ABSTRACT – Large youth unemployment rates are a structural characteristic of most economies. In Brazil this problem is accompanied by very high youth labor turnover rates, which suggests that the job destruction margin plays an important role in explaining high unemployment rates for this age group. Our main objective in this paper is to provide a first evaluation of a large subsidized training program (Lei do Aprendiz, Apprentice’s Act) targeted to young workers. The program was introduced in 2000 in Brazil with the intention to help the placement of young workers and their attachment to formal jobs. We make use of a huge longitudinal dataset (Rais, Relatório Annual de Informações Sociais), based on administrative data collected by the Labor Ministry, that contains information on the employment histories of all formal workers in Brazil from 2001 to 2006. We attempt to measure the impact of the program on five different outcomes that represent formal labor market attachment and remuneration, using other temporary workers as a control group. We employ three distinct estimation procedures to deal with self-selection in program participation that exploit a discontinuity in its age requirement: 18 year olds could not participate in the program between 2000 and 2005. We use a standard 2SLS, which we denote as parametric IV and two recently proposed estimators: i) a semi-parametric IV due to Battistin and Rettore (2008), and ii) the adjusted matching estimator proposed by Dias et al. (2010) that corrects the standard matching approach with an IV estimated correction term based on a sharp observed cutoff criterium. Our results indicate that, compared to other temporary workers, apprentices have a higher probability of getting a formal job in the years after the program and a higher probability of getting a non-temporary contract. On the other hand, our estimates suggest that treated workers get jobs with lower tenure than other temporary workers. We also find positive (and significant) effects on wages but very small in magnitude. These results are robust, holding for the whole set of estimation procedures.