In this issue of JAMA, Xie and colleagues 1 report findings from an evaluation of adverse outcomes following initial prescriptions for tramadol or codeine among residents of Catalonia, Spain. Using 10 years of linked primary care and pharmacy claims data, the investigators used propensity score matching to assemble a cohort of 368 960 individuals who received an initial prescription for either tramadol or codeine. Initial receipt of a tramadol prescription, compared with codeine, was associated with a significantly higher 1-year cumulative incidence of all-cause mortality (absolute rate difference [ARD], 7.37 [95% CI,] per 1000 personyears), cardiovascular events (ARD, 1.36 [95% CI,] per 1000 person-years), and fractures (ARD, 4.10 [95% CI,] per 1000 person-years) but no significant difference in risk of falls, delirium, constipation, opioid abuse/dependence, or sleep disorders. The study findings were robust to several sensitivity analyses, including a dose-response analysis that demonstrated higher effect sizes with higher quantities of tramadol dispensed.These findings add to a growing body of literature that reports adverse outcomes associated with tramadol use, such as hypoglycemia, seizure, and all-cause mortality. [2][3][4] Notably, the results reported by Xie et al 1 diverge from the findings of a smaller study by Zeng et al 4 of 16 992 older adults with osteoarthritis, which found no significant difference in mortality between initial tramadol prescriptions compared with codeine but increased mortality risk with tramadol compared with nonsteroidal anti-inflammatory drugs (NSAIDs) and cyclooxygenase-2 inhibitors. However, the study by Zeng et al used prescribing data from the UK at a time when tramadol and codeine were available over the counter.Inferring causal relationships from observational studies that compare different clinical treatments can be difficult because of the potential for underlying confounding by indication. 5 For example, administration of a specific medication in observational studies can be affected by the perceived clinical indication as determined by the prescribing physician. Individuals who received opioids for osteoarthritis, for example, might differ from those who received NSAIDs with respect to pain intensity, disability, and comorbidities. Although randomized treatment allocation is the ideal method to limit such confounding, conducting a randomized trial may not always be feasible, particularly for outcomes with very low incidence rates, such as those evaluated by Xie et al. In such cases, propensity score methods can be used to help reduce bias when estimating potential causal relationships. 6 Xie et al 1 used propensity score matching to adjust for initial dissimilarities between individuals receiving tramadol and codeine and eventually selected a propensity-matched cohort with respect to sex, age, socioeconomic status, existing medical diag-