Cost Effectiveness of Nivolumab in Patients with Advanced, Previously Treated Squamous and Non-squamous Non-small-cell Lung Cancer in England

The patient population, adults aged ≥ 18 years with advanced or metastatic squamous or non-squamous NSCLC after failure of prior platinum doublet-based chemotherapy [5, 6], is aligned with that of CheckMate 017 and CheckMate 057.

The intervention considered in the economic models was nivolumab, which was compared with docetaxel. In the clinical trials, nivolumab was administered at 3 mg/kg every 2 weeks (Q2W) and docetaxel was administered at 75 mg/m² every 3 weeks (Q3W) [5, 6]. However, dosing of nivolumab was updated to a flat dosing schedule of 240 mg Q2W in 2019, which is deemed to be clinically equivalent and is the recommended dose for NSCLC in the summary of product characteristics [4], subsequent to the trials being conducted. Therefore, to reflect the current clinical use of nivolumab in the economic models and facilitate comparison, we used 240 mg Q2W dosing as the base case in both the original and the updated models.

Both models were developed by using a partitioned-survival approach based on PFS and OS predictions from CheckMate 017 and CheckMate 057 to estimate the projected proportion of patients in each health state (progression free, progression, and death) throughout the model’s time horizon. Costs and health outcomes of patients with squamous and non-squamous NSCLC were assigned to the cohort based on the projected proportion of patients in each health state over time. Drug, administration and monitoring costs of nivolumab and docetaxel were based on the TTD observed in CheckMate 017 and CheckMate 057. Key elements of the model structure and input parameters are detailed in Table 1.

OS and PFS were based on parametric survival models fitted to the 60-month data cut of CheckMate 017 and CheckMate 057 clinical trial data. All analyses followed the decision support unit guidelines by fitting both standard parametric functions and spline models to the available data [17]. For the spline models, up to three knots were fitted; however, in line with the analyses for the original TA, knots were limited to two for selection of base-case distributions in order to avoid overfitting the data. For instances where a good visual fit to the Kaplan–Meier data could not be achieved with curves other than three-knot splines due to more complex shapes of the survival pattern, scenario analyses were included. Whether proportional hazard could be assumed was explored based on log-cumulative hazards and log-cumulative odds plots to determine if parallel lines were evident. In instances where proportional hazards could not be ruled out, survival models with treatment as a covariate were fitted. Models were fitted to each arm separately where proportional hazards assumption did not hold. Selection of the best fitting distribution for each outcome was informed by the Akaike information criterion (AIC), the Bayesian information criterion (BIC) and the visual fit to the data. For AIC, distributions with a difference of < 4 to the distribution with the lowest AIC were considered appropriate based on the Burnham and Anderson rule of thumb [18]. Similarly, based on Raftery’s rule of thumb, it was considered that a difference in BIC > 10 to the distribution with the lowest BIC was inappropriate [19] (see Online Resource 1, Table A-1, in the electronic supplementary material [ESM] for AIC and BIC values). Furthermore, if statistical and visual fit for both arms could be achieved by using the same distribution, use of a common distribution was preferred over different distributions between arms.

The distributions selected for the base-case analyses of each model are presented in Table 1, and Figs 1 and 2 present the predicted survival overlaid with the CheckMate 017 and CheckMate 057 clinical trial data. Distributions for OS and PFS preferred by the committee in the original submission are shown in parentheses in Table 1. All distributions considered for the updated analyses are presented in Online Resource 1, Figs A-1 to A-4 (see ESM). Long-term extrapolations of the selected distributions are presented in Online Resource 1, Figs A-5 to A-6 (see ESM). Duration of treatment data were available beyond the 2-year stopping rule agreed on in the original TA; therefore, it was modelled using the Kaplan-Meier curves for TTD, without need for extrapolation.

