In this paper, we propose and implement non-experimental regression-adjustment methods in an attempt to isolate one of the potentially important reasons for across-program differences in training effects, namely that programs differ in the mix in and assignment of different types of training to the participants in its programs. The latter source of treatment heterogeneity and its potential consequences for across-program differences in training effects has been noted by Hotz, Imbens and Mortimer (2005) and others who have evaluated the effectiveness of training programs. We show how regression-adjustment methods can be used to estimate such differential effects and how data for a control group generated by random assignment can be use to partially assess the validity of such methods. Substantively, we use these methods and tests to reanalyze data from the MDRC Evaluation of California’s welfare-to-work program in the 1990s, namely the Greater Avenues to Independence (GAIN) program. This experimental Evaluation found that, compared to other counties that stressed Human Capital Development (HCD) training strategies, Riverside County’s GAIN program, which stressed getting program participants into jobs quickly through a Labor Force Attachment (LFA) training strategy, had relatively larger effects on post-random assignment employment, labor market earnings, and welfare participation. We apply regression-adjustment methods and implement partial tests of these methods noted above to these data to directly assess the short- and long-term differential impacts of these two training strategies and, thus, previous conclusions about the reasons for the success of the Riverside GAIN program.