Axitinib, cabozantinib, everolimus, nivolumab, sunitinib and best supportive care in previously treated renal cell carcinoma: a systematic review and economic evaluation

Population

The population in all of the studies comprised patients with amRCC who had received at least one prior VEGF-targeted therapy apart from AXIS, in which only a subgroup of 54% of patients received prior VEGF-targeted therapy. 43 Data for the eligible subgroup were used for the primary outcome analyses (OS and PFS) to minimise potential bias, but data used for the secondary outcomes refer to the full AXIS population. 43 The sample size varied across the studies from 3358 to 821. 54 The sample size was generally higher among the RCTs (range 41664 to 82154) than among the observational studies (range 3358 to 45262).

All of the studies recruited adults (people aged ≥ 18 years), with the median age of patients across the studies, when age was reported at baseline, generally between 60 and 70 years. When ethnicity was reported, > 70% of the patients in the studies were white and there was a higher proportion of males than females in all studies (lowest proportion 58% in the everolimus group in the ESPN study55 and highest proportion 86% in both groups in Paglino et al. 60).

Seven out of the eight studies that did describe RCC subtype recruited a solely or primarily clear-cell population. ESPN was the only study to recruit a non-clear-cell population, or clear cell with at least 20% sarcomatoid features. 55 The type of RCC was not reported in the three retrospective studies comparing the sequence of sunitinib and sorafenib,58,60,61 or in SWITCH. 56 The stage of RCC was generally poorly reported, although all studies had inclusion criteria that patients were required to have amRCC. This inclusion criteria would suggest that all patients were a minimum of stage II, although the spread of patients over the different stages across the studies is unclear. Baseline prognostic score was not reported for all studies although, when reported, nearly all patients had a reasonably good performance status (i.e. ECOG performance status 0 or 1).

Study inclusion and exclusion criteria with regard to prior therapies varied across the studies (see Table 3). Prior therapies in the table reflect the population included in this review and not necessarily the study inclusion criteria (e.g. treatment-naive patients were recruited at the start of ESPN,55 but had all received one prior treatment in the period included in our analyses). Eight study populations had received one prior TKI treatment (or mTORi in the case of ESPN),43,55,56,58,60–63 and four included patients who had received two lines of prior therapy: RECORD-164 specified one or two prior TKIs, CheckMate 02554 and Iacovelli et al. 59 allowed two prior targeted therapies, and METEOR57 allowed any number of prior TKI treatments. Other types of prior therapy (e.g. chemotherapy, cytokines, bevacizumab) were allowed in most studies, and prior mTORi therapy was usually not permitted. Prior cytokine use was exclusionary in Vogelzang et al. 62 and Wong et al. 63 The impact of prior therapies is discussed and explored through subgroup analyses (see Subgroup analyses).

Formal statistical tests for between-treatment group differences at baseline were generally not reported. On visual inspection, the RCT populations appear to be balanced but there were notable differences between treatment groups in some of the observational studies. Imbalances include the percentage of second-line patients in Porta et al. 61 (62% sunitinib vs. 29% sorafenib); mean age in Calvani et al. 58 (70 years sunitinib vs. 61 years sorafenib); percentage male imbalances in Calvani et al. ,58 ESPN,55 and Porta et al. ;61 and differences in the distribution of MSKCC prognostic scores in Paglino et al. 60 and Porta et al. 61 Therefore, the results from the observational studies may be subject to bias when the analyses have not been adjusted for these baseline imbalances (e.g. age could have a substantial impact on OS irrespective of treatment effect).

Intervention and comparator

The RCTs evaluating all of the treatments of interest apart from sunitinib were identified. The inclusion of observational studies on axitinib, sorafenib and sunitinib also led to the inclusion of observational studies for everolimus. There were two studies that included axitinib (one RCT43 and one observational study),62 one study for cabozantinib (one RCT),57 seven studies for everolimus (three RCTs and four observational studies),53–55,59,62,63 one study for nivolumab (one RCT54) and five studies for sunitinib (five observational studies). 55,56,58,60,61 There were seven studies that also included sorafenib (one RCT,43 and six observational studies),56,58–61,63 and one that included BSC (one RCT). 53 Network diagrams for PFS (Figures 2 and 3) and OS (Figures 4 and 5) illustrate which direct treatment comparisons contributed to each MTC.

FIGURE 2.

Network diagram for PFS (primary analysis). Notes: the size of the nodes represent the number of patients on each intervention. The thickness of the lines represents the number of studies informing the direct comparison.

FIGURE 3.

Network diagram for PFS for SA1. Notes: the size of the nodes represent the number of patients on each intervention. The thickness of the lines represents the number of studies informing the direct comparison.

FIGURE 4.

Network diagram for OS primary analysis. Notes: the size of the nodes represent the number of patients on each intervention. The thickness of the lines represents the number of studies informing the direct comparison.

FIGURE 5.

Network diagram for OS SA. Notes: the size of the nodes represent the number of patients on each intervention. The thickness of the lines represents the number of studies informing the direct comparison.

The active study drug dose was not specified in all studies but, when it was specified, it was the standard licensed dose (see Chapter 1, Description of the technologies under assessment). Doses were varied according to clinician and patient factors, with limited details reported on these dose adjustments. No study reported the explicit use of any concomitant medications.

The duration of treatment was only reported in the four RCTs and one of the crossover observational studies. 43,53,54,56,57 As such, we are uncertain of the variation in treatment duration within the observational studies and whether or not it differed between them and the RCTs. The median treatment duration in the five studies, when it was reported, varied from 1.9 months [placebo (BSC) group of RECORD-1] to 8.3 months (cabozantinib group of METEOR). 57,64 Treatment was reported in the RCTs to be continued until disease progression, unacceptable toxicity or withdrawal of consent. Treatment discontinuations are discussed alongside the quality assessment in Quality assessment of studies. Median length of follow-up was also poorly reported among the included studies with only four studies reporting data. 55,57,62,63 The median length or follow-up ranged from 12.1 months to 23.6 months. 55,63 One study reported mean length of follow-up, which was 10.3 months. 56 Study duration was also not reported for all studies and varied widely when it was reported (range 22–44 months), although the longer studies relate to the crossover RCTs used as observational studies (and so the first part of the study is not relevant to this review). These data for length of study and length of follow-up should, thus, be interpreted with caution.

Details on subsequent therapies following treatment discontinuation in the included studies was limited. Most studies allowed subsequent treatment in the event of progression or intolerable toxicity on the study drug, but gave minimal detail about what was actually received and the possible impact it might have on the results. However, in RECORD-1, BSC (placebo) patients could cross over to receive open-label everolimus during the study and it was reported that 76.2% of patients did so. 53 The OS results of RECORD-1 used in the MTC are crossover adjusted in an attempt to address the potential bias introduced from the high level of crossover. 53,65 Treatment crossover was not reported to have occurred in any other studies, except when it was part of the study design. In METEOR and Checkmate 025, patients were allowed to continue the study therapy after initial disease progression if a clinical benefit was noted by the investigator. 54,57

Outcomes

The outcomes of interest to this review and reported in the included studies are listed in Table 4. Data for all of the outcomes were available, although outcome data were not available for all of the interventions. The primary outcome in the majority of the studies was PFS. The data reported in the six retrospective observational studies comprised only PFS and OS data, which as discussed earlier (see Methods for reviewing effectiveness) have only been used in SAs. 58–63 Data from the RCTs also included response rate, HRQoL (RCT data only) and AE data; however, the reporting of these outcomes varied across the studies both in terms of their presence and the type of data (e.g. the HRQoL tools used and individual AEs reported varied across studies). 43,54–57,64 The available data will be discussed in detail in the results subsections below (see Assessment of effectiveness to Subgroup analyses).

Progression-free survival

Comparative clinical effectiveness of PFS was evaluated through a MTC. The primary network generated comprised just two studies (RECORD-1 and METEOR) and provides information on three treatments: cabozantinib, everolimus and BSC (see Figure 2). 53,57 As described in Chapter 2, Comparators, the term BSC has been used to refer to placebo throughout this report and placebo is assumed to be a surrogate for BSC. In RECORD-1, PFS was defined as the time from randomisation to the first documentation of disease progression or death (from any cause) assessed via blinded independent central review. Similarly, in METEOR, PFS was defined as the time from randomisation to radiographic progression per RECIST or death from any cause. In CheckMate 025,54 PFS was also assessed using the RECIST criteria, which it has been suggested does not take into account ‘tumour flare’. Tumour flare is a result of the immune response to immunotherapies like nivolumab and may be misinterpreted as progression. However, while the evidence review groups’ (ERGs) clinical experts consider that, in theory, the RECIST criteria may be conservative for assessing PFS in patients treated with immunotherapies like nivolumab, they also consider tumour flare to be rare in clinical practice. Unfortunately, no estimates of PFS for nivolumab could be generated using MTC because the KM curves suggested that proportional hazards (PHs) did not hold for this outcome in CheckMate 025,54 which was the only study evaluating nivolumab in this review. Axitinib and sunitinib were not included in the primary analysis as these interventions were assessed in studies that could not be connected in a network of high-quality studies. SA1 included observational studies of reasonable quality in addition to the RCTs; SA2 included all relevant studies identified (i.e. RCTs and observational studies of any quality, including studies deemed to be at a critical risk of bias).

