Denosumab, raloxifene, romosozumab and teriparatide to prevent osteoporotic fragility fractures: a systematic review and economic evaluation

Other resources

In addition to database searches, the reference lists of relevant studies were checked. Identified systematic reviews were checked to identify any additional trials meeting the inclusion criteria.

Bisphosphonate studies were identified from the assessment report34 used to inform the development of NICE TA464. 9 As the searches for this TA were last updated in September 2014, more recent studies were sought from the database searches.

When data from included trials were missing, the company submissions were checked. Any academic or commercial-in-confidence data taken from a company submission were underlined and highlighted in the assessment report.

Inclusion criteria

Exclusion criteria

• Studies with patients with normal or unspecified BMD.

• Studies with patients with other indications for the same drugs. Cancer populations at risk of osteoporosis that are covered by NICE guideline (NG) 10136 and NICE CG175. 37

• Studies in which interventions were administered not in accordance with licensed indications.

• Studies in which interventions were co-administered with any other therapy with the potential to augment bone, unless concomitant treatments are specified in the SmPC.

• Studies that were considered methodologically unsound in terms of study design or the method used to assess outcomes.

• Reports published as abstracts or conference presentations only, for which insufficient details are reported to allow an assessment of study quality or results.

• Studies that were published in languages other than English.

• Studies based on animal models, and pre-clinical and biological studies.

• Narrative reviews, editorials, opinions.

Results of the risk-of-bias assessment

Non-bisphosphonates versus placebo

A summary of the Cochrane risk-of-bias assessment across the placebo-controlled non-bisphosphonate studies is presented in Figure 2.

FIGURE 2.

Cochrane risk-of-bias summary across placebo-controlled non-bisphosphonate studies. ?, Unclear-risk of bias; +, low-risk of bias; –, high-risk of bias; blank cells, not a study outcome. ACTIVE, Abaloparatide Comparator Trial In Vertebral Endpoints; ADAMO, A multicenter, randomized, double-blind, placebo-controlled study to compare the efficacy and safety DenosumAb versus placebo in Males with Osteoporosis; BRIDGE, phase III randomized placeBo-contRolled double-blind study evaluatIng the efficacy and safety of Romosozumab in treatinG mEn with osteoporosis; DIRECT Denosumab fracture Intervention RandomizEd placebo Controlled Trial; FRAME, Fracture Study in Postmenopausal Women with Osteoporosis; FREEDOM, Fracture REduction Evaluation of Denosumab in Osteoporosis every 6 Months; MORE, Multiple Outcomes for Raloxifene Evaluation.

Head-to-head non-bisphosphonates

The summary of the Cochrane risk-of-bias assessment across the head-to-head non-bisphosphonate studies is presented in Figure 3.

FIGURE 3.

Cochrane risk-of-bias summary across head-to-head non-bisphosphonate studies. ?, Unclear-risk of bias; +, low-risk of bias; –, high-risk of bias; blank cells, not a study outcome. DATA, Denosumab and Teriparatide Administration; DIRECT, Denosumab fracture Intervention RandomizEd placebo Controlled Trial; EUROFORS, European Study of Forsteo; EVA, EVista Alendronate comparison; STRUCTURE, STudy evaluating effect of RomosozUmab Compared with Teriparatide in postmenopaUsal women with osteoporosis at high risk for fracture pReviously treated with bisphosphonatE therapy.

Of the four head-to-head studies,64,66–68 three reported the method for the random sequence generation,64,66,67 and three reported that allocation was concealed. 66–68

All four studies were reported as open label and were considered to have a high risk of bias for blinding of participants and personnel. 64,66–68

All four studies reported fractures as an outcome;64,66–68 of these, two studies reported that fracture assessment was not blinded to treatment allocation. 66,67 All four studies assessed BMD and three were considered to have a low risk of bias for the blinding of BMD assessments. 64,66,67

Two66,67 of the three studies assessing fracture were considered to have a low risk of attrition bias (< 10% of participants withdrew/were not included in the analysis). 64,66,67 All four studies reported BMD outcomes;64,66–68 one of these was considered to have a high risk of attrition bias (≥ 10% of participants in both treatment groups did not complete the study) for this domain. 68 All other studies were considered to have a low risk of bias.

Three studies reporting the location of a protocol were judged to be at low risk of selective reporting bias. 64,67,68

Non-bisphosphonates versus bisphosphonates

The summary of the Cochrane risk-of-bias assessment across the non-bisphosphonate versus bisphosphonate studies is presented in Figure 4.

FIGURE 4.

Cochrane risk-of-bias summary across non-bisphosphonate vs. bisphosphonate studies. ?, Unclear-risk of bias; +, low-risk of bias; –, high-risk of bias; blank cells, not a study outcome. ARCH, Active-controlled fracture study in postmenopausal women with osteoporosis at high risk; DAPS, Denosumab Adherence Preference Satisfaction; DECIDE, Determining Efficacy: Comparison of Initiating Denosumab versus alEndronate; EFFECT, EFficacy of Fosamax versus Evista Comparison Trial; EVA, EVista Alendronate comparison; FACT, Forteo Alendronate Comparator Trial; STAND, Study of Transitioning from Alendronate to Denosumab; VERO, VERtebral fracture treatment comparisons in Osteoporotic women.

Vertebral fractures

Results for vertebral fractures reported in the included studies are presented in Appendix 5, Table 17, for the non-bisphosphonate treatments compared with placebo, non-bisphosphonate treatments compared head to head, and non-bisphosphonate treatments compared with bisphosphonates. Fracture data used in the NMAs are shown in Appendix 9.

Non-vertebral fractures

Non-vertebral fracture outcomes were reported in 28 RCTs and are shown in Appendix 5, Table 18. Hip, wrist and proximal humerus fracture outcomes were reported separately in 22 RCTS; these are shown in Appendix 5, Table 19. These fractures are also counted among the non-vertebral fracture total. Results of the NMAs for these outcomes are shown in Results of the network meta-analysis. Fracture data used in the NMAs are shown in Appendix 9.

Femoral neck bone mineral density

Results for femoral neck BMD reported by the included studies are presented in Appendix 5, Table 20, for the non-bisphosphonate treatments compared with placebo, non-bisphosphonate treatments compared head to head, and non-bisphosphonate treatments compared with bisphosphonates.

Mortality

Mortality across the included studies is presented in Appendix 5, Table 21, for the non-bisphosphonate treatments compared with placebo, non-bisphosphonate treatments compared head to head and non-bisphosphonate treatments compared with bisphosphonates. None of the included studies reported on mortality following hip fracture, mortality following vertebral fracture or mortality following any other type of fracture.

Adverse events and serious adverse events

Adverse events and serious adverse events (SAEs) reported across the included studies are presented in Appendix 5, Table 22, for the non-bisphosphonate treatments compared with placebo, non-bisphosphonate treatments compared head to head and non-bisphosphonate treatments compared with bisphosphonates.

Health-related quality of life

Five studies86,90,101,116,117 published results of reported HRQoL, measured by a validated assessment tool (see Appendix 6).

Vertebral fractures

Vertebral fracture data were available from 46 RCTs; 45 of these compared two treatments and one was a three-arm study. 79 Nineteen of these studies were included in TA46434 (including one study79 for which an additional non-bisphosphonate treatment arm was added for the current review). Two further bisphosphonate studies129,130 not already in TA464,34 and 24 non-bisphosphonate studies were included from the current review. A total of 11 interventions were assessed, including five non-bisphosphonate treatments.

The effects of each treatment relative to placebo are presented in Figure 7, and pairwise comparisons between treatments are provided in Appendix 12, Table 34. All treatments were associated with statistically significant beneficial treatment effects relative to placebo. TPTD was associated with the greatest effect (HR 0.23, 95% CrI 0.16 to 0.32), with the highest probability of being the best-ranking treatment (PB) (0.38), and was statistically significantly more effective than all active treatments apart from DEN, ROMO and ROMO/ALN (see Appendix 12, Table 34). The HR for a randomly chosen study for a new bisphosphonate is 0.47 (95% PrI 0.19 to 1.16), with the reported prediction interval allowing for both between-study and between-treatment heterogeneity.

In the network, both direct and indirect comparisons were available for 12 treatment pairs. None of the comparisons showed significant evidence of inconsistency (see Appendix 13, Table 40).

Four sensitivity analyses were conducted for the main vertebral fracture network. Treatment effects are provided in Appendix 11, Figure 14, and a summary of model fit and heterogeneity is shown in Appendix 11, Table 33.

Sensitivity analysis 1 included data reported at 12 months only. Data were available from 29 RCTs, which assessed a total of 10 interventions, including four non-bisphosphonate treatments. The main difference in the results is that RIS has a more beneficial treatment effect in the 12-month sensitivity (HR 0.44, 95% CrI 0.32 to 0.60) than in the primary analysis (HR 0.52, 95% CrI 0.42 to 0.65). In both analyses, RIS has zero probability of being the best-ranking treatment. It was concluded that the results are generally consistent with those of the primary analysis, which included the longest duration of follow-up for each study, and therefore supports the use of a constant HR.

