Rifaximin for preventing episodes of overt hepatic encephalopathy

The company conducted a systematic literature review and identified 3 relevant published studies including rifaximin. Of these, 2 studies were excluded because the doses used were different from those in the UK marketing authorisation for rifaximin. The submission therefore included 1 published study (Bass et al. 2010), which reported results from the pivotal trial, RFHE3001. This was a 6 month, international, multicentre, randomised, double‑blind trial comparing rifaximin with placebo for maintaining remission in people with recurrent, overt, episodic hepatic encephalopathy resulting from chronic liver disease. People taking lactulose at baseline were allowed to continue its use during the study period and dose changes were allowed as needed. The company also identified a second trial, RFHE3002; this was an international, multicentre, single‑arm, open‑label study that assessed the long‑term safety and tolerability of rifaximin in people with a history of hepatic encephalopathy. It was an extension to RFHE3001 for people who completed RFHE3001, and for newly enrolled people.

The company carried out a systematic literature review but stated that of the 3 studies identified, none were relevant because the population and outcomes did not match those of the current decision problem. The company stated that a comparison with neomycin was excluded because it is not routinely used in clinical practice, no clinical data are available for the use of neomycin in this indication, it is associated with risks of ototoxicity, nephrotoxicity and hepatic impairment, and it is also not licensed for the indication being appraised.

The company carried out a de novo analysis of the cost effectiveness of rifaximin plus concomitant lactulose compared with placebo plus concomitant lactulose, given that approximately 91% of people in each arm of RFHE3001 had concomitant lactulose. The 2 arms of the model are referred to in this document as the 'rifaximin' and 'lactulose' arms respectively. The company presented 4 iterations of its economic model, with its 4 submissions (referred to in this document as 'original', 'October 2013', 'December 2013' and 'November 2014'). The October 2013 submission was prepared after the first consultation, and incorporated updates to the clinical‑effectiveness inputs, mortality estimates, hospital admission rates and utility estimates. The December 2013 submission explored additional issues raised by the Committee in the second appraisal consultation document, relating to estimates of utility, the mortality benefit associated with rifaximin and the time horizon. The November 2014 submission provided further exploration of utility scores and incorporated evidence from clinical audits to inform the rate and duration of hospital admissions.

In the original submission, the company developed a Markov cohort model consisting of 5 states to reflect the clinical pathway of hepatic encephalopathy. The health states are referred to in this document as initial remission, breakthrough overt episode, subsequent remissions, subsequent overt episodes and death. The company noted (November 2014) that the remission states could also be described as covert hepatic encephalopathy, because the symptoms of hepatic encephalopathy do not completely resolve. Given that time to subsequent episodes was not available from RFHE3001, the company assumed that the risk of having a subsequent breakthrough episode was independent of the risk of preceding episodes and the time spent in the remission state. It was also assumed that the risk reduction for the first breakthrough episode could also be applied to subsequent episodes, based on clinical expert opinion, and this was assumed to be constant over time for subsequent episodes. However, in response to the Committee's considerations in the first appraisal consultation document, that the assumptions in the original model oversimplified the nature and course of the disease, the company submitted the October 2013 analysis in which the risks of subsequent episodes were based on data on time to subsequent events from RFHE3002. The original and October 2013 models had a time horizon of 5 years; the December 2013 and November 2014 analyses used a lifetime (42‑year) time horizon. The models used monthly cycles, did not include a half‑cycle correction, and included both costs and benefits discounted at 3.5%. The analyses were done from the perspective of the NHS and personal social services.

In the original submission, time to first breakthrough episode of hepatic encephalopathy was extracted and combined from the Kaplan–Meier survival curves in RFHE3001 and RFHE3002. The combined data set was extrapolated, using the log normal distribution, to estimate the effectiveness of rifaximin beyond the final 168‑day observation point in RFHE3001. However, in its October 2013 analysis the company took into account the Committee's comments that the use of a combined data set was not appropriate. It used data from RFHE3001 only to model time to first breakthrough episode of hepatic encephalopathy for the rifaximin and lactulose arms, and survival times were censored at day 170. The survival curves were then extrapolated beyond 170 days for both groups using a log normal distribution and a proportional hazards assumption. To model subsequent hepatic encephalopathy episodes (for which there were no data from RFHE3001), the company used data from only the newly enrolled people in RFHE3002 to avoid potential bias from people who had been in RFHE3001 and did not have a breakthrough overt episode in RFHE3001. Because RFHE3002 was a single‑arm study (rifaximin only), the company applied the rifaximin treatment effect from RFHE3001 to the RFHE3002 data for all subsequent events. All subsequent episodes were modelled together as a single health state. As for the first overt episode, the survival curves for subsequent overt episodes were extrapolated to 5 years using the log normal distribution and applying an assumption of proportional hazards. The company considered the log normal distribution to provide the best model fit in both cases based on the log likelihood values of the distributions tested and visual assessment.

