Primary trabeculectomy versus primary glaucoma eye drops for newly diagnosed advanced glaucoma: TAGS RCT

Primary objective

The primary objective of this trial was to compare primary medical management with primary augmented trabeculectomy (glaucoma surgery) for patients presenting with advanced glaucoma [Hodapp–Parrish–Anderson (HPA) Classification severe] in terms of patient-reported health status using the National Eye Institute’s Visual Function Questionnaire-25 (VFQ-25). 45–54

Secondary objectives

• To compare generic and vision- and glaucoma-specific patient-reported health and experiences in the short and medium term.

• To compare the costs and benefits [quality-adjusted life-years (QALYs) gained] of surgery and medication at 2 years based on responses to the (1) EQ-5D-5L, (2) the Health Utilities Index version 3 (HUI-3) and (3) the Glaucoma Utility Index (GUI). 55

• To compare clinical outcomes (VFMD changes, logMAR visual acuity changes, IOP, Esterman VF for driving vision, registered visual impairment).

• To compare the need for additional cataract surgery.

• To compare safety by comparing AEs arising from both surgical and medical interventions.

• To employ an existing discrete choice experiment (DCE) among participants with advanced glaucoma to generate a revised scoring system for the GUI that is more sensitive and specific for those with advanced disease.

• To compare long-term costs and benefits through a modelling evaluation.

Clinical centres

Twenty-seven secondary care centres with at least one consultant who subspecialises in glaucoma recruited patients for the study.

Population

Adults with advanced (severe) glaucoma in at least one eye defined as severe according to the HPA grading64 for VF loss severity were invited to participate in the study.

Primary medical management: escalating medical therapy

Participants randomised to medical management could be prescribed a variety of licensed glaucoma medications (eye drops). These eye drops were used in accordance with NICE guidelines. 65 Escalating medical management was defined as follows: study participants may be started on one or more medications at their initial visit depending upon the judgement of the treating clinician. When monotherapy is initiated this should be with a prostaglandin analogue as directed by NICE guidelines. Subsequent addition of medications was based on clinician judgement/preference. When drops failed to control IOP adequately oral carbonic anhydrase inhibitors may be used.

Primary trabeculectomy: standard trabeculectomy augmented with mitomycin C

Standard trabeculectomy was defined as the creation of a ‘guarded fistula’ by making a small hole in the eye that is covered by a flap of partial-thickness sclera, which allows aqueous humour to egress from the eye into the subconjunctival space. The operation could be performed under either local or general anaesthetic. The dose of mitomycin C in terms of exposure time and concentration was left to the discretion of the operating surgeon and decided on a case-by-case basis.

The protocol specified that all surgery be undertaken within 3 months of randomisation by a consultant who subspecialises in glaucoma or a glaucoma fellow who has performed at least 30 trabeculectomies. Where both eyes were eligible for the study, the eye with better VFMD was allocated as the index eye. An amendment to the protocol allowed the decision about which eye would undergo trabeculectomy first to be made locally when subjects were allocated to the trabeculectomy arm. 1

To ensure that recognised standard trabeculectomy procedures66,67 were being followed by all participating glaucoma surgeons, all potential surgeons completed a questionnaire about their surgical technique that was reviewed and signed off by the chief investigator. No feedback was given because all surgeons were essentially conducting the same operation.

Compliance with study treatment

We designed TAGS as a pragmatic trial and compliance with study treatment was monitored as it would be in routine clinical practice – by asking the patient if they are using their eye drops. There is currently no practical and effective method for monitoring compliance in patients taking glaucoma medications. 68 The degree of compliance feeds into the outcome measurements, as poor compliance for medications is likely to lead to further disease progression. There was no requirement for participants to return any unused eye drops.

Accountability of the study treatment

The local clinical team used a standard hospital prescription form or asked patients’ general practitioners (GPs) to prescribe the medications required in line with standard practice for that team, pragmatically reflecting standard NHS practice.

Concomitant medication

Medications as required for normal clinical care were prescribed for the participants irrespective of their randomised allocation. Throughout the study, investigators could prescribe any concomitant medications or treatments deemed necessary to provide adequate supportive care.

Primary outcomes

The primary outcome was a vision-specific patient-reported outcome QoL measurement, the vision-specific health profile (VFQ-25), which was evaluated at 24 months. The VFQ-25 is a validated questionnaire that has been widely used to evaluate visual outcomes in glaucoma. 47,69–71 In addition to eliciting information about general health and vision, it specifically addresses difficulty with near vision, distance vision, driving and the effect of light conditions on vision.

Secondary outcomes

Methods to protect against sources of bias

The allocation to treatment arms was concealed by using a central randomisation service that could be accessed only by clinical trials unit programming staff.

The evaluation of VFs collected during the study period was by CARF masked to the participant intervention. IOP was measured by two observers: one taking the reading and the other reading the IOP value to minimise the risk of measurement bias.

