieved at https://www.ine.gub.uy/web/guest/ encuesta-continua-de-hogares1. Le Barbanchon, T., D. Ubfal, and F. Araya (2023): “The Effects of Working While in School: Evidence from Employment Lotteries,” American Economic Journal: Applied Economics, 15, 383–410. Lee, D. S. (2009): “Training, Wages, and Sample Selection: Estimating Sharp Bounds on Treatment Effects,” The Review of Economic Studies, 76, 1071–1102. Lee, D. S., J. McCrary, M. J. Moreira, and J. Porter (2022): “Valid T-Ratio Infer- ence for IV,” American Economic Review, 112, 3260–3290. McKenzie, D. (2017): “How Effective are Active Labor Market Policies in Devel- oping Countries? A Critical Review of Recent Evidence,” World Bank Research Observer, 32, 127–154. Ruhm, C. J. (1997): “Is High School Employment Consumption or Investment?” Journal of Labor Economics, 15, 735–776. Schwandt, H. and T. von Wachter (2019): “Unlucky Cohorts: Estimating the Long-Term Effects of Entering the Labor Market in a Recession in Large Cross- Sectional Data Sets,” Journal of Labor Economics, 37, S161–S198. 18 Scott-Clayton, J. and V. Minaya (2016): “Should Student Employment be Subsi- dized? Conditional Counterfactuals and the Outcomes of Work-study Participa- tion,” Economics of Education Review, 52, 1–18. Sebastian, C., E. Viding, K. D. Williams, and S.-J. Blakemore (2010): “Social Brain Development and the Affective Consequences of Ostracism in Adoles- cence,” Brain and Cognition, 72, 134–145. 19 Figures Figure 1: Quarterly Earnings 2000 800 1500 600 Earnings (USD) ToT effect 1000 400 200 500 0 0 App -1y Application First Jobs (FJ) FJ +1y +2y +3y +4y +5y +6y +7y +8y ToT Effect 95% CI Complier Control Mean Complier Treat Mean Source: Administrative data and YET Application Form. Notes: This figure plots the evolution of quarterly treatment effects (left Y-axis), and of average quarterly earnings by treatment group (right axis). We use blue dots to report treatment effects, and red vertical lines for their 95 percent confidence intervals. The dashed yellow (resp. solid green) line reports quarterly earnings for the treated individuals (resp. compliers in the control group). 20 Figure 2: Treatment Effect Heterogeneity by Baseline Characteristics. Average Ef- fect Years 3 — 7 (a) By Gender (b) By Financial Vulnerability 1500 1000 800 1000 Earnings Earnings 600 500 400 200 0 Men Women Not Vulnerable Vulnerable Gender Gender (c) By Age for Academic High Schoolers (d) By Age for Technical High Schoolers 4500 4500 3000 3000 1500 1500 Earnings Earnings 0 0 -1500 -1500 -3000 -3000 -4500 -4500 16 17 18 19 20 16 17 18 19 20 Age at aplication Age at aplication (e) By Age for University Students 4500 3000 1500 Earnings 0 -1500 -3000 -4500 18 19 20 Age at aplication Source: Administrative data and YET Application Form. Notes: This figure shows treatment effects by gender, household vulnerability, and age and edu- cation at application date. Vulnerable households include households receiving a cash transfer. Treatment effects are obtained by two stage least squares regressions of Equation (1), where we further interact the treatment dummy with, respectively, a gender, vulnerability, and age dummy. Vertical lines represent 95 percent confidence intervals. 