The company also presented 3 separate analyses to justify using RFHE3002 data in the economic model. The analyses included:

• comparing time to first observed breakthrough hepatic encephalopathy episode between the RFHE3002 subgroups, that is, rollover (people previously enrolled in RFHE3001 and treated with rifaximin), crossover (people previously enrolled in RFHE3001 and treated with placebo) and newly enrolled people, from RFHE3002 baseline

• comparing time to first observed breakthrough hepatic encephalopathy episode between RFHE3001 treatment groups and RFHE3002 subgroups, from each respective study baseline

• comparing time to first breakthrough episode between newly enrolled people and the combined RFHE3001 and RFHE3002 data for rollover and crossover people.
  
  The results of these analyses showed that time to first breakthrough hepatic encephalopathy episode was comparable between people in RFHE3001 and RFHE3002 given the similarity in their baseline characteristics, except in the number of prior hepatic encephalopathy episodes before trial entry (see sections 3.2 and 3.10).

The company incorporated mortality into the economic model through the modelling of transition into the 'death' health state. In the original submission, hepatic encephalopathy‑specific mortality was estimated from other external data sources (Bustamante et al. 1999 and Shawcross et al. 2011) rather than RFHE3001. The company stated that this was because the trial did not provide evidence on people at the more severe end of the disease spectrum, even though it reflected the range of people who would present with hepatic encephalopathy in clinical practice, and that data from the trial were not sufficiently mature for an analysis of mortality. However, the company explored alternative mortality estimates in its October 2013 and December 2013 analyses.

In the October 2013 economic analysis, the company modelled mortality using data for all people in RFHE3002, and noted the small number of deaths seen. Time to death for the initial remission state was estimated from the first dose of rifaximin in RFHE3002, whereas for the subsequent remission state it was estimated from day 31 after each overt episode (assumed to be when people returned to the remission state after an overt episode). The survival curves were then extrapolated based on an assumption of proportional hazards using the log normal distribution for the first remission state and the Weibull distribution for subsequent remissions taking into consideration the log likelihood and visual assessment of the best model fit. The probability of death for the overt states was estimated as the number of deaths within 30 days of the onset of the first observed overt episode in RFHE3002 and the number of deaths within 30 days of the onset of any subsequent overt episode.

As part of the October 2013 analysis, the company also carried out a literature review of mortality in people with hepatic encephalopathy and an analysis of the mortality data for people with hepatic encephalopathy using the Clinical Practice Research Datalink (CPRD). The literature review and CPRD data showed that mortality was highest in the first 6 months after diagnosis of hepatic encephalopathy, which could be associated with the first overt hepatic encephalopathy episode. A study by Neff et al. (2012) showed that people who had lactulose alone had a higher 6‑month mortality; 40% compared with 35% for people who had rifaximin plus lactulose and 24% for people who had rifaximin alone. The company also noted from the CPRD data that people who had survived 6 months after the first overt episode had similar mortality to that seen in RFHE3001 and RFHE3002, in which people had been in remission after at least 1 overt episode in the 12 months before randomisation.

The company noted that in the base case, the risk of death differed between the 4 health states, but that these risks did not differ between the rifaximin and lactulose arms. Consequently, any apparent differences between the groups in mortality resulted from differences in the time spent in each health state (for example, delaying overt episodes), rather than an explicit mortality benefit with rifaximin. Moreover, the company noted that because the time horizon was extended to lifetime (42 years) in the December 2013 analysis, the mortality benefit for rifaximin diminished over time. In its December 2013 analysis, the company presented 3 scenario analyses to explore the impact of the differences in mortality estimates on the results of the economic model:

• In the first scenario, it was assumed that mortality was the same in both of the remission health states.

• In the second scenario, it was assumed that mortality did not increase with overt episodes (compared with the remission health state) and therefore, for each overt episode health state, mortality was the same as for the preceding remission health state.

• In the third scenario, it was assumed that mortality was the same for all 4 health states.

