Multiple Imputation for the Fatal Accident Reporting System

The Fatal Accident Reporting System (FARS) gathers data on fatal traffic accidents in the United States. Data are collected at the state level for the National Highway Traffic Safety Administration (NHTSA), which publishes results in an Annual Report (NHTSA 1985). An important methodological concern with FARS data is the high level of nonresponse for certain variables, particularly blood alcohol content (BAC), an important variable in FAILS analysis that is available only when a blood sample is taken. In 1987 NHTSA convened a panel of statisticians to provide the agency with guidance on imputation methods for FARS. The main goals articulated by the agency were a) to produce a public use data file that does not contain missing data; and b) to obtain a method for imputation of FAILS variables that can provide valid standard errors for values thus imputed. This paper proposes a multiple imputation method for FAILS data that largely meets these goals, and improves on existing practice. Some preliminary results on applying the method to 1985 FARS data are presented, and some further refinements are outlined.