In most developed countries, information on road crashes are routinely collected by the police. However, comparison of police records and hospital data underlines a deficit of the number of road accidents in the routine statistics. In La Réunion, a French overseas dependency, an epidemiological study of injuries leading to hospitalisation or deaths has been performed from June 1993 to June 1994. The comparison between hospital data and police records showed that only 37.3% of non-fatally traffic-injured in-patients were recorded by the police. Length of stay in hospital, physician in charge of the first aid, urban place of the crash, type of vehicle involved, day and time of the crash and blood alcohol concentration were significantly associated with the presence in the police file. Police overestimated the severity of the injuries. Police notified 100 deaths on the 115 counted by the study. In France, non-fatally traffic-injured should be followed 30 days to improve quality of police death records. A capture-recapture method was used to estimate the total number of injured people. The capture-recapture method consists in merging information from several sources of notification to determine the real number of cases in the population and the exhaustivity of each source. We estimated that 346 subjects were injured in one month whereas police data recorded only 87 and hospital data 137. This method seems interesting to use in routine after validation when unique personal identifiers are available.