Using Negative Control Outcomes and Difference-in-Differences Analysis to Estimate Treatment Effects in an Entirely Treated Cohort: The Effect of Ivacaftor in Cystic Fibrosis

Abstract

Abbreviations

 
• CF

  cystic fibrosis

 
• CI

  confidence interval

 
• CTE

  causal treatment effect

 
• DAG

  directed acyclic graph

 
• IV

  intravenous

 
• NCCTE

  negative-control–corrected treatment effect

 
• NCE

  negative control effect

 
• NCO

  negative control outcome

 
• NTE

  naive treatment effect

 
• ppFEV₁

  percent predicted forced expiratory volume in 1 second

 
• SWIG

  single-world intervention graph

Randomized controlled trials are the gold standard for estimating treatment effects but are typically infeasible for estimating long-term effects. Observational data provide opportunities to estimate such effects, under strong assumptions. A key assumption is positivity, meaning individuals have a probability less than 1 of receiving (or not) the treatment, given any covariates controlled in the analysis (1).

One situation where this assumption is not met is when an entire cohort of patients receives treatment. Then it is difficult to estimate a treatment effect because we do not observe contemporary untreated individuals. It may be possible to identify a comparable group who could not receive treatment, for example, historical controls prior to its availability. However, resultant analyses make the strong assumption of no differences between the control and treatment group that affect the outcome, except the treatment itself and its consequences. We consider estimation of causal treatment effects (CTEs) in this setting through an investigation of the effect of the disease-modifying treatment ivacaftor in people with cystic fibrosis (CF).

In the United Kingdom, approximately 10,500 persons have CF (2). The most seriously affected organ is the lung, with long-term deterioration in lung function observed. Ivacaftor has been available in the United Kingdom since 2012, and approximately 5% of the UK CF population—those with a particular CF-causing gating mutation—are eligible to receive it.

Randomized controlled trials have found that ivacaftor improves lung function and reduces the incidence of pulmonary exacerbations (3–5). Ramsey et al. (3) reported that in patients aged 12 years or older with the specified gating mutation, the mean percent predicted forced expiratory volume in 1 second (ppFEV₁) was 10.5 percentage points higher after 48 weeks of ivacaftor treatment versus placebo, and the ivacaftor group was 55% less likely to have a pulmonary exacerbation. Similar results have been reported in younger children (4, 5). These patients will take ivacaftor for many years, and it is hoped that it will change their slope of lung function decline (“slope-change effect”), as well as improve lung function during the initial treatment phase (“step-change effect”).

In most countries with high CF prevalence, almost all eligible patients now receive ivacaftor, meaning that observational data provide no contemporary controls. Four studies using national patient registry data compared ivacaftor users either with people not receiving ivacaftor because they did not have a gating mutation (“genotype comparison”) or people who were eligible for ivacaftor but in the time period prior to its availability (“time-period comparison”) (6–9). The results were similar to randomized controlled trial findings and also suggested longer-term benefits up to 4 years later. However, even after accounting for baseline differences between the treated and untreated, these studies were prone to bias because of people with different CF-causing mutations having different disease trajectories, or because of general improvements over time in the health of the CF population (10). Lung function decline among people with different CF genotypes has previously been found to be similar (11, 12), but even small differences could result in biased findings. The health of people with CF has been improving over time, which could affect time-period comparisons.

In this paper, we use directed acyclic graphs (DAGs) and single-world intervention graphs (SWIGs) (13) to illustrate assumptions made in the choice of control groups for the genotype and time-period comparisons. We describe how negative control outcomes (NCOs) can be used as a tool to detect bias in treatment effect estimates in this setting and used in combination with the difference-in-differences approach to obtain unbiased estimates of the CTE under weaker assumptions. The methods are applied using data from the UK Cystic Fibrosis Registry (14) to obtain more robust estimates of the effect of ivacaftor on lung function and annual rate (days/year) of intravenous (IV) antibiotic use (i.e., IV days).