21 Tables Table 1: Effect of YET on Labor Outcomes (1) (2) (3) (4) (5) Total Months with Positive Wages Months with earnings earnings earnings regular job Program year Year 0 2158.36 6.86 0.56 -20.14 7.64 (41.72) (0.08) (0.01) (3.15) (0.07) [1141.28] [2.73] [0.44] [360.58] [1.80] Post-Program years Year 1 92.87 0.01 0.05 7.64 0.42 (75.17) (0.12) (0.01) (7.48) (0.11) [2075.98] [4.35] [0.58] [421.70] [3.01] Year 2 218.96 0.05 0.02 25.65 0.35 (93.75) (0.13) (0.01) (8.49) (0.12) [2873.93] [5.35] [0.65] [481.45] [3.85] Year 3 341.43 0.16 0.01 34.59 0.36 (108.27) (0.13) (0.01) (9.16) (0.13) [3547.53] [6.00] [0.69] [534.54] [4.42] Year 4 512.31 0.33 0.03 38.31 0.55 (123.41) (0.13) (0.01) (10.41) (0.13) [4295.68] [6.56] [0.71] [596.57] [4.99] Year 5 610.84 0.32 0.02 45.90 0.61 (136.61) (0.13) (0.01) (11.41) (0.13) [4699.80] [6.74] [0.73] [636.71] [5.25] Year 6 669.94 0.36 0.01 57.77 0.55 (148.39) (0.13) (0.01) (12.59) (0.14) [5183.92] [7.06] [0.75] [669.72] [5.61] Year 7 652.00 0.31 0.02 43.71 0.58 (159.64) (0.13) (0.01) (13.42) (0.14) [5761.85] [7.46] [0.76] [708.89] [5.93] Observations 90,423 90,423 90,423 67,793 90,423 Source: Administrative data and YET Application Form. Notes: Two-stage least squares regressions where we instrument the YET participation dummy with a job of- fer dummy. Controls for lottery design (lottery and quota dummies) and number of applications are included. Covariates include gender, a dummy for age 18 or less at application, a dummy for receiving cash transfers, baseline earnings, and dummies for baseline education type. Total earnings: total labor income over 12 months, winsorized at the top 1 percent of positive values, adjusted for inflation using the CPI and converted into US dollars. Months with earnings: number of months over 12 months with positive income. Positive earnings: indica- tor for positive earnings in any month over 12 months. Wages: Total earnings divided by Months with earnings; it is missing for those who have not worked any month over the 12 months. Standard errors robust to het- eroskedasticity shown in parentheses, and control complier means in brackets. Months with regular job: number of months over 12 months with regular contract. 22 Table 2: Effect of YET on Earnings by Aggregate Sector (1) (2) (3) (4) Agriculture Manufacturing Market Services Non-Market & Energy Services Program year Year 0 -28.53 1493.12 518.11 177.35 (7.84) (38.76) (46.87) (29.05) [35.67] [188.57] [755.37] [160.05] Post-Program years Year 1 -20.78 56.90 121.87 -65.10 (13.50) (40.81) (65.15) (36.50) [64.42] [321.15] [1341.07] [348.30] Year 2 -15.65 120.96 170.82 -55.07 (19.48) (54.07) (81.00) (50.09) [74.57] [410.21] [1835.93] [550.33] Year 3 -0.93 149.99 220.46 -19.20 (22.35) (63.13) (93.70) (64.46) [90.49] [493.52] [2177.49] [781.42] Year 4 -7.31 215.50 326.64 -9.80 (26.81) (73.73) (106.72) (80.40) [104.53] [547.17] [2502.96] [1139.18] Year 5 -30.50 250.68 353.63 44.20 (26.27) (81.27) (115.99) (94.50) [128.69] [574.78] [2541.96] [1451.32] Year 6 -29.47 272.57 407.91 28.59 (30.88) (90.37) (125.15) (104.73) [130.46] [619.40] [2669.59] [1761.55] Year 7 -30.70 223.33 416.81 44.79 (31.99) (95.37) (135.23) (115.30) [143.25] [695.89] [2869.25] [2048.84] Observations 90,423 90,423 90,423 90,423 Source: Administrative data and YET Application Form. Notes: Two stage least squares regressions where we instrument the YET participation dummy with the offer to take the YET job. In Column (1), the dependent variable is earnings in firms belonging to the Agriculture sector. Columns (2) to (4) correspond to the Manufacturing & Energy Sector, Market Ser- vices, and Non-Market Services, respectively. Market Services are sectors where services are typically provided in exchange for payment under competitive market conditions. They include Whole-sale and retail trade, Transportation and storage, Accommodation and food service activities, Information and communication, Financial and insurance activities. Non-Market Services are sectors typically funded or provided by the