Previous attempts to develop engineering risk curves for mild traumatic brain injury have been hampered by a paucity of data from living humans. A recent study by Pellman et al. (2003) presented risk curves for concussion based on data from reconstructed head impacts sustained by National Football League (NFL) players. These risk curves were flawed because the data were heavily biased towards injurious impact exposures. Unbiased head impact exposure data for collegiate football players at Virginia Tech (VT) have been collected using the Head Impact Telemetry System (HITS), which estimates head acceleration based on spring-mounted helmet accelerometers. Nearly 23,000 head impacts were recorded during the 2003 – 2005 seasons, including 3 impacts in which a player sustained a concussion. A concussion risk curve was estimated using the Consistent Threshold (CT) method. Head impact exposure was modeled using a Weibull distribution and normalized on a per player per play basis. An error deconvolution technique was developed to analyze the effect of measurement error in the HITS data on the distribution of head impact exposures. The expected incidence of concussion was estimated by combining the CT risk curve with the head impact exposure data. The CT risk curve derived from the HITS data provided a far more reasonable estimate of concussion incidence than the NFL risk curve.