Sensitivity analysis 2 included outcomes assessed by clinical methods only. Data were available from 20 RCTs, which assessed a total of 11 interventions, including five non-bisphosphonate treatments. It was concluded that the results are generally consistent with those of the primary analysis, which includes both clinical and morphometric/radiographic outcomes. This supports the assumption that the treatment effect is not highly influenced by assessment method.

Sensitivity analysis 3 excluded studies for which there was a risk of bias in the reported outcomes. Four studies86,92,93,95 were excluded owing to blinding issues, two studies62,80 were terminated early and, for 10 studies,42,43,50,129,131–136 the number of events or analysis sample size was estimated from other information. Data were available from 30 RCTs, which assessed a total of 10 interventions, including five non-bisphosphonate treatments. It was concluded that the results are consistent with those of the primary analysis, which includes all studies, and therefore supports the use of the full network of 46 studies to improve the strength of the network.

Sensitivity analysis 4 excluded studies for which prior treatment with bisphosphonates was permitted. The proportion of individuals receiving prior treatment ranged from 8–73% across the studies. Data were available from 36 RCTs, which assessed a total of 11 interventions, including five non-bisphosphonate treatments. It was concluded that the results were consistent with those of the primary analysis.

Non-vertebral fractures

Non-vertebral fracture data were available from 42 RCTs; 40 of these compared two treatments, and two were three-arm studies. 79,82 Fifteen of these studies were included in TA46434 (including one study79 for which an additional non-bisphosphonate treatment arm was added for the current review), and 27 non-bisphosphonate studies from the current review were included. A total of 11 interventions were assessed, including four non-bisphosphonate treatments.

Pairwise comparisons between treatments are provided in Appendix 12, Table 35. All treatments were associated with beneficial treatment effects relative to placebo, although the results were not statistically significant for all treatments. TPTD was associated with the greatest effect (HR 0.58, 95% CrI 0.45 to 0.76), with the highest PB (0.52), although there was insufficient evidence to differentiate between TPTD and the other active treatments apart from IBN daily, DEN and RLX (see Appendix 12, Table 35). The HR for a randomly chosen study for a new bisphosphonate is 0.78 (95% CrI 0.60 to 1.08), with the reported prediction interval allowing for both between-study and between-treatment heterogeneity.

In the network, both direct and indirect comparisons were available for 14 treatment pairs. None of the comparisons showed significant evidence of inconsistency (see Appendix 12, Table 35).

Hip fractures

Hip fracture data were available from 23 RCTs; 22 of these studies compared two treatments and one was a three-arm study. 79 Eight of these studies were included in TA46434 (including one study79 for which an additional non-bisphosphonate treatment arm was added for the current review), and 15 non-bisphosphonate studies from the current review were included. A total of nine interventions were assessed, including five non-bisphosphonate treatments.

Pairwise comparisons between treatments are provided in Appendix 12, Table 36. All treatments were associated with beneficial treatment effects relative to placebo, although the comparison with placebo was not statistically significant for RLX. TPTD was associated with the greatest effect (HR 0.35, 95% CrI 0.15 to 0.73), with the highest PB (0.50), although there was insufficient evidence to differentiate between TPTD and the other active treatments (see Appendix 12, Table 36). The HR for a randomly chosen study for a new bisphosphonate is 0.64 (95% PrI 0.32 to 1.29), with the reported PrI allowing for both between-study and between-treatment heterogeneity.

In the network, both direct and indirect comparisons were available for 14 treatment pairs. None of the comparisons showed significant evidence of inconsistency (see Appendix 13, Table 42).

Wrist fractures

Wrist fracture data were available from 15 RCTs; 14 of these compared two treatments and one was a three-arm study. 79 Six of these studies were included in TA46434 (including one study79 for which an additional non-bisphosphonate treatment arm was added for the current review), and eight non-bisphosphonate studies from the current review were included. A total of eight interventions were assessed, including four non-bisphosphonate treatments.

Pairwise comparisons between treatments are provided in Appendix 12, Table 37. All treatments were associated with beneficial treatment effects relative to placebo, apart from DEN and RLX. Treatment effects for DEN are based only on one small study with two events in the ALN arm and three events in the DEN arm. 70 Treatment effects for these interventions are therefore highly uncertain.

Romosozumab was associated with the greatest effect (HR 0.12, 95% CrI 0.00 to 1.19), with the highest PB (0.88), although there was insufficient evidence to differentiate between ROMO and the other active treatments (see Appendix 12, Table 37). The HR for a randomly chosen study for a new bisphosphonate is 0.84 (95% PrI 0.29 to 2.50), with the reported PrI allowing for both between-study and between-treatment heterogeneity.

In the network, both direct and indirect comparisons were available for eight treatment pairs. None of the comparisons showed significant evidence of inconsistency (see Appendix 13, Table 43).

Proximal humerus fractures

Proximal humerus fracture data were available from 13 RCTs, each comparing two treatments. Two of these studies were included in TA46434 and 11 non-bisphosphonate studies from the current review were included. A total of eight interventions were assessed, including two bisphosphonate treatments.

Pairwise comparisons between treatments are provided in Appendix 12, Table 38. All treatments were associated with beneficial treatment effects relative to placebo, apart from RLX. Treatment effects for RLX are based on one small study77 only, with zero events in the ALN arm and one event in the RLX arm, and so treatment effects are highly uncertain. Event numbers were generally low in this network and five of the 13 included RCTs had zero counts in one of the treatments arms.

Romosozumab was associated with the greatest effect (HR 0.10, 95% CrI 0.0 to 3.66), with the highest PB (0.77), although the treatment effect was highly uncertain and there was insufficient evidence to differentiate between ROMO and the other active treatments (see Appendix 12, Table 38). Only RIS was associated with a HR that was statistically significant compared with placebo (HR 0.49, 95% CrI 0.23 to 0.96). The HR for a randomly chosen study for a new bisphosphonate is 0.47 (95% CrI 0.18 to 1.15), with the reported PrI allowing for both between-study and between-treatment heterogeneity.

In the network, both direct and indirect comparisons were available for five treatment pairs. None of the comparisons showed significant evidence of inconsistency (see Appendix 12, Table 38).

Heterogeneity in treatment effects between studies, and between bisphosphonates, is summarised in Table 4. The estimates of between-study SD suggest mild (non-vertebral) and moderate (vertebral, hip, wrist, proximal humerus, femoral neck BMD) heterogeneity in treatment effects between RCTs. The estimates of between-treatment SD indicate moderate heterogeneity in effects between treatments for all outcomes (i.e. the effects of the bisphosphonates are relatively similar).

Meta-regressions were conducted to test for different treatment effects separately, according to the mean age of participants in each study and the proportion of female participants. A common meta-regression coefficient was assumed for all treatments. 121 Based on comparison of models with and without a covariate for mean age or mean percentage of females, there was no evidence that treatment effect varied with age or sex. Meta-regression coefficients were not statistically significantly different from zero, and DIC estimates were higher, implying a less favourable model. A summary of the results is provided in Appendix 14, Table 45.

Baseline fracture risk can be used as a proxy for differences in participant characteristics across trials that may be modifiers of treatment effect, and so introduce a potential source of heterogeneity in the NMA. The effect of baseline fracture risk as a potential treatment-effect modifier was explored using the method of Achana et al. ,122 assuming a common meta-regression coefficient for all treatments (as for age and sex), and assuming that the baselines of each study follow a normal distribution with common mean and between-study variance. Based on a comparison of models with and without an adjustment for baseline risk, and inspection of the regression coefficients, there was no evidence that treatment effect varied with baseline risk for any of the fracture outcomes (see Appendix 14, Table 45).

Femoral neck bone mineral density

Femoral neck BMD data were available from 73 RCTs; 69 of these each compared two treatments, one was a four-arm study68 and three were three-arm studies. 72,78,137 Thirty-two of these studies were included in TA464. 34 Three further bisphosphonate studies129,137,138 not already in TA464,34 and 38 non-bisphosphonate studies, were included from the current review. A total of 12 interventions were assessed, including five non-bisphosphonate treatments. The network is shown in Figure 6.

The effects of each treatment relative to placebo are presented in Figure 8. Pairwise comparisons between treatments are provided in Appendix 12, Table 39. All treatments were associated with statistically significant beneficial treatment effects relative to placebo. ROMO/ALN was associated with the greatest treatment effect (MD 6.08, 95% CrI 4.25 to 7.91), with the highest PB (0.96), and was statistically significantly more effective than all active treatments apart from ROMO (see Appendix 12, Table 39). The treatment effect for a randomly chosen study for a new bisphosphonate is 2.34 (95% PrI 1.26 to 3.28), with the reported PrI allowing for both between-study and between-treatment heterogeneity.

FIGURE 8.

Forest plot for percentage change in femoral neck BMD.

To account for differing trial durations, study duration was included as a trial-level covariate. The estimated impact on treatment effect of study duration, assuming a common relationship for each treatment, was 1.09 (95% CrI 0.73 to 1.45), indicating an increase in treatment effect with increasing duration of study, as expected.

