In this paper, we estimate returns to classroom and on-the-job firm-sponsored training in terms of value-added per worker using longitudinal linked employee-employer Canadian data from 1999 to 2006. We estimate a standard production function controlling for endogenous training decisions because of perceived net benefits and time-varying market conditions using dynamic panel GMM methods. We find that employees who undertook classroom training are 11 percent more productive than otherwise similar employees. We show that returns to on-the-job training are on average lower (3.4 percent). We provide evidence that these lower returns are due to on-the-job training being more closely related to turnover and more geared toward subjects that are less productivity-enhancing.