• seatbelts
• misclassification

In a recent article, Robertson¹ commented on our study of seat belts and death in a crash.² Robertson wrote: “What is not explained adequately by the theory [about misclassification of seat belt use] is the sudden gap in police reported use by the dead and survivors that appeared in the mid-1980s”.

Robertson’s criticism seems misplaced, as we offered no theory to explain changes in the prevalence of belt use. We reported that among front seat occupant pairs in which one or both died, the prevalence of belt use decreased from 12% in 1975 to 4% in 1980, and then rose to 40% in 1998.² Explaining these changes, however, was not the focus of our paper. Using matched cohort methods, we noted that the risk ratio for death, comparing belted with unbelted occupants, was 0.59 using data from 1975–83, and 0.39 using data from 1986–98. We examined theories that might explain why these risk ratio estimates changed over time.³ We presented evidence against the theory that seat belts have become truly more effective and against the theory that estimates changed because of changes in crash characteristics. The observed change in risk ratio estimates could be explained by either, or both, of two theories:

1. Differential misclassification. Seat belt misclassification is differential when the proportion misclassified is related to the outcome (death). Risk ratio estimates could move away from their true value and toward 0 if, over time, an increasing proportion of crash survivors were classified as belted, when they were not, or an increasing proportion of those who died were classified as unbelted, when they were: this possible mechanism is illustrated with hypothetical data in the top half of table 1. (For simplicity, the table ignores the matching used in our published analysis.)

2. Non-differential misclassification. Without regard to death or survival, some belt users could be classified as not belted, or some non-users as belted, or both. Non-differential misclassification of a binary variable tends to bias risk ratio estimates toward 1.⁴ If non-differential error decreased over time, more recent risk ratio estimates could be less subject to this bias; they could move away from 1 toward their true value. However, even if non-differential error was constant over time, more recent risk ratios might also tend to be less biased, because of the influence of changing seat belt prevalence: bottom half of table 1.

For both differential and non-differential misclassification, the size and direction of any change over time in risk ratio estimates will be related to the size and direction of the errors and changes in the prevalence of seat belt use. The observed changes in risk ratio estimates alone cannot tell us which estimates are least subject to bias.

One of us has reported that there is some degree of both differential and non-differential misclassification of belt use; but the amount of error in recent data suitable for a matched-cohort analysis was so trivial, and biases toward 1 and toward 0 so balanced, that the misclassification did not appreciably influence the risk ratio estimate.⁵ Robertson interpreted these results as showing only that trained crash investigators were as prone to differential misclassification as police investigators.¹ Whatever the correct interpretation, we and Robertson agree that additional measures of seat belt use would be useful. We hope that information from electronic crash recorders will be added to publicly available data, such as the Crashworthiness Data System (CDS). It might be feasible for the CDS to assess some crashes with a second investigator assigned to determine belt use only by vehicle inspection, without knowledge of occupant outcomes or the police report. To minimize costs, this additional investigation could be reserved for those crashes with front seat occupant pairs among whom at least one died. This would allow a matched cohort analysis to compare risk ratio estimates using three sources of belt information: (1) police reports; (2) the usual CDS investigation; and (3) an investigator who could not be biased by knowledge of the outcome.