The overt states also included hospital admissions caused by hepatic encephalopathy episodes. In RFHE3001, 19 of the 140 people in the rifaximin group and 36 of the 159 people in the placebo group were hospitalised; 31 people in the rifaximin group and 73 people in the placebo group had breakthrough episodes. The company therefore calculated the probability of hospital admission in people who had an episode to be 61.29% and 49.32% in the rifaximin and placebo groups respectively. In the original submission, these 6‑month probabilities of hospital admission were converted to monthly probabilities of 14.63% and 10.71% in the rifaximin and lactulose arms respectively (assuming a constant hazard over time) and applied to the people predicted to reach the overt health state. However, in the October 2013 analysis, the company applied the aggregated 6‑month probability of 52.88% to both arms of the model. The company used the aggregated 6‑month probability rather than individual probabilities of 61.29% for rifaximin and 49.32% for lactulose on the basis that the proportions of overt hepatic encephalopathy episodes that led to hospital admission in the trial were not statistically significantly different between the 2 arms. In the original, October 2013 and December 2013 submissions, the company assumed that each hospital stay lasted 5 days.

In the November 2014 submission, the company presented analyses incorporating evidence on hospital admission rates and bed stays from audits of clinical practice in the UK (see section 3.13). The company adjusted the model to simulate as closely as possible the number of hospital admissions and bed days seen in these audits. The company assumed that all episodes of overt hepatic encephalopathy led to hospital admission (that is, the probability of admission was set to 100%). It noted that this was necessary because the number of episodes predicted by the model (based on RFHE3001) was lower than seen in clinical practice. The company also adjusted the assumed length of stay for each hospital admission, based on its meta‑analysis of the audit data; it presented analyses using the mean reduction in bed days in the meta‑analysis of all 4 centres, the mean reduction when Bolton was excluded, and the upper and lower 95% confidence limits for each meta‑analysis. In the scenario based on the meta‑analysis of all 4 centres, the length of stay was assumed to be 30.6 days in the lactulose arm and 20.2 days in the rifaximin arm.

The utility values used in the original model were derived from 200 randomly selected members of the general public, using the time‑trade‑off and standard gamble approach. Noting the Committee's concerns about this approach in the first appraisal consultation document and the Committee's preference for quality‑of‑life data collected from RFHE3001, the company presented alternative results in the October 2013 analysis. The company used a validated algorithm developed by Gray et al. (2006) to map the SF‑36 utility scores collected in RFHE3001 to EQ‑5D utility values. The company also included in the analysis CLDQ scores that were derived from the quality‑of‑life study by Sanyal et al. (2011). The study reported CLDQ data only for North American and Canadian people from RFHE3001 (219 of 299 people) because there was not a validated Russian translation of CLDQ for Russian people. The company then carried out a regression analysis and a repeated measures analysis to derive and quantify the relationship between overall CLDQ scores from RFHE3001 and EQ‑5D utility. The resulting parameter estimate for CLDQ was applied to the baseline CLDQ score, the increment in the CLDQ score for the rifaximin arm while in remission and the CLDQ scores for the overt states, to give utility estimates for the remission and overt health states. The utility estimate for the initial remission health state was used for the lactulose arm in the October 2013 analysis, and a utility increment of 0.106 was applied to this estimate to give the utility associated with remission in the rifaximin arm. Utility values for the overt states were assumed to be the same for the rifaximin and lactulose arms. The utility values used in the model were designated academic‑in‑confidence by the company, and therefore they cannot be reported here.

In its December 2013 submission the company provided additional analyses exploring alternative utilities. Although it considered the utility values in its October 2013 analysis to be the preferred approach, in response to a request from the Committee the company presented a scenario analysis in which utility values were obtained by directly mapping the SF‑12 subset of SF‑36 scores from RFHE3001 to EQ‑5D utility values (that is, without the additional CLDQ steps included in the October 2013 analyses described in section 3.26). The company used a mapping method developed by Gray et al. (2006) with multiple imputation to account for missing responses. It presented unadjusted results based on all observations and observations from the remission state only, and it also presented baseline‑adjusted results. The company reported that in the unadjusted analysis utility scores for the remission state appeared to increase after baseline, and that people who had an overt episode had lower utility than people in remission. The baseline‑adjusted results suggested that utility increased in people treated with rifaximin compared with placebo during the study.