The expected attrition rate was low based on a previous glaucoma treatment RCT,75 but we allowed for a potential attrition rate of 13.5% over 24 months to accommodate this.

External validity was maximised by recruiting from multiple centres and participants were treated by multiple clinicians.

Ground rules for statistical analysis

The trial analysis followed a statistical analysis plan (see Report Supplementary Material 2), which was agreed in advance by the TSC. The main analyses were based on the intention-to-treat (i.e. analyse as randomised) principle and took place after the 24-month follow-up. Baseline and follow-up data were summarised using appropriate statistics and graphical summaries. Statistical significance was at the two-sided 5% level with corresponding CIs derived. All analyses were carried out using Stata^® version 16 software (StataCorp LP, College Station, TX, USA).

Sample size

The primary patient-reported outcome was the health status measured by the VFQ-25 assessment at 24 months. A study with 190 participants in each arm would have 90% power at a two-sided 5% significance level to detect a difference in means of 0.33 of a standard deviation (SD), which translates to 6 points on the VFQ-25 assuming a common SD of 18 points observed in previous work that has a clinically relevant effect size in patients with advanced glaucoma. 76,77 Seven points is a likely minimally important difference based on our pilot work on VFQ-25 scores in patients with glaucoma,77 but there is uncertainty, and so we opted for a more conservative 6-point difference, which is supported by the literature for another chronic eye disease, macular degeneration. 78 Assuming a drop-out rate of 13.5% as a result of declining further follow-up and death, we had to randomise a total of 440 participants to detect this difference.

For the secondary clinical outcome (VFMD), TAGS had 90% power at 5% significance to detect a 1.3 dB difference in VFMD. This is derived from a subgroup of patients with advanced glaucoma and is a clinically significant difference in the context of advanced glaucoma and predictive further visual disability. 30,43

Centres that had a throughput of at least 20 eligible patients annually were approached to participate in the trial. This enabled the development of a recruitment projection based on 20 sites recruiting approximately 9–11 patients per year, with a staggered start of recruiting sites, and allowed 440 participants to be recruited over 3 years (assuming a reduced rate during the first month of each site set-up and 50% reduction during the holiday months of August and December).

Primary/secondary outcome analysis

The primary outcome, VFQ-25 assessed at 24 months, was analysed using a heteroscedastic partially nested repeated-measures mixed-effects linear model correcting for baseline score and bilateral disease, and time as a fixed effect. 79 The repeated measures were VFQ-25 assessed at 4, 12 and 24 months. Treatment effects were estimated from time-by-treatment interactions at each time point; the primary time of interest was 24 months. This approach uses participant data from all of the time points and incorporates a random effect for centre and surgeon using restricted maximum likelihood. We used this approach to account for the potential lack of independence of outcome data for patients treated by the same surgeon. This is known as clustering or the surgeon effect, and occurs in the trabeculectomy arm only (hence the ‘nested’ approach) in TAGS. We also adjusted for heteroscedastic errors because the residuals compared with the fitted values showed evidence that the variance of errors were different between treatment arms. We did not carry out any sensitivity analysis for missing data because < 10% of VFQ-25 data were missing. Given the pragmatic nature of TAGS, there was the potential for participants to cross over from one intervention to the other. To estimate the treatment effect in the subgroup of participants who complied with allocated treatment, we used complier-average causal effect (CACE) methods using instrumental variable regression80 and a per-protocol analysis.

For the secondary outcomes, EQ-5D-5L, HUI-3, GUI, IOP and logMAR VA, VFs were analysed following the same method as the primary outcome. The utility values for the GUI were derived from the results of the DCE (which are presented in Chapter 5). For the IOP, logMAR VA and VF tests, a sensitivity analysis was performed using data from all eligible eyes by also including a random effect at the participant level to reflect the lack of independence of eyes within participants. In addition, for the VFs a further two sensitivity analyses were performed: the first included only the VF tests performed in accordance with SITA standard and the second removed participants who had only one VF test or in whom the false positive standard was > 15%. The post hoc Dunnett’s multiplicity p-values were presented for IOP and logMAR VA. Patient experience (glaucoma getting worse) was analysed using a repeated-measures mixed-effects Poisson model adjusting for bilateral disease and including a random effect for surgeon and treatment. The need for cataract surgery, visual standards for driving and safety outcomes were analysed using the same modelling approach. Given that the number of events for patients registered as sight impaired was small, we used Fisher’s exact test. Data missing at baseline were reported as such, and for the primary/secondary outcomes continuous data were imputed using the centre-specific mean of that variable.

Subgroup analyses

Planned subgroup analyses explored the potential treatment effect moderation of sex, age (< 65 years vs. ≥ 65 years), one or both eyes affected with advanced glaucoma, Index of Multiple Deprivation (quintile) and extent of VF loss at baseline (–20 dB, ≥ –20 dB) on the primary outcome. The subgroup-by-treatment interaction was assessed by including interaction terms in the models outlined above. We used a stricter level of statistical significance (two-sided 1% significance level) and 99% CIs to reflect the exploratory nature of these analyses. A post hoc subgroup analysis of IOP at diagnosis (< 21 mmHg vs. ≥ 21 mmHg) was also explored following the same analysis method as the planned subgroup analysis.

