The effects of prescribing varenicline on longer-term health outcomes

Parts of this text have been reproduced from Davies et al.16 © 2018 The Authors. Addiction published by John Wiley & Sons Ltd on behalf of Society for the Study of Addiction. This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.

Although randomised trials are the gold standard for demonstrating efficacy, they are underpowered to detect the effect of treatments on rare secondary outcomes and have a limited follow-up. For smoking cessation treatments, follow-up is generally ≤ 1 year. In this chapter, we investigate the effect of varenicline on health outcomes between 3 and 48 months after the first prescription. In the results presented below, we focus on the instrumental variable analyses because the specification tests suggested that the multivariable-adjusted and propensity score regression results may suffer from residual confounding. We present the results for the propensity score analysis in Report Supplementary Material 1, Tables S9–S11.

Population characteristics

Of the 287,079 patients who were prescribed smoking cessation treatments during the study period, 126,718 patients met inclusion criteria for assessing the health outcomes. Of these patients, most (84,976, 67.1%) were prescribed NRT and the remainder (41,742, 32.9%) were prescribed varenicline. Most (73%) of the patients who were prescribed varenicline were issued an additional prescription of varenicline in the 3 months following their first prescription. However, only 20% of patients were issued ≥ 151 tablets within the 3-month treatment window. This is the minimum number of tablets patients needed to adhere to a full course of treatment. As with the effectiveness analysis presented in Chapter 3, patients in this restricted sample who were prescribed varenicline, and had linked Hospital Episode Statistics and ONS mortality data, were healthier in almost every way that we could measure (Table 3).

Specification tests of the instrumental variable assumptions

The physicians’ previous prescriptions were strongly associated with their prescriptions to their subsequent patients. A doctor who issued varenicline to their previous patient was 24.5% (95% CI 23.3% to 25.9%) more likely to prescribe varenicline to their subsequent patients. The instrument was the number of varenicline prescriptions issued to the physicians’ previous seven patients who were receiving smoking cessation treatment. Hence, the physicians’ prescribing behaviours explain a substantial proportion of the variation in treatment rates. The partial F-statistic, a measure of instrument strength, ranged from 4045 to 7775.31 A value of the partial F-statistic ≥ 10 suggests that the analysis is unlikely to suffer from weak instrument bias.32 This value of the F-statistic does not guarantee that the estimates will be unbiased, or will be sufficiently powered to detect effect sizes of interest. However, it does mean that the coverage of the estimates (CIs) should be correct. Values of the F-statistic < 10 can result in overly precise CIs and over-rejection of the null hypothesis. This problem is a particular concern in analyses with more than one instrument. This statistic indicates that the instrument was very strong. Furthermore, the potential confounders were more balanced across the instrument than across the patients’ prescriptions (Figure 8 and also see Report Supplementary Material 1, Table S8). These results mean that even after taking account of the strength of the physicians’ preferences (the instrument), the bias for the instrumental variable analysis (attributable to all of the observed potential confounders except BMI, the number of prior GP visits and deprivation) was likely to be lower than for the patients’ prescriptions. The estimated instrumental variable bias terms for BMI, number of GP visits and deprivation were similar to the linear regression bias terms.

FIGURE 8 The bias terms for linear and instrumental variable regression. These plots indicate the bias caused by omitting a single observed covariate from a multivariable-adjusted or instrumental variable regression analysis. This analysis treats the omitted covariate as an ‘outcome’ and the exposure or instrumental variable (prior prescription) as the dependent variable. Relative linear regression and instrumental variable bias component terms are indicated by squares (▪) and triangles (▲) respectively. 95% CIs are plotted. The patients’ actual prescriptions more strongly associated with the measured covariates than the proposed instrument, even after taking into account the instrument strength. This statistic suggests that the instrumental variable estimates are likely to be less biased than the linear regression estimates. CIs allow for clustering between physicians. We adjusted each bias term except sex for year, sex and age and no other covariates. The instrumental variable was defined using seven prior prescriptions. Reproduced from Davies et al.16 © 2018 The Authors. Addiction published by John Wiley & Sons Ltd on behalf of Society for the Study of Addiction. 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/. The figure includes minor additions and formatting changes to the original text. (Figure not reproduced here; see the original article.)