DATA

We use UK Cystic Fibrosis Registry data from 2008–2016. The registry has been described elsewhere (14). Briefly, patients undergo an annual assessment, which captures measures of current health status, events that have occurred since the last assessment, and treatments used. Our analyses include 8,444 people: 467 with a gating mutation, and therefore eligible to receive ivacaftor after its introduction, and 7,977 with a nongating mutation (see Web Figure 1, available at https://doi.org/10.1093/aje/kwab263).

We define 2008–2012 as the pre-ivacaftor period and 2013–2016 as the post-ivacaftor period. Outcomes measured in 2012 were made prior to ivacaftor’s becoming available to eligible patients. Key data for this analysis include treatment (ivacaftor), genotype (gating mutation or nongating mutation), calendar period (pre- or post-ivacaftor period), and outcomes (lung function, annual number of IV days). For lung function, the main analyses use ppFEV₁. We also present results for 2 other measures of lung function (percent predicted forced vital capacity and percent predicted forced midexpiratory flow). Adjusted analyses use several demographic and clinical variables, as measured in 2008 for analyses in the pre-ivacaftor period and in 2012 for the post-ivacaftor period (see “Analysis” section).

We divide individuals into 4 groups defined by genotype and time period (Table 1). Ivacaftor is only available in group B, this being the group with an eligible genotype (G = 1) in the post-ivacaftor period (P = 1). The other 3 groups do not receive ivacaftor, either because it is not yet available (A) or because they have an ineligible genotype (D) or both (C). Some individuals appear in both groups A and B or both groups C and D.

METHODS

Causal treatment effect

The CTE of interest is the effect of ivacaftor on an outcome |$Y$| in persons who actually receive the treatment (⁠|$X=1$|⁠). Let |${Y}^{X=0}$| denote the potential value of Y had, contrary to fact, a patient not received ivacaftor, and let |${Y}^{X=1}$| denote the potential value of |$Y$| had a patient received ivacaftor. The CTE is the average treatment effect in the treated, and for a continuous outcome (ppFEV₁) this is measured using the mean difference:

Naive treatment effect

Figure 1

Directed acyclic graphs (DAGs) and single-world intervention graphs (SWIGs) illustrating scenarios 1 (first column), 2 (second column), and 3 (third column). Panels A–C are DAGs showing 3 possible causal pathways between ivacaftor use (⁠|$X$|⁠), genotype (⁠|$G$|⁠), time period (⁠|$P$|⁠), measured covariates of health at baseline (⁠|$H$|⁠), and the outcome (⁠|$Y$|⁠). Panels D–F are corresponding SWIGs for the intervention world in which |$X$| is set to 0.

Open in new tabDownload slide

Scenario 1: G and P are conditionally independent of Y given X

In DAG A (Figure 1A), receiving treatment |$X$| depends deterministically on |$G$| and |$P$|⁠, but |$Y$| is independent of |$G$| and |$P$| conditional on |$X$|⁠. From the corresponding SWIG (Figure 1D), we see that |${Y}^{X=0}$| is independent of |$G$| and |$P$|⁠. Thus, |$E({Y}^{X=0}|P=1,G=1)=E({Y}^{X=0}|P=p,G=g)$| for any |$g,p$|⁠. In particular, |$E({Y}^{X=0}|P=1,G=1)=E({Y}^{X=0}|P=0,G=1)=E({Y}^{X=0}|P=1,G=0)$|⁠. In group A |$(P=0,G=1)$| and group D |$(P=1,G=0)$|⁠, |${Y}^{X=0}=Y$|⁠; therefore, |$E({Y}^{X=0}|P=1,G=1)=E(Y|P=0,G=1)=E(Y|P=1,G=0)$|⁠. It follows that |${\mathrm{NTE}}_P$| and |${\mathrm{NTE}}_G$| correspond to the CTE.