Sensitivity analysis 1 incorporated the two RCTs from the primary analysis and five observational studies. 55,56,58,62,63 The network created by these additional studies facilitated the inclusion of one additional RCT (AXIS – prior sunitinib subgroup) and provides PFS estimates for axitinib and sunitinib. 43 A network diagram for SA1 is provided in Figure 3.

Sensitivity analysis 2 included the eight studies in SA1 as well as the two observational studies60,61 rated being at a critical risk of bias because of confounding from significant imbalances of patient characteristics at baseline. 60,61 The two additional studies in SA260,61 provided additional data on sunitinib and sorafenib with the network otherwise remaining the same as for SA1.

Everolimus was chosen as the baseline treatment for the MTCs because of the comparatively large number of studies available for analysis.

The results of the primary analysis for PFS are presented in Table 7. Cabozantinib (HR 0.17, 95% CrI 0.12 to 0.24) and everolimus (HR 0.33, 95% CrI 0.25 to 0.43) showed a statistically significant PFS benefit compared with BSC, and cabozantinib showed a benefit over everolimus (HR 0.51, 95% CrI 0.41 to 0.63).

The results of SA1 were consistent with that of the primary analysis and provided estimates for additional treatment comparisons (see Appendix 4). Everolimus (HR 0.33, 95% CrI 0.25 to 0.43), cabozantinib (HR 0.17, 95% CrI 0.12 to 0.24), axitinib (HR 0.31, 95% CrI 0.214 to 0.44) and sunitinib (HR 0.27, 95% CrI 0.17 to 0.40) all showed a benefit on PFS compared with BSC. Cabozantinib has significantly better PFS than all other treatments: everolimus (HR 0.51, 95% CrI 0.41 to 0.63), sunitinib (HR 0.63, 95% CrI 0.44 to 0.95), axitinib (HR 0.54, 95% CrI 0.40 to 0.76) and BSC (HR 0.17, 95% CrI 0.12 to 0.24). Differences in PFS between sunitinib, everolimus and axitinib were not statistically significant. Based on sampling from the MTC, cabozantinib has a 99% probability of being the most effective treatment for improving PFS compared with the other treatments included in the analysis.

The results of SA2 (see Appendix 4) were similar to those of SA1 and the primary analysis, with no change in the statistical significance of any of the results.

The residual deviance was similar to the number of unconstrained data points in all of the MTC analyses for PFS except for SA2, for which the residual deviance was slightly lower than the number of unconstrained data points (8 points vs. 10 points, respectively for SA2). These results suggest that the fixed-effects MTC model was a good fit for the primary analysis and SA1. The random-effects model was deemed unsuitable as discussed in Methods of data synthesis.

The direct and indirect estimates of the HRs generated for the interventions in the connected loops in SA1 were compared to assess possible inconsistency in the MTC. There were two loops in SA1: loop 1, consisting of everolimus, axitinib and sorafenib, and loop 2, consisting of everolimus, sorafenib and sunitinib. The results of the inconsistency assessments demonstrated no evidence of significant inconsistency (p < 0.05, see Appendix 5).

Overall survival

The primary analysis for OS included three RCTs53,54,57 covering four interventions: cabozantinib, everolimus, nivolumab and BSC. CheckMate 025 was included in the MTC analysis for OS because inspection of the KM curves suggested that the assumption of PHs holds from 6 weeks onwards. 54 No high-quality studies connect axitinib or sunitinib to the network for the primary analysis. As 6 weeks is a small proportion of time in the analysis of OS, the pragmatic decision was made to include CheckMate 025 in the MTC, particularly given the absence of alternative sources of data for nivolumab. 54 The data in the MTC analysis for OS from RECORD-1 was adjusted for crossover using the rank-preserving structural failure time model (RPSFTM), as this was expected to give a less biased estimate in the presence of crossover than the unadjusted data. 53 A SA for OS was conducted that included observational studies as well as the RCTs from the primary analysis, which enabled comparison with axitinib. Unfortunately, no RCT or observational studies were identified that reported OS data on sunitinib and so it has not been possible to provide an estimate of its effect on OS. An additional four studies,43,59,62,63 which included the prior sunitinib subgroup from the AXIS RCT, were suitable for inclusion in the SA for OS. 43 Network diagrams for the primary analysis and SA for OS are presented in Figures 4 and Figure 5, respectively.

The results of the MTC primary analysis for OS did not show statistically significant benefits of any treatment over BSC, but all point estimates were in favour of the active treatment (Table 8). Cabozantinib and nivolumab led to longer OS than everolimus (HR 0.66, 95% CrI 0.53 to 0.82; and HR 0.73, 95% CrI 0.60 to 0.89; respectively, see Table 8); however, the difference between nivolumab and cabozantinib was not statistically significant (HR 1.12, 95% CrI 0.82 to 1.49). Cabozantinib was associated with the highest probability of being the most effective treatment for prolonging OS:

• cabozantinib, 72.31%

• nivolumab, 24.26%

• everolimus and BSC, 3.43%.

The results of the SA for OS, including data to compare treatments with axitinib, were in keeping with those of the primary analysis (i.e. no statistically significant benefits of treatments over BSC, nivolumab and cabozantinib benefit over everolimus; see Appendix 4). Everolimus, cabozantinib and nivolumab all showed longer OS than axitinib (HR 0.74, 95% CrI 0.56 to 0.99; HR 0.48, 95% CrI 0.34 to 0.71; and HR 0.54, 95% CrI 0.38 to 0.77, respectively).

The residual deviance was similar to the number of unconstrained data points in the MTC primary analysis for OS (3 points vs. 3 points, respectively; see Appendix 5) suggesting a good model fit. However, the residual deviance was considerably higher than the number of unconstrained data points in the SA (13 points vs. 7 points, respectively). These findings suggest that the results of the SA for OS should be interpreted with caution owing to the poor fit of the MTC model. There was one loop of three studies in the SA: everolimus, axitinib and sorafenib. Investigation of potential inconsistency in the data loop present in the MTC SA for OS suggested that the results were statistically inconsistent (p > 0.05; see Appendix 5).

Response rate

The MTC for objective response rate (ORR) comprised three RCTs53,54,57 and allowed only the comparison of cabozantinib, everolimus, nivolumab and BSC (see Figure 4). Response data reported in AXIS could not be connected to the network and so have been reported narratively. Response rate was also reported in two observational studies but SAs including non-RCTs were not planned for this outcome. 55,56 The active treatments all showed statistically significant improvements in ORR compared with BSC (Table 9). In addition, cabozantinib and nivolumab resulted in statistically significant improvements in ORR compared with everolimus (OR 6.67, 95% CrI 3.28 to 12.78; and OR 6.18, 95% CrI 3.75 to 9.84, respectively). There was no statistically significant difference in ORR between nivolumab and cabozantinib (OR 1.05, 95% CrI 0.41 to 2.18). It should be noted that CheckMate 025 was rated as being at a high risk of bias as a result of the absence of blinding of outcome assessors for response and METEOR was rated as being at an unclear risk of bias for missing data, but evidence for this outcome was otherwise rated as being at a low risk of bias. 54,57 The impact of these potential biases on the overall direction of treatment effects is unknown. Data from AXIS that could not be connected to the MTC showed higher ORR in axitinib-treated (23%) than sunitinib-treated patients (12%), which was similar when ORR was reviewed independently (19% vs. 9%). 43,66

There were two response rate categories for which it was deemed there were sufficient data for clinically meaningful analyses: stable disease and progressive disease. These analyses used the same three studies as the primary analysis for ORR. 53,54,57 Consistent with the results for ORR, all treatments lowered the odds of having progressive disease and increased the odds of having stable disease, compared with BSC. Cabozantinib significantly lowered the odds of having progressive disease compared with everolimus (OR 0.39, 95% CrI 0.25 to 0.58) and nivolumab (OR 0.27, 95% CrI 0.16 to 0.45); cabozantinib also improved the odds of stable disease compared with nivolumab (OR 2.70, 95% CrI 1.79 to 4.17), but not everolimus (OR 1.18, 95% CrI 0.85 to 1.61). Everolimus lowered the odds of progressive disease compared with nivolumab (OR 0.70, 95% CrI 0.53 to 0.96). All data are shown in Table 10. Data from AXIS that could not be connected in the MTC showed similar rates of independently reviewed progressive disease (22% axitinib vs. 21% sunitinib) and higher rates of prolonged stable disease (≥ 20 weeks) with axitinib (27%) than sunitinib (21%). 43,66 Rates of stable disease for < 20 weeks assessed by the investigator were in favour of sunitinib (23% axitinib and 33% sunitinib).

