Preclinical work suggested GLP-1 drugs might reduce opioid craving, and observational studies have reported fewer overdoses. A target trial emulation in Veterans Health Administration data tested it again in people carrying both type 2 diabetes and opioid use disorder, who started semaglutide or tirzepatide against one of five other diabetes medications between 2020 and 2024 [1]. The trial evidence in this area is covered separately in the alcohol trials, which asked a related question with a randomized design.
Against insulin, the hazard ratio for all-cause overdose over 12 months was 0.32 (95% CI 0.14–0.71). Against SGLT2 inhibitors it was 0.25 (95% CI 0.09–0.68). Both are large reductions and both are statistically significant.
Against the other three comparators, nothing. Metformin gave 0.75 (95% CI 0.38–1.45), sulfonylureas 0.82 (95% CI 0.33–1.96), and DPP4 inhibitors 1.20 (95% CI 0.52–2.78), an estimate that sits on the wrong side of 1 entirely. Same drug, same outcome, same database, same 12 months.
The intervals themselves say something worth noticing. An interval running from 0.09 to 0.68 is not a precise estimate; it is a small number of rare events in a matched set of a few hundred people. Overdose is uncommon even in this population, and the width of every interval in the study reflects that. The authors conclude by calling for randomized controlled trials, which is the correct reading of their own result, and puts them in the same position as the team in the Crohn's disease cohort who published a dramatic number and then told readers not to believe it.
None of this makes the hypothesis wrong. A real effect on craving could plausibly show up first against the comparators with the most room to move, and two of five estimates landing below 1 with three inconclusive is not evidence of absence. It is evidence that the question is still open, and that a headline built from the insulin comparison alone would be describing a choice of control group rather than a property of the drug. The same reasoning applies whenever a study picks its own comparator, as in the tirzepatide network comparison and in the carpal tunnel cohort, where two comparison groups gave two different answers about the same drug.