government, non-profit organizations, or institutions where users do not pay directly. They include Public Administration and Defense, Education, Human Health and Social Work Activi- ties. Controls for lottery design (lottery and quota dummies) are included. Covariates include gender, a dummy for age 18 or less at application, a dummy for receiving cash transfers, baseline earnings and dummies for baseline education type. Earnings are winsorized at the top 1 percent of positive values, adjusted for inflation using the CPI and converted into US dollars. Robust standard errors shown in parenthesis and control complier means in brackets. 23 Online Appendix The Lasting Effects of Working while in School: A Long-Term Follow-Up Mery Ferrando Noemi Katzkowicz Thomas Le Barbanchon Diego Ubfal* August 15, 2025 The Online Appendix includes two sections. Section A contains additional tables and figures. Section B provides further details about the work-study program. * Ferrando: ´ Tilburg University. Katzkowicz: Universidad de la Republica. Le Barbanchon: Bocconi University. Ubfal: World Bank. 1 A Additional Tables and Figures Table A1: Main Features of the Program by Edition Edition 1 2 3 Application Date May 2012 May 2013 May 2014 Applications 46,544 43,661 31,990 Applicants 46,008 42,643 30,969 Job Offers Made 754 981 955 Jobs Started 592 754 718 Jobs Completed 549 686 660 Sector: Civil 0.82 0.73 0.70 Sector: Industry/Trade 0.02 0.04 0.04 Sector: Banking 0.16 0.23 0.26 Localities 51 64 67 Source: Administrative data and YET Application Form. Notes: There is a downward trend in applications over time, probably due to the program spending more resources on advertising in the first two editions, and due to longer lottery registration time windows in the first two editions. However, we do not see any notable trend in applicants’ characteristics over time. 2 Table A2: Balance Between Treatment and Control Groups (1) (2) (3) (4) (5) Control Offered Mean S.D. Mean S.D. p-value Panel A. Demographic Female 0.58 0.49 0.60 0.49 0.15 Aged 16-18 0.71 0.45 0.72 0.45 0.88 Aged 19-20 0.29 0.45 0.28 0.45 0.88 Montevideo (Capital City) 0.49 0.50 0.55 0.50 . Panel B. Education and Social Programs Year -1 Enrolled in Academic Secondary Education 0.49 0.50 0.48 0.50 0.51 Enrolled in Technical Secondary Education 0.22 0.41 0.22 0.42 0.56 Enrolled in University 0.15 0.36 0.16 0.36 0.32 Enrolled in Tertiary Non-University 0.01 0.11 0.01 0.10 0.68 Enrolled in Out-of-School Programs 0.02 0.14 0.02 0.14 0.54 Highly Vulnerable HH (Food Card Recipient) 0.10 0.30 0.09 0.29 0.25 Vulnerable Household (CCT recipient) 0.27 0.45 0.27 0.44 0.72 Panel C. Labor Outcomes Year -1 Earnings (winsorized top 1%, USD) 228.13 800.69 200.21 757.38 0.20 Positive Earnings 0.15 0.36 0.15 0.35 0.84 Months with Positive Earnings 0.71 2.13 0.62 1.97 0.12 Months with Regular Job 0.41 1.63 0.36 1.49 0.09 textbfPanel D. Aggregate orthogonality test for panels A-C p-value (joint F-test) 0.54 Observations 87,737 2,686 90,423 Source: Administrative data and YET Application Form. Notes: The p-value reported in Column 5 is obtained from a regression of each variable on a YET job offer dummy with robust standard errors, controlling for lottery design (lottery and quota dummies) and number of applications. We do not test for differences in means for Montevideo since the lottery was randomized within each locality and we control for lottery design in all our specifications. Vul- nerable households include households receiving a cash transfer and/or a food card (labelled as Highly Vulnerable). We code Enrolled in University by using two indicators available in the administrative data: “entering a new program that year” or “taking at least two exams that year”, for the first edition we do not have data on Year -1 and we use the value as self-reported by participants in the application form. p-value (joint F-test): corresponds to the orthogonality test in a regression of the YET job offer dummy on covariates; the regression also controls for lottery design and number of applications (coefficients not included in the F-test). 