Health-related quality of life

All four of the included RCTs43,53,54,57 reported data on HRQoL, but the data were not suitable for combining in a MTC. Studies varied in the measures used and analyses undertaken, and they did not create a connected network. The number of available data decreased significantly over the course of the studies owing to progression and, to a lesser extent, questionnaire completion rates. All four studies projected change in HRQoL using repeated measures mixed-effects models to impute data for participants who were not available at each time point. The AXIS, CheckMate 025 and RECORD-1 studies also conducted alternative analyses using pattern mixture models to explore the possibility that the number of, and reasons for, missing data were related to patients’ health state (i.e. not missing at random). 43,54,64

A summary of HRQoL results from the four included RCTs is given in Table 11. HRQoL scores were similar between axitinib and sorafenib in AXIS,43 regardless of the measure used. Results favoured nivolumab over everolimus in CheckMate 025,54 a result that was seemingly consistent across the various measures used and type of analysis undertaken. Results in RECORD-1 favoured BSC over everolimus, although this effect was only apparent if models were used to account for data not missing at random. 53 METEOR results favoured everolimus over cabozantinib on a measure of disease-specific quality of life,57 but scores were similar on two measures of general HRQoL. Completion rates, scales used, analyses undertaken and effect sizes are described in more detail for each study below.

Across studies, the proportion of randomised patients in either group with baseline HRQoL measurements was high, ranging from 86.4% (everolimus group of CheckMate 02554) to 95.8% (axitinib group of AXIS43). METEOR did not state the proportion of the population with baseline measurements. 57 Available measurements as a proportion of the total randomised populations dropped sharply through the course of the studies, primarily owing to deaths and disease progression [axitinib 4.2% and sorafenib 1.9% in AXIS by 19 months (21 cycles),43 nivolumab 4.9% and everolimus 2.3% in CheckMate 02554 at 24 months, and everolimus 19% and BSC 4% in RECORD-153 by 8 months]. When reported, scale completion rates, defined as the number of completed scales out of the total patients available at each time point, were fairly good; rates remained > 75% in METEOR over 48 weeks,57 71–89% in the nivolumab group and 60–90% in the everolimus group in CheckMate. 54 AXIS also took HRQoL measurements at the point when patients’ stopped treatment,43 which were available for 45.2% of the axitinib group and 52.8% of the sorafenib group.

AXIS reported the mean difference (MD) between axitinib- and sorafenib-treated patients on four measures of HRQoL. 43 Patients’ baseline Functional Assessment of Cancer Therapy-Kidney Symptom Index-Disease Related Symptoms (FKSI-DRS) subscale scores were described as comparable to the general population in both the axitinib and sorafenib groups of AXIS. 43 Mean scores while on treatment were not significantly different between axitinib and sorafenib, and neither patient group showed a substantial decline during treatment. None of the results from the repeated measures mixed-effects models showed a statistically significant difference between groups [EuroQoL visual analogue scale (EQ-VAS), p = 0.645; EuroQol-5 Dimensions (EQ-5D) index score, p = 0.19; Functional Assessment of Cancer Therapy Kidney Cancer Symptom Index (FKSI)-15, p = 0.483; and FKSI-DRS, p = 0.675]. The quality-of-life scores at the point that patients discontinued treatment also did not differ between groups but were significantly below the baseline means, attributed to disease progression. Results from the SAs using pattern-mixed models were not reported in full but were deemed similar to those of the standard mixed-effects model. 67

RECORD-1 measured HRQoL using the FKSI-DRS subscale and two subscales of the European Organisation for Research and Treatment of Cancer Quality of Life Questionnaire (EORTC QLQ-C30). 53 Pattern-mixed models showed that EORTC QLQ-C30 global health status [HR 0.85, 95% confidence interval (CI) 0.75 to 0.96; p = 0.006] and physical functioning (HR 0.84, 95% CI 0.75 to 0.94; p = 0.001) deteriorated more quickly in patients taking everolimus than those on BSC. 71 Analyses not fitting these models did not show a difference between groups in rate of decline on these measures64 and the difference between groups was not significant for the FKSI-DRS regardless of which analysis was used.

CheckMate 025 measured HRQoL using the FKSI-DRS subscale, EQ-VAS and EQ-5D index score. 54 For each measure, results were reported for the difference in mean change from baseline up to week 84 (repeated measures mixed-effects models), proportion of patients with meaningful improvement and median time to improvement. The prespecified end point was the number of people in each group with a meaningful improvement of ≥ 2 points on the FKSI-DRS; other analyses were post hoc. Mixed models comparing change from baseline to week 84 favoured nivolumab over everolimus for the FKSI-DRS (MD 1.7, 95% CI 1.2 to 2.1; p < 0.0001), EQ-5D index score (MD 0.04, 95% CI 0.02 to 0.07; p = 0.0003) and EQ-VAS (MD 5.7, 95% CI 3.8 to 7.7; p < 0.0001). The results were deemed consistent when pattern-mixed models were used instead, but the results are not presented, and FKSI-DRS scores, analysed non-parametrically, showed a similar pattern of nivolumab benefit. In the dichotomous analyses using scale cut-off points, more people in the nivolumab group than the everolimus group had clinically significant improvements (an increase on the scale of at least 2 points) over the course of the study on the FKSI-DRS (55% vs. 37%; p < 0.001) and the EQ-VAS (53% vs. 39%; p = 0.0001); equivalent results for the EQ-5D did not show a difference between groups. 69 The median time to improvement analyses favoured nivolumab, but some CIs were not estimable.

METEOR, comparing cabozantinib with everolimus, reported the EQ-VAS, EQ-5D index score and FKSI-19. 57 Patients randomised to cabozantinib deteriorated more than the everolimus group on the FKSI-19 (mean change from baseline –3.48 cabozantinib vs. –2.21 everolimus; p < 0.0001). 28 The scores taken at end of treatment were around 7 points lower than baseline in each group which was attributed to disease progression. Mean change from baseline was similar between the cabozantinib and everolimus groups on the EQ-VAS (MD –0.051; p = 0.921) and the EQ-5D index score (MD –0.002; p = 0.825).

Baseline prognostic score

The prognostic scores used at baseline varied across the four included RCTs and included ECOG performance status, MSKCC and Karnofsky performance status. RECORD-1 and METEOR were the only studies that reported subgroup data on PFS by baseline MSKCC prognostic score. 53,57 The MSKCC groups with data were defined as favourable, intermediate and poor prognosis. The results of the MTC analyses for PFS demonstrate consistent treatment benefit with both cabozantinib and everolimus compared with BSC across all three MSKCC categories (favourable, intermediate and poor). PFS subgroup data in METEOR that could not be combined in a MTC showed a larger effect in favour of cabozantinib than everolimus for people with an ECOG score of 0 at baseline (HR 0.46, 95% CI 0.36 to 0.59) than those with a score of 1 (HR 0.64, 95% CI 0.46 to 0.90). 57

Overall survival data for the baseline prognostic score MTC subgroup analysis by MSKCC risk were from two studies (CheckMate 025 and METEOR). 54,57 These results suggested a trend in favour of cabozantinib and nivolumab over everolimus irrespective of MSKCC baseline score, although the results failed to reach statistical significance in some subgroups. However, this is to be expected because the individual studies in the MTC were not powered sufficiently for these subgroup analyses. OS subgroup data in METEOR that could not be combined in a MTC showed a slightly larger effect in favour of cabozantinib over everolimus for people with an ECOG score of 0 at baseline (HR 0.65, 95% CI 0.49 to 0.87) than those with a score of 1 (HR 0.72, 95% CI 0.51 to 1.02), but CIs were overlapping. 57

Prior therapy

The only data available for analyses by number of prior therapies were data on patients with one prior TKI and data on people with more than two prior TKIs. Two studies (METEOR and RECORD-1) were included in the MTC of PFS. The MTC consistently demonstrated better PFS with cabozantinib than with everolimus and either of the active treatments compared with BSC, irrespective of number of prior therapies (see Appendix 7). 57,74 OS subgroup data in RECORD-1 that could not be combined in a MTC showed very similar effects of everolimus over placebo (BSC) for people previously treated with sorafenib only, sunitinib only, or both (HR 0.29, 0.30 and 0.28, respectively, CIs not given). 64

There were only two studies (CheckMate 025 and METEOR) reporting suitable data for analysis of OS by number of prior TKI therapies. 54,57 The MTC results for OS indicate no statistically significant difference in efficacy between cabozantinib and nivolumab irrespective of the number of prior TKI therapies (see Appendix 7). Treatment with cabozantinib and nivolumab both resulted in a longer OS than with everolimus, but the differences were only statistically significant after one prior TKI.

Search results

The systematic review identified 633 potentially relevant citations after deduplication was carried out. The titles and abstracts of the papers were reviewed independently by two reviewers (FS and PC); conflicts were resolved by discussion between the reviewers and, when necessary, by a third reviewer (MB). A total of 464 papers were excluded based on title and abstract.

