There is a large difference between the rates of observed seat belt use by the general public and belt use by motor vehicle occupants who are fatally injured in crashes. Seat belt use rates of fatally injured occupants, as reported in the Fatality Analysis Reporting System (FARS), are much lower than the use rates found in observation surveys conducted by the states. A series of mathematical models describing the empirical relationship between FARS and observed rates were explored. The initial model was a 'straw man' and used two simplifying assumptions: (a) belt users and nonusers are equally likely to be involved in 'potentially fatal collisions', and (b) belts are 45% effective in preventing deaths. The model was examined by comparing each state's FARS use rate with the predicted rate. The model did not fit the state data points even when possible biases in the data were controlled. We next examined the assumptions in the model. Changing the seat belt effectiveness parameter provided a reasonable fit, but required an assumption that seat belts are 67% effective in preventing fatalities. The inclusion of a risk coefficient for non-belted occupants also provided a reasonable fit between the model and data. A variable risk model produced the best fit with the data. The major finding was that a model consistent with the data can be obtained by incorporating the assumption that nonusers of seat belts have a higher risk of involvement in potentially fatal collisions than do seat belt users. It was concluded that unbelted occupants are over-represented in fatal collisions for two reasons: (a) because of a greater chance of involvement in potentially fatal collisions in the first place, and (b) because they are not afforded the protection of seat belts when a collision does occur.