Cost-Effectiveness of Pazopanib Versus Sunitinib for Renal Cancer in the United States

We write to comment on a recently published study by Delea et al. in the January 2015 issue of JMCP that evaluated the cost-effectiveness (CE) of sunitinib (SU) versus pazopanib (PAZ) as first-line treatment for metastatic renal cell carcinoma (mRCC) from a U.S. third-party payer perspective.¹ This analysis was based on COMPARZ and PISCES, clinical trials that compared SU and PAZ^2,3 and led the authors to conclude that PAZ is cost-effective (in fact, dominant, according to the base-case results) compared with SU. Such assessment of economic value is clearly important for deciding between therapies to ensure fair access; therefore, we welcome a comparative evaluation of SU and PAZ. However, we believe that some of the key assumptions and inputs used in the model by Delea et al. render their results and conclusions invalid.

Best practice requires that results from a health economic model should reflect the most likely outcomes based on sound methodology and robust evidence for its inputs, as recommended by the International Society of Pharmacoeconomics and Outcomes Research (ISPOR).⁴ Here, we focus on 2 key areas (utilities and survival modeling) where, in our view, the analysis by Delea et al. falls short of this standard, and a third area (treatment costs) where the basis for the data derived is unclear.

Utilities

The CE analysis by Delea et al. sourced utility values (as measured by the EuroQol 5-dimensions [EQ-5D] questionnaire) from the PISCES study, a blinded study that mandated switching between SU and PAZ before disease progression. The aim was to assess patient preferences with regard to tolerability and safety under the explicit assumption of equal efficacy.³ However, for the reasons stated below, the PISCES study could not, by design, provide a valid measure of utility—a crucial input for the type of CE analysis conducted by Delea et al.

First, PICSES did not provide EQ-5D data from patients while on active treatment with PAZ or SU before the switch to the subsequent agent. Instead, EQ-5D was assessed at baseline (i.e., before any treatment), at the end of the 2-week washout period before the switch (week 12), and after the switch at the end of treatment (week 22). In the case of SU, the PISCES study design took into account the drug’s intermittent 4/2 dosing schedule (i.e., 4 weeks on treatment followed by 2 weeks off) by correctly conducting the postwashout assessment at the end of 2 complete 6-week treatment cycles. As a result, the mean EQ-5D estimate for SU-first patients at washout was 0.8103, a 0.0478 improvement from their mean predose EQ-5D of 0.7625 for SU.⁵ For the corresponding PAZ-first group, the mean EQ-5D utility score at the end of washout, 0.7595, was much lower than for SU and perhaps unchanged from the mean predose level of 0.7664. However, since there is no 2-week pause in the treatment regimen for PAZ, it is unclear what these data mean.

Faced with this paucity of data, Delea et al. imputed a weighted average of the mean EQ-5D estimates—but only for SU—before the switch to PAZ from 1 arm (0.8103) and after the switch from PAZ from the other arm (0.6325), thereby yielding an EQ-5D estimate of 0.6918, which is significantly lower than the mean baseline value of 0.7625. For PAZ, only the EQ-5D estimate after the switch from SU was used (0.7487). However, it is clearly not meaningful to use the data after the switch or to intermix utility values before and after a mandated switch. In the real world, where disease progression or tolerability issues motivate a switch in treatment,⁶ utility values may fully reflect the trade-offs that patients and physicians make, but that is not the case here.

Of note, unequal disease progression rates in PAZ and SU observed in PISCES before the switch (20% vs. 11%, respectively) and a high degree of overall dropouts (~37%)⁵ led to noncomparable patient populations before and after the switch, resulting in potential bias and difficulty in interpreting the EQ-5D data.

It is also notable that the EQ-5D data were not included in the peer-reviewed primary PISCES publication.³ Indeed, interpretation of the EQ-5D results (which trend in the opposite direction to patient preference data in PISCES) has previously been considered “highly problematic” by GlaxoSmithKline, the trial’s sponsor.⁵

Importantly, too, the imputed EQ-5D data in the Delea et al. study are not consistent with more robust published data. In Cella et al. (2008), EQ-5D scores for SU as measured in the pivotal phase III trial of SU in mRCC with a much longer treatment period (11 cycles or 66 weeks), a larger patient population, and no switch prior to progression were stable throughout treatment and equivalent to scores at baseline (~0.76).⁷ We therefore believe that not making use of the results in Cella et al. is a major limitation of the Delea et al. analysis, since it contradicts ISPOR recommendations to identify and incorporate all relevant evidence.^4,7 Furthermore, the difference in EQ-5D results between the “on” and “off” treatment periods for SU is rendered unrealistically high in the data imputed by Delea et al., compared with the measured data reported from the same SU trial (0.1778 vs. 0.03, respectively).^7,8 The fact that the 0.1778 difference in health utility is as great as that between pre- and postprogression in many different cancers makes this an implausible input.^9-12

