Despite over two decades of U.S. experience in operating large scale subsidized training programs for low income and unemployed workers, the effects of these programs are still highly controversial. The controversy arises from the difficulty of specifying the model of participant outcomes in the absence of training that is necessary in any nonexperimental program evaluation. In this paper we suggest that some of these difficulties may be overcome by focusing on a very simple measure of outcomes: namely, the probability of employment. We present a variety of estimates of the effect of training on the probability of employment for the 1976 cohort of adult male participants in the Comprehensive Employment and Training Act (CETA) program. Our methods range from a simple comparison of pre- and post-training employment probabilities between trainees and nonparticipants, to a fully specified first-order Markov model of employment probabilities with individual heterogeneity. There is consistent evidence across methods that CETA participation increased the probability of employment in the three years after training by from 2 to 5 percentage points. Classroom training programs appear to have had significantly larger effects than on-the-job programs, although the estimated effects of both kinds of programs are positive. We also find that movements in and out of employment for trainees and nonparticipants are reasonably well described by a first-order process, conditional on individual heterogeneity. In the context of this model, CETA participation appears to have increased both the probability of moving into employment, and the probability of continuing employment.