Scenario 2: G and P are conditionally independent of Y given X and H

Scenario 3: G and P are not conditionally independent of Y

In DAG C (Figure 1C), there is dependence between |$Y$| and |$G,P$| conditional on |$X$|⁠. In an extended version of DAG C in Web Figure 2, we add covariates |$H$|⁠. The main issue is encompassed in DAG C (Figure 1C), and we focus on this here. The corresponding SWIG (Figure 1F) shows that |${Y}^{X=0}$| is not independent of |$G$| and |$P$|⁠. Now, the unadjusted NTEs (equations 4 and 5) do not correspond to the CTE. If there remains dependence between |$Y$| and |$G,P$| after conditioning on |$H$| as well as |$X$|⁠, the adjusted NTEs (equations 8 and 9) also do not correspond to the CTE.

NCOs are tools for detecting bias due to unmeasured confounding and other sources in observational studies, and they are defined as outcomes that are not affected by the treatment but have the same associations with other variables as the true outcome of interest (15). The difference-in-differences approach can also be used to estimate treatment effects in the presence of unobserved confounding. Sofer et al. (16) showed the link between NCOs and difference-in-differences analysis, using pretreatment outcome as the NCO. Here we use these tools to detect bias in the NTEs and estimate the CTE under certain assumptions. While these tools have previously been discussed primarily in the context of addressing unmeasured confounding, we use them instead to address bias due to dependence between |${Y}^{X=0}$| and |$G,P$|⁠. By noting that |${\mathrm{NTE}}_G$| can equivalently be written as |${\mathrm{NTE}}_G=\mathrm{E}(Y\vert P=1,X=1)-\mathrm{E}(Y\vert P=1,X=0)$|⁠, it is clear that |$G$| is uncontrolled in this difference, and similarly that |$P$| is uncontrolled in |${\mathrm{NTE}}_P$|⁠. This can be considered a form of unmeasured (or uncontrolled) confounding, as there are backdoor paths from |$X$| to |$Y$| through |$G$| and |$P$| that cannot be blocked using a standard analysis because of lack of positivity. The uncontrolled confounder |$G$| or |$P$| takes the role of an unmeasured confounder (⁠|$U$| in Sofer et al. (16)). We consider 2 NCOs that detect the bias in |${\mathrm{NTE}}_G$| and |${\mathrm{NTE}}_P$| due to uncontrolled confounding.

A nonzero |${\mathrm{NCE}}_G$| would indicate that the estimate of |${\mathrm{NTE}}_G$| is due not only to treatment but also to dependence between |${Y}^{X=0}$| and |$G$|⁠. As in the description of Lipsitch et al. (15), this NCE uses the treatment |$X$| and assesses its association with the NCO.

We have shown how a difference-in-differences approach to estimating the CTE corresponds to using an NCO to detect bias in the NTE when there is a violation of the positivity assumption. The NCCTEs correspond to the CTE under weaker assumptions than the NTEs. In our situation there are 2 possible NCOs, corresponding to different difference-in-differences formulae for the CTE (equations 13 and 18). Figure 2 shows reformulations of DAG C (Figure 1C), such that the outcome is shown separately by time period or genotype group. Figure 2A corresponds to the DAG of Sofer et al. (16), with |$G$| playing the role of the unmeasured (or uncontrolled) confounder |$U$|⁠. This illustrates that the outcome in period 0 is not affected by treatment but the 2 outcomes share the same association with |$G$|⁠. Our second NCO is illustrated in Figures 2B and 2C, where the outcome in group |$G=0$| (Figure 2B) is not affected by treatment but the outcomes in the 2 genotype groups share the same association with |$P$|⁠, which plays the role of the unmeasured confounder in this case. The model for |${Y}^{X=0}$| in equation 14, with |${\gamma}_{PG}=0$|⁠, is similar to Sofer et al.’s (16) model for pre- and postexposure outcomes (their equation 3).

Figure 2

Reformulation of the directed acyclic graph (DAG) in Figure 1C showing the negative control outcomes (NCOs). A) Use of the outcome in period 0 as the NCO. Panels B and C show the DAGs in genotype groups |$G=0$| and |$G=1$|⁠, respectively, when using the outcome in group |$G=0$| as the NCO. In panel A, |${Y}_{P=0}$| and |${Y}_{P=1}$| denote the pre- and posttreatment outcomes, respectively, and treatment |$X$| occurs after |${Y}_{P=0}.$| In panel A, |$G$| is an uncontrolled confounder and is equivalent to the unmeasured confounder |$U$| in the work of Sofer et al. (16). In panel C, |$P$| is the uncontrolled confounder.