A total of 169 full-text papers were considered potentially relevant and were reviewed independently by two reviewers. Conflicts were resolved by third-party arbitration with a third reviewer. After review of the studies, 11 studies were included. Out of the 11 included citations, 10 were economic evaluations and one was a HRQoL paper. No UK costing studies were identified. The remaining 158 studies were excluded for the following reasons:

• abstract with insufficient methodological details and no additional details published elsewhere (n = 104)

• non-UK costing study (n = 15)

• wrong population (n = 13)

• no relevant outcomes reported (n = 6)

• irretrievable (n = 6)

• study not available in English (n = 4)

• duplicate (n = 5)

• wrong study design (i.e. not economic evaluation, costing study or quality-of-life study) (n = 3)

• letter/commentary (n = 1)

• systematic review (n = 1).

Non-UK costing studies were excluded given that the review of cost-effectiveness evidence presented in previous NICE TA on second-line treatment of RCC was deemed sufficient to capture the resource use in advanced RCC in the UK.

The PRISMA flow diagram for the search is reported in Figure 6. 75 The full references of the studies excluded after review of full papers are reported in Appendix 10.

FIGURE 6.

The PRISMA flow diagram.

Narrative description of published cost-effectiveness studies

Multiple technology appraisal number 178: bevacizumab (first line), sorafenib (first and second line), sunitinib (second line) and temsirolimus (first line) for treatment of advanced and/or metastatic renal cell carcinoma

For the purposes of this HTA report, the review of TA178 will only focus on the areas relating to second-line treatments of RCC; that is, the evaluation of sorafenib (second line only) and sunitinib. We begin by presenting a brief description of the company’s submission on the different drugs, followed by the AG analysis.

Single technology appraisal number 219: everolimus for the second-line treatment of advanced renal cell carcinoma

The holder of the marketing authorisation for everolimus (Novartis Pharmaceuticals, Basel, Switzerland) presented a cost–utility analysis assessing the cost-effectiveness of everolimus plus BSC compared with BSC alone in patients with advanced RCC. The company developed a Markov model consisting of four health states: stable disease without AEs, stable disease with AEs, PD and death. The model had 8-weekly cycles and a time horizon of 144 weeks, which was justified as being the maximum life expectancy of the population in an analysis of the RECORD-1 trial, used as the main source of clinical evidence for the economic evaluation. 93 The target population of the model was defined as adults with advanced RCC whose cancer had progressed within 6 months of receiving sunitinib and/or sorafenib, and had demographic characteristics reflecting those of the RECORD-1 trial. Patients in the trial were allowed previous therapy with a cytokine or bevacizumab.

Transition probabilities from the initial state of stable disease without AEs were calculated using the incidences of grade 3 and 4 AEs, treatment withdrawal, disease progression and death obtained from the RECORD-1 trial. As the trial provided data only up to the seventh cycle in the model (i.e. 56 weeks), the company assumed that the event rates remained constant after this cycle until the end of the time horizon. Transition probabilities to stable disease with AEs and PD were calculated directly from event frequencies observed in each arm of the trial. For transitions to death, this method was used only for the everolimus plus BSC arm of the model, while a HR was estimated and applied to calculate the transition probabilities for the BSC arm. Owing to the presence of crossover between the everolimus and placebo arm in the RECORD-1 trial, the company adjusted the HR for OS using the inverse probability of censoring weighted (IPCW) method. The rationale for applying a HR to estimate BSC transition probabilities for death, rather than using the trial data directly in the same way as the everolimus arm, was to keep a constantly higher relative risk of mortality for any given cycle in the BSC model arm compared with the everolimus arm.

The HSUVs were taken from the sunitinib company submission in a previous MTA in the same disease area (TA178). The sunitinib model included utilities for a PFS state and a PD state. In the everolimus model, the PFS utility was applied to the ‘stable disease without AEs’ health state and the PD utility score was used for the ‘progressed disease’ state. For the stable disease with AEs, a disutility of 0.05 was assumed, regardless of which AE(s) occurred. This disutility was applied for the first cycle only when patients entered the stable disease with AEs, as AEs were assumed to be resolved within one cycle.

Treatment costs were included in the model as per the proposed PAS, whereby the first treatment pack (30 tablets, 10 mg each) was free to the NHS while subsequent treatment packs were discounted by 5% to £2822. Costs associated with BSC, monitoring and AEs were taken from TA178. 29 Subsequent treatment costs were also included after everolimus treatment had ended, including sunitinib, sorafenib and bevacizumab. Results of the model were later updated with a revised PAS which was designated as CiC. This report presents the initial and revised ICERs.

In the original analysis, the base-case ICER was estimated to be £51,613 per QALY gained (see Appendix 12). OWSAs showed that the ICER was most sensitive to the OS estimate in the BSC arm. The PSA estimated that the probability that everolimus plus BSC was cost-effective at a threshold of £50,000 per QALY gained was 40%. An additional analysis was performed by the company using the ITT OS HR (i.e. not adjusted for crossover). In this SA, the HR changed from 0.55 in the base-case analysis to 0.87, resulting in an ICER equal to £91,256 per QALY gained.

The ERG commented that the clinical effectiveness evidence was of good quality, albeit from just one RCT being found through the systematic review. The ERG found that the OS estimate was the main factor affecting cost-effectiveness results, and it agreed that it was important to adjust the data for confounding due to crossover in the trial. The ERG identified and corrected two technical errors in the estimation and application of the HR for mortality in the model and in how discounting was applied. Correcting the two errors increased the company’s base-case ICER of £51,613 to £65,231 per QALY gained. The ERG noted that other methods used to adjust for treatment crossover such as the RPSFTMs could have been investigated to assess the impact of adjusting for crossover in the economic results.

The company produced analyses using the RPSFTM in response to the ERG’s comments, which resulted in an ICER of £53,128 per QALY gained (with all other base-case assumptions unchanged). However, the ERG felt that the mortality risk in the BSC arm had been overestimated due to an extrapolation based on a single data point. The ERG felt that more data should have been used in the RPSFTM analysis.

Following the revision to the PAS and the ERG’s comments, the company provided an updated cost-effectiveness analysis, incorporating the ERG’s assumptions in the RPSFTM analysis, more recent trial data (November 2008 cut-off point) and a longer time horizon. The ICER from this analysis was £49,272 per QALY gained. The ERG considered the changes reasonable but reiterated their concerns about the wide CIs for the OS HR, particularly as OWSAs showed that the ICER was associated with substantial uncertainty.

A PSA was performed, which resulted in a mean ICER of £50,047 per QALY gained. The company believed that the wide CI for the OS HR was not clinically plausible, and performed an additional analysis with an adjusted range for the PSA. This resulted in an ICER of £47,811 per QALY gained.

The ERG noted that it was unclear whether or not all sources of uncertainty had been included in the company’s PSA. When the ERG re-ran the PSA, keeping the adjusted CI for the OS HR, it estimated the mean ICER to be £49,479 per QALY gained. With the original CI, the ICER increased to £51,661 per QALY gained in the ERG analysis. The results of the final ERG base-case analysis are shown in Appendix 12.

Single technology appraisal number 333: axitinib for treating renal cell carcinoma after failure of prior systemic treatment

The holder of the marketing authorisation for axitinib (Pfizer) presented a cost–utility analysis that evaluated axitinib compared with BSC in people with advanced RCC after failure of treatment with sunitinib or a cytokine. The analysis used data from three RCTs of second-line treatments for amRCC: AXIS, RECORD-1 and TARGET. 43,64,104 The AXIS trial compared axitinib with sorafenib, TARGET compared sorafenib with placebo, and RECORD-1 included a placebo plus BSC arm following first-line sunitinib treatment. These studies were identified by systematic review and were used to form an indirect comparison of axitinib with BSC due to the lack of head-to-head RCT evidence. Subgroup analyses by prior treatment were carried out.

For the prior-cytokine therapy group, a MTC was performed using a Markov chain Monte Carlo simulation in WinBUGS to estimate the relative treatment efficacy of axitinib and BSC, using data from the placebo arm of TARGET, under the assumption of equivalent effectiveness to BSC following cytokine therapy. HRs for both PFS and OS from AXIS and TARGET were used in a fixed-effects model assuming PHs. Point estimates and 95% CrI were estimated and the results showed a PFS of 11 months for the axitinib group compared with 3.5 months in the BSC group (median HR 0.25, 95% CrI 0.17 to 0.38). The axitinib group showed a median OS of 33.5 months compared with the 23.5 months in the BSC group (median HR 0.63, 95% CrI 0.41 to 0.99). The HR for OS was based on data censored at crossover.

The company performed a simulated treatment comparison (STC) for the first-line sunitinib arm, as there was no direct or indirect trial evidence This compared PFS with OS for axitinib versus everolimus and BSC using data from the AXIS43 and RECORD-193 trials. Owing to several differences between the patients in the trials, two approaches were tested to simulate the treatment relative effectiveness. The first was by comparing the ITT placebo group in RECORD-1, which included patients who have received sunitinib and/or sorafenib, with the sunitinib-refractory patients in AXIS. This implicitly assumes that the patients who received prior sunitinib and prior sorafenib have similar characteristics. The second approach compared the axitinib post-sunitinib group with the everolimus post-sunitinib group and then applied the RPSFTM-adjusted HR for everolimus to BSC to create a modelled prior-sunitinib group.

