Abstract Background The results of observational studies evaluating and comparing the cardiovascular safety of glitazones, metformin and sufonylureas are inconsistent.To conduct and evaluate heterogeneity in a meta-analysis of observational studies on the risk of acute myocardial infarction (AMI) or stroke in patients with type 2 diabetes using non-insulin blood glucose–lowering drugs (NIBGLD). Methods We systematically identified and reviewed studies evaluating NIBGLD in patients with type 2 diabetes indexed in Medline, Embase, or the Cochrane Library that met prespecified criteria. The quality of included studies was assessed with the RTI item bank. Results were combined using fixed- and random-effects models, and the Higgins I 2 statistic was used to evaluate heterogeneity. Sensitivity analyses by study quality were conducted. Results The summary relative risk (sRR) (95 % CI) of AMI for rosiglitazone versus pioglitazone was 1.13 (1.04–1.24) [I 2 = 55 %]. In the sensitivity analysis, heterogeneity was reduced [I 2 = 16 %]. The sRR (95 % CI) of stroke for rosiglitazone versus pioglitazone was 1.18 (1.02–1.36) [I 2 = 42 %]. There was strong evidence of heterogeneity related to study quality in the comparisons of rosiglitazone versus metformin and rosiglitazone versus sulfonylureas (I 2 ≥ 70 %). The sRR (95 % CI) of AMI for sulfonylurea versus metformin was 1.24 (1.14–1.34) [I 2 = 41 %] and for pioglitazone versus metformin was 1.02 (0.75–1.38) [I 2 = 17 %]. Sensitivity analyses decreased heterogeneity in most comparisons. Conclusion/interpretation Sulfonylureas increased the risk of AMI by 24 % compared with metformin; an imprecise point estimate indicated no difference in risk of AMI when comparing pioglitazone with metformin. The presence of heterogeneity precluded any conclusions on the other comparisons. The quality assessment was valuable in identifying methodological problems in the individual studies and for analysing potential sources of heterogeneity.