Open in new tabDownload slide

We have discussed NCCTEs in the context of DAG C (Figure 1(C)). In practice, it is not known which scenario we are in. NCOs are a way of investigating the validity of the NTE as an estimate of the CTE and, combined with the difference-in-differences analysis, of correcting for bias in the NTE. In scenarios 1 and 2, the NCE is null.

An extended scenario of interest is one in which the effects of |$G$| and |$P$| on |$Y$| are partially mediated through measured covariates |$H$| (Web Figure 2). The arguments using NCOs and difference-in-differences analysis can be extended to incorporate adjustment for |$H$| (Web Appendix 2). Adjusting for |$H$| could result in more efficient estimates of the CTE.

ANALYSIS

Our aim is to estimate the causal effect of ivacaftor on those eligible to receive it in the post-ivacaftor period using longitudinal data from the UK Cystic Fibrosis Registry. We estimate NTEs, NCEs, and NCCTEs using the time-period and genotype comparisons. We focus here on the continuous outcome ppFEV₁. For the count outcome, IV days, the treatment effect is quantified by rate ratios after 1, 2, and 3 years of treatment (Web Appendix 1).

Nonparametric bootstrapping with 1,000 resamples was used to obtain 95% confidence intervals (CIs) and P values.

RESULTS

Table 2 summarizes baseline characteristics by group. Comparing groups B and D, a higher proportion of people in the ivacaftor group had an infection (88.2% vs. 80.6%) and White ethnicity (98.2% vs. 95.5%). Comparing groups A and B, the mean age was higher in group B (22.4 years vs. 20.4 years), which also had a slightly higher proportion of infections (88.2% vs. 81.9%), CF-related diabetes (22.7% vs. 15.8%), and use of mucolytic treatment (66.5% vs. 51.0%). Characteristics were otherwise similar across groups. Out of 14,594 people, many counted twice, 7,933 (54.4%) had the maximum of 4 visits postbaseline, with only 1,667 (11.4%) having just 1 visit.

Results for ppFEV₁ are shown in Figure 3 (Web Table 1). We focus on the adjusted analysis. The unadjusted results are qualitatively similar, with wider 95% CIs. First consider the step-change effect. In the time-period comparison, the NTE (⁠|${\mathrm{NTE}}_P^{\mathrm{Con}}$|⁠) estimates a 7.27–percentage-point absolute increase in ppFEV₁ (95% CI: 5.87, 8.57) in the ivacaftor group. The corresponding NCE (⁠|${\mathrm{NCE}}_P^{\mathrm{Con}}$|⁠; see Web Appendix 2) estimates a 0.77–percentage-point increase (95% CI: 0.44, 1.08), indicating a small improvement in mean absolute ppFEV₁ in the post-ivacaftor period in the |$G=0$| group. The resulting |${\mathrm{NCCTE}}_P^{\mathrm{Con}}$| estimates a 6.50–percentage-point increase (95% CI: 5.06, 7.85) in ppFEV₁. In the genotype comparison, |${\mathrm{NTE}}_G^{\mathrm{Con}}$| is estimated to be 6.22 (95% CI: 5.17, 7.24) and |${\mathrm{NCE}}_G^{\mathrm{Con}}$| is estimated to be −0.37 (95% CI: −1.36, 0.65), resulting in an |${\mathrm{NCCTE}}_G^{\mathrm{Con}}$| estimate of a 6.59–percentage-point increase in ppFEV₁ (95% CI: 5.22, 7.90). The NCE indicates that in the pre-ivacaftor period, mean ppFEV₁ was slightly lower in persons with the ivacaftor-eligible genotype than in those ineligible.