As for fracture outcomes, there was no evidence that treatment effect varied with age, sex or baseline response (see Appendix 14, Table 45).

Inclusion/exclusion criteria

Studies were included in the review if they reported full economic evaluations comparing DEN, RLX, ROMO or TPTD with each other, with bisphosphonates or with no treatment. Studies were included if any of the population considered would be eligible for risk assessment as per CG146. 8 For example, studies of postmenopausal women were included whether or not they specified that the women had risk factors, as those aged > 65 years would be eligible for risk assessment under CG1468 even without risk factors being present. 8 Studies that did not assess outcomes using QALYs or that did not report the incremental cost per QALY of alternative treatment strategies were excluded. Studies that did not assess the cost-effectiveness in a UK setting were excluded, to ensure consistency with the NICE reference case. 139 Studies that assessed the cost-effectiveness of treatment at non-licensed doses were also excluded, as were studies that used treatments for other indications such as the treatment of Paget’s disease or metastatic bone disease. Studies published prior to 2006 were included when identified in existing NICE appraisals or published systematic reviews, as described previously. Studies were included only if they were reported as full papers; conference abstracts were excluded from the review as they present insufficient detail to allow for a rigorous assessment of study quality. Studies not reported in the English language were also excluded. De novo economic analyses reported in the consultee submissions were included if they met the inclusion criteria of the review.

Review methods

The results of the economic searches described above were combined with the results of the searches conducted for the HRQoL review (see Appendix 11) and a combined sift was conducted to pick up any cross-relevant papers. The combined database was sifted by title and abstract by one reviewer. The full papers of studies that potentially met the inclusion criteria were retrieved for further inspection by the same reviewer. Studies included in the systematic review were examined to determine whether or not they met the NICE reference case. 139 We stated in our protocol that we would critically appraise the included cost-effectiveness analyses using the checklist published by Philips et al. ,140 but this was not done owing to time constraints.

Quantity of evidence identified

The database search identified 3853 citations across the combined cost-effectiveness and HRQoL searches. Three additional articles141–143 were identified from the reference list of published reviews. None of the consultee submissions identified any published analyses not already picked up by the systematic search, but two reported de novo economic analyses, which were included, giving a total of 3858 citations. Of these, 3837 were excluded at the title and abstract stage and a further 11 were excluded at the full-paper stage; the most common reasons for exclusion were that they were non-UK studies or conference abstracts with limited data presented. Appendix 15 provides the reasons for exclusion for those papers that were included during the title and abstract sift, but were later excluded after considering the full paper.

A total of 10 articles20,34,100,141–147 were included; however, one paper, Kanis et al. ,142 reported a previous version of the model reported by Stevenson et al. ,143 and was therefore not separately extracted, and two articles provided the Evidence Review Group’s summary of the company submission for TA204. 145,147 Therefore, the review included eight unique cost-effectiveness analyses. Additional documents related to TA20410 were downloaded from the NICE website to allow a full examination of this model [note that this model is referred to as ‘Waugh et al. 147’ to avoid confusion with the Amgen submission for the current multiple technology appraisal (MTA)]. The model described in the Amgen submission for the current MTA100 was an adaptation of the model described in the company submission for TA204,145,147 but these were separately extracted owing to differences between the decision problems.

Although the assessment report for TA464 by Davis et al. 34 did not strictly meet the inclusion criteria for this review, as it did not include any non-bisphosphonate interventions, it has been included as it was stated in the protocol for this MTA that, to ensure consistency across related appraisals, the economic analysis conducted to inform TA464 was intended to be used as the starting point for any cost-effectiveness analysis conducted by the assessment group (AG). Therefore, it was necessary to compare this model with the relevant published analyses to identify any significant areas of difference.

Study characteristics

The characteristics of the included studies are summarised in Table 5. Here we describe the key differences between the models in terms of their population, structure and assumptions.

Consistency with the National Institute for Health and Care Excellence reference case

All of the included studies measured direct health effects for patients, and none included any benefits for carers. All of the studies reported using published estimates of utility following fracture from studies that had measured utility using the EQ-5D using the UK general population valuation set. There was some inconsistency in the approach taken to estimating utility following nursing home admission, with one study147 reporting no additional disutility, one study143 reporting using a value based on an expert panel, one study34 reporting a value based on the EQ-5D and several studies not reporting the approach taken to estimating utility values for nursing home admission. 20,100,141,144,146

One study141 based its effectiveness estimate on a single RCT and reported a comparison only between the interventions included in the RCT (RLX vs. no treatment). The other studies all sourced their effectiveness estimates from a systematic review and meta-analysis, although only the three most recent models20,34,100 used NMA to estimate the relative treatment between active comparators. One study147 used the method published by Bucher et al. 148 to conduct an indirect comparison. Two studies143,146 present incremental analyses that appear to be based on naive indirect comparisons based on equivalent outcomes for patients receiving placebo. The remaining study144 provided comparisons only against no treatment.

Five studies explicitly reported using an NHS and Personal Social Services (PSS) perspective. 20,34,100,143,147 Three studies reported taking a health-care perspective,141,144,146 but two of these144,146 also included nursing home costs, which are likely to fall under PSS rather than NHS in a UK context, although some may also fall under societal costs if families pay privately for nursing home care. Discounting consistent with the current NICE reference case (3.5% for both costs and QALYs)139 was applied in all but two studies,141,143 in which discounting was applied at rates consistent with previous NICE methods guidance (6% for costs and 1.5% for QALYs). The time horizon is not explicitly stated for the 2005 publication by Kanis et al. ,141 but, otherwise, all of the included economic evaluations incorporated a lifetime horizon, although, in the analysis by Stevenson et al. ,143 the Markov model was used for the first 10 years and then additional calculations were used to estimate QALYs gained over the remaining lifetime.

Quality and applicability of studies

The only analyses considered to be broadly consistent with the NICE reference case were the models described in the submissions by UCB S.A. 20 and Amgen Inc. 100 and the analysis by Davis et al. ,34 which informed TA464. 9 None of the other models provided an incremental analysis informed by a systematic review and NMA, which is a significant deviation from the NICE reference case, and may be a potential source of bias. However, it is noted that the analysis by Davis et al. 34 was not relevant to the decision problem; it was included purely to allow comparisons to be made between the published models and the model we intended to adapt for this appraisal.

Study conclusions

Owing to the concerns regarding applicability to the decision problem and consistency with the NICE reference case, for several of the studies34,141,144–147 included in the review, results are summarised here only for the UCB S.A. 20 and Amgen Inc. 100 submissions.

In the Amgen Inc. company submission,100 which investigated the cost-effectiveness of DEN in a population of patients with a 10-year fracture risk of 20%, DEN was found to be associated with an incremental cost-effectiveness ratio (ICER) of £27,792 per QALY, compared with RLX, and an ICER of £27,363 per QALY compared with no treatment. At the same risk of facture, DEN was also found to dominate ZOL.

In the UCB S.A. submission,20 which investigated the cost-effectiveness of a treatment sequence of 1 year of ROMO followed by 4 years of ALN (ROMO/ALN), in a population of postmenopausal women with a 10-year fracture risk of 30%, ROMO/ALN was found to be associated with an ICER of (confidential information has been removed) per QALY compared with ALN alone, and (confidential information has been removed) per QALY compared with no treatment. The UCB S.A. submission20 also presented scenario analyses comparing ROMO/ALN with RIS, ZOL, RLX, DEN and TPTD (administered for 18 months and 24 months). The ICERs for ROMO/ALN when compared with these alternative comparators were (confidential information has been removed) and dominating (ROMO/ALN had more QALYs and a lower cost than TPTD for both the 18- and 24-month treatment durations), respectively, when using the Patient Access Scheme (PAS) price for ROMO.

Review conclusions

The review has identified that there are no published cost-effectiveness studies that compare all of the interventions and comparators specified in the scope of this appraisal across the broad population specified in the scope, which is patients eligible for risk assessment under CG146. 8 Although the Amgen Inc. 100 and UCB S.A. 20 submissions provide an incremental analysis for the majority of the interventions and comparators specified in the scope (neither compared with i.v. IBN), their analyses are restricted to high-risk subgroups of the population. However, this review was useful in identifying areas where the model used in TA46434 differed from the models included in the review. These are discussed further in Independent economic assessment, where we describe the changes made to the model reported by Davis et al. 34

Model structure

The ScHARR osteoporosis model (used in TA46434) is a discrete event simulation (DES), which simulates the clinical events occurring over the lifetimes of individual patients with heterogeneous characteristics. A patient-level simulation approach was chosen to allow the future events experienced by patients to be affected by prior events such as incident fractures. We chose to model a heterogeneous population because we anticipated that certain patient characteristics, such as age, would be non-linearly related to cost-effectiveness. For example, older patients are more likely to have experienced a prior fracture, and may therefore have a lower quality of life at baseline, and they are also more likely than younger patients to be admitted to a long-term nursing or residential care home following a hip fracture. Both of these factors will influence the costs and QALYs that can be gained from avoiding a fracture. In this situation, the cost-effectiveness for a patient with average characteristics is not the same as the average cost-effectiveness when taking into account the distribution of that characteristic across the population.

