The CBD Cessation Trial’s Missing-Data Analysis Relied on a Missing-at-Random Assumption
The CBD cannabis-use trial used Bayesian multiple imputation for missing outcomes. These model-generated values were not additional observed urine samples or participant reports.
The analysis retained randomized participants
The investigators specified intention-to-treat analyses and generated missing outcomes within their Bayesian modeling framework. They assumed data were missing at random, conditional on the information represented in the model. The outcome models also incorporated time, baseline measurements and participant-level differences.
An assumption is not a laboratory observation
Imputation allows incomplete observations to contribute under stated statistical assumptions. It does not recover the actual concentration that an uncollected specimen would have contained. If unmeasured reasons for missingness were strongly related to the missing outcome, the model could represent those outcomes imperfectly. That is an interpretive limit, not proof that the assumption was violated.
Report estimates with their uncertainty
The trial’s CBD-versus-placebo findings were estimates from an adaptive dose-finding study, rather than complete measurements from every person at every visit. Credible intervals and the prespecified decision criteria should accompany them. Imputed values should not be counted as independent biological confirmations of abstinence, and the model should not be converted into a prediction for an individual patient. Its purpose was to estimate group treatment effects while acknowledging incomplete follow-up information.
Related analysis: the trial’s main results and limitations.
