Agitation and neuropsychiatric symptoms in dementia have few good pharmacological options, which is why cannabinoids keep being proposed. This systematic review searched to November 2025 and pooled nine randomised trials with 334 participants, using both frequentist and Bayesian random-effects models, RoB 2 for bias and GRADE for certainty.
Nothing worked. Agitation on the Cohen-Mansfield Agitation Inventory showed a standardised mean difference of -0.58 (95% CI -1.71 to 0.55, I² = 84%); Neuropsychiatric Inventory total -0.02 (-1.00 to 0.96); its agitation subscale -0.44 (-1.45 to 0.57); Mini-Mental State Examination 0.86 (-16.33 to 18.06, I² = 96%). Bayesian posterior estimates sat close to zero, which strengthens the null rather than merely failing to reject it. Leave-one-out analysis reduced heterogeneity without changing conclusions. Certainty was moderate for behavioural outcomes and low for cognition. Overall adverse events matched placebo, but somnolence was twice as common (risk ratio 2.03, 1.29 to 3.20).
The honest summary for a family asking about medical cannabis is that nine trials have now looked, none has shown benefit for agitation or cognition, and the one reliable effect is sedation. Sedation is not nothing — a sleepier patient is a quieter patient, and that is precisely how an ineffective drug acquires a reputation for working in this population. The alternative remains unglamorous and effective: identify the trigger, treat pain, treat constipation and urinary retention, review the drug list, and address the environment before reaching for any molecule.
- Do not offer cannabinoids for agitation or cognition in dementia; nine trials show no benefit.
- Be explicit with families that sedation is the observed effect, and that quieter is not better.
- Look for pain, constipation, retention and infection first — these cause most acute agitation.
- Review anticholinergic and sedative burden before adding anything.
- Non-pharmacological approaches remain first line and are the only ones with a favourable balance here.
The statistics, in plain English
I² of 84% and 96% means these trials disagreed profoundly, so the pooled figures summarise a scattered literature rather than a consistent one. A Bayesian posterior close to zero adds something a P value cannot: it estimates how likely a real effect is, and here it supports genuine absence rather than mere failure to detect. A confidence interval running from -16.33 to 18.06 on cognition is not a null result so much as no information at all.
Read the rest in the app
You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.
Scan to keep reading on your phone. No account needed to start.
Tomorrow morning, before your first patient
One edition a day for geriatrics, written by the desk, every claim tied to its paper. Six minutes.
Get the app — free