In its December 2013 analysis, the company presented a scenario in which people who had rifaximin had a utility increment compared with those in the lactulose arm, based on the utility increment seen in the baseline‑adjusted results. For this scenario, the utility scores in the overt states were derived from the study of the general population presented in the company's original model, using the standard gamble methodology. The estimated utility during the remission states for people treated with lactulose was 0.568. Episodes of overt hepatic encephalopathy were associated with a decrease in utility of 0.286 compared with remission. The utility increment associated with rifaximin treatment (compared with lactulose) was 0.032, and this was applied in all 4 health states.

In its November 2014 analysis, the company stated that clinical opinion suggests rifaximin is associated with an observable improvement in quality of life. The company considered that this shows rifaximin gives an improvement at least as large as the minimum important difference (MID) for utility scores based on EQ‑5D. It therefore presented evidence on the MID for EQ‑5D. It noted that Walters and Brazier (2005) reported a pooled estimate for the MID for EQ‑5D of 0.074 (based on estimates ranging from −0.011 to 0.139) and that Coretti et al. (2014) reported estimates ranging from 0.03 to 0.52. The company presented a scenario analysis in which the utility increment associated with rifaximin in the remission states was equal to the average MID for EQ‑5D estimated by Walters and Brazier (that is, 0.074).

The average monthly cost of rifaximin (£281.80) used in the model was based on the recommended dosing schedule of 550 mg twice daily at a unit price of £259.23 per 56‑tablet pack. In the December 2013 model, average monthly lactulose costs (£8.15 and £9.09) for the rifaximin and lactulose groups respectively were obtained from the BNF 63, based on the mean dose of concomitant lactulose in RFHE3001 (3.14 cups for the rifaximin group and 3.51 cups for the placebo group). No administration costs were included in the model because rifaximin and lactulose are taken orally; therefore, the total drug costs per month in the model were £289.95 for the rifaximin arm compared with £9.09 for the lactulose arm. The company estimated a total monthly cost of £58.76 associated with the remission states, reflecting outpatient visits every 3 months. It estimated costs associated with the overt states using the assumptions that a proportion of episodes led to hospital admission, and that if admitted to hospital people would stay for a given period. In the October 2013 and December 2013 submission, the cost associated with the overt states was £1,040.77, reflecting the assumption that 53% of people were admitted to hospital and stayed for 5 days (see section 3.24); in the November 2014 submission, the cost was adjusted using the revised hospital admission assumptions (see section 3.25). These costs were applied to both the rifaximin and lactulose arms. No costs were included in the model for adverse events because there were no statistically significant differences between the adverse events reported with rifaximin and placebo in RFHE3001 and there was limited evidence available for disutilities associated with the adverse events. This approach was validated by the company's clinical experts.

The company's original base‑case analysis showed that rifaximin was associated with an incremental cost‑effectiveness ratio (ICER) of £23,186 per quality‑adjusted life year (QALY) gained. In the October 2013 analysis, the base‑case analysis showed that rifaximin was associated with an ICER of £20,799 per QALY gained. The company stated that the mortality predicted by the October 2013 model at 6 months (rifaximin: 6.68%; lactulose: 10.28%) reflected the trial data and CPRD data better than did the predictions from the original model. In the October 2013 model, the mortality increased to 47.66% and 55.53% respectively after 5 years. In a scenario analysis exploring the use of alternative distributions for extrapolating the time to first and subsequent hepatic encephalopathy episodes, it was noted that varying the distributions for the first episode produced a small range of ICERs (£18,909 to £19,687 per QALY gained), whereas varying the distributions for subsequent episodes produced a wider range of ICERs (£13,779 to £21,380 per QALY gained).

In the company's December 2013 base‑case analysis, which comprised the October 2013 base case with a lifetime time horizon, rifaximin was associated with an ICER of £17,834 per QALY gained, compared with lactulose. In the scenario in which utility values were estimated by directly mapping SF‑36 results from RFHE3001 to EQ‑5D, the ICER was £29,076 per QALY gained (lifetime time horizon). In the scenarios exploring mortality (which used a lifetime time horizon and the utility values from the October 2013 analysis, that is, developed from both SF‑36 and CLDQ data; see section 3.23), the ICERs ranged from £21,797 per QALY gained (mortality in overt states equal to preceding remission state) to £31,916 per QALY gained (mortality equal in all health states).

The company presented probabilistic and deterministic sensitivity analyses for its December 2013 analysis. The deterministic sensitivity analysis showed that the results were most sensitive to the incremental utility benefit associated with rifaximin. In the December 2013 base case (lifetime time horizon), the probabilistic ICER for rifaximin compared with lactulose was £18,144 per QALY gained. The company stated that the probability that rifaximin was cost effective was 95.5% if the maximum acceptable ICER were £30,000 per QALY gained. For the scenario analyses, the probabilistic ICERs were similar to the deterministic ICERs, differing by £300 to £600 per QALY gained.