3 Table A3: Effect of YET Offer on YET Participation (First Stage) (1) (2) (3) (4) YET Participation All Editions Edition 1 Edition 2 Edition 3 Won Lottery 0.77 0.79 0.77 0.77 (0.01) (0.01) (0.01) (0.01) Observations 90,423 36,181 30,410 23,832 Source: Administrative data and YET Application Form. Notes: OLS regressions of YET participation in Year 0 on the offer to take the YET job (winning the lottery). Controls for lottery design (lottery and quota dummies) and number of applications are included. Covariates include gen- der, a dummy for age 18 or less at application, a dummy for receiving cash transfers, baseline earnings and dummies for baseline education type. Ro- bust standard errors shown in parenthesis. Results for the first edition are obtained with the same method used to select unique applications as in the other editions. Results are almost identical if we keep the first application. 4 Table A4: Returns to Having a Regular Job. Control Group (1) Total earnings Program year Year 0 428.25 (1.79) [462.39] Post-Program years Year 1 459.73 (1.82) [892.37] Year 2 494.76 (1.97) [1238.76] Year 3 533.06 (2.14) [1511.34] Year 4 579.65 (2.36) [1760.67] Year 5 626.53 (2.57) [1852.16] Year 6 666.95 (2.77) [1893.23] Year 7 697.54 (2.98) [1993.21] Observations 87,737 Source: Administrative data and YET Application Form. Notes: The independent variable is defined as number of months with a regular contract. Co- variates include gender, a dummy for age 18 or less at application, a dummy for receiving cash transfers, baseline earnings and dummies for baseline education type. Robust standard errors shown in parenthesis and control means in brackets. 5 Table A5: Effects on Graduating from University (1) Graduated from University Treated -0.00 (0.01) CCM 0.060 Observations 90,423 Source: Administrative data from main university (UDELAR) and YET Application Form. Notes: Two stage least squares regressions where we instrument the YET participation dummy with a job offer dummy. Controls for lottery design (lottery and quota dummies) are included. Covariates include gender, a dummy for age 18 or less at application, a dummy for receiving cash transfers, baseline earnings and dummies for baseline education type. Robust standard errors shown in parenthesis. Table A6: Treatment Effect Heterogeneity by Gender (1) (2) (3) (4) Program Year Post-Program years Year 0 Year 1 Year 2 Years 3-7 Treated (T) 1973.21 19.74 238.99 889.80 (70.59) (134.30) (170.54) (218.28) T * Female 302.29 119.39 -32.70 -542.84 (87.12) (160.39) (201.71) (258.31) Female -268.77 -559.24 -709.43 -903.60 (14.50) (20.34) (24.69) (30.94) CCM Male [1325.65] [2446.60] [3332.39] [5140.69] CCM Female [1024.63] [1841.74] [2584.28] [4418.18] p-value T+T*Female=0 0.00 0.11 0.06 0.01 Observations 90,423 90,423 90,423 90,423 Source: Administrative data and YET Application Form. Notes: Two stage least squares regressions where we instrument the YET participation dummy, and its interaction with a female dummy with a job offer dummy and the cor- responding interaction. Controls for lottery design (lottery and quota dummies) are in- cluded. Covariates include gender, a dummy for age 18 or less at application, a dummy for receiving cash transfers, baseline earnings and dummies for baseline education type. Robust standard errors shown in parenthesis. p-value: p-value of the test that the treat- ment effect for females is zero (sum of the treated and interaction coefficients). 