Finally, the large imputed difference in EQ-5D for SU in the Delea et al. analysis is not consistent with evidence from the COMPARZ trial,² which found there was unlikely to be a clinically meaningful difference in quality of life, as measured by the Functional Assessment of Cancer Therapy–Kidney Symptom Index–19 item (FKSI-19) questionnaire. This is despite the fact that no assessment using FSKI-19 was included during the break in the SU dosing schedule (a potential source of bias against the drug^8,13). Therefore, to impute large clinically meaningful differences in EQ-5D, despite the weight of evidence from Cella et al. and COMPARZ, is not justifiable.^2,7

In summary, data from PISCES used to derive EQ-5D utility, one of the key inputs driving the results of the analyses, were not suitable for use in the CE model by Delea et al.

Survival Modeling

We also have doubts about the model inputs and assumptions used in estimating survival in Delea et al.’s analysis. This was, for instance, based on a 37.5-month time horizon, rather than using a lifetime perspective—the standard requirement in health economics analysis to guarantee that the model is “long enough to capture relevant differences in outcomes across strategies.”⁴ Because of the short time horizon, it excluded the overall survival (OS) benefit for SU that occurs after the follow-up, as seen in Figure 1B in Delea et al.’s article—a treatment effect that would otherwise reduce or possibly eliminate the life-year gain (LYG) for PAZ in Table 2. Indeed, sensitivity analysis (Figure 3) suggests that a shorter time horizon benefits PAZ.

The Kaplan-Meier curves for OS are dependent on post-progression treatment patterns that are very complex. For instance, the COMPARZ study report (Table 17, page 69)¹⁴ shows that, beyond a median per-protocol treatment duration of approximately 8 months, more than half of the patients were treated with a variety of other anticancer agents, with clear differences between the arms. For example, 27% of the PAZ group received SU and 14% of the SU group received SU as follow-up therapy while, in contrast, only 3% of PAZ patients and 8% of SU patients received PAZ as follow-up therapy.¹⁴ OS reflects the efficacy of the different treatment arms, including first-line and subsequent lines of therapies. Therefore, any attribution of OS to the original treatment groups is highly problematic, unless the subsequent lines of treatment are similar between the treatment arms, which is not the case in the COMPARZ study. Consequently, it may not be valid to attribute any computed LYG advantage to PAZ, as is done in Delea et al.

Furthermore, Delea et al. chose to use investigator-assessed progression-free survival (PFS) from COMPARZ, arguing that this may better reflect real-world practice, which is a reasonable assumption in general. But, in this case, investigator-assessed PFS favored PAZ, whereas the primary endpoint, PFS as assessed by the independent review committee (IRC), actually favored SU. Apart from the selective use of data, this raises the possibility of bias, since COMPARZ was an open-label study in which only the IRC assessments were blinded.

In summary, the selective model inputs for survival modeling in Delea et al.’s analysis may have also led to invalid CE results for PAZ and SU.

Treatment Costs

Costs associated with SU and PAZ in the model analyses, based on the health care resource utilization (HCRU) data from COMPARZ, appear questionable, since they are inconsistent with data in a companion article by Hansen et al. (2015),¹⁵ the source cited by Delea et al. for corresponding detailed cost analysis and results. In Hansen et al. (Figure 2A), the derived total average medical and ambulatory costs for PAZ- and SU-related treatment were $2,151 versus $3,153, respectively, over a duration of approximately 10.6 months; however, in Delea et al. (Table 2), the same costs over a slightly longer time period (~13 months by the estimated progression-free life-years in Table 2), were around 4 times higher ($9,147 vs. $12,381, respectively). While such a difference between the drugs in this item may appear small compared with the overall costs, it is still important to consider. Where treatments have similar derived efficacy in a CE analysis (as suggested by available data for SU and PAZ), the incremental cost-effectiveness ratio becomes very sensitive to cost differences, so otherwise unimportant cost differences may drive results.

Finally, it is worth noting that COMPARZ was a noninferiority study, and CE analyses do not distinguish evidence from such studies from that provided by superiority trials. A potentially small cost difference in favor of PAZ, as suggested by Delea et al., must therefore be viewed in the context of possibly lower efficacy for PAZ, particularly given that noninferiority criteria in COMPARZ were only met for the intent-to-treat analysis but not for the per-protocol analysis (rather than for both analyses, as required to confirm a drug’s noninferiority).^16-18

In conclusion, to argue that PAZ is dominant or cost-effective compared with SU on the basis of data used by Delea et al. is not justified. Available clinical trial evidence (including studies run by the drug’s sponsor) appears to be overlooked in their study. Also, the EQ-5D and survival analysis estimates that drive the QALY estimates lack credibility and validity. Making essential corrections for the base-case analysis would result in a change in the main results. CE models, which in many countries may impact patient access to cancer treatments, need to be based on more reliable data than those used by Delea et al.