In the company's November 2014 submission, it presented scenario analyses developed from the December 2013 base‑case analysis. In the scenario in which the utility increment associated with rifaximin was matched to the MID for EQ‑5D (that is, 0.074; see section 3.29), rifaximin was associated with an ICER of £21,331 per QALY gained, compared with lactulose. When the hospital admission rate and length of stay were adjusted based on the company's meta‑analysis (mean difference, all 4 centres), the ICERs for rifaximin compared with lactulose were £7,205, £8,630 and £11,654 per QALY gained (assuming utility increments with rifaximin of 0.106, 0.074 and 0.032, respectively). In the corresponding scenarios using the meta‑analysis results when the Bolton data were excluded, the ICERs were £10,063, £12,053 and £16,276 per QALY gained for rifaximin compared with lactulose (utility increments of 0.106, 0.074 and 0.032 respectively).

The ERG stated that it was likely that the company's systematic review, updated after clarification, contained all the relevant studies. The ERG identified a 6‑month trial of rifaximin compared with neomycin reported by Miglio et al. (1997), which was not presented in the company's submission. In its clarification response, the company stated that this trial had been identified but excluded because it was not considered appropriate for this appraisal in terms of the population, treatment regimens and outcomes included. The ERG acknowledged the company's justification for excluding the study. However, it remained concerned about excluding neomycin from the analysis because clinical expert opinion indicated that it works in the same way as rifaximin and is sometimes used in clinical practice, especially for people not having a liver transplant, even though it is not as well tolerated as rifaximin.

The ERG stated that the evidence submitted by the company generally reflected the decision problem adequately even though the population and comparator differed from those specified in the final scope. Whereas the scope referred to adults who have had episodes of hepatic encephalopathy, the company's submission only considered adults with chronic liver disease, excluding hepatic encephalopathy caused by acute liver disease. People with more severe liver disease (MELD score of 25 or more) were also excluded from the analysis. However, the ERG noted that the company stated that the results would apply to this population as well. The ERG was concerned about the validity of this assumption, given that the treatment effect of rifaximin compared with placebo was not statistically significant in the subgroup with MELD scores of 19 to 24 (the more severe MELD score category in the trial), although the company emphasised the small numbers in this group. In addition, it noted that the study by Hassest et al. (2001), which was provided as evidence for the effectiveness of rifaximin in people with MELD scores of 20 or more, was a poor‑quality descriptive study. The final scope referred to a comparison of rifaximin with lactulose, neomycin or neomycin plus lactulose, but the ERG noted that the analysis presented by the company was based on rifaximin plus concomitant lactulose compared with placebo plus lactulose, in line with the pivotal clinical trial and UK clinical practice. The ERG stated that their clinical expert agreed that UK current practice involved using concomitant lactulose.

The ERG indicated that RFHE3001 was a high‑quality study. It noted that the treatment groups were similar in terms of baseline characteristics, although there were more men in the placebo group (67%) than in the rifaximin group (54%). The ERG stated that the outcomes assessed were appropriate and in line with those specified in the scope. Overall, the ERG considered the clinical‑effectiveness data and the statistical approaches in the company's submission to be of good validity. However, it recognised that RFHE3002 was unpublished at the time of the original submission, and although assessments of Conn scores and asterixis scores were conducted to monitor hepatic encephalopathy status, only summary statistics were presented for these analyses.

The ERG reviewed the evidence from audits of clinical practice and the company's meta‑analysis of this evidence, presented in the company's November 2014 submission. It noted that the audits had some limitations, such as small sample sizes, wide confidence intervals and the potential influence of time and context (given they were observational studies). Moreover, the characteristics of the people in the audits were not known, and the ERG was uncertain how well the populations would match the people who would be treated in clinical practice, if rifaximin were recommended for widespread use in the NHS. The ERG was also uncertain whether the hospital admissions were related to hepatic encephalopathy or to any cause. While reviewing the company's meta‑analysis, the ERG identified some limitations in the analysis, including discrepancies within the data, the assumption of a constant event rate, and the use of a random‑effects analysis. It considered that a fixed‑effects analysis might be more conservative given the small sample sizes. The ERG agreed with the company's decision to exclude multicentre data from Patel et al. (2014) and to carry out a sensitivity analysis in which the Bolton audit was excluded. It considered that the analysis in which the Bolton audit was excluded provides a better estimate for the mean difference in bed days between treatment with and without rifaximin. The ERG considered the results from the multicentre study to be published by Orr et al. in 2015; the results of this study were provided as academic‑in‑confidence and so cannot be reported here. The ERG highlighted that this study provided evidence from a larger population than the other audits. However, it also noted that there was limited information available on how the multicentre data had been collated and so the validity of the analysis was unknown.