Criteria for the termination of the trial

Owing to the staggered nature of recruitment and, therefore, the measurement of the primary outcome at 24 months, we did not anticipate that the trial would be terminated early for benefit. We proposed one main effectiveness analysis at the end of the trial. During the trial, safety and other data were monitored by reports prepared for the Data Monitoring Committee (DMC). The DMC decided to meet every 6 months until June 2018.

Study Management Group

The Study Management Group (SMG) was responsible for the day-to-day management of the trial. This group consisted of the chief investigator, trial manager, senior trial manager, data co-ordinator, statistician and a sponsor representative. Members of the SMG are listed in the Acknowledgements.

Project Management Group

The Project Management Group (PMG) was responsible for overseeing the management of the trial. This group consisted of the SMG plus grant applicants, a public and patient involvement (PPI) representative and a senior programmer. Members of the PMG are listed in the Acknowledgements.

Trial Steering Committee

The Trial Steering Committee (TSC) was responsible for monitoring and supervising the progress of TAGS. The committee met seven times between April 2014 and December 2019 at agreed intervals. The TSC consisted of independent experts, a PPI member, the chief investigator and key members of the PMG. Members of the TSC are listed in the Acknowledgements.

Data Monitoring Committee

The DMC was independent of the trial and was responsible for monitoring safety and data integrity. The committee met eight times between April 2014 and August 2018, approximately every 6 months at intervals agreed by the committee. The trial statistician provided the data and analyses requested by the DMC prior to each meeting. The committee consisted of three independent experts. Members of the DMC are listed in the Acknowledgements.

Subgroup analysis

Figures 4a–c show the prespecified subgroup analyses at 4, 12 and 24 months for VFQ-25 score, sex, age (< 65 vs. ≥ 65 years), VFs (–20 dB vs. ≥ –20 dB), advanced glaucoma in both eyes, Index Multiple Deprivation quintile (most deprived to least deprived) and IOP at diagnosis (< 21 mmHg vs. ≥ 21 mmHg). There was no evidence of any treatment effect heterogeneity at 4 or 12 months. At 24 months, a potential moderating effect of age is emerging; however, there is considerable uncertainty around the estimate of the interaction effect.

FIGURE 4.

Subgroups for trabeculectomy vs. medical management: (a) 4 months, (b) 12 months and (c) 24 months. First quintile, most deprived; fifth quintile, least deprived. Boxes indicate mean differences. Solid black line indicates 99% CI. Solid vertical line indicates no effect. Dashed vertical line indicates overall effect.

EQ-5D-5L

The mean baseline EQ-5D-5L score was 0.844 in the trabeculectomy arm and 0.837 in the medical management arm. At the follow-up time points, there was no evidence of a difference in the mean EQ-5D-5L score, and at 24 months the MD was 0.016 (95% CI –0.021 to 0.053; p = 0.405) (Table 6).

Health Utility Index version 3

The mean baseline HUI-3 score was 0.814 in the trabeculectomy arm and 0.809 in the medical management arm, with no evidence of a difference at the follow-up time points. At 24 months, the mean score was 0.786 in the trabeculectomy arm and 0.751 in the medical management arm, with a MD of 0.036 (95% CI –0.006 to 0.078; p = 0.094) (see Table 6).

Glaucoma Utility Index

The baseline mean GUI score was 0.884 in the trabeculectomy arm and 0.863 in the medical management arm. These mean scores decreased in both arms over the course of the study, but there was no evidence of a difference at any time point. At 24 months, the mean score was 0.849 in the trabeculectomy arm and 0.830 in the medical management arm, with a mean difference of 0.011 (95% CI –0.017 to 0.039; p = 0.434) (see Table 6).

Patient perception: glaucoma getting worse

At baseline, 95 out of 208 (45.7%) participants in the trabeculectomy arm and 76 out of 209 (36.4%) participants in the medical management arm perceived that their glaucoma was getting worse. At the later follow-up time points, this proportion decreased in both study arms. At 12 months, 38 out of 207 (18.4%) participants in the trabeculectomy arm and 57 out of 199 (27.3%) participants in the medical management arm perceived that their glaucoma was getting worse [relative risk (RR) 0.59, 95% CI 0.38 to 0.91; p = 0.018]. At 24 months, the RR was 0.70 (95% CI 0.46 to 1.07; p = 0.099) (see Table 6).