6 Table A7: Treatment Effect Heterogeneity by Household Vulnerability (1) (2) (3) (4) Program Year Post-Program years Year 0 Year 1 Year 2 Years 3-7 Treated (T) 2074.53 56.81 166.59 502.74 (49.46) (90.12) (111.69) (144.58) T * Vulnerable 311.54 134.02 194.62 202.78 (91.08) (160.79) (203.41) (253.40) Vulnerable -120.81 -124.06 -271.02 -820.49 (14.60) (20.76) (25.21) (31.03) CCM Non-Vulnerable [1223.60] [2190.00] [3057.78] [5050.76] CCM Vulnerable [917.65] [1766.47] [2374.90] [3739.71] p-value T+T*Vulnerable=0 0.00 0.15 0.03 0.00 Observations 90,423 90,423 90,423 90,423 Source: Administrative data and YET Application Form. Notes: Two stage least squares regressions where we instrument the YET participation dummy, and its interaction with a vulnerability dummy with a job offer dummy and the correspond- ing interaction. Vulnerable households include households receiving a cash transfer. Controls for lottery design (lottery and quota dummies) are included. Covariates include gender, a dummy for age 18 or less at application, a dummy for receiving cash transfers, baseline earn- ings and dummies for baseline education type. Robust standard errors shown in parenthesis. p-value: p-value of the test that the treatment effect for individuals in vulnerable households is zero (sum of the treated and interaction coefficients). 7 Figure A1: MVPF over the Life Cycle ∞ >5 4 MVPF 3 2 1 0 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 (Projection) Age Source: Administrative and household survey data. Notes: This figure plots the MVPF by age. The vertical line indicates the last age for which the MVPF is based on observed data. 8 Figure A2: MVPF for YET and Job Training Programs. 8- and 21-year Horizon ∞ YET Year Up JOBSTART >5 MVPF (21-Year Horizon) 4 3 2 1 NSW Youth Job Corps 0 <-1 JTPA Youth <-1 0 1 2 3 4 >5 ∞ MVPF (8-Year Horizon) Source: Authors’ calculations based on administrative and household survey data, and estimates from Hendren and Sprung-Keyser (2020). Notes: This figure plots the MVPF at 21 years after the programs against the MVPF 8 years after the programs. The black circles represent job training programs, based on Table C.I of the Online Appendix in Hendren and Sprung-Keyser (2020) (the 8-year estimate for Job Corps is our own calculation based on the study’s replication file). The red diamond represents estimates for YET, the program we study. 9 A.1 Robustness Checks In this section, we test the robustness of our main results to alternative specifications. Specifically, we show that our main results in Table 1 are robust to omitting controls (see Table A8), clustering the standard errors at the locality level (see Table A9), and to not winsorizing the earning variables (see Table A10). Additionally, we provide intention-to-treat estimates (ITT) that do not rely on the ex- clusion restriction, and we obtain consistent results (see Table A11). We also explore an alternative definition of treatment that allows us to estimate a parameter that may be closer to the effect of working while in school, but relies on stronger assumptions. Un- der this alternative specification, we define treatment as working in any firm while being enrolled in school during the program year. Results are even stronger, and overall con- sistent with our main estimates (see Table A12). This alternative specification assumes that the type of in-school job has no effect on future labor and educational outcomes. In particular, it assumes that there are similar effects of program jobs and of the potential control jobs students would have accepted if they had not been offered a program job. Since program jobs are well-paid temporary jobs, we see this alternative specification as less appropriate. Finally, we provide bounds for the ITT effects on monthly wages to account for selection into employment (see Table A13). 