Implications for service provision
Our results are consistent with large randomised trials and meta-analyses of smaller trials, suggesting that varenicline is the most efficacious and effective method of smoking cessation.7,10 There may be a need to update clinical guidance in the light of recent evidence.
Suggested research priorities
Trends in smoking cessation prescriptions
More evidence is needed about why prescribing of NRT and varenicline has fallen over the last decade. One explanation may be because of increased use of e-cigarettes and over-the-counter use of NRT. Alternatively, it may be because of changes to the funding and provision of addiction services. Our results suggest that this decline is not due to the general decline in smoking rates because the rate at which smokers receive NRT or varenicline prescriptions has fallen.
Long-term follow-up
Varenicline has only been available in the UK since 2006. This means that we have only a limited observational follow-up, which, in our case, is a maximum of 9 years. Many of the health benefits of smoking cessation are likely to accrue much later in life. It could be informative to prospectively follow up patients treated with varenicline to compare their outcomes with patients treated with NRT for much longer follow-up periods (10–20 years). It would be of methodological interest to compare prospective estimates with retrospective estimates derived from cross-sectional and cohort studies.
Survival analysis with instrumental variables
Recent methodological papers have described estimators for using instrumental variables with survival outcomes. Using survival outcomes could substantially increase the power of the study and could be more informative than the additive structural mean models used in this report. Furthermore, it would be informative to have more information about how these estimators perform on an applied example with a large data set.
Comparing instrumental variable and multivariable-adjusted regression
One of the most significant limitations to observational analyses of electronic medical records is residual confounding. Outside RCTs, patients are rarely treated at random. Researchers are rarely able to measure all indications for treatment, which can result in bias. Instrumental variable analysis can provide an alternative source of evidence based on different assumptions.36 However, instrumental variable analysis can provide contradictory evidence to multivariable-adjusted regression using the same data. During this project, we developed some tools that can potentially be used to assess the relative plausibility of different methods.37 These methods use measured covariates to assess the relative bias of instrumental variable and linear regression estimates using intuitive graphic plots. However, more research is needed to guide clinicians and regulators about when estimates from different methods should be used and preferred.