Intraocular pressure

At baseline, the mean IOP was 19.40 mmHg in the trabeculectomy arm and 19.05 mmHg in the medical management arm. At later time points, the mean IOP decreased, with lower IOP in the trabeculectomy arm at all follow-up time points. Figure 5 shows the mean and SD over time. At 24 months, the mean IOP was 12.40 mmHg in the trabeculectomy arm and 15.07 mmHg in the medical management arm (mean difference –2.75, 95% CI –3.84 to –1.66; p < 0.001) (see Table 6). The Dunnett’s multiplicity p-value was also < 0.001 for all time points The sensitivity analysis incorporating data from the non-index eye for participants with bilateral disease (see Appendix 3, Table 30) showed similar results.

FIGURE 5.

Mean (SD) IOP by arm over time. The SD values are at the end point of the vertical lines.

LogMAR visual acuity

At 4 and 24 months, there was evidence of a difference in logMAR VA in favour of medical management (MD at 24 months 0.07, 95% CI 0.02 to 0.11; p = 0.006) (see Table 6). The Dunnett’s multiplicity p-value was 0.002 for 4 months, 0.483 for 12 months and 0.061 for 24 months. The sensitivity analysis incorporating data from the non-index eye for participants with bilateral disease showed similar results (see Appendix 3, Table 30).

Visual fields mean deviation

At 24 months, the mean VFMD was –15.15 in the trabeculectomy arm and –15.42 in the medical management arm, with no evidence of a difference between arms [mean difference 0.18, 95% CI –0.58 to 0.94; p = 0.645) (see Table 6). Two sensitivity analyses were carried out. The first included only SITA standard VF (excluding SITA fast) and the second removed participants who had only one VF or for whom the false positive standard was > 15% required for reliability. Overall, the results were similar (see Appendix 3, Table 31). The sensitivity analysis incorporating data from the non-index eye for participants with bilateral disease showed similar results (see Appendix 3, Table 30).

Need for cataract surgery

Over the course of the trial, 28 out of 222 participants (12.6%) in the trabeculectomy arm and 27 out of 221 participants (12.2%) in the medical management arm needed cataract surgery, with no evidence of a difference between arms (RR 0.98, 95% CI 0.50 to 1.95; p = 0.963) at 24 months (see Table 6).

Visual standards for driving (pass/no defects)

At 24 months, 167 out of 187 participants (89.3%) in the trabeculectomy arm and 168 out of 188 participants (89.4%) in the medical management arm passed the visual driving standards, with no evidence of a difference between arms (RR 1.01, 95% CI 0.81 to 1.25; p = 0.951) (see Table 6).

Registered as sight impaired at 24 months

At 24 months, 4 out of 186 participants (2.2%) in the trabeculectomy arm and 0 out of 184 participants (0%) in the medical management arm were registered as sight impaired, with the difference between the arms having a p-value of 0.123 (see Table 6). In addition, we also recorded whether or not participants were eligible to be registered (see Table 6 and Appendix 3, Table 32, for baseline data). At baseline, 10 participants (94.3%) in the trabeculectomy arm were sight impaired and three participants (1.3%) were severely sight impaired. In the medical management arm, 12 participants (5.3%) were sight impaired at baseline and two participants (0.9%) were severely sight impaired. At 24 months, the results were similar.

Safety

A safety event is defined as either a SAE or an AE.

Antiglaucoma medication

Table 9 shows the prescribed antiglaucoma medication agents as allocated. The mean number of glaucoma eye drop agents at 4 months was 0.43 (SD 0.79) in the trabeculectomy arm and 1.77 (SD 0.92) in the medical management arm. At 12 months, the mean was 0.34 (SD 0.79) and 1.78 (SD 1.01) and at 24 months the mean was 0.47 (SD 0.92) and 1.64 (SD 1.15) for the trabeculectomy and medical management arm, respectively. The majority of participants used prostaglandin analogue and beta-blocker.

Table 10 shows the antiglaucoma medication agents prescribed by as treated. At 24 months, the mean number of antiglaucoma eye drops was 0.29 (SD 0.72) for those who received trabeculectomy and 1.96 (SD 0.98) for those who received medical management.

Appendix 3, Table 36, shows the antiglaucoma medication prescribed for the non-index eye.

Appendix 3, Tables 37 and 38, show the number of different bottles of medication as allocated and as treated, respectively.

Other trabeculectomy interventions

Table 11 shows details of other trabeculectomy interventions for those who received a trabeculectomy and were followed up. The total number of participants who received any intervention was 123 out of 200 (61.5%) in the trabeculectomy arm and 26 out of 39 (66.7%) in the medical management arm. The median time to the first intervention was 8 [interquartile range (IQR) 6–27] days in the trabeculectomy arm and 11 (IQR 7–19) days in the medical management arm, with the majority of participants receiving a massage or releasable release.

Non-index eye

A post hoc analysis looked at IOP (mmHg), logMAR and VFMD for the non-index eye (see Appendix 3, Table 35).