The simulated relative treatment effectiveness was performed by analysing IPD from the axitinib arm of AXIS to derive parametric survival functions incorporating baseline predictors of the two main trial outcomes (i.e. PFS and OS). Five different distributions were examined, but only the two best-fitting models were used for both PFS and OS, which were the log-normal and Weibull distributions. The results of the analysis showed a benefit for axitinib relative to BSC and everolimus in both PFS and OS, when log-normal and Weibull distributions were used. However, the data relating to the estimated increase in PFS and OS in the ITT population and prior-sunitinib subgroup were CiC.

The PFS HR of 0.34 for the prior-sunitinib group and the adjusted OS HR of 0.53 for the ITT group of RECORD-1 were applied to the everolimus STC curves to generate AXIS-like prior-sunitinib PFS and OS curves for BSC. This resulted in an estimated median PFS of 1.7 months for the ‘axitinib-like patients’ if they had received placebo, compared with a median of 5.8 months if they had received axitinib (HR not reported). The median OS estimated for the same patients was 8.3 months for placebo compared with 15.2 months for axitinib (HR not reported).

The company also provided an additional analysis that used retrospective observational data for patients who received first-line sunitinib followed by sorafenib or BSC from a Swedish database, Renal Comparison (RENCOMP), to estimate the OS HR for people who received sorafenib or BSC following first-line sunitinib treatment. 105 A multivariate Cox PH regression analysis was performed using variables with significance at the 5% level to account for any bias resulting from confounding. The results of this analysis showed a median OS HR of 0.62 (95% CI 0.41 to 0.94; p = 0.023). The results from RENCOMP were used in an indirect comparison with the results from the first-line sunitinib group in the AXIS trial, which compared axitinib with sorafenib (median PFS HR 0.74, 95% CI 0.57 to 0.96; median OS HR 0.997, 95% CI 0.78 to 1.27), to estimate indirect HRs for axitinib compared with BSC in the prior-sunitinib group. The resulting HR was 0.62 (95% CI 0.38 to 0.997).

The company developed a Markov model for the cost–utility analysis consisting of three health states: PFS, PD and death. All patients were assumed to enter the model in the PFS state and could transition to any of the other states or remain in the current state after each 4-week cycle. The time horizon of the model was 10 years.

Parametric survival functions were used to calculate the proportion of patients in each health state at each point in time. For the first-line cytokine group who received axitinib at second line, Weibull models were used to extrapolate PFS and OS as this was considered to have the best fit among the alternatives tested. Log-logistic and Gompertz distributions were also tested in SAs for OS, and the log-normal and Gompertz models were used alternatively in SAs for PFS. For the purpose of estimating PFS and OS in the BSC arm, HRs estimated from the indirect comparison analysis were applied to the parametric survival functions used for the axitinib arm.

Utilities were taken from the AXIS trial using the EQ-5D questionnaire analysis undertaken in the trial. The analysis was based on the full population, as there was no statistically significant difference in utilities between patients receiving different types of first-line therapies. The utility for the PFS health state was 0.69 and was 0.61 for the PD state. The PFS utility was derived from the average EQ-5D index value at each time point, weighted by the number of people still on treatment, whereas the PD score was taken as the weighted average of the mean utility at the end of treatment. These utilities were also assumed to incorporate implicitly the AE profile of the treatments in the trial. A systematic review of the literature did not find any sources of utilities for patients with amRCC receiving BSC after sunitinib treatment had failed. Therefore, the utilities from AXIS were also assumed to apply to BSC. Utilities from previous NICE TAs were explored as part of the SAs.

Costs of treatment were based on the proposed PAS, which was CiC. Costs were adjusted for dosing intensities and discontinuation probabilities were applied at each cycle during treatment (0.8% and 1.26% for prior-cytokine and prior-sunitinib, respectively) while AE rates were assumed to be the same between subgroups. AE-related costs were included in the PFS health state only and assumed to be equal across the two prior treatment subgroups. Only grade 3 and 4 AEs that occurred in > 5% of the population were included. For axitinib, hypertension had a cost of £424.00 per episode and diarrhoea had a cost of £544.00 per episode. In the BSC arm, anaemia was included at a cost of £2068.47 per episode as a SA.

No administration costs were incurred and no drug costs for the BSC arm were assumed to apply. Costs of routine monitoring were based on TA178 and were validated with expert clinical opinion to ensure consistency with current clinical practice. 29 These costs were assumed to apply to both arms of the model equally. The total cost per cycle for the PFS state was £109.69 based on one GP visit, one tumour scan per three cycles and one blood test per cycle. For the PD state, the cost per cycle was £319, based on one GP visit, three visits by a specialist community nurse every two cycles and 28 vials of pain medication per cycle. A SA explored changing the GP visit to an oncologist visit, which changed the costs for the respective health states to £176.69 and £386.00.

The updated economic analysis increased the time horizon to 15 years after the ERG’s concern that 10 years might not have reflected a lifetime horizon in the economic analysis. First-line cytokine and first-line sunitinib subgroup specific utilities and relative dose intensities (RDIs) were applied rather than estimates for the ITT population used in the original model. The percentage of people with hypertension was reduced from 2% to 0% as the percentage of hypertension in TARGET was < 1%. 104 The PSA was updated to include SEs rather than SDs for the PFS and PD HSUVs, RDI and the cost of death. An error in the STC was also corrected, which had only a marginal effect on the results.

The results of the first-line cytokine subgroup analysis showed an ICER of £55,284 per QALY gained. A range of OWSAs were performed using the ranges of the 95% CIs, and these showed that the results were most sensitive to the HR of OS and the PD utilities. ICERs ranged from £40,000 to £100,000 per QALY gained for changes in utilities and up to £350,000 per QALY for changes in the OS HR. The results of changing the method of extrapolation for OS varied from £21,959 per QALY gained for the log-logistic method to £72,537 per QALY gained for the Gompertz method. Other scenario analyses resulted in ICERs similar to the base-case analysis.

For the first-line sunitinib subgroup the base-case ICER was £33,538 per QALY gained. OWSAs showed that the results were most sensitive to changes in the parameters for the survival functions for PFS and OS, with ICERs ranging from around £25,000 to £48,000 per QALY gained. The ICER was also sensitive to the PD HSUVs for both arms and for PFS in the axitinib arm. The resulting ICERs ranged from around £29,000 to £40,000 per QALY gained. Other OWSAs had very little impact on the results. The ICER was most sensitive to different survival distributions using the method of comparison based on RENCOMP data, with use of the Weibull and Gompertz distributions resulting in ICERs of £47,515 and £39,479 per QALY gained, respectively. Reducing the RDI to 80% resulted in an ICER of £27,324 per QALY gained. The company also applied an assumption of no QALY or survival gain post progression, which resulted in an ICER of £52,850 per QALY gained.

Before the end of the STA process, the scope of the appraisal was updated by NICE after an appeal hearing, with sunitinib and pazopanib added as comparators for the prior-cytokine subgroup. The company conducted an additional systematic review to search for relevant evidence to inform the new comparisons. As no head-to-head trials were found, indirect comparisons and naive indirect comparisons (assuming trial groups are homogenous) were used to estimate relative treatment effects. In addition to AXIS, two trials were identified. The first trial compared pazopanib with placebo106 while the second was a single-arm trial assessing the efficacy safety of sunitinib. 107 A naive comparison between axitinib and sunitinib showed that axitinib increased PFS by 3.3 months and OS by 5.5 months. A naive comparison between axitinib and pazopanib showed that axitinib increased PFS by 4.7 months and OS by 6.7 months. An indirect comparison showed that PFS was longer for axitinib compared with pazopanib (median HR 0.47, 95% CrI 0.26 to 0.85). The company did not conduct an indirect comparison for OS owing to the crossover between placebo and treatment groups in the axitinib and pazopanib trials.

The company did not provide a full incremental analysis with the new comparators but instead provided a naive economic comparison based on the base-case analysis with the PAS (with an ICER of £55,284 per QALY gained). The naive treatment comparison showed an increase in median OS for BSC (24 months compared with 23.9 months and 22.7 months for sunitinib and pazopanib, respectively), so the company used the lower 95% CI of 17.6 months for a more realistic estimate. The recalculated ICER for axitinib versus BSC was £36,493 per QALY gained and an upper limit of £55,000 per QALY gained was estimated for the ICERs comparing axitinib with either sunitinib or pazopanib.

The company also provided an additional analysis that used data from a retrospective cohort study,108 showing correlation between tumour shrinkage and OS, to weight the median OS estimates in the BSC arm of RECORD-1 by multiplying the median overall survival by the proportion of patients. The resulting weighted median OS for BSC was 8.2 months (95% CI confidential information has been removed) which the company believed was consistent with the 8.3 months OS result from the STC. The company concluded that the results were consistent with the results of the base-case analysis using the STC, which resulted in an ICER of £33,538 per QALY gained.

