Parts of this text have been reproduced from Davies et al.13 Published by the BMJ Publishing Group Limited. For permission to use (where not already granted under a licence) please go to http://www.bmj.com/company/products-services/rights-and-licensing/. This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/4.0/.
Research objectives
- To investigate the long-term smoking abstinence of patients prescribed either varenicline or NRTs in the Clinical Practice Research Datalink (CPRD) and whether or not the effects differ by area-level deprivation.
- To investigate the causal effects of prescribing smoking cessation therapies on:
- all-cause mortality and death due to chronic lung disease, lung cancer, coronary heart disease, pneumonia, cerebrovascular disease, diabetes and external causes
- all-cause and incident cause-specific hospitalisation (specifically chronic lung disease, lung cancer, coronary heart disease, pneumonia, cerebrovascular disease, diabetes and external causes)
- incident primary care diagnosis of respiratory illness
- incident primary care diagnosis of myocardial infarction
- incident primary care prescription of a hypnotic or anxiolytic
- the frequency of attendance at a general practice or a hospital
- weight in kg
- incident registration of (1) any malignant neoplasm and (2) lung malignancy.
Study design
This was a prospective longitudinal cohort study of electronic medical records of patients prescribed smoking cessation treatments in primary care. The CPRD’s Independent Scientific Advisory Committee approved this study (protocol number 15_107). The protocol has been published.13
Data extraction
We extracted data from patients aged ≥ 18 years who were prescribed smoking cessation treatment (see British National Formulary, chapter 4.10.214) in ‘up to standard’ primary care centres contributing to the CPRD after 1 September 2006, when varenicline became available in the UK. ‘Up to standard’ practices are those that provided consistent data to the CPRD. Inclusion criteria included having an ‘acceptable’ record and having at least 1 year of registration at their practice before their first smoking cessation prescription to allow for high-quality assessment of confounders. The CPRD defines patients’ data as acceptable if they have valid information, including information regarding age and sex.15 The CPRD linked data from the Office for National Statistics (ONS) mortality records and the Hospital Episode Statistics for around half of the practices in the study. We defined the health outcomes mainly from these linked data sets. For this reason, we restricted the evaluation of the health impact of smoking cessation treatments to linked practices. Figure 2 shows a flow chart of the inclusion and exclusion of patients for the effectiveness and morbidity and mortality aspects of the study.
FIGURE 2 Flow chart of inclusion and exclusion from the study.16,17 (Figure not reproduced here; see the original article.)
Intervention
The intervention was the first prescription of either smoking cessation treatment in the eligible period. We compared the effect of prescribing varenicline with the effect of prescribing NRT. We focus on the first recorded treatment because subsequent treatment decisions are likely to be highly non-random. We extracted the date of the first prescription from the patients’ records.
Follow-up
Follow-up started on the date that each patient received their first prescription and ended when they had an event, died or left the practice.
Outcomes
We assessed all outcomes at 3, 6, 9, 12, 24 and 48 months after the first prescription. We took the date of diagnosis from the first record that indicated an outcome after the first prescription. We excluded patients who were prevalent cases at the time of their first prescription from these analyses.
Smoking cessation
We defined smoking cessation using GP records that indicated the patients’ smoking status. Each patient had multiple records that stated their smoking status, which can be indicated as current, former or never smoker. GPs were incentivised to record their patients’ smoking status as part of the Quality and Outcomes Framework.18 We defined smoking cessation as the status indicated by the most recent record within each follow-up period. If patients had no record that indicated their smoking status but were prescribed smoking cessation medication, we assumed that they had relapsed. We conducted sensitivity analyses to investigate whether or not our results were affected when we assumed that all unobserved smokers relapsed.
Mortality
We defined mortality using the linked ONS mortality data. This data set gave the date and primary causes of death. We investigated the following secondary outcomes: mortality due to chronic lung disease [International Statistical Classification of Diseases and Related Health Problems, Tenth Revision19 (ICD-10) codes = J40–44], lung cancer (ICD-10 code = C34), coronary heart disease (ICD-10 codes = I21–25), pneumonia (ICD-10 codes = J12–18), cerebrovascular disease (ICD-10 codes = I60–69), diabetes (ICD-10 codes = E10–14) and external causes (ICD-10 codes = V01–Y98).
