Cabozantinib for previously treated advanced renal cell carcinoma

In the intention-to-treat population of METEOR (December 2015 data cut-off):

• Progression-free survival improved with cabozantinib compared with everolimus (median 7.4 and 3.9 months respectively; hazard ratio 0.51; 95% confidence interval [CI] 0.41 to 0.62; p<0.0001).

• Overall survival improved with cabozantinib compared with everolimus (median 21.4 and 16.5 months respectively; hazard ratio 0.66; 95% CI 0.53 to 0.83; p=0.00026).
  
  The committee also noted the updated survival data from METEOR, presented by the company in response to the first consultation (based on a cut-off date of October 2016), and welcomed the availability of new, more mature data. It concluded that cabozantinib was more effective than everolimus in METEOR.

Because there were no head-to-head trials comparing cabozantinib with axitinib or nivolumab, the company did a network meta-analysis to compare the treatments indirectly. The original network linked 6 trials, including TARGET, which compared sorafenib with placebo. Although sorafenib was not a comparator for cabozantinib in this appraisal, the company included TARGET to link together treatments. The committee was concerned about including this trial for 2 reasons. First, none of the patients had previously had VEGF-targeted therapies. Second, the company used immature data from the trial, which censored patients who switched from placebo to sorafenib. This was likely to have underestimated the effect of sorafenib because the placebo data reflected patients whose disease responded relatively well (who were therefore not censored), and this would in turn have underestimated the effect of axitinib. In response to the first consultation, and in line with the committee's preference, the company submitted a revised network that excluded TARGET. This assumed that axitinib was as effective as everolimus in terms of overall and progression-free survival. The committee concluded that the company's simplified network reduced the potential bias associated with using immature data from TARGET.

The committee understood that, to estimate long-term outcomes, the company used a 'family' of related survival curves for cabozantinib and for all of the comparator treatments. The company chose the curves based on how well, on average, they fitted the data on overall or progression-free survival for all the treatments in the network. The committee noted that, because of this simplification, the parametric distribution chosen by the company for both progression-free and overall survival (log-normal for both endpoints) did not fit the data for each individual treatment well. In response to the first consultation, the company used fractional polynomial modelling, as described by Janssen et al. (2011), to fit survival curves. The new method also used a family of related survival curves for all the treatments. However, the committee agreed that it was a more flexible family, which improved the curve fits to the Kaplan–Meier data on overall and progression-free survival for all treatments in the network compared with the original parametric modelling using the log-normal distribution. The committee appreciated that the fractional polynomial modelling did not fit data in the extrapolation period. It noted that the evidence review group (ERG) considered that estimating survival based on the 'average fit' across the network (as opposed to the fit for each individual treatment) was less of an issue with fractional polynomial models than with parametric curve fitting. The committee was satisfied that the company's revised modelling of overall and progression-free survival was more appropriate than the original parametric modelling.

The committee considered the company's revised modelling in response to the first consultation. It noted that, to estimate overall and progression-free survival for cabozantinib and its comparator treatments, the company extrapolated the curves based on fitting fractional polynomial models (see section 4.10) up to the end of the time horizon. As such, to estimate overall and progression-free survival for cabozantinib and its comparator treatments, the company used fractional polynomial modelling during both the trial follow-up and extrapolation. Hereafter, this analysis will be referred to as 'the company's revised base case'.

In its revised base case (see section 4.13), the company predicted that 5% of people in the everolimus arm would be alive 5 years after starting treatment. The committee compared this estimate with 2 sources of observational data submitted during the first consultation:

• Registry-based pharmaco-epidemiological data from a publication by Ruiz-Morales et al. (2016) submitted by the company. These data came from the International Metastatic renal cell carcinoma Database Consortium (IMDC) reflecting people initially treated with either pazopanib or sunitinib. Some people then had second-line treatment. The company presented data for people who had second-line treatment after sunitinib (n=2,667) because this group was larger than the group that had second-line treatment after pazopanib (n=260). It noted that, in this group, about 10% were alive 5 years after starting treatment.

• Audit data from the Christie Hospital (Manchester, UK) submitted by a clinical expert. These data showed that, among people who had axitinib or everolimus as a second-line treatment (n=282), around 6% were alive 5 years after starting treatment.

The committee discussed whether the Ruiz-Morales et al. (2016) data on were generalisable to patients who would be offered everolimus in the UK. It observed that:

• Ruiz-Morales et al. did not include patients from the UK. The committee acknowledged that the company considered that this study included people with similar characteristics at baseline to patients in METEOR, and that the countries from which these people were included had similar socio-economic profiles and health systems to the UK.

• As second-line treatment, only 45% of people had everolimus, and some had treatments that were not available in the NHS.

• Ruiz-Morales et al. did not report information on the third-line treatments; these treatments may not be available in the NHS, and may have biased the effect of second-line treatment.
  
  For these reasons, the committee agreed that the survival estimates from Ruiz-Morales et al. were likely to overestimate the survival of patients who have everolimus in the NHS. It considered the 5‑year survival estimate from the Christie Hospital audit to be unreliable because the numbers were small and there were no observations beyond 3 years 3 months after starting treatment. The committee concluded that survival in the UK was likely to have been overestimated in Ruiz-Morales et al., but it did provide useful data with which to compare the survival prediction of the company's model.