Figure 3

Estimated naive treatment effect (NTE), negative control effect (NCE), and negative-control–corrected treatment effect (NCCTE) of ivacaftor on percent predicted forced expiratory volume in 1 second (ppFEV₁) among persons with cystic fibrosis, using the time-period comparison (unadjusted: |$\mathrm{NT}{\mathrm{E}}_P$|⁠, |$\mathrm{NC}{\mathrm{E}}_P$|⁠, and |$\mathrm{NCCT}{\mathrm{E}}_P$|⁠; adjusted: |$\mathrm{NT}{\mathrm{E}}_P^{\mathrm{Con}}$|⁠, |$\mathrm{NC}{\mathrm{E}}_P^{\mathrm{Con}}$|⁠, and |$\mathrm{NCCT}{\mathrm{E}}_P^{\mathrm{Con}}$|⁠) and the genotype comparison (unadjusted: |$\mathrm{NT}{\mathrm{E}}_G$|⁠, |$\mathrm{NC}{\mathrm{E}}_G$|⁠, and |$\mathrm{NCCT}{\mathrm{E}}_G$|⁠; adjusted: |$\mathrm{NT}{\mathrm{E}}_G^{\mathrm{Con}}$|⁠, |$\mathrm{NC}{\mathrm{E}}_G^{\mathrm{Con}}$|⁠, and |$\mathrm{NCCT}{\mathrm{E}}_G^{\mathrm{Con}}$|⁠), UK Cystic Fibrosis Registry, 2008–2016. A) Absolute step change in ppFEV₁ (⁠|${\beta}_{\mathrm{St}}$|⁠); B) absolute change in the annual ppFEV₁ slope (⁠|${\beta}_{\mathrm{Sl}}$|⁠). In the adjusted analysis, results were adjusted for the baseline variables: sex, age, ethnicity, smoking status, cystic-fibrosis–related diabetes, ppFEV₁, annual number of days of intravenous (IV) antibiotic use (i.e., IV days) (including an indicator of a nonzero count and a linear term for the nonzero counts), use of mucolytic treatment, and bacterial infection. Bars, 95% confidence intervals (CIs).

Open in new tabDownload slide

Slope-change effect estimates suggest a small improvement in lung function decline due to ivacaftor, but 95% CIs for NCCTE estimates include 0. In the time-period comparison, the |${\mathrm{NTE}}_P^{\mathrm{Con}}$| estimates a 0.68–percentage-point absolute improvement in the annual rate of ppFEV₁ decline (95% CI: 0.11, 1.32). The corresponding estimated |${\mathrm{NCE}}_P^{\mathrm{Con}}$| is 0.20 (95% CI: 0.04, 0.37), giving an |${\mathrm{NCCTE}}_P^{\mathrm{Con}}$| estimate of 0.49 (95% CI: −0.15, 1.13). In the genotype comparison, |${\mathrm{NTE}}_G^{\mathrm{Con}}$| is estimated to be 0.67 (95% CI: 0.27, 1.10) and |${\mathrm{NCE}}_G^{\mathrm{Con}}$| is estimated to be 0.19 (95% CI: −0.36, 0.70), giving an |${\mathrm{NCCTE}}_G^{\mathrm{Con}}$| estimate of a 0.49–percentage-point improvement (95% CI: −0.14, 1.13).

Results for IV days are shown in Figure 4 (Web Table 2). We focus on the adjusted estimates. In the time-period comparison, according to the |${\mathrm{NTE}}_P^{\mathrm{Con}}$| the rate of IV days was estimated to decrease by 58% (95% CI: 46, 71), 72% (95% CI: 59, 79), and 69% (95% CI: 57, 81) after 1, 2, and 3 years of treatment, respectively. Corresponding estimates of |${\mathrm{NCE}}_P^{\mathrm{Con}}$| were 23% (95% CI: 17, 26), 25% (95% CI: 20, 30), and 26% (95% CI: 21, 33), indicating reductions in IV days in the post-ivacaftor period in the |$G=0$| group. The resulting |${\mathrm{NCCTE}}_P^{\mathrm{Con}}$| estimates of the percentage reduction in the rate of IV days were 45% (95% CI: 31, 64), 63% (95% CI: 45, 72), and 58% (95% CI: 42, 74) after 1, 2, and 3 years of treatment, respectively. Results from the genotype comparison were similar.