The effects of varenicline on mortality and morbidity after 2 years

Patients prescribed varenicline had slightly higher rates of all-cause mortality than those who were prescribed NRT (risk difference per 100 patients treated = 0.67, 95% CI –0.11 to 1.46) (Figure 9). From the instrumental variable analyses, there were few detectable differences in cause-specific mortality, all-cause or cause-specific hospitalisation, or primary care diagnoses of myocardial infarction or COPD (see Figure 9). Our results suggest that positive or negative effects > 2.5 per 100 patients treated are unlikely for all-cause hospitalisation. Varenicline is unlikely to cause positive or negative effects > 1 per 100 patients on the other outcomes, including a lung cancer diagnosis.

FIGURE 9 The effects of prescribing varenicline vs. NRT on mortality and morbidity in the 2 years following the first prescription. The estimates are calculated from instrumental variable regression. CIs allow for clustering between physicians. Instrumental variable results use seven prior prescriptions and adjust for the year of first prescription, sex and age. Reproduced with modification Davies et al.16 © 2018 The Authors. Addiction published by John Wiley & Sons Ltd on behalf of Society for the Study of Addiction. 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/. The figure includes minor additions and formatting changes to the original text. (Figure not reproduced here; see the original article.)

Consistent with the broader literature, the multivariable-adjusted results suggested that after 2 years, patients who were prescribed varenicline had substantially lower mortality and morbidity than patients who were prescribed NRT (Figure 10). Patients who were prescribed varenicline had lower rates of all-cause and cause-specific mortality, except lung cancer (Figure 11). After fully adjusting for all the baseline confounders listed in Table 3, the estimates modestly attenuated towards the null; however, substantial differences remained. The multivariable-adjusted estimates systematically differed from the instrumental variable estimates of the effect of varenicline after 2 years on all-cause mortality and hospital admission (p_Hausman = 4.14 × 10^–9 and 6.35 × 10^–8 respectively). The Hausman test is a test for the difference between the ordinary least squares and instrumental variable estimates. There were systematic differences in estimated effects on cause-specific mortality: chronic lung disease, coronary heart disease, pneumonia and diabetes (p_Hausman = 0.0006, = 0.005, = 0.001 and = 0.04 respectively). The instrumental variable results were inconsistent with the reduction in risk of hospitalisation for coronary heart disease estimated via multivariable adjustment (p_Hausman = 0.006). The effects of varenicline on other cause-specific hospitalisation, weight and the number of GP visits were similar for both multivariable-adjusted regression and instrumental variable regression.

FIGURE 10 The effect of prescribing varenicline on (a) all-cause, (b) chronic lung disease-related (ICD-10 J40–44), (c) lung cancer (ICD-10 C34), (d) coronary heart disease (ICD-10 I21–25), (e) pneumonia-related (ICD-10 J12–18) and (f) cerebrovascular disease-related (ICD-10 I60–69) mortality in the 4 years following first prescription as indicated via ONS mortality records. Linear regression is indicated by squares (▪), the solid line adjusts for basic confounders, the dashed line adjusts for all confounders listed in Table 3, and the instrumental variable results are indicated by triangles (▲). 95% CIs are indicated. CIs allow for clustering between physicians. The multivariable-adjusted results adjust for all the covariates listed in Table 1. Instrumental variable results use seven prior prescriptions. Reproduced from Davies et al.16 © 2018 The Authors. Addiction published by John Wiley & Sons Ltd on behalf of Society for the Study of Addiction. 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/. The figure includes minor additions and formatting changes to the original text. (Figure not reproduced here; see the original article.)

