Theories of innovation and technical change posit that inventions that combine knowledge across technology domains have greater impact than inventions drawn from a single domain. The evidence for this claim comes mostly from research on patented inventions and ignores failed patent applications. We draw on insights from research into institutional gatekeeping to theorize that, to be granted, patent applications that span technological domains must have higher quality than otherwise-comparable, narrower applications. Using data on failed and successful patent applications, we estimate an integrated, two-stage model that accounts for this differential selection. We find that more domain-spanning patent applications are less likely to be approved, and that controlling for this differential selection reduces the estimated effect of knowledge recombination on innovative impact by about half. By conceptualizing the patentapproval process as a form of institutional gatekeeping, this paper highlights the institutional underpinnings of and constraints on the innovation process.