Stage 1: identification of attributes and levels

The attributes and levels used in the DCE were informed by the GPI developed by Burr et al. 55 The work of Burr et al. 55 involved using a DCE to estimate the preference weights for the glaucoma utility index (GUI). 55 Briefly, Burr et al. 55 considered a wide range of potentially relevant attributes and levels that could be used to define a range of glaucoma health state profiles. These were identified using established glaucoma QoL instruments and expert opinion. These attributes and levels were refined using qualitative research methods (focus groups) with individuals with glaucoma. The attributes identified as important to individuals with glaucoma were central and near vision, lighting and glare, mobility, activities of daily living, eye discomfort, and other effects of glaucoma and its treatment. Each of these six attributes had the same four levels: no difficulty, some difficulty, a lot of difficulty or severe difficulty. 55 These six attributes with their corresponding levels were used to inform the health state profiles valued in the DCE.

Stage 2: creating an efficient design

The combinations of all attributes (n = 6) and levels (n = 4) from the GPI generates 4096 possible health state profiles (47). In the DCE, participants were asked their preference between two of these health state profiles (a choice-set question); however, asking participants to choose between all possible combinations of these health state profiles would be too onerous and not feasible. As with all DCEs, an experimental design was used so that the number of choice-set questions being valued was manageable and the main effects and higher-order interactions could still be estimated in the logistic regression. 86 The experimental design used by Burr et al. 55 was not used in the DCE conducted as part of TAGS because the design methodology has developed since that work was conducted. 55 For the current study, the experimental design followed best practice guidelines and a D-efficient design was used to identify the most efficient combination of choice-set questions. 86,87 The D-efficient design was produced using Ngene design software (ChoiceMetrics Pty Ltd, Sydney, NSW, Australia) and the design chosen was the best design that minimised the standard errors. 86

The design was split into four blocks to maximise variance in the data, and each participant was randomly allocated to one block and presented with 15 choice-set questions. Only participants who had not withdrawn from TAGS follow-up or who had not died during the study were included in the block randomisation. Figure 6 is an illustrative example of a choice-set question that participants were asked to answer.

FIGURE 6.

Example of a choice-set question.

Stage 3: data collection

The paper-based questionnaire and guide to completing the DCE were sent to participants, who were randomly allocated to one of the four blocks at 27 months post randomisation, 3 months after their final study visit. The randomisation was adjusted to ensure that equal numbers of respondents were randomised to each block at the end. A freepost envelope was provided for participants to return the completed questionnaire. After 3 weeks, a reminder was sent to participants who had not returned the questionnaire.

Stage 4: data analysis and interpretation

The data were analysed in Stata^® version 15 (StataCorp LP, College Station, TX, USA). Descriptive statistics were used to describe the demographic characteristics of the full sample, responders and non-responders to identify any potential differences between responders and non-responders. Descriptive statistics were also used to identify any potential differences in demographic and socioeconomic characteristics (i.e. income and education) between participants in each of the four blocks.

The DCE was analysed under a random utility framework. Logistic regression techniques, specifically a conditional regression logistic model, were used to analyse participants’ stated preference for each choice-set question and quantify the relative importance of each attribute and attribute level. 84 A base category (i.e. severe difficulty) was selected for each attribute and against which the other levels of that attribute were compared; hence, 18 explanatory variables (six attributes × three levels) were included in the model. It was hypothesised that each attribute level would be positive compared with the severe difficulty level and that the coefficients would increase with improvements in difficulty. Given that all six attributes had the same levels, comparisons could be made across the attributes and the attribute levels could be ranked in order of preference.

Participant characteristics

Table 12 presents a summary of the demographic characteristics of the participants in the overall sample, and Appendix 4, Table 39, presents a summary of the sociodemographics of those who responded to the questionnaire and the difference in characteristics between blocks. Participants who responded to the DCE questionnaire had a mean age of 68 years. The majority of these participants were white (91%) and male (67%), and almost half (45%) had attended college or university. Each of the income brackets was equally represented across the four blocks, with nearly half of the participants who responded to the DCE questionnaire (48%) reporting an income of ≥ £20,000. The proportion of non-responders was similar across the four blocks.

Regression model results

The Walt test determined that the explanatory variables were not equal to zero and should, therefore, be included in the model (p < 0.01). The results of the conditional logit model for the full sample are presented in Table 13. The levels for four of the attributes (i.e. central and near vision, lighting and glare, mobility, and activities of daily living) and the level ‘no difficulty’ for the attribute ‘eye discomfort’ were positive compared with the base level (i.e. severe difficulty), as expected. These results suggest that participants have a preference for health state profiles that report lower levels of difficulty in these attributes. At a 95% confidence level, only three of the attributes [i.e. central and near vision (all levels), mobility (excluding the level quite a lot of difficulty) and activities of daily living (all levels)] were statistically significant, suggesting that participants had a stronger preference for improvements in these attributes than in the other attributes.

Counterintuitively, two of the levels for the attribute ‘eye discomfort’ and three levels for the attribute ‘other effects of glaucoma and its treatment’ had negative coefficients on average, suggesting that there was a negative preference (disutility) associated with improvements in these attributes compared with the base level (severe difficulty). However, there was no evidence that the preference for these attribute levels differed to the base level (p-values were 0.180 and 0.756, respectively).