In general, in a DES model, a patient’s progression through the model is determined by the events that occur, rather than by the health states they occupy. Figure 10 shows the clinical events that can occur during a patient’s lifetime, with the arrows showing which events can occur following other events (note that this is not a state-transition diagram, as patients do not reside in the state defined by the most recent event until the next event is experienced). In the ScHARR osteoporosis model, the main clinical events were fracture, death and new admission to residential care. Fractures at different sites were processed using separate fracture events for hip; wrist; vertebral and proximal humerus. These are the sites most strongly associated with osteoporosis and these are the fracture sites included by both the QFracture and FRAX risk calculators. Fractures at additional sites (e.g. femoral shaft, humeral shaft, pelvis, scapula, clavicle, sternum, ribs, tibia and fibula) have been incorporated by increasing the incidence of these four event types, rather than by adding additional competing events.

FIGURE 10.

Clinical events that can occur during a patient’s lifetime in the DES.

In a DES, no changes are made to a patient’s attributes between events, but the event list that determines the future events experienced can be resampled each time an event occurs to incorporate any changes in patient characteristics. Dummy events were included in the model to ensure that patient attributes were updated at 1 year after the start of the model, at the end of treatment, at the end of the offset period, at 5 years, at 10 years and 1 year after each incident fracture. Linear approximation is used to adjust for age-related changes in utilty between events.

Utility in the model is based on a combination of sex, age, fracture history and residential status (community dwelling or institutionalised). Separate utility multipliers and costs are applied to the first and subsequent years after fracture to reflect the differences between the acute and chronic impact of fracture. The chronic cost is set to the maximum chronic cost for all fracture events experienced so far. Therefore, the maximum chronic cost for any individual is the cost for institutionalised patients. Drug costs are applied from the start of the simulation until the end of the treatment period and are assumed to accrue at a constant rate across time. Death does not incur any additional costs in the model, but the acute cost of fracture is incurred for both fatal and non-fatal fractures.

The model also incorporates the following structural assumptions:

• There are no restrictions on the sequence of fractures that can be experienced.

• The maximum number of fractures that can be experienced is limited to one per bone (i.e. two hip fractures), with an additional limit of four vertebral fractures, four rib fractures and two pelvic fractures.

• Death attributable to fracture occurs 3 months after fracture, with other fracture events possible during this period, but no mortality from non-fracture-related causes.

• Incident fractures increase the risk of future fractures.

• A fracture event occurring < 1 year after a previous event supersedes the dummy event used to update patient attributes 1 year after fracture, thus reducing the acute period for the earlier fracture.

• Nursing home admission can occur only following fracture; therefore, patients who are community dwelling at the start of the simulation do not transfer to nursing home care as they age unless this is simulated to occur following a fracture.

A brief overview of the key features of the ScHARR osteoporosis model used in TA46434 is provided in Table 6, alongside a description of the key changes to the model since TA464. The only deviation from the NICE reference case to note is that the utility estimates for ONJ have been valued using the US rather than the UK valuation set for the EQ-5D.

Population

The population is patients eligible for risk assessment under CG146,8 as per the final NICE scope (see Chapter 1, Measurement of disease). It should be noted that this includes both men and women, those with and those without a prior fracture, those with steroid-induced osteoporosis, those with secondary osteoporosis and those with other risk factors for fragility fracture. CG146 recommends that either FRAX30 or QFracture31,163,164 be used to assess the absolute risk of fracture. To explore whether or not the most cost-effective treatment varies for patients at different levels of absolute fracture risk, we report the variation in incremental net monetary benefit (INMB) across risk using two approaches. First, we report outcomes for 10 risk categories, based on deciles of absolute fracture risk. Second, we use regression to determine the relationship between INMB and absolute risk as a continuous variable. These steps are undertaken for absolute risk assessed by FRAX and for absolute risk assessed by QFracture.

Interventions and comparators

The treatment strategies modelled (and the intended treatment durations) were as follows:

• oral ALN (5 years)

• oral RIS (5 years)

• oral IBN (5 years)

• i.v. IBN (5 years)

• i.v. ZOL (3 years)

• RLX (5 years)

• DEN (10 years)

• TPTD (2 years)

• ROMO (1 year) followed by ALN (4 years).

These were all compared with a strategy of no treatment to estimate the incremental costs, incremental QALYs and INMB relative to no treatment. We note that, in the base-case analysis, the actual treatment duration modelled is determined by the duration of treatment persistence rather than the intended treatment duration, but it is necessary to specify the intended treatment duration for the scenario analysis assuming full persistence.

The intended treatment durations for bisphosphonates (3 years for ZOL and 5 years for all others) are based on the assumption made in TA464. 34 For the sequence of ROMO followed by ALN, the 1-year treatment duration for ROMO is based on the anticipated marketing authorisation. However, the anticipated marketing authorisation also states that ROMO should be followed by an anti-resorptive agent, but does not specify the duration for anti-resportive treatment. In the ARCH trial,83 patients in both arms received open-label ALN after the 1-year double-blind phase. In the clinical study report20 for the ARCH trial, the mean duration of ALN exposure after the 1-year double-blind phase is (confidential information has been removed) in both arms, but the maximum treatment exposure is between (confidential information has been removed) years across the two trial arms. To have the same overall intended treatment duration as the ALN strategy, we decided to model the ROMO/ALN strategy as including 4 years of ALN. For DEN, we have assumed an intended treatment duration of 10 years, as this is what was assumed in the Amgen Inc. submission,100 in which it was argued that there are data from the FREEDOM study104 on the efficacy and safety of up to 10 years of DEN treatment.

Treatment persistence

In the AG model, we have assumed that costs and benefits are linearly related to the duration of treatment persistence; therefore, the individual-level variation in persistence does not need to be modelled. The assumption was found to be reasonable in sensitivity anslyses reported by Davis et al. 34 Therefore, the variable that needs estimating to inform the model is the mean treatment persistence and standard error of the mean, which describes the uncertainty around the mean for the probabilistic sensitivity analysis (PSA).

In the model that informed TA464, Davis et al. 34 used published estimates of treatment persistence from observational cohort studies, with separate estimates applied for oral bisphosphonates, based on a systematic review by Imaz et al. ,165 and for i.v. bisphosphonates, based on a US study of Medicare patients (Curtis et al. 166). Davis et al. 34 applied the mean persistence reported in these studies to all patients receiving treatment, rather than modelling individual-level heterogeneity in treatment persistence. The model in the Amgen Inc. submission100 used persistence data from a retrospective analysis of a large UK primary care database (the CPRD) (Amgen Inc.,100 data on file). The proportion persisting with treatment over 5 years was estimated from these data and extrapolated beyond 5 years in the model based on the last year of data. The model in the UCB S.A. submission20 used published estimates for treatment persistence for bisphosphonates and RLX from a UK GPRD study and data from a non-UK registry study for DEN. Unpublished data were cited by UCB S.A. 20 as the source for TPTD and ZOL persistence. For the sequence of ROMO followed by ALN, the model submitted by UCB S.A. 20 assumed that 90% of patients would persist with ROMO up to 1 year, based on experience from clinical trials, and that, once patients switched to ALN, the treatment persistence would be 85% of that observed for DEN – the treatment with the highest persistence rate, based on the published estimates. Strom et al. 146 used persistence data for oral bisphosphonates from a UK CPRD study (Li et al.;167 similarities suggest that this is the same study cited by UCB S.A.) to model persistence over time for the first 3 years and then assumed that all patients reaching 3 years would continue on oral bisphosphonates. Strom et al. 146 used a non-UK randomised crossover comparison study109 to model treatment persistence with DEN. Kanis et al. 144 assumed that 50% of patients receiving oral bisphosphonates persist up to 3 months and the rest persist up to the intended treatment duration, based on the assumption used in the analysis that informed TA160 and TA161. It is not clear what assumption was made by Kanis et al. 144 regarding treatment persistence for RLX. In the model based on the MORE study,141 patient compliance was not taken into account, but it was noted in the discussion that 92% of patients took > 80% of their study medication. In the model submitted by Amgen Inc. for TA204,147 treatment persistence was assumed to be 100% for all treatments in the base-case analysis, but a lower rate of treatment persistence for oral bisphosphonates was applied in a sensitivity analysis based on data from the GPRD (GPRD is the previous name of the CPRD, but the data used here appear to be from a different study to that used in the current Amgen Inc. submission100).

Both of the company submissions used data from the same large UK primary care database (GPRD/CPRD). The published analysis by Li et al. 168 gave median durations of persistence for oral bisphosphonates ranging from 5 to 7 months across the more commonly used weekly and monthly preparations, whereas the more recent, but unpublished, analysis cited in the Amgen Inc. submission100 had a lower median persistence of (confidential information has been removed) months for all oral bisphoshonates. However, the AG notes that the data from Li et al. 168 suggest that the time-to-discontinuation curve has a long tail, so mean persistence will be longer than median persistence.