Figure 4

Estimated naive treatment effect (NTE), negative control effect (NCE), and negative-control–corrected treatment effect (NCCTE) of ivacaftor on annual number of days of intravenous (IV) antibiotic use (i.e., IV days) among persons with cystic fibrosis, using the time-period comparison (unadjusted: |$\mathrm{NT}{\mathrm{E}}_P$|⁠, |$\mathrm{NC}{\mathrm{E}}_P$|⁠, and |$\mathrm{NCCT}{\mathrm{E}}_P$|⁠; adjusted: |$\mathrm{NT}{\mathrm{E}}_P^{\mathrm{Con}}$|⁠, |$\mathrm{NC}{\mathrm{E}}_P^{\mathrm{Con}}$|⁠, and |$\mathrm{NCCT}{\mathrm{E}}_P^{\mathrm{Con}}$|⁠) and the genotype comparison (unadjusted: |$\mathrm{NT}{\mathrm{E}}_G$|⁠, |$\mathrm{NC}{\mathrm{E}}_G$|⁠, and |$\mathrm{NCCT}{\mathrm{E}}_G$|⁠; adjusted: |$\mathrm{NT}{\mathrm{E}}_G^{\mathrm{Con}}$|⁠, |$\mathrm{NC}{\mathrm{E}}_G^{\mathrm{Con}}$|⁠, and |$\mathrm{NCCT}{\mathrm{E}}_G^{\mathrm{Con}}$|⁠), UK Cystic Fibrosis Registry, 2008–2016. A) Year 1 (exp(⁠|${\gamma}_{X1}))$|⁠; B) year 2 (exp(⁠|${\gamma}_{X2}))$|⁠; C) year 3 (exp(⁠|${\gamma}_{X3}))$|⁠. The adjusted analysis adjusted for the following baseline variables: sex, age, ethnicity, smoking status, cystic-fibrosis–related diabetes, percent predicted forced expiratory volume in 1 second, IV days (including an indicator of a nonzero count and a linear term for the nonzero counts), use of mucolytic treatment, and bacterial infection. Bars, 95% confidence intervals (CIs).

Open in new tabDownload slide

Results for 2 other measures of lung function (percent predicted forced vital capacity, percent predicted forced midexpiratory flow) (Web Tables 3 and 4) were similar to those for ppFEV₁. We conducted sensitivity analyses restricting the |$G=0$| genotype group to patients who were either heterozygous or homozygous for the phenylalanine 508 deletion (F508del) (Web Tables 5–10) or only those patients who were homozygous for F508del (Web Tables 11–16), corresponding to 90% and 54% of the original |$G=0$| group, respectively. The results showed no substantial differences.

DISCUSSION

We have shown how NCOs can be used to assess whether a control group is suitable for estimating the treatment effect in a group where everyone receives treatment, and how they can be used in combination with the difference-in-differences approach to provide a more robust treatment effect estimate. Previous descriptions of NCOs have focused on unmeasured confounding bias (15–19). Potential bias in our situation is due to an inability to block all paths from |$X$| to |$Y$| using a standard analysis, which could be considered a form of unmeasured confounding. A key assumption of our methods is that there is no genotype × period product term in the model for the counterfactual outcome under no treatment (equation 14). This is a strong assumption that is not verifiable using the data, though it is weaker than the assumptions made when using NTEs. NCCTEs also provide unbiased treatment effect estimates in further scenarios—for example, allowing for unmeasured variables |$U$| affecting |$H$| and |$Y$| (Web Figure 3). It is of interest to investigate how the assumption made in our difference-in-differences analyses using NCOs could be relaxed. One approach could be the use of synthetic control methods, which make use of pre- and postintervention observations in the group receiving the intervention, and observations in multiple time periods for groups that have not received the intervention (20, 21).

