BACKGROUND: The "oldest old" are now the fastest growing section of most western populations, yet there are scarcely any data concerning even the common problem of falls amongst the very old. Prospective data collection is encouraged as the most reliable method for researching older people's falls, though in clinical practice guidelines advise taking a history of any recalled falls. This study set out to inform service planning by describing the epidemiology of falls in advanced old age using both retrospectively and prospectively collected falls data. DESIGN: Re-survey of over-90-year-olds in a longitudinal cohort study - cross-sectional interview and intensive 12-month follow-up. PARTICIPANTS AND SETTING: 90 women and 20 men participating in a population-based cohort (aged 91-105 years, in care-homes and community-dwelling) recruited from representative general practices in Cambridge, UKMeasurements: Prospective falls data were collected using fall calendars and telephone follow-up for one year after cross-sectional survey including fall history. RESULTS: 58% were reported to have fallen at least once in the previous year and 60% in the 1-year follow-up. The proportion reported to have fallen more than once was lower using retrospective recall of the past year than prospective reports gathered the following year (34% versus 45%), as were fall rates (1.6 and 2.8 falls/person-year respectively). Repeated falls in the past year were more highly predictive of falls during the following year - IRR 4.7, 95% CI 2.6-8.7 - than just one - IRR 3.6, 95% CI 2.0-6.3, using negative binomial regression. Only 1/5 reportedly did not fall during either the year before or after interview. CONCLUSION: Fall rates in this representative sample of over-90-year-olds are even higher than previous reports from octogenarians. Recalled falls last year, particularly repeated falls, strongly predicted falls during follow-up. Similar proportions of people who fell were reported by retrospective and prospective methods covering two consecutive years. Recall methods may underestimate numbers of repeated falls and the extent of recurrent falling. Professionals caring for people of advanced age can easily ask routinely whether someone has fallen at all, or more than once, in the past year to identify those at high risk of subsequent falls.