The AG estimated mean time on treatment from the Kaplan–Meier estimates published by Li et al. 168 by crudely estimating the area under tha Kaplan–Meier curve, assuming linear changes between the estimates reported. The data from the more recent analysis presented in the Amgen Inc. submission100 were considered less mature than the data presented by Li et al. 168 Mean durations of persistence in the first 5 years after starting treatment were estimated to be 1.7 years, 1.5 years and 1.4 years for ALN, RIS and RLX, respectively. Estimates for oral IBN were not possible owing to missing data at 5 years. Although separate estimates of persistence are provided for ALN and RIS, in the absence of any data demonstrating that treatment persistence differs significantly between different oral bisphosphonates, we decided to apply the average persistence data from ALN and RIS to all oral bisphosphonates. We note that mean treatment persistence is approximately three times longer under this assumption than assumed previously in the model that informed TA464. 34

The AG was not convinced that data from a primary care database, as used in the Amgen Inc. model,100 would be generalisable to i.v. bisphosphonates (and likewise TPTD) as these are usually prescribed in secondary care. Given this concern, and in the absence of any other alternative data sources, the AG decided to use the same estimates of treatment persistence for i.v. bisphosphonates as assumed in the model that informed TA464. 34

The evidence on the long-term persistence with DEN appears to be very limited, with most studies reporting a maximum of 24 months’ follow-up. 109,169–171 It is difficult to estimate the mean or median duration of treatment from studies that are limited to 2 years when persistence is high at 2 years and it is possible for DEN to be given long term. The analysis of CPRD data presented in the Amgen Inc. submission100 presents data beyond 2 years, but these were described as exploratory analyses only. The AG were concerned about whether or not the analysis of CPRD data presented by Amgen Inc. would accurately capture DEN persistence as, although DEN may sometimes be administered in primary care, treatment is usually initiated in secondary care. Therefore, any estimate of persistence derived solely from primary care records may fail to accurately capture treatment discontinuation in the transition between secondary and primary care. The persistence data used for DEN in the UCB S.A. submission20 match the cited source (Karlsson et al. 170) up to 24 months, but beyond that they have simply assumed a fixed proportional decrease in the numbers that are persistent, based on a comparison between the 18-month and 24-month persistence rates. The AG decided to estimate the mean treatment persistence from the CRPD data presented by Amgen Inc. in their model. The estimates of persistence appear to be very uncertain beyond 4 years, but there appears to be a constant risk of discontinuation from years 2 to 4. The AG decided to use the rate of discontinuation between years 2 and 4 to estimate the proportionate decrease in persistence experienced thereafter. From this, the mean treatment persistence over 10 years was estimated to be (confidential information has been removed). The AG notes that these estimates are uncertain owing to the exclusive use of primary care records and the need for an assumption to be made to extrapolate persistence up to 10 years because of the low proportion of patients captured in the analysis beyond 2 years (confidential information has been removed).

Several sources of persistence data were identified for TPTD. As stated above, the estimates based on UK primary care databases were discounted based on the fact that TPTD is usually prescribed in secondary care. However, two published articles172,173 were identified from ad hoc literature searches that described persistence in UK patients in real clinical practice based on data from the main home care provider of TPTD in the UK. Both these studies were conducted before the maximum duration of treatment in the marketing authorisation was extended from 18 to 24 months, but they show high levels of persistence at 18 months of 79%172 and 74%173 for women and men, respectively. However, these estimates were based on Kaplan–Meier data taking into account the censoring of patients who were still on treatment at the longest follow-up. Data from the European Extended Forsteo Observational Study (ExFOS),174 which was a large European real-life clinical practice study of TPTD use after the licence was extended to 24 months, showed a mean treatment duration of 20.7 months, despite 29% of patients residing in countries where the licence remained restricted to 18 months. All three papers show a fairly linear fall-off in persistence, although a more rapid fall in persistence was seen in the ExFOS study at 18 months in the countries with 24-month reimbursement, which could be explained by a lack of uptake of the longer dosing schedule. We decided to use the data from UK women to estimate the average duration of treatment. To do this, we assumed a constant rate of discontinuation from 0 to 24 months, based on the rate observed over 18 months by Arden et al. ,172 giving an estimated mean persistence time of 1.72 years (20.6 months), which is reasonably consistent with the estimate from ExFOS, which had a mean treatment duration of 20.7 months. We decided to take the standard error of the mean (0.14 months) from the ExFOS study as the measure of uncertainty for the estimate applied in the model. When sampling this parameter in the PSA, the maximum number of doses was capped at 24, as per the SmPC for TPTD. 25

For ROMO, the manufacturer claimed that 90% of patients persisted to 12 months, based on data from the clinical trials. The AG used data on doses received in the ARCH study83 to estimate mean persistence with treatment, and found that this agreed with patients being treated for a mean of (confidential information has been removed), although it noted that only (confidential information has been removed) of patients received all 12 doses of ROMO. When sampling this parameter in the PSA, the maximum number of doses was capped at 12, as per the draft SmPC for ROMO provided in the UCB S.A. submission. 12 For the sequence of ROMO followed by ALN, we have assumed that treatment persistence with ALN is the same as for the ALN-only strategy.

Effectiveness data

The HRs estimated in the NMA (see Figure 6) were applied in the model for the duration of treatment, with a linear increase to a HR of 1 (i.e. no treatment effect) during the offset period. For the treatment sequence of ROMO followed by ALN, the HR for ROMO followed by ALN was applied during both the ROMO and the ALN treatment periods, as the HR estimate in the NMA was based on fractures occurring during both treatment phases. The NMA requires a single estimate of treatment effect for each study; therefore, it would not have been possible to generate separate estimates of treatment efficacy for the ROMO and ALN parts of the treatment sequence.

When data on fracture outcomes were lacking for i.v. IBN, the AG used the NMA estimate for daily oral IBN, as the marketing authorisation for i.v. IBN was based on studies demonstrating that i.v. IBN had superior BMD outcomes compared with daily oral IBN. It is noted that this is potentially unfavourable to i.v. IBN if superior BMD outcomes translate into superior fracture prevention outcomes. However, this is consistent with the approach taken in TA464. 34

For vertebral fracture, we have used the outputs of the base-case NMA, which included studies reporting morphometric fractures. This is because the outcome of morphometric fracture was more widely reported, and the NMA sensitivity analysis that excluded studies that reported only morphometric fractures, leaving just those studies reporting clinical vertebral fracture, was found to produce results that were consistent with the base-case analysis.

In the model that informed TA464,34 it was possible to use the bisphosphonate class effect estimate when data on individual bisphosphonates were lacking. In the updated networks described in Chapter 3, Network meta-analysis, no hip fracture data were available for i.v. IBN and monthly oral IBN, but data were available for all non-bisphosphonates. We decided to apply the bisphosphonate class effect estimate for i.v. IBN and monthly oral IBN when data were lacking for hip fracture. We note that the class effect for bisphosphonates was very similar to the estimates for ALN, RIS and ZOL, and so this was not considered to unfairly bias the cost-effectiveness analysis.

In the analysis that informed TA464,34 the data were considered too sparse for the outcome of proximal humerus fracture, so the non-vertebral NMA estimates were used instead. In the NMAs conducted for the current MTA, the networks were sparsely populated for non-bisphosphonates for the outcomes of both wrist fracture and proximal humerus fracture. The AG decided to use the NMA estimates from the non-vertebral fracture NMA for both wrist and proximal humerus fractures as this allowed a single network to be used to estimate HRs for all interventions. This was considered preferable to using data from different networks for bisphosphonates and non-bisphosphonates, as the wrist and proximal humerus fracture estimates would be more uncertain than the non-vertebral fracture estimates.

In the base-case analysis, the convergence diagnosis and output analysis (CODA) samples from the NMA were used, as these preserve the underlying joint distribution of the HRs, but, in the deterministic analyses, the median HR was used.

Offset period

The AG used a review by Idolazzi et al. 175 and papers cited in the company submission to identify relevant studies that could be used to inform the assumptions regarding the appropriate offset periods for the different treatments modelled.

For ALN, the key study was considered to be the Fracture Intervention Trial Long-term Extension (FLEX),176,177 as this provided comparative data on both fracture risk and BMD for patients remaining on, or stopping treatment with, ALN after 5 years of treatment. This study found that it took 5 years for total hip BMD to return to pre-treatment levels when treatment with ALN was discontinued after 5 years. This was supported by no separation of the time-to-event curves for non-vertebral fractures for patients remaining on treatment compared with those stopping treatment. There was some evidence of a continued treatment effect for lumbar spine BMD, and a continued reduction in vertebral fracture risk was observed (RR 0.45, 95% CI 0.24 to 0.88) for patients who continued ALN compared with those who discontinued ALN.

For RIS, two studies were identified. 178,179 Watts et al. 179 reported the outcomes of patients randomised to either placebo or RIS in the year after discontinuing the study drug. Eastell et al. 178 reported the outcomes of patients in the year after completing the Vertebral Efficacy with Risedronate Therapy-Multinational (VERT-MN) study, in which patients were randomised to either RIS or placebo for 3 years, followed by a 2-year open-label extension on the allocated study drug, followed by 2 years of open-label RIS in both groups. Both studies reported that BMD gains at the hip were lost in the 1 year following treatment discontinuation, although Watts et al. 179 observed smaller losses in lumbar spine BMD and reported a statistically significant reduction in vertebral fracture incidence between those previously treated with RIS and those previously treated with placebo, in the year after treatment discontinuation.