For the NICE appraisals, health-state utility values were taken from the pre-progression and post-progression utility values from the EQ-5D preference-based health-state utility questionnaire (EQ-5D utility index) and visual analogue scale for overall health status data collected in CheckMate 017 and CheckMate 057. The EQ-5D data from the trials were converted to health-state utilities using the published UK value set [20] resulting in utility values of 0.750 and 0.739 for the progression-free health state and 0.592 and 0.688 for the progressed-disease health state in squamous and non-squamous NSCLC, respectively [21, 22]. These values were deemed too high by the evidence review group (ERG) on the basis that the trial population was not representative of NSCLC patients in the UK and values from the published literature were argued to be more appropriate. For the squamous appraisal, the ERG suggested a utility value of 0.65 for the progression-free health state and 0.43 for the progressed-disease health state from Nafees et al. [23]. For the non-squamous appraisal and in order to account for low completion rates of EQ-5D questionnaires, the ERG used a subset of early EQ-5D responses from the CheckMate 057 trial to estimate a progression-free health-state value of 0.713. Further, the ERG suggested using a utility value of 0.545 from van den Hout et al. [24] and applied a terminal care disutility, which resulted in a final utility value of 0.476 for the progressed-disease health state. The committee felt that the health-state utility values proposed by the ERG were too low, those proposed by the company were too high, and the true value would be somewhere in between the two. The base-case utility values in the squamous model reflect the committee’s preferred values, in between the trial utility values and those provided by the ERG from the literature. The base-case utility values in the non-squamous model reflect the committee-preferred assumption for the progression-free state and the trial value in the progressed-disease health state (Table 1). Utility values based fully on the trial were included in a scenario analysis.

Data on the safety outcomes for each arm were captured in the CheckMate 017 and CheckMate 057 clinical trials. Grade 3–4 adverse events that were reported in ≥ 5% of patients and ≥ 2% of patients in CheckMate 017 and CheckMate 057, respectively, were included in the analyses. Adverse events were accounted for in the model by applying a unit cost and a disutility value for each event, as shown in Online Resource 1, Tables A-2 and A-3 (see ESM).

Drug acquisition unit costs are shown in Table 2. The list price for nivolumab is shown; however, a confidential patient access scheme (PAS) was applied in the model and is reflected in the results presented. As described earlier, the drug, administration and monitoring costs of both nivolumab and docetaxel were based on the TTD from the CheckMate 017 and CheckMate 057 clinical trials. This was done to adequately reflect that some patients could be treated beyond progression or discontinue treatment prior to progression. Further, a 2-year stopping rule was applied to the nivolumab drug costs in both models to reflect NICE recommendations [8, 9]. In the base case, treatment effect of nivolumab following the 2-year stopping rule was assumed to be durable over the lifetime of the patients. A continued treatment effect of nivolumab for 3 years following treatment discontinuation was explored in scenario analyses. This scenario is applied such that after 3 years following discontinuation, patients in the nivolumab arm switch to the docetaxel hazard.

Healthcare resource use was taken from previous NICE appraisals in NSCLC and were validated with a UK clinician [25, 26]. For each health state, the number of each element of resource use required per 4-week period was estimated, and unit costs were applied to these values. Unit costs for each element of resource use were estimated based on UK sources and are shown in Online Resource 1, Tables A-2 and A-3 (see ESM).

The cost of subsequent treatments was included in the analyses. Drug acquisition unit costs of subsequent therapies are included in Table 2. The proportion of patients in each arm receiving subsequent therapy and the type of subsequent therapy that they received were estimated from the CheckMate 017 and CheckMate 057 clinical trials.

Finally, a terminal cost reflecting end-of-life care was included in the models. So that the impact of only the newly collected data could be assessed, costs were set to be consistent with the original TA submissions. Therefore, the cost-year was 2015, and costs quoted for other cost-years were inflated to 2015 using inflation indices from the Personal Social Services Research Unit [27].

The analyses that were originally submitted during NICE TA483 (squamous) and TA484 (non-squamous) [15, 16] were conducted with a confidential NHS England (NHSE) PAS discount specific to the CDF, as well as the original dosing of nivolumab, 3 mg/kg every 2 weeks. Therefore, the cost-effectiveness analysis deemed plausible under the appraisal committee’s preferred assumptions cannot be easily compared with the updated analyses.

To isolate the impact of the additional data and to facilitate comparison, the original analyses were conducted with the current confidential NHSE PAS discount and with the nivolumab dosing of 240 mg every 2 weeks. Thus, the change in ICER resulting from additional information and reduction of uncertainty provided during the CDF data collection period can then be assessed.

Beyond the base-case analysis, several scenarios were explored to investigate areas of particular uncertainty. The scenarios explored included the following:

To fully characterise uncertainty around the results of both analyses, probabilistic and one-way sensitivity analyses were carried out. Tested ranges in the one-way sensitivity analysis were based on 20% increase or decrease on the mean value. The probabilistic sensitivity analyses were conducted by using second-order Monte Carlo simulation for 1000 iterations, sufficient to reach a stable probabilistic result. Model validation was performed in alignment with best practices [28] and included face validity and internal validity (verification of input data and coding). The model also was externally reviewed by the ERG, per NICE’s TA process [21, 22].