Eighty-two studies across a field this scattered will produce results pointing in several directions at once, and this review does. The cleanest way to read it is to separate claims about preventing a condition from claims about treating one [1]. The dementia version of the same split is in the dementia risk score.
On Parkinson’s disease both appear, and they disagree. People taking these drugs were less likely to develop Parkinson’s, at a pooled hazard ratio of 0.70 (95% CI 0.53–0.92). Among people who already had it, motor symptoms did not improve — a mean difference of -3.29 points on the standard rating scale with a confidence interval running from -7.84 to 1.26, which crosses no effect. Nor did non-motor symptoms, quality of life or dyskinesia.
The substance use results split the same way and carry a further problem. Cannabis use disorder risk was lower at a pooled hazard ratio of 0.55 (95% CI 0.39–0.77), while opioid use disorder showed nothing (HR 0.61, 95% CI 0.24–1.52). The heterogeneity statistics are the part to notice: 71% for cannabis and 91% for opioids. An I² of 91% means the contributing studies disagree so profoundly that the pooled number describes none of them — a related caution to the comparator problem in the overdose comparison.
One result is cleaner than the rest. Liraglutide reduced binge frequency in binge eating disorder against placebo, by 1.28 episodes (95% CI -1.67 to -0.89) with moderate heterogeneity. That is a randomized comparison, a specific drug, and a countable outcome, which makes it the most actionable line in the review.
The review applied GRADE to assess certainty across outcomes, which is more rigor than most syntheses in this field manage. What it leaves is a field where prevention signals keep appearing in observational data and treatment effects keep failing to appear in trials — a shape also visible in the effort and reward trial and the bipolar cohort.