FIGURE 11 The effect of prescribing varenicline on (a) all-cause, (b) chronic lung disease-related (ICD-10 J40–44), (c) lung cancer (ICD-10 C34), (d) coronary heart disease (ICD-10 I21–25), (e) pneumonia-related (ICD-10 J12–18) and (f) cerebrovascular disease-related (ICD-10 I60–69) hospital inpatient admission in the 4 years following first prescription as indicated via ONS death records. Linear regression is indicated by squares (▪), the solid line adjusts for basic confounders, the dashed line adjusts for all confounders listed in Table 3 and the instrumental variable results are indicated by triangles (▲). 95% CIs are indicated. CIs allow for clustering between physicians. The fully multivariable-adjusted results adjust for all the covariates listed in Table 1. Instrumental variable results use seven prior prescriptions. Reproduced from Davies et al.16 © 2018 The Authors. Addiction published by John Wiley & Sons Ltd on behalf of Society for the Study of Addiction. 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/. The figure includes minor additions and formatting changes to the original text. (Figure not reproduced here; see the original article.)

The effects of varenicline versus nicotine replacement therapy over time

Figure 10 illustrates the effects of prescribing varenicline versus NRT on rates of all-cause and cause-specific mortality over time. The navy line indicates the effects as estimated via instrumental variables and the solid light blue line indicates the minimally adjusted multivariable-adjusted regression. The dashed light blue line indicates the fully adjusted multivariable-adjusted regression estimates. The instrumental variable estimates provide little evidence that varenicline affects mortality at any follow-up period. The partially adjusted multivariable estimates suggest that all-cause and cause-specific mortality risks generally drop linearly after the first prescription. The exception to this is lung cancer mortality, which drops more quickly in months 3–6 than months 6–48. Adjustment for the full range of covariates reported in Table 1 resulted in little change in the estimates. One explanation of the difference between the instrumental variable and multivariable-adjusted results is that the latter suffer from residual confounding.

Figure 11 illustrates the effects of prescribing varenicline over 4 years on all-cause and cause-specific hospitalisation. The instrumental variable estimates suggest a modest reduction in all-cause admissions in the 3–9 months following the first prescription. There is little difference in rates of hospitalisation between months 12 and 24, and an imprecise positive effect at 48 months. The instrumental variable estimates of the effect of varenicline are too imprecise to distinguish them from the null hypothesis or the multivariable-adjusted estimates. The partially adjusted multivariable estimates of the effect on all-cause hospitalisation suggest a substantial protective effect that peaks 12 months after the first prescription, before modestly attenuating by 48 months. The cause-specific estimates for chronic lung disease, pneumonia and cerebrovascular disease suggest a protective effect. There was little evidence of a difference in the rates of admission for lung cancer or (after full adjustment) for coronary heart disease. These differences attenuated towards the null after adjustment but remained detectable.

The instrumental variable estimates suggest that prescribing varenicline caused an increase in weight, on average 1.14 kg (95% CI 0.09 to 2.20 kg) after 2 years; this effect was detectable across the entire follow-up period (Figure 12). The partially multivariable-adjusted results suggested a similar increase in weight across the entire follow-up period. The fully adjusted results were mildly attenuated towards the null. There were little differences between the instrumental variable and multivariable-adjusted estimates (p_Hausman for all follow-up periods > 0.01). The instrumental variable estimates suggest that prescribing varenicline caused a 19.16% (95% CI 11.17% to 27.15%) reduction in the number of primary care attendances. These estimates were consistent with the partially multivariable-adjusted estimates, but these effects largely attenuated after adjusting for all the baseline covariates listed in Table 3.

FIGURE 12 Effect of prescribing varenicline on (a) weight (kg) and (b) frequency of attendance to primary care in the 4 years following initial prescription. Linear regression is indicated by squares (▪), the solid line is adjusted for basic confounders, the dashed line adjusts for all confounders listed in Table 3 and the instrumental variable results are indicated by triangles (▲). 95% CIs are indicated. Reproduced from Davies et al.16 © 2018 The Authors. Addiction published by John Wiley & Sons Ltd on behalf of Society for the Study of Addiction. 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/. The figure includes minor additions and formatting changes to the original text. (Figure not reproduced here; see the original article.)