10 Table A8: Effect of YET on Labor Outcomes. No controls (1) (2) (3) (4) (5) Total Months with Positive Wages Months with earnings earnings earnings regular job Program year Year 0 2137.31 6.83 0.56 -32.26 7.62 (45.12) (0.08) (0.01) (3.18) (0.07) [1162.33] [2.76] [0.44] [372.70] [1.81] Post-Program years Year 1 70.89 -0.02 0.05 3.26 0.41 (78.11) (0.12) (0.01) (7.77) (0.11) [2097.95] [4.38] [0.58] [426.09] [3.02] Year 2 200.06 0.03 0.02 22.63 0.35 (96.60) (0.13) (0.01) (8.80) (0.13) [2892.83] [5.36] [0.65] [484.47] [3.85] Year 3 327.46 0.15 0.01 32.76 0.37 (111.77) (0.13) (0.01) (9.56) (0.13) [3561.50] [6.00] [0.69] [536.37] [4.42] Year 4 503.80 0.33 0.03 37.31 0.57 (128.16) (0.13) (0.01) (10.86) (0.14) [4304.19] [6.56] [0.71] [597.58] [4.98] Year 5 607.17 0.32 0.02 45.17 0.63 (141.89) (0.13) (0.01) (11.92) (0.14) [4703.47] [6.74] [0.73] [637.44] [5.24] Year 6 672.52 0.36 0.01 59.44 0.57 (154.72) (0.13) (0.01) (13.20) (0.14) [5181.34] [7.06] [0.75] [668.05] [5.59] Year 7 660.41 0.31 0.02 44.56 0.61 (167.31) (0.13) (0.01) (14.16) (0.14) [5753.44] [7.45] [0.76] [708.05] [5.91] Observations 90,423 90,423 90,423 67,793 90,423 Source: Administrative data and YET Application Form. Notes: Replicates Table 1 without including control variables. 11 Table A9: Effect of YET on Labor Outcomes. Clustering at Locality Level (1) (2) (3) (4) (5) Total Months with Positive Wages Months with earnings earnings earnings regular job Program year Year 0 2158.36 6.86 0.56 -20.14 7.64 (186.61) (0.36) (0.04) (8.84) (0.26) [1141.28] [2.73] [0.44] [360.58] [1.80] Post-Program years Year 1 92.87 0.01 0.05 7.64 0.42 (79.49) (0.13) (0.01) (5.80) (0.09) [2075.98] [4.35] [0.58] [421.70] [3.01] Year 2 218.96 0.05 0.02 25.65 0.35 (65.79) (0.09) (0.01) (7.07) (0.08) [2873.93] [5.35] [0.65] [481.45] [3.85] Year 3 341.43 0.16 0.01 34.59 0.36 (78.69) (0.12) (0.01) (5.21) (0.12) [3547.53] [6.00] [0.69] [534.54] [4.42] Year 4 512.31 0.33 0.03 38.31 0.55 (120.05) (0.13) (0.01) (8.80) (0.17) [4295.68] [6.56] [0.71] [596.57] [4.99] Year 5 610.84 0.32 0.02 45.90 0.61 (88.21) (0.08) (0.01) (7.35) (0.09) [4699.80] [6.74] [0.73] [636.71] [5.25] Year 6 669.94 0.36 0.01 57.77 0.55 (111.67) (0.12) (0.01) (8.08) (0.10) [5183.92] [7.06] [0.75] [669.72] [5.61] Year 7 652.00 0.31 0.02 43.71 0.58 (116.75) (0.14) (0.01) (7.96) (0.09) [5761.85] [7.46] [0.76] [708.89] [5.93] Observations 90,423 90,423 90,423 67,793 90,423 Source: Administrative data and YET Application Form. Notes: Replicates Table 1, but clustering the standard errors at the locality level. 12 Table A10: Effect of YET on Labor Outcomes. No Winsoring (1) (2) (3) (4) (5) Total Months with Positive Wages Months with earnings earnings earnings regular job Program year Year 0 2157.97 6.86 0.56 -20.68 7.63 (43.53) (0.08) (0.01) (3.34) (0.07) [1153.62] [2.74] [0.44] [362.12] [1.80] Post-Program years Year 1 116.16 0.01 0.05 10.69 0.42 (79.18) (0.12) (0.01) (8.04) (0.11) [2087.92] [4.35] [0.58] [423.34] [3.01] Year 2 252.09 0.05 0.02 29.63 0.35 (99.32) (0.13) (0.01) (9.24) (0.12) [2890.35] [5.35] [0.65] [483.65] [3.85] Year 3 377.10 0.16 0.01 38.87 0.36 (114.09) (0.13) (0.01) (9.96) (0.13) [3556.65] [6.00] [0.69] [535.61] [4.43] Year 4 555.88 0.33 0.03 43.53 0.55 (130.75) (0.13) (0.01) (11.48) (0.13) [4294.38] [6.56] [0.71] [596.20] [4.99] Year 5 679.03 0.32 0.02 53.38 0.61 (147.25) (0.13) (0.01) (12.80) (0.13) [4693.68] [6.74] [0.73] [636.16] [5.25] Year 6 735.40 0.36 0.01 64.94 0.55 (159.02) (0.13) (0.01) (13.97) (0.14) [5177.61] [7.06] [0.75] [669.12] [5.61] Year 7 720.48 0.31 0.02 50.82 0.58 (173.79) (0.13) (0.01) (15.14) (0.14) [5755.39] [7.46] [0.76] [708.37] [5.93] Observations 90,423 90,423 90,423 67,793 90,423 Source: Administrative data and YET Application Form. Notes: Replicates Table 1, without winsorizing the dependent variables used in Column (1) and Column (4). 13 Table A11: Effect of YET on Labor Outcomes. ITT Effects (1) (2) (3) (4) (5) Total Months with Positive Wages Months with earnings earnings earnings regular job Program year Year 0 1669.87 5.31 0.44 -17.32 5.91 (36.11) (0.08) (0.01) (2.72) (0.08) [1321.06] [3.07] [0.46] [380.18] [2.06] Post-Program years Year 1 71.85 0.01 0.04 5.88 0.33 (58.22) (0.09) (0.01) (5.77) (0.09) [2260.56] [4.57] [0.60] [438.53] [3.20] Year 2 169.41 0.04 0.02 19.74 0.27 (72.66) (0.10) (0.01) (6.54) (0.10) [3005.40] [5.42] [0.66] [495.25] [3.93] Year 3 264.16 0.13 0.01 26.92 0.28 (83.92) (0.10) (0.01) (7.14) (0.10) [3684.42] [6.03] [0.69] [550.99] [4.50] Year 4 396.36 0.26 0.02 29.88 0.43 (95.76) (0.10) (0.01) (8.13) (0.10) [4406.61] [6.53] [0.71] [611.72] [4.99] Year 5 472.60 0.25 0.01 36.12 0.47 (106.00) (0.10) (0.01) (8.99) (0.10) [4868.15] [6.73] [0.72] [657.42] [5.24] Year 6 518.32 0.28 0.01 45.45 0.43 (115.16) (0.10) (0.01) (9.92) (0.11) [5312.69] [6.97] [0.74] [692.67] [5.51] Year 7 504.44 0.24 0.02 34.25 0.45 (123.86) (0.10) (0.01) (10.53) (0.11) [5766.69] [7.24] [0.75] [725.81] [5.77] Observations 90,423 90,423 90,423 67,793 90,423 Source: Administrative data and YET Application Form. Notes: Replicates Table 1, but presents ITT effects rather than ToT effects. Control means are presented in brack- ets. 14 Table A12: Effect of Working and Studying During Program Year (1) (2) (3) (4) (5) Total Months with Positive Wages Months with earnings earnings earnings regular job Program year Year 0 3787.65 12.04 0.99 -97.66 13.40 (85.69) (0.20) (0.02) (15.80) (0.23) [-118.74] [-0.68] [0.01] [349.87] [-1.16] Post-Program years Year 1 162.97 0.01 0.09 17.62 0.74 (131.59) (0.20) (0.02) (17.24) (0.20) [1279.84] [2.87] [0.43] [390.14] [1.84] Year 2 384.25 0.09 0.04 51.36 0.62 (164.00) (0.22) (0.02) (16.99) (0.22) [2182.46] [4.41] [0.57] [441.17] [3.09] Year 3 599.17 0.28 0.02 66.62 0.64 (189.62) (0.23) (0.02) (17.71) (0.23) [2794.00] [5.16] [0.62] [483.85] [3.67] Year 4 899.04 0.58 0.05 72.40 0.97 (216.38) (0.23) (0.02) (19.72) (0.23) [3533.76] [5.86] [0.65] [548.70] [4.26] Year 5 1071.96 0.57 0.03 85.02 1.07 (239.73) (0.23) (0.02) (21.22) (0.23) [3905.02] [6.15] [0.68] [581.14] [4.61] Year 6 1175.67 0.63 0.01 106.06 0.97 (260.82) (0.23) (0.02) (23.26) (0.24) [4338.97] [6.55] [0.71] [604.80] [5.03] Year 7 1144.19 0.54 0.04 79.27 1.02 (280.25) (0.22) (0.02) (24.38) (0.24) [5145.99] [7.15] [0.74] [659.00] [5.47] Observations 90,423 90,423 90,423 67,793 90,423 Source: Administrative data and YET Application Form. Notes: Two stage least squares regressions where we instrument a dummy variable taking the value of one if youth work (positive yearly earnings) and study (enrolled at any level) during the program year with the offer to take the YET job. Controls for lottery design (lottery and quota dummies) are included. Covariates include gender, a dummy for age 18 or less at application, a