A key question facing conservation biologists is whether declines in species' distributions are keeping pace with landscape change, or whether current distributions overestimate probabilities of future persistence. We use metapopulations of the marsh fritillary butterfly Euphydryas aurinia in the United Kingdom as a model system to test for extinction debt in a declining species. We derive parameters for a metapopulation model (incidence function model, IFM) using information from a 625-km2 landscape where habitat patch occupancy, colonization, and extinction rates for E. aurinia depend on patch connectivity, area, and quality. We then show that habitat networks in six extant metapopulations in 16-km2 squares were larger, had longer modeled persistence times (using IFM), and higher metapopulation capacity (lambdaM) than six extinct metapopulations. However, there was a > 99% chance that one or more of the six extant metapopulations would go extinct in 100 years in the absence of further habitat loss. For 11 out of 12 networks, minimum areas of habitat needed for 95% persistence of metapopulation simulations after 100 years ranged from 80 to 142 ha (approximately 5-9% of land area), depending on the spatial location of habitat. The area of habitat exceeded the estimated minimum viable metapopulation size (MVM) in only two of the six extant metapopulations, and even then by only 20%. The remaining four extant networks were expected to suffer extinction in 15-126 years. MVM was consistently estimated as approximately 5% of land area based on a sensitivity analysis of IFM parameters and was reduced only marginally (to approximately 4%) by modeling the potential impact of long-distance colonization over wider landscapes. The results suggest a widespread extinction debt among extant metapopulations of a declining species, necessitating conservation management or reserve designation even in apparent strongholds. For threatened species, metapopulation modeling is a potential means to identify landscapes near to extinction thresholds, to which conservation measures can be targeted for the best chance of success.