Most of the literature on the evaluation of training programs focuses on the effect of participation on a particular outcome (e.g. earnings). The “treatment” is generally represented by a binary variable equal to one if participation in the program occurs, and equal to zero if no participation occurs. While the use of a binary treatment indicator is attractive for ease of interpretation and estimation, it treats all exposure the same. The extent of exposure to the treatment, however, is potentially important in determining the outcome; particularly in training programs where a main feature is the varying length of the training spells of participating individuals. In this paper, we illustrate how recently developed methods for the estimation of causal effects from continuous treatments can be used to learn about the consequences of heterogeneous lengths of enrollment in the evaluation of training programs. We apply these methods to data on Job Corps (JC), America’s largest and most comprehensive job training program for disadvantaged youth. The length of exposure is a significant source of heterogeneity in these data: while the average participation spell in JC is 28 weeks, its standard deviation and interdecile range are 27 and 62 weeks, respectively. We estimate average causal effects of different lengths of exposure to JC using the “generalized propensity score” under the assumption that the length of the individual’s JC spell is randomly assigned, conditional on a rich set of covariates. Finally, using this approach, we document important differences across different spell lengths and across three racial and ethnic groups of participants (blacks, whites and Hispanics) that help understand why the benefits these groups receive from JC are so disparate from estimates derived using traditional methods.