Reliable information on demography is necessary for conservation of albatrosses, the most threatened family of pelagic birds. Albatross survival has been estimated using mark-recapture data and the Cormack-Jolly-Seber (CJS) model. However, albatross exhibit skipped breeding, violating assumptions of the CJS model. Multistate modeling integrating unobservable states is a promising tool for such situations. We applied multistate models to data on Grey-headed Albatross (Talassarche chrysostoma) to evaluate model performance and describe demographic patterns. These included a multistate equivalent of the CJS model (MS-2), including successful and failed breeding states and ignoring temporary emigration, and three versions of a four-state multistate model that accounts for temporary emigration by integrating unobservable states: a model (MS-4) with one sample per breeding season, a robust design model (RDMS-4) with multiple samples per season and geographic closure within the season, and an open robust design model (ORDMS-4) with multiple samples per season and staggered entry and exit of animals within the season. Survival estimates from the MS-2 model were higher than those from the MS-4 model, which resulted in apparent percent relative bias averaging 2.2%. The ORDMS-4 model was more appropriate than the RDMS-4 model, given that staggered entry and exit occurred. Annual survival probability for Greyheaded Albatross at Marion Island was 0.951 ± 0.006 (SE), and the probability of skipped breeding in a subsequent year averaged 0.938 for successful and 0.163 for failed breeders. We recommend that multistate models with unobservable states, combined with robust-design sampling, be used in studies of species that exhibit temporary emigration.