Data sources used for motor vehicle injury research frequently contain considerable inaccuracies when describing crash characteristics. Sixty-three MVC patients transported to a university hospital emergency department via ambulance and also reported on an in-depth crash investigation record (CIR) were identified during the period January, 1993 to December, 1995. The accuracy of the police report (PR), ambulance report (AR) and emergency department record (EDR) for each patient was compared to the CIR by computing percent agreement and kappa for each data source. Fifty-one cases (81%) had at least one episode of discordance. The mean percent agreement for the variables studied was PR .929, AR .807, and EDR .897. Mean values for kappa were PR .889, AR .706, and EDR .846. The inaccuracy between data sources for determining MVC characteristics varies. Injury control researchers need to consider how to increase accuracy for data sources.