Single technology appraisal number 417: nivolumab for previously treated advanced renal cell carcinoma

The holder of the marketing authorisation for nivolumab (Bristol-Myers Squibb, NY, USA) submitted a cost–utility analysis evaluating the cost-effectiveness of nivolumab in comparison with everolimus, axitinib and BSC. A six-state transition model was developed with health states defined as: PFS on treatment, PFS off treatment, post-progression survival on treatment, post-progression survival off treatment, terminal care (TC) and death. All patients started in the PFS on treatment state and could only transition to death via the tunnel state TC, which they were assumed to stay in for 8 weeks before death. The time horizon was set at 30 years and weekly cycles were used.

Model parameters were informed by data from the CheckMate 025 trial,54 a Phase III multicentre open-label RCT comparing nivolumab with everolimus in patients with histologically confirmed advanced/metastatic RCC who have received one or two previous antiangiogenic agents. A MTC including nine studies was performed to estimate the treatment effects on OS and PFS between nivolumab and axitinib, and nivolumab and BSC. 54,66,91,93,109–113

To extrapolate survival for nivolumab and everolimus, a generalised gamma model was used, which was considered to give plausible results based on expert clinical opinion. The relative effectiveness of everolimus and BSC was incorporated by applying the crossover-adjusted HR from the network meta-analysis (NMA) to the OS curve of the everolimus arm data from CheckMate 025. It was assumed that BSC would be as effective as placebo.

The company considered that standard parametric models did not fit the PFS data sufficiently well so more flexible spline-based models were explored. Models from Royston and Parmar114 were applied to the PFS data from CheckMate 025. The PFS data for axitinib and BSC were estimated by applying the HR from the NMA analysis to the everolimus survival curve, assuming that BSC was as effective as placebo.

The company used the same spline-based survival analysis approach to model time-to-discontinuation (TTD) data for nivolumab and everolimus as for PFS. In the absence of TTD data for axitinib, the company assumed that treatment was continued until disease progression.

Pharmacological resource use was based on the treatment indications and adjusted for dosage, with 92% and 94% of the planned dosage being costed for nivolumab and everolimus, respectively, based on the CheckMate 025 trial. For axitinib, the actual usage was 102% of that planned based on the AXIS trial, as reported in TA333. 25 Resource use in the PFS and post-progression survival (PPS) states was assumed to be the same as that in TA333. The cost of TC was considered in the model and based on a paper by Addicott and Dewar on improving choice at the end of life. 115

The company only included serious grade III/IV treatment-related AEs experienced by ≥ 1% of patients in either arm of the CheckMate 025 trial. These were pneumonitis, anaemia, diarrhoea and pneumonia. Rates and duration of these AEs were based on the trial observations for both nivolumab and everolimus, and management costs were applied weekly in the model. Owing to a lack of data, the cost of managing AEs for patients treated with axitinib was assumed to be equal to that of everolimus treatment. The weekly cost of AEs was £0.35 for nivolumab and £1.31 for axitinib and everolimus.

The HSUVs were applied to each health state and were derived from EQ-5D data obtained from the CheckMate 025 trial. A mixed-effects model was used to analyse the data, with fixed effects for progression status and treatment allocation; a variable effect for the interaction between treatment arm and progression status; and a random effect for the subject. For patients receiving axitinib, HSUVs were derived from EQ-5D data from the AXIS trial, reported in TA333. It was assumed that patients receiving BSC would experience the same quality of life as patients receiving axitinib. The HSUVs before progression were 0.80, 0.76, 0.69 and 0.69 for nivolumab, everolimus, axitinib and BSC, respectively. For PPS, the HSUVs were 0.73, 0.70, 0.61 and 0.61 for these treatments, respectively.

The results of the company’s model showed an estimated survival benefit of 16, 11 and 24 months (undiscounted) for nivolumab compared with axitinib, everolimus and BSC, respectively. Nivolumab increased mean (discounted) QALYs by 1.07, 0.63 and 1.43 compared with axitinib, everolimus and BSC, respectively. The company estimated pairwise ICERs of £42,417, £83,829 and £56,427 per QALY gained for nivolumab compared with axitinib, everolimus and BSC, respectively.

Single technology appraisal TA463: cabozantinib for treating advanced renal cell carcinoma in adults who have received at least one prior vascular endothelial growth factor-targeted therapy

The holder of the marketing authorisation for cabozantinib (Ipsen, Paris, France) submitted a de novo economic model that evaluated cabozantinib in two separate cost–utility analyses. The first, a trial-based analysis comparing cabozantinib with everolimus, using effectiveness data obtained solely from the METEOR trial, a Phase III open-label RCT comparing cabozantinib and everolimus in patients with advanced RCC who progressed after previous VEGF receptors TKI treatment. 57 The second was a pairwise analysis of cabozantinib compared with everolimus, axitinib, nivolumab and BSC based on effectiveness data derived from a MTC of trials identified in a systematic review of the literature. 54,66,93,104

The two analyses used the same partitioned survival model structure, composed of three health states: PFS, PD and death. Patients entered the model in the progression-free state and could transition to PD or death. From the PD state, patients could transition to the death state at each future cycle. The time horizon was set at 30 years and 4-weekly cycles were used.

A log-logistic distribution was chosen to extrapolate observed OS and PFS for both cabozantinib and everolimus in the METEOR trial in the trial-based analysis. In the MTC-based economic evaluation, parametric survival curves were estimated and extrapolated based on regenerated KM data from CheckMate 025, AXIS, RECORD-1 and TARGET, in addition to KM data from METEOR. 54,57,66,93,104 The NMA estimated the parameters of independently fitted survival curves for each treatment group in each trial in the network and the parameters were adjusted to the everolimus group of the METEOR trial. 57 A single family of distributions was fitted to each group in each trial. The log-normal distribution was deemed to be the best fit for both OS and PFS in the NMA-based analysis and was used in the company’s base case for the curves of each comparator.

The TTD data were used to estimate pharmacological costs in both analyses. In the trial-based analysis TTD KM data from METEOR were used and extrapolated using a log-normal distribution. In the NMA-based analysis regenerated KM data from trials were used to estimate the parameters of the best-fitting curve and adjust them to the everolimus group of the METEOR trial. KM data for axitinib were not available and PFS data were used as a proxy for TTD.

Relative dose intensities were applied to estimate the true cost of the treatments compared and were taken from the respective trial data. Treatment administration costs were only applied for nivolumab as it is administered intravenously, with no wastage assumed. Disease management costs were estimated based on the expert clinician opinion of oncologists practising in the UK, and varied according to health state. Costs related to AEs were limited to those events that occurred in ≥ 5% of patients in each group of the relevant trials. Treatment-emergent adverse events (TEAEs) were used in the models for all comparators except nivolumab, as only treatment-related adverse event (TRAE) data were identified for it.

The HSUVs were estimated based on a regression analysis of EuroQol-5 Dimensions, five-level version (EQ-5D-5L) collected in the METEOR trial; utility decrements for AEs were applied. The HSUVs used in the model for patients regardless of treatment arm/analysis were 0.817 prior to progression and 0.777 after progression. A disutility of 0.055 was applied to patients when experiencing AEs regardless of treatment arm.

The results of both analyses are not shared publicly, and remain confidential at the time of writing this report. A range of OWSAs and scenario analyses were performed as well as a PSA to test the impact of uncertainty of all relevant parameters on the model results.

Survival modelling

The first stage in generating the survival curves was to obtain the data points from published KM survival curves. The images of the KM plots from publications identified in the systematic review were digitised using g3data software (version 1.5.1; www.frantz.fi/software/g3data.php), which enabled time points and survival probabilities to be generated and saved as a comma-separated text file. These data, along with the numbers of patients at risk at various time points, was then loaded into R version 3.3.2 (The R Foundation for Statistical Computing, Vienna, Austria) to perform the analysis.

The next stage of the process was to generate pseudo IPD using the algorithm derived by Guyot et al. 116 This algorithm was implemented using the R code in the pre-release version of the survHE package, in which this algorithm had previously been applied. The pseudo IPD data were then inputted into the survival functions built into the flexsurv package of R, to assess and generate fitted survival functions. 117

Overall survival

The following distributions were fitted to the pseudo IPD data generated from the OS KM plots from the CheckMate 025 trial: exponential, Weibull, Gompertz, log-normal, log-logistic, generalised gamma, generalised F and two hazard-based spline functions with 1 knot and 2 knots, respectively. The Akaike information criterion (AIC) and Bayesian information criterion (BIC) were calculated using the standard stats package included in R and, in combination with a visual plot of the resulting curves, were used to determine the best-fitting distributions for both treatment groups. A single distribution was chosen for all treatment groups following the advice outlined in NICE Technical Support Document 14. 118 The chosen distribution was also restricted by those that allow proportionality of hazards (i.e. exponential, Weibull, Gompertz and the hazard-based spline functions), in order to appropriately apply the HRs derived in the clinical analysis to produce estimated OS curves for cabozantinib and BSC relative to everolimus. Axitinib was assumed to have the same OS curve as everolimus as a reliable analysis could not be conducted to estimate a HR for OS.