In its review of the additional evidence on the long‑term effectiveness of rifaximin and the effect on mortality (November 2014 submission), the ERG noted that the company did not describe any details of a systematic search to identify this evidence. It highlighted that some of the studies of long‑term effectiveness identified by the company were not relevant because they did not include people with hepatic encephalopathy. The ERG considered that the study by Irimia et al. (2012) provided relevant evidence, although it noted the small population size in this study. Overall, the ERG considered that the company had provided limited additional evidence on the long‑term effectiveness of rifaximin. Similarly, for the evidence on mortality, the ERG noted that 2 of the 4 identified studies were not appropriate sources for decision‑making because they did not include people with hepatic encephalopathy. The ERG stated that the other 2 studies – Sharma et al. (2013) and Neff et al. (2012) – provided evidence to support a mortality benefit associated with rifaximin for up to 3 years.

The ERG was satisfied with the company's modelling approach and agreed that the health states in the model appropriately captured disease progression over time. Clinical opinion obtained by the ERG confirmed that the company's assumption of concomitant lactulose use in both arms of the model was appropriate, but emphasised that the effectiveness results were based on 91.3% of people using rifaximin with concomitant lactulose and only 8.7% taking rifaximin alone. The ERG stated that a 5‑year time horizon was appropriate in the company's original model to capture the relevant costs and benefits, because at that time there was no robust evidence of a mortality effect; however, it stated that in the October 2013 analysis a lifetime time horizon would have been appropriate given that approximately 52% of people in the rifaximin arm and 44% of people in the lactulose arm were alive after 5 years. In an exploratory analysis carried out in response to the company's October 2013 analysis (after the first consultation), the ERG increased the time horizon to lifetime (40 years, when all people in the model had died) and this resulted in an ICER of £22,069 per QALY gained. In its critique of the company's analysis including the lifetime time horizon presented in the December 2013 analysis (after the second consultation), the ERG stated that the company's approach was appropriate but corrected a small mistake in the company's model that reduced the ICER to £17,681 per QALY gained.

The ERG reviewed the company's October 2013 analysis and stated that the company used different time points to censor survival times when it modelled breakthrough overt hepatic encephalopathy episodes; data censoring took place at day 168 in the company's original analysis and at day 170 in the October 2013 analysis. The ERG highlighted that censoring people at day 170 resulted in different regression coefficients when extrapolating RFHE3001 data, leading to different transition probabilities. It noted, however, that censoring at day 168 rather than day 170 in the updated analysis would only increase the base‑case ICER to £21,329 per QALY gained. The ERG stated that the company should have carried out some validity tests to assess the assumption of a proportional treatment effect between rifaximin and lactulose when fitting the parametric distributions to the trial data; the company presented a test of the proportional hazards assumption in its response to the second consultation, stating that there was no evidence of a violation of the proportional hazards assumption for treatment effect. The ERG noted that assuming a proportional treatment effect implied that people will continue to have the drug until death, and this was confirmed by the ERG's clinical experts who indicated that people were generally kept on rifaximin until death or liver transplant.

The ERG considered that the company's use of data only from the newly enrolled people in RFHE3002 to avoid potential enrichment bias when modelling subsequent overt episodes (see section 3.18) was reasonable. The ERG also stated that clinical opinion indicated that the company's approach of applying the treatment effect from RFHE3001 to model subsequent overt episodes (see section 3.18) was reasonable. The ERG stated, based on clinical expert opinion, that the company's original assumption of a constant probability of subsequent episodes over time did not reflect reality. The ERG stated that the company's approach of modelling subsequent episodes as dependent on previous episodes was a significant improvement from the original analysis. However, it was concerned that combining all subsequent episodes into 1 health state does not take into account the number of episodes and does not fully address the issues around the complexity of disease progression. The ERG noted from the scenarios presented in the company's October 2013 analysis (see section 3.31) that the choice of distribution for extrapolating hepatic encephalopathy episodes had less impact on the resulting ICERs in the October 2013 analysis than it did in the original model. However, it stated that the company should have taken further steps to assess the goodness of fit of the distributions used and that the gamma distribution should have been included for completeness; the company presented an assessment of goodness of fit in its response to the second consultation.

