This paper provides a cautionary note on evaluating subsidy programs. Using data on applicants, we analyze the subsidy program for SMEs in Japan (the “Monozukuri Subsidy” program). When evaluating subsidy programs, researchers often use causal-inference methods such as regression discontinuity design that compares recipients with scores slightly higher than the threshold to non-recipients with scores slightly lower than the threshold. If researchers fail to consider the possibility that non-recipients may receive subsidies from the same program in the future, they will underestimate its effects. We find that in the Monozukuri Subsidy program of FY2015, non-recipients near the threshold had a higher probability of receiving the subsidy in the future than recipients near the threshold. This result indicates that researchers should be aware of this point when evaluating subsidy programs. We discuss several ways to cope with this problem.