Causal peer influence on marijuana use is difficult to identify, largely due to selection—marijuana users tend to select marijuana users to be friends. This study takes advantage of a natural experiment of randomly assigned college roommates (n = 1953) in the United States to investigate peer influence on marijuana use. The randomized roommate assignment eliminates selection bias. Compared to behaviors such as drinking and smoking, marijuana use is less legally and socially acceptable. Therefore, peer influence on marijuana use may not be pervasive. To identify the potential peer effects, we considered both heterogeneity in respondents and heterogeneity in peers. We found that peer influence depended on respondent's predisposition to marijuana use. Results showed that peer effects only existed among those who used marijuana before college. For respondents who did not use marijuana before college, monthly marijuana use in college was close to zero and no peer effect was found. Further analysis revealed that marijuana use was contingent on roommate's history of marijuana use. For respondents who used marijuana before college, their college marijuana use decreased substantially if they were assigned roommates who desisted from using marijuana (i.e., used marijuana before college but stopped using in college) than if they were assigned roommates who never used marijuana, always used, or started using after entering college.