The ERG stated that the company did not justify using different approaches to modelling mortality, that is, using survival analysis to model mortality in the remission states and using simple proportions to estimate the probability of death in the overt states. It also stated that the company did not adequately explain how it calculated the probability of death in the overt states. It noted that the Kaplan–Meier curves were not provided for visual comparison of the different distributions used to model mortality in the remission states and stated that the company should have taken further steps to assess the goodness of fit of the distributions used to extrapolate mortality; the company provided the Kaplan–Meier curve in its response to the second consultation. The ERG also highlighted the inconsistency in the use of RFHE3002 data in the model, in which data for all people in RFHE3002 were used to model mortality, whereas only data for newly enrolled people in RFHE3002 were included in the modelling of the time to subsequent episodes.

The ERG compared mortality from RFHE3001, RFHE3002, the CPRD data and the October 2013 model. It noted that the model overestimated mortality for the lactulose arm at 6 months (10%) compared with RFHE3001 (7%), whereas mortality estimated for the rifaximin arm at 6 months was similar to that in RFHE3001 (7%). It also noted that 2‑year mortality for the rifaximin arm of the model was similar to that seen in RFHE3002 and that 5‑year mortality estimated from the model showed an incremental survival benefit of 8% for rifaximin compared with lactulose. The ERG stated that clinical opinion indicated that, although the survival benefit predicted was reasonable, the overall mortality seemed low for people in both arms of the model, and also that a higher number of liver transplants would be expected among people with hepatic encephalopathy (only 1 liver transplant was reported in RFHE3001; the company considered that this was not unexpected given the exclusion criteria in the study). The ERG explored the impact of mortality on the model results, by excluding the survival benefit from the remission states only, the overt states only, and all 4 states together. This resulted in ICERs of £22,700, £26,120 and £30,200 per QALY gained respectively, which shows that mortality is still a significant driver of the results.

In its critique of the company's exploration of mortality in the December 2013 analysis, the ERG stated that it was generally satisfied with the company's approach of adopting a lifetime time horizon to incorporate a diminishing mortality benefit for rifaximin. The ERG stated that it was reasonable and clinically plausible that rifaximin affects mortality. However, it noted concerns about the clinical plausibility of the different mortality rates in the 4 health states. The ERG considered it plausible that people have a higher mortality risk during overt episodes. However, although it is also plausible that people have higher mortality risk immediately after overt episodes (that is, at the start of the subsequent remission health state), the ERG queried whether the company's approach of applying a higher mortality risk throughout the subsequent remission state was appropriate. Also, the ERG noted that the company had assumed a lower mortality risk in subsequent overt episodes compared with the initial overt episode, but considered that it would be more plausible that people have an increasing mortality risk with increasing numbers of overt episodes. The ERG highlighted a small error in the company's economic model, and correcting this error decreased the ICERs in the company's scenario analyses by approximately £200 to £300.

The ERG was satisfied with the company's use of the 6‑month probabilities to model hepatic encephalopathy‑related hospital admissions rather than the monthly probabilities used in the original analysis. However, it stated that clinical opinion did not support the use of the same (aggregated) probability for the rifaximin and lactulose arms, even though the company showed that the differences between the individual probabilities were not statistically significant. The ERG stated that the more conservative approach of using the individual probabilities should have been taken and, when this was explored in a scenario analysis, the ICER increased slightly to £21,389 per QALY gained. The ERG noted that, although more people had an episode of hepatic encephalopathy in the lactulose arm than in the rifaximin arm, the rate of hospital admission for those who had an episode was higher in the rifaximin arm. Therefore, the overall effect of the costs of hospital admission on the ICER was neutralised to some extent, indicating that this was not a key driver of the cost‑effectiveness results.

The ERG noted that in the November 2014 submission, the company incorporated hospital admission data from its meta‑analysis of clinical audit data. The ERG highlighted that the length of stay per hospital admission in this analysis was substantially longer than what has been reported in other published studies. The ERG carried out an exploratory analysis using the results from the multicentre study to be published by Orr et al. in 2015. In this analysis, using a lifetime time horizon and assuming a utility increment with rifaximin of 0.032, rifaximin was associated with an ICER of £19,791 per QALY gained; when the utility increment with rifaximin was assumed to be 0.106, rifaximin was associated with an ICER of £12,139 per QALY gained.