Our NTE estimates for the effect of ivacaftor are similar to results from previous studies, which found that ivacaftor results in a step-change absolute improvement in ppFEV₁ of 3.2–8.2 percentage points and a decrease in the rate of annual ppFEV₁ decline of approximately 0.8 percentage points (6, 9). However, these are only unbiased estimates of the treatment effect if the assumptions that |$G$| and |$P$| are conditionally independent of |$Y$| given |$X$| (or |$X$| and |$H$|⁠) are valid.

The time-period comparison NCE showed a 0.77–percentage-point absolute increase in ppFEV₁, indicating a small, non–clinically significant improvement in population average lung function since the introduction of ivacaftor. This means that the NTE slightly overestimates the ivacaftor effect. In the genotype comparison, the NCE was negative, indicating slightly lower ppFEV₁ in the eligible genotype group versus the ineligible group in the pre-ivacaftor period, but the NCE was also small, as it was in the time-period comparison. This resulted in NCCTE estimates of step-change improvement in ppFEV₁ of 6.50% and 6.59%, which are similar to the NTE estimates (7.27% to 6.22%) but have wider CIs, correctly reflecting uncertainty in the comparability of the groups.

When considering ivacaftor’s effect on the slope change of lung function, the NCE suggested that some of the NTE estimate was due not to ivacaftor but to general improvements in lung function decline over time. The NCCTE suggests a beneficial effect of ivacaftor, with an estimated absolute improvement in annual rate of decline of 0.49%, though with 95% CIs including 0.

Findings for the ivacaftor effect on rate of IV days were similar to those from previous studies, indicating a treatment benefit that persists at least up to 3 years. The NCE results estimated that in the absence of treatment, the rate of IV days was slightly lower in the |$G=1$| group versus the |$G=0$| group and in the later time period. NCCTE estimates were therefore slightly attenuated in comparison with the NTEs. Our results support evidence of a long-term clinical benefit of ivacaftor.

ACKNOWLEDGMENTS

Author affiliations: Department of Medical Statistics, Faculty of Epidemiology and Population Health, London School of Hygiene and Tropical Medicine, London, United Kingdom (Simon J. Newsome, Ruth H. Keogh); Division of Population Medicine, School of Medicine, Cardiff University, Cardiff, United Kingdom (Rhian M. Daniel); National Heart and Lung Institute, Imperial College London, London, United Kingdom (Siobhán B. Carr, Diana Bilton); and Royal Brompton Hospital, London, United Kingdom (Siobhán B. Carr, Diana Bilton).

This work was supported by the Cystic Fibrosis Trust (Strategic Research Centre grant “CF-EpiNet: Harnessing Data to Improve Lives” (to S.B.C.)); UK Research and Innovation (Future Leaders Fellowship MR/S017968/1 (to R.H.K.)); the Medical Research Council (Methodology Fellowship MR/M014827/1 (to R.H.K.)); and the Wellcome Trust and the Royal Society (Sir Henry Dale Fellowship 107617/Z/15/Z (to R.M.D.)).

This work used anonymized data from the UK Cystic Fibrosis Registry, which has research ethics approval (Research Ethics Committee reference number 07/Q0104/2). Use of the data was approved by the Registry Research Committee. Data are available following application to the Registry Research Committee (https://www.cysticfibrosis.org.uk/the-work-we-do/uk-cf-registry/apply-for-data-from-the-uk-cf-registry).

We thank the Cystic Fibrosis Epidemiological Network (CF-EpiNet) Strategic Research Centre data group for contributions to data preparation and cleaning.

R.H.K. has received funding through a Vertex Circle of Care Award and personal fees from Vertex Pharmaceuticals, Inc. (Boston, Massachusetts). D.B. was the principal investigator in a UK Cystic Fibrosis Registry–based pharmacovigilance study of ivacaftor funded by Vertex Pharmaceuticals. S.B.C. reports receiving speaker’s fees, honoraria, and advisory board fees from Chiesi Ltd. (Manchester, United Kingdom), payment to her institution for advisory boards from Vertex, and payment for consultation work from Profile Pharma Ltd. (Chichester, United Kingdom), Vertex Pharmaceuticals, and Insmed Inc. (Bridgewater Township, New Jersey) outside the scope of this work.

REFERENCES