dummy for receiving cash transfers, baseline earnings and dummies for baseline education type. Robust standard errors shown in parenthesis and control complier means in brackets. The control complier mean is obtained as the difference between the average outcome for compli- ers offered a YET job and the estimated local average treatment effect. To recover the former from the data we assume that the average outcome for and the share of always takers are the same among those offered and not offered a YET job. 15 Table A13: Bounds for the ITT Effects on Monthly Wages (Post-program Years) (1) (2) (3) (4) ITT effect Lee bounds Imbens and Manski on wages on wage effects 95% Confidence Interval Lower Upper Lower Upper Year 1 5.88 -27.66 28.93 -35.86 38.18 (5.77) (4.99) (5.62) [438.53] Year 2 19.74 17.90 29.07 7.27 39.73 (6.54) (6.46) (6.48) [495.25] Year 3 26.92 25.78 31.83 14.10 43.51 (7.14) (7.10) (7.10) [550.99] Year 4 29.88 6.98 41.00 -5.19 54.30 (8.13) (7.40) (8.08) [611.72] Year 5 36.12 33.78 41.72 19.13 56.49 (8.99) (8.91) (8.98) [657.42] Year 6 45.45 43.59 52.30 27.40 68.58 (9.92) (9.84) (9.89) [692.67] Year 7 34.25 2.25 47.28 -13.30 64.59 (10.53) (9.45) (10.53) [725.81] Observations 90,423 Source: Administrative data and YET Application Form. Notes: This table presents bounds on causal effect on wages for the “always employed” (indi- viduals who would be employed regardless of whether they are offered the program job or not) based on the procedure described in Lee (2009). To obtain the upper bound, we trim the sam- ple of observed wages in the offered group with the p percent lower wages, where p is the ratio of the ITT effect on employment over the employment rate on the offered group. The lower bound is the symmetric case where we trim the p percent higher wages. Robust standard er- rors shown in parenthesis and control means in brackets. We follow Imbens and Manski (2004) to construct confidence intervals for the bounds. 16 B Institutional Details: The Work-Study Program The work-study program “Yo Estudio y Trabajo” (YET) operates in 77 localities across Uruguay, covering nearly all major cities. Applicants to the program can complete the application online or at an employment center. To participate, selected applicants must present proof of enrollment from an educational institution showing at least 240 hours of attendance, a valid national ID, and, if over 18, an electoral card. Eligibility is verified by cross-checking social security data to confirm the applicant is not formally employed. Students must also provide updated enrollment proof every three months. Those aged 16–17 receive guidance on how to obtain a work permit. As of January 2016, participants earn a fixed monthly salary of 13,360 pesos (equivalent to four times the minimum tax unit) for 30 hours of work per week. Pregnant women and mothers with children under age 4 — who make up about 4 percent of lottery applicants — receive 50 percent higher wages. Students can reapply in future rounds under specific rules. Those who begin a job through the program cannot apply again, while those who are selected but do not take up a position may reapply, but do not receive any priority. 17 References Hendren, N. and B. Sprung-Keyser (2020): “A Unified Welfare Analysis of Government Policies,” The Quarterly Journal of Economics, 135, 1209–1318. Imbens, G. and C. Manski (2004): “Confidence Intervals for Partially Identified Param- eters,” Econometrica, 72, 1845–1857. Lee, D. S. (2009): “Training, Wages, and Sample Selection: Estimating Sharp Bounds on Treatment Effects,” The Review of Economic Studies, 76, 1071–1102. 18