The ERG considered the modelling of utility values in its critiques of the company's October 2013 and December 2013 analyses. In response to the company's October 2013 analysis, the ERG stated that it was unclear why the company used the EQ‑5D utility values estimated from the condition‑specific CLDQ in the model given that it was possible to incorporate values directly from mapping the SF‑36 utilities onto the EQ‑5D. It was concerned that the estimated quality‑of‑life increment with rifaximin compared with lactulose in the remission states was in contrast to the company's original analysis in which the impact on quality of life was linked to movement between the overt and remission health states only. The company considered the October 2013 approach more appropriate, based on its re‑examination of quality‑of‑life data from RFHE3001 after the first consultation. The ERG stated that there was uncertainty in the validity of the utility increment estimated for the rifaximin arm given that the company appeared to have measured the value in centimetres directly from the study by Sanyal et al. (2011) rather than using the actual published value of difference in least square means. It also highlighted that applying the increment to the remission states only further increased the uncertainty because the incremental value reported in the study represents the CLDQ data collected for the whole duration of treatment in the trial. The ERG stated that the utility values used in the model were a key driver of the results, given that excluding the utility increment associated with rifaximin in the remission states increased the ICER to £59,421 per QALY gained.

The ERG noted that, in the December 2013 submission, the company presented a scenario analysis in which utility values were mapped directly from SF‑36 data. The ERG was generally satisfied with the methods used to prepare the unadjusted utilities, although it noted a potential bias in the baseline populations (that is, people who had an overt episode during the study had lower quality of life at baseline than those who did not) and highlighted that many of the apparent differences seen may not have been statistically significant. Moreover, the ERG highlighted a number of concerns about the baseline‑adjusted utilities. The ERG was not entirely clear how the adjusted analyses had been conducted, particularly regarding any differentiation between overt and remission states, and the statistical significance of the reported results was unknown. It also stated that it considered the face validity of the adjusted results to be somewhat questionable. The ERG highlighted specific concerns about each of the 3 key utility values used in the company's December 2013 analysis:

• The origin of the utility value for the remission states in the lactulose arm was unclear.

• The estimate for the difference in utility between the remission and overt states was taken from a study of the general public, rather than from RFHE3001.

• A utility improvement (increment) was applied to the rifaximin arm relative to the lactulose arm, and the ERG was concerned that this was applied across all health states (not just the remission states); it was also unclear on the origin and statistical validity of the estimated increment.
  
  The ERG presented exploratory analyses to assess the effect of altering the utility assumptions. It noted that removing the utility increment associated with rifaximin in the remission states had the largest effect, increasing the ICER for rifaximin compared with lactulose by £10,379 per QALY gained; removing the utility increment associated with rifaximin from the overt states increased the ICER for rifaximin compared with lactulose by £61 per QALY gained. The ERG also examined the sensitivity of the model to the utility decrement associated with overt states compared with the remission states (that is, how much quality of life deteriorates during an overt episode). It found that reducing this decrement by approximately 40% to 70% increased the ICER by £100 to £200 per QALY gained. The ERG emphasised that the utility values used in the economic model, and in particular the utility increment associated with rifaximin in the remission states, were key drivers of the cost‑effectiveness results.

The ERG reviewed the evidence for the minimum important difference (MID) in utility scores and considered the company's analysis in which the rifaximin utility increment was the same as the MID. It noted that studies included in the analysis of MID by Walters and Brazier (2005) did not include people with liver disease or hepatic encephalopathy, and so the reported average MID might not apply for people with this condition. The ERG also emphasised the wide range of values for the MID reported by this study (see section 3.29). It noted that an increment of 0.032 falls within this range, and therefore stated that it could not conclude which utility increment would be the most appropriate based on this evidence.

The ERG was satisfied with the calculation of costs in the model. It noted differences in the cost of outpatient attendance for hepatic encephalopathy events in people who did not need hospital admission between the original and October 2013 analyses, but the company explained that these resulted from an error in the original submission that was corrected in the October 2013 analysis. The ERG noted that the company did not include costs for adverse events because they were comparable between the rifaximin and placebo groups in the pivotal trial, but stated that including relative risks (or risk difference) and 95% confidence intervals for each adverse event would have strengthened this justification. However, the ERG also stated that including adverse events in the costs and QALY calculations would not have a significant impact on the ICER.

Full details of all the evidence is in the committee papers.