The data identified for oral IBN were limited to those from 1-year post-trial follow-up from an early dose-finding study,180 which included the 2.5-mg daily dose that has been shown in non-inferiority bridging studies to be equivalent to the 150-mg monthly dose that is now licensed. 181 This study180 appears to show a similar pattern to that seen for RIS, in that hip BMD appears to return to pre-treatment levels in the year after treatment, with a slightly slower return for lumbar spine BMD. However, as the duration of treatment was only 1 year, it is not clear whether the offset time is 1 year regardless of treatment duration, or whether it would increase in proportion to treatment duration.

For oral bisphosphonates, the AG decided to keep the assumption made previously in the model that informed TA464,34 which was that treatment effect falls to zero over a period equal to the initial treatment duration for all oral bisphosphonates, as this was accepted previously by the NICE Appraisal Committee. However, in a sensitivity analysis, we have also explored the possibility of a fixed 1-year offset time for RIS and oral IBN.

For i.v IBN, no studies were identified that explored BMD or fracture outcomes following treatment discontinuation. Therefore, we assumed that the offset period would be the same as for oral IBN and set it equal to treatment duration, with a fixed 1-year offset explored in a sensitivity analysis.

For i.v. ZOL, data from the Health Outcomes and Reduced Incidence with Zoledronic Acid Once Yearly (HORIZON) – Pivotal Fracture Trial extension study are provided by Black et al. 182 In the extension study, patients who had received 3 years of ZOL were randomised to receive either ZOL or placebo for a further 3 years. At the end of the study, femoral neck BMD had declined in those who switched to placebo, but not to baseline levels, suggesting an offset period that is longer than the treatment duration when measured based on BMD changes. This suggests a slightly longer tailing-off of treatment effect than observed for ALN in the FLEX study. There was, however, no statistically significant difference in non-vertebral fractures between placebo and ZOL in the extension phase. Similar to the results from the FLEX study, further gains were made in lumbar spine BMD after discontinuation, and there was a statistically significant difference in new vertebral fractures in the extension stage of HORIZON.

For i.v. ZOL, the AG decided to keep the assumption made previously in the model that informed TA464,34 which was that treatment effect falls to zero 10 years after the start of a 3-year treatment period. For patients stopping treatment early, the offset duration was assumed to decrease proportionately. A sensitivity analysis assuming an offset period equal to treatment duration was also conducted.

For TPTD, data on treatment in women were identified from the FPT follow-up study,183,184 which followed up patients for a median duration of 30 months after the RCT phase of the study. The RCT phase was terminated early (owing to concerns regarding the safety of long-term use); the median treatment duration was 20 months. During the follow-up study, patients were treated according to local standards and a high proportion (i.e. 56.9% of those randomised to the licensed dose of TPTD in the RCT phase) received other osteoporosis interventions. To account for this, results were presented for the subgroup with no further osteoporosis intervention, in addition to the analysis for all patients. Statistically significant reductions in vertebral fractures were reported by Lindsay et al. 183 in the 18 months following discontinuations, and not all of the lumbar spine BMD gained during treatment had been lost by 18 months. For non-vertebral fractures, statistically significant differences were not found for the licensed dose compared with placebo at the longer follow-up point of 30 months post discontinuation when adjusting for usage of other osteoporosis medications. Furthermore, the gains in femoral neck and total hip BMD appeared to be lost by 18 months in the group not receiving other osteoporosis interventions. A second smaller study185 in men with a shorter follow-up time had similar findings. Based on these two studies, we decided to assume an offset period equal to the treatment duration.

For RLX, two relevant studies were identified. One compared continuation with RLX with discontinuation in patients previously treated for 96 weeks. 186 Although there were some baseline differences in BMD, the percentage change in lumbar spine BMD from baseline was no longer statistically significant at 144 weeks in the group that had discontinued at 96 weeks, whereas the benefit in lumbar spine BMD was maintained in those continuing RLX up to 192 weeks from baseline. A second RCT extension study,187 which examined 1-year outcomes in patients discontinuing after 5 years of RLX, oestrogen or placebo, found that BMD values were significantly lower 1 year after discontinuing than at the end of treatment therapy for both lumbar spine and femoral neck BMD. Although these data are from a small study, they support a rapid loss of efficacy in the year after treatment even for patients treated for > 2 years. Based on these two studies,186,187 we decided to apply a 1-year offset period for RLX.

For DEN, two papers188,189 reporting outcomes from a single study were identified. The paper reporting 2 years’ follow-up post discontinuation in patients allocated to either 2 years of DEN or 2 years of placebo found that gains in both lumbar spine BMD and total hip BMD were lost in the first year after discontinuation, suggesting that an offset period of 1 year would be reasonable for DEN. A third paper,190 presenting an analysis of post-trial outcomes of patients from the FREEDOM study, was also identified, which described a rapid fall in BMD in the first year after discontinuation, even after treatment lasting 10 years. Although this analysis was limited to 12 women from a single site, and can therefore be considered as only weak evidence, this analysis is supportive of a fixed offset period of 1 year, rather than one that varies with treatment duration. Therefore, for DEN we have assumed a fixed offset period equal to 1 year (or, when this is < 1 year, the treatment duration).

For ROMO, no data were identified in the published literature on the treatment effect following discontinuation. In sequences in which ROMO is followed by ALN, we have assumed an offset period equal to the total duration of the treatment sequence, with efficacy during the offset linearly declining from the efficacy observed across the treatment sequence. This is consistent with the assumption applied by UCB S.A. 20

Drug costs

For drugs with multiple preparations, the cost was based on the lowest cost preparation available. For drugs administered in primary care, the costs were taken from the NHS drug tariff. 191 For drugs administered in secondary care, the electronic market information tool (eMIT) database192 was used for generic preparations (i.v. bisphosphonates) and the NHS drug tariff191 price was used when no generic preparation was listed as being available (i.e. for TPTD and DEN). For ROMO, the annual costs for both the list price and the PAS price were taken from the company submission. The PAS price was used in the AG’s base-case analysis. The price used for TPTD was based on the branded formulation (Forsteo), as no prices were available for the biosimilar versions (Movymia and Terrosa)22,23 when this report was prepared.

The dosing, cost per item and annual cost for each treatment strategy are summarised in Table 7.

Treatment initiation, administration and monitoring

Six of the studies assumed that patients would undergo DXA every other year while on treatment. 20,100,141,144,146,147 Stevenson et al. 143 assumed that patients would undergo DXA at years 2 and 5. Davis et al. 34 did not include any DXA to monitor treatment with bisphosphonates. Not all of the papers were explicit about whether or not patients were assumed to have undergone DXA before starting treatment, but, in Davis et al. ,34 all costs that related to risk assessment, which may include DXA for some patients, were considered to have been already inccurred prior to treatment choice, as these were included in the cost-effectiveness analysis for risk assessment in CG146. 8 The AG considered that the inclusion of routine DXA in the model was problematic as the approach taken may differ depending on the baseline risk of the patient and the treatment being administered. For example, CG146 does not recommend that DXA is performed routinely as part of the risk assessment of patients. 8 Therefore, it is reasonable to assume that many patients may be started on the current first-line therapy, which is oral bisphosphonates, without DXA, and this is consistent with the approach recommended in the NICE-accredited NOGG guideline. 14 However, the NOGG also recommends that FRAX with BMD is used to reassess patients at the end of 5 years of bisphosphonate therapy (3 years for ZOL). On this basis, we decided to assume that patients undergo DXA when they reach the end of the intended treatment duration. We made an exception for DEN, as the intended treatment duration is much longer than for other therapies, so we assumed that DXA is undertaken every 5 years. This was based on advice from one of our clinical experts that patients receiving DEN in primary care would be likely to be reviewed in specialist care at 3 or 5 years. For the treatment sequence of ROMO followed by ALN, we assumed that a patient would undergo DXA once at the end of the 1 year of ROMO and once at the end of the 4 years of ALN. Because treatment duration in the model is based on average treatment persistence rather than the distribution of persistence across the population, the AG incorporated the cost of DXA as an annualised cost; otherwise, no DXA costs would be applied, as the average patient never reaches the intended treatment duration. This is consistent with the assumption that costs and benefits are linearly related to the duration of treatment persistence and, therefore, the individual-level variation in persistence does not need to be modelled. The cost applied for DXA is based on the NHS reference cost for direct-access DXA (£68.29 for RD50Z). 193