The committee noted that the company presented a scenario analysis to align the revised base-case predictions (see section 4.14) and the data from Ruiz-Morales et al. (2016). In this, the company did not change the modelling of progression-free survival, that is, it continued to use fractional polynomial modelling across the entire time horizon. For overall survival, it used fractional polynomial modelling during the trial follow-up period (as per the revised base case), but used parametric modelling choosing the log-normal distribution during the extrapolation period. The committee noted that this scenario aligned the model's predictions of survival with data from Ruiz-Morales et al. However, the committee did not consider that it was appropriate to base the extrapolation on meeting the 5‑year death rate observed in Ruiz-Morales et al. The committee recalled that survival among people having everolimus was likely to have been overestimated in Ruiz-Morales et al. It concluded that it preferred the company's revised base case, which used fractional polynomial modelling across the entire time horizon for both overall and progression-free survival.

The committee noted that both the company and the ERG assumed that the effect of cabozantinib continued beyond the trial follow-up, even after the disease progressed or patients stopped treatment, but the committee was not presented with evidence to support this. The clinical experts considered that it was not clear whether a survival benefit would continue after stopping treatment. They explained that, in clinical practice, some patients have stable disease for 2 to 3 years after stopping treatment, whereas the disease progresses more quickly in others. Also, some patients have a prolonged response after a short length of treatment and others do not. The committee concluded that assuming the effect of cabozantinib continues for up to 30 years, based on a trial with a median follow-up of under 2 years for overall survival, was highly uncertain.

The committee noted that, for nivolumab, the company's revised base case (see section 4.13) estimated a longer progression-free survival than overall survival. It understood that, in the model, disease progression could occur until the point where overall and progression-free survival curves cross (around 5 years after starting treatment), after which people whose disease had not progressed followed the overall survival curve. This meant that the company assumed that the disease would never progress at that point, instead people would die of causes other than their cancer. It also meant that they would accrue the utility associated with pre-progressed disease during their remaining time in the model. The committee was not presented with any evidence that people who are alive and on treatment 5 years after starting treatment remain progression-free until they die. The company did not consider it plausible that progression-free survival would be longer than overall survival, and conducted a scenario analysis. In this, it continued to use fractional polynomial modelling for overall survival across the entire time horizon as in the revised base case. For progression-free survival, it used parametric modelling using the log-normal distribution during both the trial follow-up and extrapolation periods. The committee recalled that the log-normal distribution did not fit the data for the individual treatments well (see section 4.10). Because of this, the committee did not consider this analysis further. It recognised the uncertainties in the company's revised base case with respect to the modelling of nivolumab, but concluded it could use it for decision-making.

The committee noted that the company presented a further scenario analysis that, as in the previous scenario (see section 4.18), used the log-normal distribution to model progression-free survival across the entire time horizon. However, it differed in that of those who were alive 5 years after starting treatment and still having nivolumab, the company assumed that half had the same mortality as the age-matched general population. The committee recalled that using the log-normal distribution to model progression-free survival did not produce robust estimates (see section 4.18). Furthermore, the committee noted that this scenario had little impact on the mean overall survival associated with nivolumab, which it did not expect. The company suggested that this may have been because the risk of death estimated by the log-normal curve was similar to that of the general population. The committee recalled from the NICE technology appraisal on nivolumab that the committee preferred to base its decision on a mixed model that relied 50% on a single generalised gamma model and 50% on a model that assumed a greater long-term survival benefit than in the single generalised gamma model for nivolumab. The committee concluded that the scenario analysis was unrealistic, but appreciated that the company had explored predictions of better survival for nivolumab.

The committee was aware that METEOR collected health-related quality-of-life data using the EQ‑5D‑5L measure, which the company adjusted for age, as requested by the committee, and used in its revised base case. The committee considered these data, together with data from other studies, including those used in previous appraisals of renal cell carcinoma. It noted that the available utility values varied widely, particularly those used for the post-progression state. The ERG explained that the utility values collected from METEOR were higher than those clinicians would expect to see in clinical practice and, notably, the utility value before disease progression was higher than that of the age-matched general population. Because of this, the ERG explored using utility values from the AXIS trial. The committee accepted that the new, more detailed version of the EQ‑5D (EQ‑5D‑5L) used in METEOR could explain the relatively high utility values reported in this trial. It also heard from the company that there was evidence showing that utility estimates were higher using EQ‑5D‑5L. Another possible explanation was greater attrition bias in METEOR, in which unhealthy people were less likely to continue filling in quality-of-life questionnaires. The committee was also aware that AXIS and METEOR differed in whether they allowed patients to switch between treatment arms, the number and type of therapies that patients took before enrolling in the trial or after the disease progressed during the trial, and the prognostic scores at baseline of the study populations. In response to the second consultation, the company argued that the utility from AXIS was not the most relevant because patients in the trial had had either a cytokine or a VEGF-targeted therapy, and it was unclear what the effect was of having previous cytokines on the patient's health. The committee generally preferred sourcing utility and effectiveness from the same trial. However, it agreed that some of the utility values from METEOR appeared high, particularly the utility value before disease progression. The committee concluded that it would take into account both sets of utility values in its decision-making.

The committee considered the cost-effectiveness results incorporating the revisions to the models in response to the first consultation, the new data from METEOR (cut-off of October 2016), and the confidential discounts for all technologies applied by the ERG. It was presented with results with and without everolimus included in the analysis but, because the committee concluded that everolimus was not a relevant comparator for cabozantinib, it considered only the analyses excluding everolimus. In its consideration of the cost-effectiveness estimates, the committee took into account:

• the company's revised base case (see section 4.10)

• the company's scenario analysis using fractional polynomial modelling during the trial follow-up period, and parametric modelling using the log-normal distribution during the extrapolation period (see section 4.16)

• the ERG's revised base case (which reflected minor changes with minimal impact on the results compared with the company's revised base case)

• the ERG's scenario analysis exploring utility values from AXIS.