The resulting survival curves along with the KM data are presented in Figures 8 and 9 for nivolumab and everolimus, respectively, and the AIC and BIC statistics are given in Appendix 15.

FIGURE 8.

Nivolumab KM plot and fitted curves for OS.

FIGURE 9.

Everolimus KM plot and fitted curves for OS.

For nivolumab, the results were unequivocal in favour of the Weibull curve, which had the lowest AIC and BIC statistics, as well as being supported by clinical expert opinion. However, for everolimus, the two-knot spline had the lowest AIC value and the log-normal had the lowest BIC value. Within the subgroup of curves in which hazards may be proportional (exponential, Weibull, Gompertz and splines), the two-knot spline, in contrast to having the lowest AIC, had the highest BIC, so the AG considered that other curves with higher AIC but lower BIC may be equally good fits or potentially better. Given that the Gompertz had higher AIC and BIC values than both the Weibull and the exponential curves, this was considered a less preferable model. The exponential and Weibull models had contrasting statistics in that the AIC for the exponential was higher than that for the Weibull but the BIC value was lower. On visual inspection of these curves, the AG considered the Weibull curve to follow a more plausible projection from the KM data than the exponential curve, and this choice was also in line with the best-fitting distribution for the nivolumab treatment group; thus, the Weibull curve was chosen for nivolumab and everolimus for OS (Figure 10). This was supported by expert clinical opinion.

FIGURE 10.

Overview of selected OS curves for all treatments.

Progression-free survival

The same approach taken for OS was applied to the PFS data from CheckMate 025. The KM plots with fitted curves are given in Figures 11 and 12 for nivolumab and everolimus, respectively, and the goodness-of-fit statistics are given in Appendix 15.

FIGURE 11.

Nivolumab KM plot and fitted curves for PFS.

FIGURE 12.

Everolimus KM plot and fitted curves for PFS.

The results of the goodness-of-fit tests clearly indicated that the two-knot spline model had the best fit for both everolimus and nivolumab when assessing the AIC and BIC statistics (Figure 13). This curve was validated by clinical experts as having a plausible extrapolation and so this was chosen to model PFS in the base case.

FIGURE 13.

Overview of selected PFS curves for all treatments.

Time to treatment discontinuation

Treatment discontinuation KM plots were identified from the CheckMate 025 trial for nivolumab and everolimus, and from the METEOR trial for cabozantinib and everolimus. 54,57 The CheckMate 025 trial was used as the source of TTD data for the everolimus group to align with the data used for PFS and OS. However, the AG noted the differences between the KM plots for the everolimus groups in the METEOR trial and the CheckMate 025 trial, and considered the appropriateness of using independently fitted curves for the CheckMate 025 trial groups and the cabozantinib group of the METEOR trial, or if adjustment of the cabozantinib curve would be more appropriate. A method of adjustment was used in the submission for the cabozantinib NICE TA (TA463)28 by performing a NMA to fit independent curves, with a treatment group adjustment applied to the distribution parameters. However, in the AG’s opinion, this method may overestimate the rate of discontinuation for the treatment group that is adjusted (in this case cabozantinib) and, therefore, underestimate the treatment costs. This is because the discontinuation of a treatment (e.g. cabozantinib in the METEOR trial) is, in part, caused by intolerable toxicity caused by the drug, which may not necessarily be reduced if it were given to the population with a different prognosis in another trial (e.g. CheckMate 025). Therefore, the AG chose to fit independent curves to the TTD data to avoid underestimating the treatment costs. This approach may overestimate the treatment costs for cabozantinib if the population characteristics that differ between CheckMate 025 and METEOR are correlated with the time of treatment discontinuation.

The approach taken for PFS and OS was, therefore, also applied to the TTD KM data from the CheckMate 025 trial and the cabozantinib group of the METEOR trial. The regenerated KM plots with fitted curves are given in Figures 14–16 for everolimus, nivolumab and cabozantinib, respectively. The goodness-of-fit statistics are given in Appendix 15.

FIGURE 14.

Everolimus KM plot and fitted curves for TTD.

FIGURE 15.

Nivolumab KM plot and fitted curves for TTD.

FIGURE 16.

Cabozantinib KM plot and fitted curves for TTD.

The two-knot spline had the best fit for the everolimus group with the lowest AIC and BIC, while for nivolumab the two-knot spline had the lowest AIC and was very close to the lowest BIC. For cabozantinib, the two-knot spline did not have as good a fit, with four other models having a better fit based on AIC and BIC. The AG chose to use the two-knot spline for all models to keep the same distribution type across treatment groups and considered this reasonable owing to the relatively small difference in the AIC and BIC statistics for the two-knot spline compared with the best-fitting log-normal for the cabozantinib group. The clear difference in AIC and BIC statistics in favour of the two-knot spline in the everolimus group also supported this decision (curves shown in Figure 17). The log-normal distribution was tested as scenario analysis, which is presented in Results.

FIGURE 17.

Overview of TTD curves for all treatments.

For axitinib, TTD data were not available, so treatment was assumed to discontinue at the point of progression. This is likely to underestimate the treatment cost for axitinib, as the marketing authorisation allows for treatment beyond progression. A scenario analysis assuming that treatment discontinues at the point of progression for all treatments was performed to assess the model results without a discrepancy between treatment assumptions. The results of this are given in Results.

Systematic review of existing health-related quality-of-life data

A systematic review was carried out in July 2016 to identify relevant published HRQoL evidence to populate the economic model. The following databases were searched:

• MEDLINE (via Ovid MEDLINE 1946 to present)

• EMBASE (via EMBASE 1974 to week 27 2016)

• Central Register of Controlled Trials (via CENTRAL, The Cochrane Library)

• HTA database (via The Cochrane Library)

• NHS EED (via The Cochrane Library).

The search strategy for all databases combined disease terms and quality of life terms to capture RCC and HRQoL. The search terms that were used are presented in Appendix 1. In addition, reference lists of studies identified in the search were scanned to identify potentially relevant papers. No restrictions were applied to any of the searches. Studies were assessed for inclusion based on the criteria outlined in Appendix 2. Papers reporting on quality of life of patients with RCC receiving first-line therapy were only included if no relevant studies were identified for patients receiving second-line treatment.

A total of 2200 studies were identified in the database search. The titles and abstracts were assessed by two reviewers. Out of these, 148 were identified as duplicates and 1968 studies were excluded on the basis of title and abstract. A total of 36 studies were identified from the abstract as reporting either condition-specific measures of HRQoL or generic non-preference-based measures of HRQoL. There were 26 studies reporting HRQoL in patients receiving first-line treatment for RCC (23 using generic preference-based measures). Therefore, 59 studies were provisionally included (the full papers of first-line treatment for RCC were only to be reviewed if no studies reporting the use of generic preference-based measures of HRQoL in patients receiving second-line treatment were identified). In cases for which type of measure used or line of treatment was unclear from the abstract, the paper was conservatively considered ‘Q1’ as described in Appendix 2 and was ordered for review. The PRISMA diagram for the search is presented in Figure 18.

FIGURE 18.

The PRISMA flow diagram for HRQoL search.

A total of 25 studies were deemed potentially relevant and were reviewed in full. Out of those, 18 studies were excluded for the following reasons: ‘secondary sources of utility values’, n = 5;76,78,80,81,84 ‘first-line’, n = 4;119–122 ‘no generic preference-based measures used’, n = 4;43,123–125 ‘irretrievable’, n = 1;126 ‘line of treatment not stated’, n = 2;127,128 ‘insufficient data’, n = 1;79 and ‘intervention not of interest, n = 1. 129

Six studies were included from the search. 67,82,92,104,130,131 A targeted search aiming to identify additional HRQoL studies in patients receiving nivolumab was carried out after the initial database search. One study was identified and included in this review which was published after the search was carried out. 68 The included studies are described in the following subsection.

Narrative description of published cost-effectiveness studies

Health-related quality-of-life data selected for the economic analysis

In the baseline analysis, the HSUVs reported in TA333 are used for everolimus, axitinib, cabozantinib and BSC. These values were chosen because the methods used in TA333 to estimate HRQoL were clearly reported and were derived from EQ-5D data collected in the AXIS trial whose population reflects RCC patients encountered in UK clinical practice. Therefore, the values used in the model are 0.69 and 0.61 for PFS and PD, respectively. 89

Patients in the model receiving everolimus, axitinib, cabozantinib and BSC are assumed to experience the same quality of life, which only differed according to progression status. Patients receiving nivolumab are assumed to enjoy a superior quality of life before progression, as the CheckMate 025 trial EQ-5D results show that patients’ quality of life continues to improve throughout the trial period while on treatment. 68 These assumptions have also been validated by the AG’s clinical experts, who have agreed that patients on immunotherapy do experience a higher quality of life than patients on the other types of RCC treatments. Furthermore, the AG’s clinical experts confirmed that patients on BSC may experience a similar quality of life as patients on active treatments because they do not suffer from the AEs associated with drugs. The impact of TRAEs on quality of life is assumed to be captured in the HSUVs used in the model and, therefore, is not modelled separately.