Four of the studies assumed that patients would attend annual general practitioner (GP) appointments to monitor treatment. 20,141,146,147 Amgen Inc. 100 assumed the same for treatments given in primary care (which included oral bisphosphonates and DEN), but assumed secondary care follow-up appointments for i.v. bisphosphonates. Kanis et al. 144 assumed one GP appointment to initiate treatment. Stevenson et al. 143 assumed two GP appointments per annum, whereas Davis et al. 34 did not include any GP appointments for monitoring. There is now a NICE QS13 that states that patients having bone-sparing treatments should have medication reviews to discuss AEs and adherence, but the frequency of the reviews is not specified. We have assumed that patients will have annual reviews and that those reviews will occur in primary care for oral bisphosphonates and RLX. For this, we applied the cost per average GP patient contact (£38.00 per 9.22 minutes). 16 For DEN, we were advised that patients would be reviewed in secondary care every 3–5 years, so we have assumed that one in four annual reviews will occur in secondary care. For i.v. bisphosphonates, ROMO and TPTD, we have assumed that the annual review occurs in secondary care as an outpatient endocrinology appointment. The cost (£150.38) for a consultant-led, non-admitted, face-to-face follow-up attendance at endocrinology outpatient has been applied [Healthcare Resource Group (HRG) currency code WF01A, service code 302]. 193

As noted previously, none of the studies identified in the review included any costs for the administration of oral therapies; this was the assumption applied in our model. UCB S.A. 20 also assumed no administration costs for s.c. therapies (i.e. DEN, TPTD and ROMO). In the Amgen Inc. submission for this MTA,100 it was assumed that DEN would be given by a general practice nurse, whereas, in the Amgen Inc. submission for TA204,147 they assumed that one injection would be administered during the annual GP visit, and therefore one additional GP appointment was required per annum for the second injection. For DEN, we assumed that patients would initiate treatment in secondary care, with the first two doses being given as an outpatient procedure using the same HRG codes as applied for i.v. IBN. Thereafter, it was assumed that DEN would be administered under a shared care agreement, with a primary care nurse providing future doses during a 15.5-minute appointment at a cost of £10.85 (based on £42.00 per hour for general practice nurse contact time). 16 This was based on advice from our clinical experts that, ideally, only the first one or two doses would be given in secondary care, although they also noted that there is significant variation in practice surrounding shared care agreements, with some local areas having a poor uptake of primary care administration.

Stevenson et al. 143 do not describe any additional administration costs for TPTD. Waugh et al. 147 included one additional GP appointment for initiation of TPTD. The AG did not consider that any additional costs were necessary for the administration of TPTD, given that it is self-administered, and an annual secondary care review has already been included for TPTD, as described previously.

Davis et al. 34 assumed that i.v. IBN is delivered during an outpatient endocrinology appointment and that i.v. ZOL is delivered as a day-case procedure using the HRG code for administration of a simple parenteral chemotherapy (SB12Z). UCB S.A. 20 assumed that administration of i.v. ZOL occurred in secondary care, but the exact source of the cost applied is unclear. In the Amgen Inc. submission for TA204, administration of i.v. bisphosphonates was assumed to occur in secondary care under the same HRG code as used by Davis et al. 34 for i.v. ZOL. However, in the Amgen Inc. submission for the current MTA,100 it was argued that the use of an oncology HRG was inappropriate; instead, the cost was based on day case and elective inpatient spells averaged over nine HRG codes related to non-inflammatory bone and joint disorders and pathological fractures. The AG was already aware of a study that compared the cost of secondary care infusion of ZOL with a home-care delivery model in a UK NHS setting. 194 In correspondence with the study author (Opinder Sahota, Nottingham University Hospitals NHS Trust, 2018, personal communication), it was stated that the reference cost, including the drug costs, for this activity was £300 per patient (£14,980 per 50 patients), and this included acquisition of the drug at a discounted (undisclosed) cost from the manufacturer. However, the income for the activity based on the tariff was much lower, at £143 per patient, which also includes the cost of drug acquisition. Based on these figures, we felt that the estimates provided by Amgen Inc. were probably too high and we decided to use the HRG codes applied in the model that informed TA464,34 but updated to the latest reference costs,193 giving a cost of £253 for day-case infusion of i.v. ZOL (day case, SB12Z delivery of simple parenteral chemotherapy at first attendance).

For i.v. IBN, no alternative estimates of administration costs were identified from the studies included in the review. Therefore, we decided to assume the same resource use as in the model used to inform TA46434 (one outpatient endocrinology follow-up appointment), but we updated the unit cost to reflect the latest reference costs,193 giving a cost of £150.

Adverse events

For oral and i.v. bisphosphonates, the AG decided not to change the approach to modelling AEs that was adopted in TA464,34 as there was no new evidence on which to base alternative assumptions identified from the review of cost-effectiveness studies.

The AG decided to include serious (i.e. leading to hospitalisation) cellulitis as an AE for DEN because it had been included in the model that informed TA204,147 although it was noted that the 10-year results of the FREEDOM study41,104 suggest that the incidence rate of cellulitis is low, at ≤ 0.2% in each of the study years. The HRG cost for a non-elective inpatient spell for minor skin conditions with interventions ranges from £2588 to £7764, depending on the level of complications and comorbidities, with a weighted average of £4467. 193 Assuming an incidence rate of 0.2% per annum and applying this weighted cost to the incident population would increase the cost of DEN by £8.93 per annum. The AG identified a paper that had estimated the QALY loss of cellulitis as 0.005 QALYs (reduction in EQ-5D score by 26.3% for 7 days), based on a comparison of EQ-5D scores in a prospective RCT of antibiotics versus placebo to prevent recurrent cellulitis. 195 This is equivalent to a loss in INMB of £0.20 per annum. As the duration of treatment persistence with DEN in the model is (confidential information has been removed) years, this would suggest that the total impact of cellulitis is a reduction in INMB for DEN of the order of (confidential information has been removed). Costs and QALY losses for cellulitis per year of exposure to DEN have been included in the base-case model.

The AG notes that the Medicines and Healthcare products Regulatory Agency (MHRA)/Commission on Human Medicines (CHM) has issued advice regarding the risk of atypical femoral fractures for both DEN and bisphosphonates,28 but this advice states that these events are rare and that they are primarily related to long-term use. The AG decided not to include atypical femoral fractures as a separate AE in the model. This was, first, because the HRs for fractures estimated from the clinical trials would already include any impact of the drug on atypical femoral fractures, and including them as a separate event may result in these outcomes being double-counted in the model. The AG accepts that atypical femoral fractures may not have been captured in the trials if they occur only after long-term use of osteoporosis treatment. However, the AG notes that the base-case scenario incorporates real-world treatment persistence, which is much shorter than the intended treatment duration for both bisphosphonates and DEN, making these AEs that occur with long-term use less relevant to these treatments as they are modelled.

The AG notes the MHRA/CHM advice regarding the risk of ONJ in patients receiving bisphosphonates. 28 The advice states that the risk is considered to be substantially higher in those receiving i.v. bisphosphonates in the treatment of cancer than in those receiving i.v. bisphosphonates for the treatment of osteoporosis, and the risk is said to be related to cumulative dose. Similarly, the MHRA/CHM advice on DEN states that it is a common side effect for those patients receiving DEN for the treatment of cancer and recommends dental examination and preventative dentistry treatment in all patients starting DEN for cancer. 28 It should be noted that the dose for cancer is 120 mg monthly, rather than 60 mg every 6 months, and, in the context of using DEN to prevent osteoporotic fracture, such precautions are recommended by the MHRA/CHM only for those with risk factors. 28 The AG also notes that a systematic review by Boquete-Castro et al. 196 states ‘it should be stressed that most of the adverse effects of DEN appear with doses of 120 mg. Adverse effects with doses of 60 mg are directly related to the duration of treatment.’. Although there appears to be less concern regarding ONJ in patients receiving anti-resportive agents for osteoporosis than for ONJ in patients receiving anti-resportive agents for cancer, the AG decided to incorporate this AE in the model to establish the likely impact on the cost-effectiveness estimates.

The AG examined a systematic review reported by Khan et al. ,197 which was conducted to inform an international consensus statement on ONJ. Khan et al. 197 conclude from their review that ‘the incidence of ONJ in the osteoporosis patient population appears to be very low, ranging from 0.15% to < 0.001% person-years of exposure and may be only slightly higher than the frequency observed in the general population’. For oral bisphosphonates, the review by Khan et al. 197 identified a UK (Scottish) prospective case series that reported an incidence for ONJ of one case per 4545 drug patient-years (0.022%) for patients exposed to ALN. 198 This was within the incidence range of 1.04–69 cases per 100,000 patient-years reported by the other studies identified in the review by Khan et al. 197 It should be noted that Lo et al. 199 found, in a cross-sectional survey conducted in the USA, that prevalence of ONJ was related to duration of exposure, with estimated prevalences of 0%, 0.05% and 0.21% in patients exposed for < 2 years, 2 to just under 4 years and ≥ 4 years. For i.v. bisphosphonates, Khan et al. 197 reported an incidence range of 0–90 per 100,000 patient-years. The incidence estimated across five RCTs is given by Khan et al. 197 as < 1 in 14,200 patient-years of exposure (< 0.007%). For DEN, Khan et al. 197 reported that the estimates of incidence ranged from 0 to 30.2 per 100,000 patient-years. However, more recent data from the 10-year follow-up of the FREEDOM trial41,104 gave an exposure-adjusted incidence of ONJ of 5.2 per 10,000 participant-years (0.052%). The SmPC for DEN states that the incidence is related to the duration of exposure. 200 Given that there is a lack of comparative data on the incidence of ONJ across the different forms of anti-resportives, and that the estimates for the different anti-resportive drugs all relate to different periods of exposure, we have decided to assume the same incidence per year of drug exposure across all anti-resportives. This was based on the estimate from the prospective case series in Scotland. 198 This was because this estimate fell within the range provided by Kahn et al. 197 for each type of anti-resportive (oral bisphosphonates, i.v. bisphosphonates and DEN) and was based on the average duration of use in clinical practice; therefore, it would be more applicable to the duration of treatment persistence modelled in this analysis.