Hospitalisation
We identified episodes of hospitalisation using the linked Hospital Episode Statistics data set, which provides the date and cause of hospitalisation for all linked patients. We investigated both all-cause hospitalisation and cause-specific hospitalisation using the codes described above for mortality.
Primary care diagnosis and attendance
Primary care diagnoses of myocardial infarction, chronic obstructive pulmonary disease (COPD) and depression, and prescriptions of hypnotics and anxiolytics were identified using validated code lists. The frequency that each patient attended primary care was identified from their records.
Weight
The patients’ weight in the months following their initial prescription was identified using their electronic medical records.
Potential confounders
The potential confounders included the following variables measured at baseline: age; body mass index (BMI); alcohol drug misuse, defined using Read codes; SEP measured by the Index of Multiple Deprivation (IMD); an indicator for having more than five primary care consultations in the year before first prescription; indicators for year of prescription; any previous prescriptions of hypnotics; any previous prescriptions of antipsychotics; any previous prescriptions of antidepressants; diagnoses of self-harm; diagnoses of myocardial infarction; diagnoses of COPD; any psychiatric disease; or any serious comorbidity (a non-zero Charlson Comorbidity Index). We accounted for missing data by using imputation by chained equations, because some patients were missing BMI and SEP data.
Statistical methods
We estimated the effects of varenicline and NRT using three statistical methods: multivariable-adjusted logistic regression, propensity score matched regression and instrumental variable regression. We did not use Cox survival models because the proportional hazards assumption did not hold. Our basic adjusted models adjust for sex, year of first prescription and age. Our fully adjusted results also adjust for all confounders described above. We used propensity score and the covariates described above to match patients who were treated with varenicline to other similar patients prescribed NRT using one-to-one matching with no calliper. There was a complete overlap in the propensity scores between smokers who were prescribed varenicline and who were prescribed NRT. The propensity score depends on similar assumptions to multivariable-adjusted regression (conditional exchangeability). Both approaches could suffer from residual confounding. We addressed residual confounding using instrumental variable analysis.
Instrumental variable analysis requires a variable, the ‘instrument’ that is defined by the following three assumptions: (1) the instrument associates with the exposure of interest, (2) the instrument does not associate with unobserved confounding factors and (3) the instrument affects the outcome only via the exposure.20 A widely used instrument in pharmacoepidemiology is physicians’ prescribing preferences for different medications.21 We cannot directly measure GPs’ preferences, but we can use their prescriptions to previous patients as a proxy for their preferences. We investigated whether or not the instrumental variable assumptions held by (1) testing the association of their prior prescriptions to their previous seven patients with their subsequent prescriptions and (2) investigating the association with potential confounders. We estimated the effects of varenicline and NRT using additive structural models for binary outcomes and two-stage least squares for continuous outcomes.
All standard errors account for the clustering of patients at the practice level. We published all analytic scripts used to generate the data to GitHub.22,23
Changes to the protocol
Survival analysis
Our original protocol proposed to use survival analysis to estimate hazard ratios of events. However, there was evidence that the proportional hazards assumption was violated; therefore, we dichotomised the outcomes instead.
Definition of mortality and hospitalisation outcomes
Peer reviewers of Davies et al.16 noted that our original definition of our cause-specific outcomes was relatively general (diseases of the respiratory system, ICD-10 = J00–99 and cardiovascular disease, ICD-10 = I00–52) and recommended reporting more specific causes of death. As a result, we amended our protocol to include the following eight outcomes defined by ICD-10 code: (1) chronic lung disease (ICD-10 = J40–44), (2) lung cancer (ICD-10 = C34), (3) coronary heart disease (ICD-10 = I21–25), (4) pneumonia (ICD-10 = J12–18), (5) cerebrovascular disease (ICD-10 = I60–69), (6) diabetes (ICD-10 E10–14) and (7) external causes (ICD-10 V01–Y98). We used the same set of codes to define cause-specific mortality and hospital admission.
Cancer registry outcomes
Our protocol originally intended to use data from the National Cancer Registry. However, unfortunately, the CPRD was not able to provide access to this data set. To overcome this, we revised our protocol to use data from primary and secondary care to identify cancer cases. These alternative sources of information are reliable indicators of cancer diagnosis date and type.24