Jorge Rodriguez, Fernando Saltiel, and Sergio Urzuar "Dynamic Treatment Effects of Job Training", Journal of Applied Econometrics, Vol. 37, No. 2, 2022, pp. 242-269. To recover training histories and associated labor market outcomes, we construct a novel database that merges three different sources of information. First, we take advantage of administrative records from Franquicia Tributaria. Using this data source, we construct workers' training histories by observing their participation in FT-subsidized courses from 1998 through 2010. The data do not include information on the types of courses taken by workers. As such, we are unable to directly model the impact of different course types and intensities on labor market outcomes. We analyze labor market outcomes using information from Chile's Unemployment Insurance (UI) system. UI data registers workers' monthly earnings and the firm of employment for all workers with formal sector contracts. We focus on the worker's main employment stint in each quarter and examine earnings in the first quarter of the year following the training event. Our final source of information comes from performance in a college-entry examination (PSU), which is a mandatory test for all students who wish to enter a post-secondary institution. We observe PSU scores for all high school graduates who took the test between 2000 and 2007. Using individual identifiers, we recover PSU scores of workers to supplement our data of labor market outcomes and training choices. We work with standardized PSU test scores (computed separately by year). The PSU database also includes information on student's observable characteristics, such as gender, age, parental education, family size, and parental employment at the time of the test. To circumvent threats to identification and for computational tractability, we restrict our sample in several ways. First, we focus our attention on the returns to multiple job training courses for young workers who are first-time labor entrants. We impose this restriction as we do not observe training histories before 1998; if training choices depend on prior training decisions, then we would be omitting a relevant variable---past training---in the choice equation. We use the sample of young workers, identify their first year of employment, and follow their labor market history thereafter---by definition, their job training history in our first period is zero. Second, for tractability, we restrict our analysis to training stints during their first two years in the labor force and examine extensive-margin training decisions on a yearly basis. As a result, workers are trained at most twice during our period of interest. Third, we restrict the sample to individuals who are eligible to participate in training financed by FT---that is, individuals who work in the formal sector. As our analysis of worker self-selection into training requires workers to be able to take part in courses each year if they want to, we limit the sample to individuals who are employed for at least nine months in each of their first two years in the labor force in firms with at least ten employees. Since UI data indicates that 90% of formal-sector employment in Chile is at firms with at least ten employees, this restriction is not necessarily binding. In this way, we analyze a group of workers who are effectively eligible for training each year. By doing so, we abstract from analyzing effects of training on employment. Our choice of sample selection limits the scope of our results: by focusing on young workers starting their careers, our results on static and dynamic returns to job training might not generalize to human capital investments later in life. Overall, we focus on labor market entrants from years 2003-2008 and their training choices two years after entry. Our final sample consists of 37,089 workers who meet all of the above criteria. See the main text for summary statistics. ***** The data used in this paper are the property of the Ministry of Finance (Chile) who do not permit open access, hence the data cannot be lodged here. Nonetheless, access to the data is not exclusive. Permission to use this data source should be directed to the Chile Ministry of Finance: consultores@hacienda.cl. Access is granted after submitting a formal data request (see pdf file "Protocolo de Acceso y Seguridad al Servidor Externo del Ministerio de Hacienda") The data are held on the Ministry's server, to whom enquiries concerning access should be addressed. Information concerning the Archive holding can be obtained using the following link: https://www.hacienda.cl/consultas-web Our analysis considers three sources of information: "Franquicia Tributaria" (FT): It contains individual-level information on FT's courses from 1998 through 2010. The file's name is acc_participante_dv_2.dta. "Seguro de Cesantia": UI individual-level records containing monthly earnings for the period 2003 to 2020. Each month the data gets updated, which implies individual-level information from each month is added to the master files. This implies that complete labor market histories can be obtained from the latest version of the UI data. At the time of writing the latest year held is 2013 and we used data from 2003 to 2010. The names of the relevant files are: afiliados_11_2013.dta cuentas_11_2013.dta empleadores_11_2013.dta remuneraciones_11_2013.dta "Proceso de Admisión a Universidade" (PSU). College admission test scores containing individual-level results for the period 2000 and 2007. (Note: This information was also published in newspapers and it can also be accessed through the National Library's archives: https://www.bibliotecanacional.gob.cl/sitio/) The variables used in our study and sample selection rules are described above. The folder "Model_Estimates" contains the estimated model estimates. The folder "Simulation_Codes" contains our main results (including do files). See master.do for the sequence of statistical procedures (and associated codes). The folder "Model_Codes" contains the FORTRAN codes used to generate the main results in the paper. We include codes and doc illustrating the estimation of the model presented in Appendix D of our paper. data_setup.do generates the simulated data. The implementation of the model is described in the accompanying PDF document. In that document, we show convergence of the factor loadings in the model. The model was compiled in Visual Studio 2019 on Windows 10. Similar estimation and computational strategies were used in Hansen, Heckman and Mullen (2004); Cunha, Heckman and Navarro (2005); Heckman, Stixrud and Urzua (2006); Urzua (2008); Prada and Urzua (2017). All these files are zipped in the file sru-files.zip. It also includes the file "Protocolo de Acceso y Seguridad al Servidor Externo del Ministerio de Hacienda". Jorge Rodriguez Fernando Saltiel Sergio Urzua