The values used for nivolumab in the economic analysis are 0.73 and 0.65 for PFS and PD, respectively. These values were estimated by applying a quality-of-life increment of 0.036 to the HSUVs for other treatments. The increment used was taken from the Cella et al. 68 analysis, which is the difference in EQ-5D utility values according to treatment arm in the CheckMate 025 trial. These values were validated by expert clinical opinion.

In summary, the HSUVs used in the AG’s economic model, presented in Table 14, were chosen for the following reasons.

• Expert clinical opinion (Dr Lisa Pickering, personal communication; Professor Martin Gore, personal communication; and Dr Amit Bahl, personal communication) considered the AXIS trial population reflective of patients seen in UK clinical practice. Expert clinical opinion sought by the AG (Dr Lisa Pickering, personal communication; Professor Martin Gore, personal communication; and Dr Amit Bahl, personal communication) also indicated that the CheckMate 025 population was less reflective of RCC patients in the UK and could potentially reflect a population with better prognosis, which could bias the quality of life experienced by patients at baseline.

• The HRQoL analysis in TA333 was robust and reported in detail in the ERG report.

Drug acquisition and administration costs

The doses considered in the model were axitinib (5 mg, twice daily), cabozantinib (60 mg, once daily) everolimus (10 mg, once daily) and nivolumab (3 mg/kg, every 2 weeks). A summary of the doses and acquisition costs of the treatments included in the model is presented in Table 15.

All of the drugs are administered orally except for nivolumab. The administration cost for nivolumab was assumed to be £185.53 (NHS Reference Costs 2014–15,136 Outpatient, Simple parenteral chemotherapy, Currency code SB12Z).

Disease management costs

There were no UK costing studies for RCC identified in the systematic literature review carried out to identify costing studies. Therefore, the estimates used in the model were based on those used in NICE TA33389 and TA10078,137 complemented by expert clinical opinion sought by the AG.

Before progressing, patients are assumed to receive CT scans every 3 months and to be seen by a consultant at the time of the scans, as the scans need clinical revision for signs of progression. Patients also were assumed to have a blood test every month within their progression-free period.

After progressing, patients are assumed to have daily pain medication and 20 community nurse visits per year. A summary of the resource use and costs estimated for each health state is presented in Table 16.

Terminal care costs

The estimated cost of TC is £7713. All patients who die in the model incur this cost. The cost estimate is taken from a paper published by the Nuffield Trust estimating costs associated with end-of-life care in the UK. Estimates for patients with a history of cancer were presented separately in the paper. 139 A summary of the cost components of TC used in the paper is presented in Table 17.

Adverse event costs

Resource use that was assumed for the management of TRAEs in the model is summarised in Table 18. This was validated by the AG’s clinical experts.

Subsequent therapy costs

Currently only nivolumab is approved as third-line treatment option for RCC in the UK. However, patients in all the three pivotal trials could receive a variety of further therapies after discontinuing treatment; therefore, the effectiveness estimates from the trials also included the impact of subsequent therapies.

Patients in the axitinib, everolimus and nivolumab arms of the model receive one further treatment line after discontinuing treatment. The proportions of patients who went on to receive subsequent therapies, and their distribution across various therapies, are based on the numbers reported in the trials as presented in Table 19. The proportions for subsequent therapy after nivolumab and everolimus are based on proportions observed in CheckMate 025, reweighted after the removal of bevacizumab and sorafenib, which are currently not treatment options for RCC in the UK at any line. The proportions assumed for after discontinuing axitinib and cabozantinib are based on the AXIS and METEOR trials as reported in the NICE TA463, reweighted after removing sorafenib.

The treatment duration for subsequent therapy is assumed to be 15.82 weeks, which is in line with the median duration of treatment reported in a trial assessing third-line treatment in RCC. 142 The dosage and costs of axitinib and everolimus as third-line treatments are the same as those used to model them as second-line treatments as reported in Drug acquisition and administration costs. The dosage and costs of pazopanib and sunitinib are summarised in Table 20. The cost of subsequent therapy is applied as a one-off cost as soon as patients discontinue treatment.

Miscellaneous costs

According to the AG’s clinical experts, around 10% of patients on BSC receive palliative radiotherapy for bone pain due to metastasis. The outpatient costs of preparation for simple radiotherapy and delivery of a fraction of treatment on a megavoltage machine has been included for patients in the BSC arm of the model, which is in line with the NHS clinical commissioning policy for palliative radiotherapy for bone pain. 143 The costs were estimated using the unbundled Healthcare Resource Group codes SC47Z for planning and SC22Z for delivery, which give a total cost of £419, which is applied to 10% of patients receiving BSC in the model. 136

Probabilistic analysis

A PSA was performed to assess the impact of parameter uncertainty on the results of the base-case analysis. This was performed using 10,000 samples from distributions fitted to each uncertain parameter that was included in the PSA. When measures of uncertainty were not available, the SD was set at 25% of the mean value of the parameter to allow sensitivity to be captured for all relevant parameters.

Parameters for costs, resource use and dose adjustments were fitted with gamma distributions to ensure non-negative sampled values. The only costs with a measure of uncertainty were those from NHS Reference Costs,136 which provided a mean value as well as an upper and lower quartile. These data were used to estimate the parameters of a gamma distribution by minimising the sum squared error (SSE) between the predicted quartiles and the reported quartiles. This was performed using the Solver tool in Microsoft Excel to vary the beta parameter until the minimum SSE had been achieved [based on the generalised reduced gradient non-linear solving method]. The distribution parameters for all other costs, resource usage and dose adjustment values were calculated directly from the mean and SD (assumed to be 25% of the mean).

Adverse event probabilities and HSUVs were fitted with beta distributions to constrain values between 0 and 1. The distribution parameters were calculated using the numbers of patients experiencing the event and the number of patients not experiencing the event in the relevant trials for each treatment group.

For efficacy of treatments, survival curves for OS, PFS and TTD were simulated in R using the mvrnorm function from the MASS package. This generates a specified number of random samples of the coefficients from the fitted survival models, given the arguments for the vector of coefficients and the covariance matrix of these coefficients. The covariance matrix was derived using the vcov function in the standard stats package included in R. Samples of HRs calculated in the NMA were provided by exporting the CODA (Convergence Diagnosis and Output Analysis) from WinBUGS.

A full summary of the distributions applied to the model parameters is given in Appendix 16.

Base-case results

A summary of disaggregated discounted costs is given in Table 22. This shows a breakdown of costs by health state for treatment acquisition, disease management both on and off treatment, subsequent therapy and end of life. A summary of discounted and undiscounted QALYs is given in Table 23 and Appendix 17, respectively, showing total QALYs for health state and treatment discontinuation status.

An incremental analysis of the results, both discounted and undiscounted, is given in Table 24 and Appendix 17, respectively.

Results of scenario analyses

The results of the scenario analyses are shown in Appendix 18.

Results of deterministic sensitivity analyses

Deterministic OWSAs were performed for every parameter that was varied in the PSA, using an upper and lower limit for each parameter based on the 95% CI for the distributions fitted to each parameter in the PSA. In addition, the discount rate was tested from 0% to 6%. Given the large number of parameters in the model, the results presented in Appendix 19 are restricted to those analyses in which either the total costs changed by at least £2000 for at least one treatment group or the total QALYs changed by at least 0.01 for at least one treatment group. In addition, given the number of comparators, the results summary contains the change in costs and QALYs for each treatment group but no ICERs are provided. However, the net monetary benefit (NMB) of each treatment is given also in Appendix 19 for the OWSAs, for which the ranking of NMB changed from the base-case results. This is given at both the £20,000 and £30,000 per QALY threshold.

Results of the probabilistic sensitivity analysis

The mean results from the 10,000 probabilistic samples is given in Table 25. A plot of costs and QALYs for all the samples in each treatment group is given in Figure 19.

FIGURE 19.

The PSA sampled results for all treatments.

MEDLINE (via Ovid) – epub ahead of print, In-Process & Other Non-Indexed Citations, MEDLINE(R) (via Ovid) Daily and MEDLINE(R) (via Ovid) 1946 to present

Date range searched: inception to 22 June 2016.

EMBASE (via Ovid)

Date range searched: inception to 22 June 2016.

The Cochrane Library, CENTRAL, DARE, NHS EED

Date range searched: inception to 9 June 2016.

MEDLINE(R) In-Process & Other Non-Indexed Citations (via Ovid) and MEDLINE(R) (via Ovid) 1946 to present

Date range searched: inception to 13 January 2016.

EMBASE (via Ovid) 1974 to week 2 2016

Date range searched: inception to 13 January 2016.

Ovid MEDLINE(R) 1946 to present

Date range searched: inception to 18 February 2016.

EMBASE 1974 to present

Date range searched: inception to 18 February 2016.

The Cochrane Library

Date range searched: inception to 18 February 2016.

MEDLINE

Date range searched: inception to 18 July 2016.

EMBASE 1974 to present

Date range searched: inception to 18 July 2016.

CENTRAL

Date range searched: inception to 18 July 2016.

NHS EED and HTA

Date range searched: inception to 18 July 2016.