Multiple imputation of missing marijuana data in the Fatality Analysis Reporting System using a Bayesian multilevel model.

BACKGROUND

The Fatality Analysis Reporting System (FARS) provides important data for studying the role of marijuana in motor vehicle crashes. However, marijuana testing data are available for only 34% of drivers in the FARS, which represents a major barrier in the use of the data.

METHODS

We developed a multiple imputation (MI) procedure for estimating marijuana positivity among drivers with missing marijuana test results, using a Bayesian multilevel model that allows a nonlinear association with blood alcohol concentrations (BACs), accounts for correlations among drivers in the same states, and includes both individual-level and state-level covariates. We generated 10 imputations for the missing marijuana-testing data using Markov chain Monte Carlo simulations and estimated positivity rates of marijuana in the nation and each state.

RESULTS

Drivers who were at older age, female, using seatbelt at the time of crash, having valid license, or operating median/heavy trucks were less likely to test positive for marijuana. There was a reverse U-shaped association between BACs and positivity of marijuana, with lower positivity when BACs < 0.01 g/dL or ≥0.15 g/dL. The MI data estimated a lower positivity rate of marijuana in the nation and each of the state than the observed data, with a national positivity rate of 11.7% (95% CI: 11.1, 12.4) versus 14.8% using the observed data in 2013.

CONCLUSIONS

Our MI procedure appears to be a valid approach to addressing missing marijuana data in the FARS and may help strengthen the capacity of the FARS for monitoring the epidemic of drugged driving and understanding the role of marijuana in fatal motor vehicle crashes in the United States.