A Machine-Learning Study Predicted Reported Driving After Cannabis, Not Measured Impairment
A cannabis-driving prediction study classified reports of driving within three hours after use. That operational definition is not the same as a measured impairment diagnosis or proof of a legal violation.
The outcome came from a questionnaire
The analysis used young-adult survey data from Washington State. The cannabis models included 4,891 respondents reporting past-month cannabis use. Participants reported whether they had driven within three hours after cannabis consumption; any reported occurrence was coded positive. No blood THC measurement or observed driving-performance test defined that outcome.
Two algorithms examined many predictors
Regularized regression and random forests considered numerous candidate variables. Cannabis-use frequency, cannabis-related memory problems, simultaneous alcohol–cannabis use and reported increased tolerance were among important predictors. Agreement between algorithms supports stability within the dataset, but does not establish that these characteristics caused the reported driving.
A ranking is not an individual rule
Predictor importance depends on the available variables, correlations among them and the outcome definition. Frequent users also have more opportunities to drive soon after use. A high ranking for use frequency therefore should not automatically become a statement about impairment from a particular dose.
Limits on forensic application
Although the broader survey program follows cohorts, this analysis used one baseline survey per participant and was repeated cross-sectional. Self-report, one-state recruitment and unmeasured circumstances restrict generalization. The models may help identify prevention priorities, but they do not establish an individual toxicology threshold, actual crash risk or impairment at a disputed time.
Related analysis: evaluating cannabis tolerance claims.
Primary research
Calhoun BH; Hultgren BA; McCabe CJ; Rhew IC; Larimer ME; Kilmer JR; Guttmannova K. Using machine learning to identify unique predictors of alcohol and cannabis impaired driving. Alcohol, clinical & experimental research (2026). DOI: 10.1111/acer.70245.