Figure 13 illustrates the effects of varenicline on diabetes-related mortality, diabetes-related hospital admission, myocardial infarction and COPD. For diabetes-related mortality, the instrumental variable analysis provides little evidence of an effect in the year after the first prescription, but shows a rapid increase in risk by 24 and 48 months. The partially and fully adjusted multivariable-adjusted regression results suggest that prescribing varenicline reduces the risk of diabetes-related mortality. There was little evidence from either the instrumental variable or the multivariable-adjusted regression that prescribing varenicline affected likelihood of diabetes-related hospital admission. The instrumental variable point estimates suggested an increased likelihood of primary care diagnosis of myocardial infarction and COPD in months 12–48 after the first prescription for myocardial infarction. However, these effects were imprecise and consistent with the null and the protective effects implied by the multivariable-adjusted estimates. The instrumental variable analysis suggested that varenicline caused an increase in the risk of a diagnosis of COPD after 48 months, but little evidence of an effect at other follow-up periods. The multivariable-adjusted estimates implied a reduction in risk of COPD at all follow-up periods; these differences attenuated after adjustment for baseline confounders. For primary care diagnosis of depression, the instrumental variable analysis provides little evidence of an effect in any period after first prescription. The fully adjusted multivariable-adjusted regression results suggest that prescribing varenicline is associated with a lower risk of depression in the year after the first prescription, but little evidence of differences after ≥ 2 years. For primary care prescription of hypnotics or anxiolytics, the instrumental variable analysis provided little evidence of an effect at any point after the first prescription. The fully adjusted multivariable-adjusted regression analysis suggests that patients prescribed varenicline had a lower risk of being prescribed hypnotics or anxiolytics in the 2 years following the first prescription of smoking cessation therapy, but there was little difference in risk after 4 years.

FIGURE 13 Effect of prescribing varenicline on diabetes-related (ICD-10 E10–14) (a) mortality and (b) inpatient admission, (c) primary care diagnosis of myocardial infarction, (d) COPD, (e) primary care diagnosis or prescription for depression, and (f) prescription of hypnotics or anxiolytics in the 4 years following initial prescription. Linear regression is indicated by squares (▪), the solid line adjusts for basic confounders, the dashed line adjusts for all confounders listed in Table 1 and the instrumental variable results are indicated by triangles (▲). 95% CIs are indicated. CIs allow for clustering between physicians. The fully multivariable-adjusted results adjust for all the covariates listed in Table 1. Instrumental variable results use seven prior prescriptions. Reproduced from Davies et al.16 © 2018 The Authors. Addiction published by John Wiley & Sons Ltd on behalf of Society for the Study of Addiction. 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/. The figure includes minor additions and formatting changes to the original text. (Figure not reproduced here; see the original article.)

Figure 14 compares the experimental estimates reported in the literature with our estimates from the CPRD. We compared the estimates 1 year after the first prescription for comparability with the trials, which had limited follow-up. Mills et al.34 conducted a network meta-analysis of data from RCTs. We assumed an incidence rate of 0.2% to convert the risk ratios they report to the risk difference scale. They found little evidence of a difference in risk of major cardiovascular events. However, their estimates were relatively imprecise and consistent with the null. The CATS trial (Cardiac Assessments Following Different Treatments of Smoking Cessation Medications in Subjects With and Without Psychiatric Disorders)33 reported the frequency of major adverse cardiovascular outcome but was less precise than the network meta-analysis above and again consistent with the null. Both estimates are also consistent with our multivariable-adjusted and instrumental variable estimates. We included Mills et al.34 in this figure because it was a large network meta-analysis on the cardiovascular risk caused by varenicline, and we included the CATS trial in this figure because it was a large RCT published after Mills et al.34

FIGURE 14 The effect of varenicline compared with NRT estimated using RCTs and observational studies. The Mills et al.34 estimate assumes an incidence rate of 0.2% in those allocated to NRT.34 The CATS trial estimates are taken from reported proportions of major adverse cardiovascular events in the varenicline and NRT arms of the trial. The asymmetric CIs on the Mills et al.34 result are because we converted the risk ratios to risk differences. Adapted from Davies et al.16 © 2018 The Authors. Addiction published by John Wiley & Sons Ltd on behalf of Society for the Study of Addiction. 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/. The figure includes minor additions and formatting changes to the original text. (Figure not reproduced here; see the original article.)