We analyzed the role of the range of variation in training exemplars as a contextual variable influencing the effects of in vivo versus simulation training in producing generalized responding. Four mentally retarded adults received single case instruction, followed by general case instruction, on washing machine and dryer use; one task was taught using actual appliances (in vivo) and the other using simulation. In vivo and simulation training were counterbalanced across the two tasks for the 2 subject pairs, using a within-subjects Latin square design. With both paradigms, more errors were made after single case than after general case instruction during probe sessions with untrained washing machines and dryers. These results suggest that generalization errors were affected by the range of training exemplars and not by the use of simulated versus natural training stimuli. Although both general case simulation and general case in vivo training facilitated generalized performance of laundry skills, an analysis of training time and costs indicated that the former approach was more efficient. The study illustrates a methodology for studying complex interactions and guiding decisions on the optimal use of instructional alternatives.