Given that all of the attributes had the same levels, comparisons could be made across the attributes to identify the order of preference for the attributes and attribute levels. Improvements in ‘central and near vision’, ‘activities of daily living’ and ‘mobility’ yielded the most utility for participants. The attribute level with the greatest utility gain was ‘no difficulty’ with ‘central and near vision’.

Preference-based weights

Similarly to Burr et al. ,55 preference-based weights were estimated for each of the attribute levels so that the health state profiles generated by the GPI could be assigned a QoL weight that could be used to estimate QALYs. 55 These health-related quality-of-life (HRQoL) weights, based on responses to the GPI administered at baseline and at 1, 3, 6, 12, 18 and 24 months post randomisation, were used in a sensitivity analysis in the economic evaluation (further details are shown in Appendix 5, Table 45).

It was assumed that attribute levels that were negative and not statistically significant had a value of zero, indicating that there was no utility to be gained from having reduced difficulty in these attributes. This assumption was made because these results did not adhere to theoretical validity (i.e. there was a negative preference associated with an improvement in the attribute level compared with severe difficulty) and there was no evidence that this value was different from zero, based on the significance levels estimated from our model. Table 14 summaries the HRQoL weights for each attribute level and provides an illustrative example of how to estimate a value that can be used to estimate QALYs for a health profile based on the HRQoL weights.

Sensitivity analyses

A sensitivity analysis adopting the methods outlined by Burr et al. ,55 combining the levels of attribute levels that were not statistically significant, was undertaken. In this analysis, the levels of ‘lighting and glare’, ‘eye discomfort’ and ‘other effects of glaucoma and its treatment’ were each combined into one variable and compared with the base level for that attribute, ‘severe difficulty’. In this analysis, the statistical significance of these attribute levels did not improve. However, the level ‘quite a lot of problems’ for the attribute ‘mobility’ did reach statistical significance at the 5% level. There was still a negative preference for the attribute ‘other effects of glaucoma and its treatment’ despite the combination of the three levels.

The model was re-estimated on a subset of the data, including only participants who completed all 15 choice-set questions (n = 283). This analysis produced similar results to the base-case analysis in terms of the importance of the attributes and levels and the statistical significance of attributes and levels did not change in the complete-case analysis.

Introduction

An economic evaluation is defined as a comparative analysis in terms of both costs (resource use) and consequences (outcomes and effects). 89 For this trial, the economic evaluation was to be carried out in two parts. The first part was a ‘within-trial’ economic evaluation that was conducted as part of the RCT to compare the costs and benefits of both medical and surgical pathways of glaucoma management. This within-trial analysis assessed the costs and benefits at the end of the 24-month trial period based on the trial data. The second part of the economic evaluation (discussed later in this chapter) is a model-based economic evaluation, which uses existing literature and the trial results to extrapolate the results beyond the end point of the trial.

There is evidence that, for advanced presentations of glaucoma, keeping IOP lower than 17.5 mmHg and stable has a beneficial effect on visual outcome. 34,90 However, the relative effectiveness and cost-effectiveness of medical and surgical pathways of glaucoma management are unclear. Chapter 4 describes the relative effectiveness of treatments. This section reports the methods and results of the within-trial economic evaluation (the model-based economic evaluation that extrapolated beyond the results of the trial is reported in the second part of this chapter).

The primary measures of effectiveness used in the economic evaluation are based on responses to the EQ-5D-5L, HUI-3 and GUI. The analysis was carried out from a health-care (NHS) perspective and from a patient perspective. NHS costs include the costs borne by the health service over the course of the trial and PSS costs (i.e. costs for local social services). Patient costs were estimated as well, and these included time and travel costs borne by the patient and those assisting participants to attend appointments in addition to any private health-care costs.

Data collection

Inpatient costs

Participants in the trial were randomised into the trabeculectomy or medical management arm of the trial. If the participants were in the trabeculectomy arm, the baseline CRF asked details about the trabeculectomy procedure. This includes the time spent in the operating theatre, the equipment used and the staff present. Two kinds of admissions were associated with the trabeculectomy procedures in this trial: a day-case admission and an overnight inpatient admission. Costs for both kinds of procedure are available in the NHS Reference Costs 2017/1891 and are attributed to each participant who completed the surgical CRF. If the admission type was not stated, the day-case rate was assumed, as this was the most common admission for the procedure according to the NHS Reference Costs 2017/18. 91

Those who were allocated to the medical management arm of the trial were not allocated an inpatient cost at baseline, but both treatment arms were asked about any inpatient stays or procedures on each subsequent CRF form. The patient was asked about any inpatient stays associated with a list of procedures and, if appropriate, provided admission and discharge dates so that the duration of any inpatient stay could be calculated. Both the cost of the hospital and the cost of the procedure were taken from NHS Reference Costs 2017/18. 91 Further details of unit costs for hospital appointments are provided in Appendix 5, Table 40.