A paper201 measuring health utility in patients with ONJ was identied using ad hoc searches of Google Scholar. It reported utility measured by the EQ-5D in 34 cancer patients with bisphosphonate-associated ONJ. However, it should be noted that it was not compliant with the reference case in several ways. First, although the pateints had all themselves experienced ONJ, they were asked to value clinical vignettes describing different stages of ONJ in patients who also have cancer, rather than being asked to value their own health state. Second, the utilty weights applied were from the US, rather than the UK, valuation set. However, given the lack of alternative estimates, we calculated the average utility decrement based on the utility decrements (relative to patients with cancer but without ONJ) for stages 2 and 3 (–0.33 and –0.61, respectively) and the distribution of ONJ stages (two were stage 3 and nine were stage 2) across the UK prospective case series reported by Malden and Lopes. 198 This gave an average utility decrement of –0.38. The mean time from diagnosis to healing (6.5 months) was taken from the same study198 to give an average QALY loss of 0.206 per case of ONJ. The NHS reference cost for a minor outpatient oral surgical procedure was applied (HRG code CD03A, £166)193 to account for the cost of surgical management, as most patients in the Malden and Lopes198 case series had some form of surgical management, with debridement being the most common procedure. We note that the Malden and Lopes198 case series may have missed less severe cases of ONJ, which would be classed as stage 1. However, as cancer pateints with stage 1 ONJ were found not to have EQ-5D values significantly different from those of cancer patients without ONJ (Miksad et al. 201), and patients with stage 1 would be more likely to be managed conservatively,197 we felt that exclusion of this group was unlikely to significantly bias the estimates of costs and QALYs resulting from ONJ, provided they are excluded from both the incidence estimates and the estimates of costs and QALYs per case. Costs and QALY losses per year of exposure to DEN, oral bisphosphoantes and i.v. bisphosphonates have been included in the base-case analysis, but we note that their impact is very small owing to the extremely low incidence.

Kanis et al. 141 applied HRG costs and a utility loss in the year after VTE, but not beyond. The utility decrement was based on an assumption, as no estimate was identified from the literature. No other models identified in the literature review included VTE as an adverse outcome. Rather than extend the AG model to incorporate the competing risk of VTE in patients at risk of fracture, the AG decided to estimate the average discounted lifetime cost and QALY loss attributable to VTE using a published model (Pandor et al. 202). As this model was constructed to estimate the costs and benefits of thromboprophylaxis, the AG removed all costs and QALY losses attributable to the thromboprophylaxis itself, including the increased risks of bleeding during the prophylaxis, thereby reducing the model to a comparison of two groups whereby the only difference between them is their risk of VTE. All conseqeunces related to asymptomatic VTE were removed from the model as these were not considered relevant, as it is only symptomatic VTE that has been recorded as an adverse outcome. The AG then compared costs, QALYs and the number of symptomatic VTEs for the strategies of prophyalixs for all and prophyalixs for none. These figures were used to estimate the average discounted lifetime cost and QALY loss per symptomatic VTE, which were estimated to be £1890 and 0.77 QALYs for a patient with a starting age of 50 years.

The largest RCT reporting VTE as an adverse outcome for RLX was the MORE study,51,102 which reported that 25 out of 2557 patients receiving RLX experienced VTE, whereas eight out of 2576 patients receiving placebo experienced VTE. Based on the increased incidence observed in the MORE study, the excess rate of VTE attributable to RLX was estimated to be 0.67% over the 3-year study period. Ettinger et al. 51 did not report the proportion of these events that were PEs, but did say that a mixture of PE and DVT events were observed. The study by Silverman et al. 50 did report the breakdown by type of VTE: four of the 12 VTE events in the RLX-treated arm were PEs. It should be noted that, in the model by Pandor et al. ,202 30% of symptomatic VTE events are PEs, which is reasonably consistent with the ratio of PE to DVT observed in the RLX arm of the study by Silverman et al. 50

By applying the estimates of costs and QALYs per symptomatic VTE derived from Pandor et al. 202 to the excess incidence observed in the MORE study, we estimated a reduction in INMB of £116 per patient enrolled in the MORE study when valuing a QALY at £20,000 (and assuming that VTE occurred at 50 years of age). Given that the average duration of persistence in the model for treatment with RLX is 1.38 years, if we assume that the absolute risk is proportional to the time spent on treatment, the INMB loss attributable to VTE would be of the order of £53 per patient started on treatment (cost of £5.80, QALY loss of 0.00237). It should be noted that the QALY losses would be fewer for older patients experiencing VTE, as much of the QALY loss is attributed to long-term sequelae that have a greater impact on patients with a higher life expectancy. However, when assuming a start age of 75 years, the INMB loss attributable to VTE per patient started on RLX was estimated to be £47 (compared with £53 for patients aged 50 years), so the error associated with applying costs and QALYs as estimated for a patient aged 50 years was not considered likely to have resulted in a large bias. The average cost and QALY loss attributable to excess VTE were applied to each patient initiating treatment with RLX, with the risk proportional to time spent on treatment, such that they have a bigger impact in the sensitivity analysis assuming full treatment persistence.

Disease costs

The costs of fracture in the TA46434 model were based on a UK resource use study reported in two papers by Gutiérrez et al. ,203,204 which used a general practice database (The Health Improvement Network) to estimate resource use for those who fractured compared with matched controls. Unit costs from the 2013/14 reference costs205 and the 2014 PSSRU unit costs206 were then applied to this resource use to estimate the total cost in the year of fracture and in the subsequent years following fracture. None of the studies included in the review provided a more recent source of resource use. Two studies20,100 reported using costs based on Gutiérrez et al. ,203,204 and five141,143,144,146,147 used estimates from the literature from less recent publications.

The AG identified two additional relevant UK studies in the systematic database search conducted to identify published cost-effectiveness analyses. Lambrelli et al. 207 used a methodology similar to that employed by Gutiérrez et al. , but using an alternative primary care database (the CPRD), with linkage to a secondary care database [Hospital Episode Statistics (HES)]. Lambrelli et al. 207 reported costs in the year following hip fracture of £7359. Leal et al. 208 reported higher costs, of £14,163, based on an analysis of HES data alone. This analysis excluded activity in primary care and was focused solely on patients admitted to hospital following fracture. For comparison, the estimate used in TA464,34 based on the data from Gutiérrez et al. ,203,204 when excluding the costs of home help, was £6274. The AG decided to use the data from TA464,34 and to adjust it using 2017 PSSRU inflation indices,16 as the two studies by Gutiérrez et al. 203,204 provided a consistent methodology for estimating both hip and non-hip fractures, included activity in both primary and secondary care settings and incorporated prescription costs.

Costs for home help and residential care/nursing home admission were estimated by uplifting the estimates used in TA46434 using PSSRU inflation indices. 16

The costs applied in the first and subsequent years following fracture are summarised in Table 8.

Health-related quality of life

We conducted a rapid update of the systematic review of HRQoL studies conducted for TA464. 34 This comprised a systematic serach for studies reporting EQ-5D utility data for the year post fracture. Further details on the review methods and findings can be found in Appendix 16. In summary, the review identified four papers all reporting outcomes from the International Costs and Utilities Related to Osteoporotic Fractures Study (ICUROS). 209,210,212,213 This study was previously identified in the review conducted for TA464. 34 However, the four new papers identified reported additional data. ICUROS was an international multicentre study; two of the papers210,212 reported outcomes from specific countries that formed subgroups of the overall ICUROS population. The other two papers reported longer-term follow-up from the overall international data set. One of these papers213 restricted its analysis to those patients with complete follow-up on both the EQ-5D and the EQ-VAS, which resulted in a smaller population available for analysis. The paper reporting outcomes from the international cohort without restricting to patients who also reported EQ-VAS was chosen, as it was the larger data set. 209 This paper reported utility multipliers for the year following fracture and subsequent years for hip, wrist and vertebral fractures. The multipliers presented in the paper were applied directly in the model. However, no data were presented in this paper for proximal humerus fractures. The only paper reporting outcomes following proximal humerus fracture was the one reporting outcomes for the Australian subpopulation of ICUROS. 210 Although these data were specific to a different country, results were presented in an appendix using the UK time trade-off tariff for the EQ-5D. From these data, we calculated utility multipliers for the year following humerus fracture and subsequent years, using the same methodology as employed in the international paper for the other fracture types. The utility values applied are summarised in Table 8.