Most studies examining the effectiveness of seatbelts in reducing injury due to automobile accidents have heen based on police-level data. Due to the circumstances surrounding the officer's investigation of the crash, such data generally contain misclassification errors relating to belt use and injury sustained, which can seriously bias any effectiveness estimates derived from that data.
In this paper, a methodolory for analyzing general categorical data with misclassification errors is described and the procedure applied to the seatbelt effectiveness question. The trchnique utilizes an original large sample based on police-reported accidents together with a relatively small supplementary sample that is cross-classified by the police and by a more reliable classification mechanism.
The procedure is illustrated using police-reported North Carolina accidents for the first 8 months of 1975 as the original sample. The true classification of the supplementary sample of accidents is assumed to be obtained through hospital reports for injured occupants and through telephone interviews for the non-injured. Comparisons are then made of the belt-associated relative risks thus obtained.