Outpatient costs

Secondary care outpatient costs were elicited from the CRF data collected from the 4-month visit onwards. This included both visits to the outpatient ophthalmology clinic and any necessary ophthalmological outpatient procedures. The CRF asked about outpatient-specific procedures including:

• massage

• adjustment/suture lysis/releasable release

• 5-FU injection

• steroid injection

• needling and 5-FU injections

• bleb resuturing

• anterior chamber reformation

• bleb revision

• phacoemulsification and intraocular lens.

There was also an option to include a different procedure to any of the ones named above. The number of ophthalmology outpatient appointments and the dates of attendance were collected on each CRF form. The procedures were costed using the NHS National Reference Costs 2017/18. 96 Further details of unit costs for procedures are provided in Appendix 5, Table 41.

Medication costs

Medications were also captured in the CRFs. Details on the use of the following glaucoma medications were collected:

• prostaglandin analogues

• beta-blockers

• carbonic anhydrase inhibitor

• pilocarpine

• α-agonists

• other.

In the trabeculectomy arm of the trial, the use of pre-operative eye drops was recorded on the CRFs and costed. Medication costs were taken from the 2018 British National Formulary. 93 Drug start dates and stop dates were reported on the CRFs. If a participant started a drug before the 2-year trial period and was still taking the medication at the end of the trial, it was assumed, for costing purposes, that they were taking the drug for the entire trial period. If, however, the participant started or stopped the medication during the course of the trial, the length of time that the participant was on the drug was calculated in months and then the cost of the drug was calculated using this duration. Further details of unit costs for medications are provided in Appendix 5, Table 42.

Primary care costs

Patients in the trial were asked to complete a participant questionnaire at baseline and at 1, 3, 4, 6 and 12 months. As part of the 4- and 12-month questionnaire, participants were also asked about the health-care resources that they utilised over the trial period, for example:

• GP appointments in practice

• GP appointments at home

• GP telephone appointments

• community optometry appointments

• district nurse appointments

• practice nurse appointments.

Costs for these appointments were taken from the Unit Costs of Health and Social Care 2018. 97 The number of appointments for each of the types above was multiplied by the unit costs for each appointment. Further details of unit costs for primary care are provided in Appendix 5, Table 43.

Costs to participants and their main caregiver

This study measured the travel costs that were incurred by participants when they travelled to appointments. This was measured using a dedicated time and travel questionnaire that the participant completed during the course of the study. It was originally planned to administer the questionnaire at 18 months, but, because of a change to the protocol, the questionnaire was administered at 30 months. The questionnaire asked participants to provide the average cost of a typical journey to:

• outpatient appointment

• GP appointment

• community optometrist appointments

• inpatient hospital stays.

If the journey was undertaken using public transport, the fare was used to represent travel costs. If a journey was undertaken by private car, a fuel rate of £0.45 per mile was applied based on the business and self-employed expense rate per mile. 98 Finally, if hospital transport was used, travel costs were estimated from the Unit Costs of Health and Social Care 2018. 97

The costs of the time commitments of the participants and their companions were also measured in this evaluation; this was also measured in the time and travel questionnaire. Participants were asked to choose which best described how they would spend their time if they were not attending appointments to manage their glaucoma. The participant chose from one of the following:

• paid work

• homemaking

• child care

• caring for a friend/relative

• retired

• full-time education

• unemployed

• voluntary work

• leisure activities

• other (please specify).

Paid work, child care, caring for a relative or friend and voluntary work were valued at £13.88 per hour, which is the average hourly wage according to the Office for National Statistics. 95 Housework and leisure activities were valued at £10.10 per hour. The time spent in unemployment, retirement or full-time education was valued at £6.04 per hour. Finally, private out-of-pockets expenses reported by participants were included in the patient time and travel costs.

Estimation of quality-adjusted life-years for the cost–utility analysis

Given that the within-trial economic evaluation is a cost–utility analysis (CUA), outcomes were measured in QALYs. QALYs were estimated from three sources: the EQ-5D-5L,99 HUI-3 (because the HUI-3 does not a have a UK value set, the existing Canadian value set was used100) and GUI. 14 The utility values for the EQ-5D-5L were mapped to an existing value set for the validated data set of results for the EQ-5D-3L. 101 During the course of the study, a UK value set for the EQ-5D-5L was produced,102 but, given that this is currently not validated, the cross-walked values from the EQ-5D-3L data set were used in the base case and the EQ-5D-5L values were examined in a sensitivity analysis. The EQ-5D-5L instrument was administered at seven time points during the trial: baseline and 1, 3, 6, 12, 18 and 24 months. The complete-case analysis included all participants who had completed the EQ-5D-5L at all seven time points. Those who had completed the EQ-5D-5L at five of the seven time points were included (as long as one of those points was baseline) with imputed data. Some participants completed an additional EQ-5D-5L before their trabeculectomy. These values are summarised in Appendix 6, Table 47.

