In its 2010 Report to Congress, NHTSA claimed that it would cost approximately US$1 billion per year to actually count and classify all crashes.2 US$1 billion was deemed to be too expensive so instead of that, "... NHTSA devised a method that "utilizes an efficient combination of census, sample-based, and existing state files to provide nationally representative traffic crash data..." and it's been doing that since 1975. NHTSA claims it only costs US$30 million per year and is just as good.
According to NHTSA, an estimated 16 million crashes occur annually3 and of that total, only 6.2 million of those crashes are ever reported to the police. FARS data is collected on a purely voluntary basis through cooperative agreements between NHTSA and each of the 50 states, the District of Columbia, and Puerto Rico.4
Koehler, Steven A.; Brown, Peggy A. (2009). Forensic Epidemiology. International Forensic Science and Investigation. Vol. 19. CRC Press. p. 135. ISBN 1-4200-6327-8. Retrieved 2011-04-05. 1-4200-6327-8 ↩
"Report to Congress NHTSA's Crash Data Collection Programs" (PDF). http://www-nrd.nhtsa.dot.gov/Pubs/811337.pdf ↩