Background Varenicline is a medication which has been shown to be more efficacious for smoking cessation than nicotine replacement products in large randomised controlled trials. Concerns have been raised about possible adverse effects of varenicline which has led the FDA to issue a black box warning. However, there is limited real-world evidence about the effects of varenicline on clinical outcomes, such as mortality, incidence of cardiovascular and respiratory disease, and primary and secondary care attendance.

Methods We investigated this hypothesis using a prospective cohort study of all patients issued a first smoking cessation prescription in the Clinical Practice Research Datalink (CPRD) between the 1^st of September 2006 and the 31^st of March 2014. We used linked Office of National Statistics mortality data and secondary care attendance through the Hospital Episodes Statistics dataset. Our primary outcome was all-cause mortality within two years of first prescription. We used multivariable adjusted and propensity score matched Cox regression, and included a range of baseline confounders. However, we were concerned that our results could suffer from residual confounding, as patients prescribed varenicline were substantially less likely to be older or have comorbidities. To overcome this we applied an instrumental variable analysis using physicians prescribing preferences. We tested for heterogeneity between our instrumental variable and conventional results using Hausman tests. Our study protocol is registered: NCT02681848.

Results There were 126,718 patients who met our inclusion criteria. The multivariable adjusted results suggested that patients prescribed varenicline had lower mortality (adjusted hazard ratio (HR) = 0.72, 95% confidence interval (CI): 0.64, 0.80), lower hospital admission rates (HR = 0.84, 95% CI: 0.92, 0.86) and less frequent attendance to primary care (mean difference in attendance over two years per patient treated = 0.29 95% CI: 0.11, 0.47). The instrumental variable analysis found little evidence that varenicline affected mortality (risk difference per 100 patients treated = 0.13, 95% CI: −0.65, 0.90). This suggests that the reduction in mortality found in the conventional multivariable adjusted and propensity score matched regressions may be due to residual confounding (p_Hausman = 0.002).

Conclusion We found little evidence that patients prescribed varenicline were at increased the risk of adverse outcomes. We found some evidence that they were less likely to attend primary care in the two years following their first prescription. However, apparent protective effects on mortality found in conventional analysis are likely to be due to residual confounding. These results provide reassurance to patients and clinicians about the safety of varenicline.