The QALYs gained was also calculated using the scores that were derived from the responses from the HUI-3 complete cases were considered to be those who had all seven time points and values were imputed for those who had four time points or more. The utility values for the GUI were derived from the results of the DCE (which are presented in Chapter 5).

Costs

A total NHS cost per participant was calculated using data from the trial. A seemingly unrelated regression (SUREG) was used to identify any difference between the trabeculectomy and the medical management arm of the trial while controlling for any modifying factors, such as the participant’s age and their baseline utility score. In a further analysis, the SUREG was repeated with the inclusion of patient time and travel costs.

Quality-adjusted life-years

As with the costs, a total QALYs gained per participant was calculated using trial data. A SUREG was used to calculate the differences between the costs and to control for modifying factors, such as baseline utility score and treatment arm.

Missing data

Cost data were reported as missing either if the sections of the CRF reporting medications taken and procedures undertaken were completely blank (no values, positive or negative, were given in either section) or if the total costs on the CRF were reported as zero and the answer to the question ‘Has participant completed the TAGS Participant Questionnaire?’ was ‘no’, as this indicates incomplete data.

With respect to the estimation of QALYs, those who had completed four of the seven data collection surveys were included for analysis. First, to account for the missing data points, it was assumed that the previous utility remained stable. This meant that the weighted average of the two utility scores around the missing values was used to calculate the missing data. Second, multiple imputation was used to estimate missing utility values for QALY scores. No patterns were observed in the missing data, so the data were imputed randomly. 103

Stochastic sensitivity analysis

To assess the robustness of the study sampling, non-parametric bootstrapping was carried out. Bootstrapping is a statistical procedure that resamples a single data set to create many simulated samples to assess statistical precision. This can model the difference in net benefit if the sampling process could be repeated many times. In this example, 1000 iterations of the bootstrapping procedure were performed. This simulation process created a sample of bootstrapped means for costs and QALYs, with distributions for each. The means and other parametric statistics were then calculated for the bootstrap distribution. Bootstrapped estimates of the difference in costs and QALYs between the experimental and the control arm were used to populate the cost-effectiveness plane (the horizontal axis represents the difference in QALYs between two interventions and the vertical axis represents the corresponding difference in costs).

A cost-effectiveness acceptability curve (CEAC) quantifying the probability that an intervention is cost-effective [based on the decision-maker’s maximum willingness to pay over a range of values (e.g. £0–100,000 per QALY)] was derived from the results of the non-parametric bootstrapping.

Deterministic sensitivity analysis

Deterministic sensitivity analysis was carried out to assess the variability of different parameters on the outcomes of the economic evaluation. The QALYs were recalculated using the utility values generated from both the HUI-3 and the GUI QoL tools to see if this changed the results. Patient time and travel costs were included to assess how these contacts have an impact on potential conclusions.

Response rates

The response rates for the data collection instruments used in the economic evaluation are described in Table 15. Data completion in this trial was generally very high.

Total resource use

The total costs for the different areas of resource use are summarised in Table 16. The largest cost drivers for total costs were ophthalmology outpatient appointments, prostaglandin analogue medications and, principally, the trabeculectomy procedure itself. The total costs are summarised in Table 17 and there was evidence of a higher cost in the first year for the trabeculectomy arm, which is largely because of the costs of the trabeculectomy procedure in the first year. In the second year, there was no evidence of a difference in the cost between the arms.

The results of the analysis of QALYs gained using different QALY measures are summarised in Table 18. Based on the EQ-5D-5L, the QALY gain at 24 months was slightly higher in the trabeculectomy arm than in the medical management arm based on the unadjusted means. Neither the HUI-3 nor the GUI provided any evidence of a difference between arms in QALYs gained.

Cost–utility analysis using the EQ-5D-5L

The results of the cost–utility analysis are presented in Table 19 and Figures 7 and 8. The results of the SUREG, reported in Table 19, show a small incremental QALY gain for surgery compared with medical management, but this small gain does not appears to offset by the higher costs of surgery. The effectiveness plane demonstrates that the difference in costs and QALYs for surgery compared with medical management is almost entirely in the quadrant that represents greater effect at greater cost (see Figure 8). The CEAC in Figure 7 shows that surgery is unlikely to be considered cost-effective over the range of values that society may be willing to pay for a QALY.

FIGURE 7.

Cost-effectiveness curves for the trabeculectomy arm and the medical management arm using the results from the EQ-5D-5L and multiply imputed data.

FIGURE 8.

Cost-effectiveness plane for the trabeculectomy arm vs. the medical management arm: adjusted bootstrapped replications for CUA for EQ-5D-5L results.

Sensitivity analysis

Appendix 5, Tables 44 and 45, and Figures 15–18 display the results of the sensitivity analysis when the HUI-3 and GUI QoL instruments are used in place of the EQ-5D-5L. Although the incremental cost-effectiveness ratios (ICERs) for the HUI-3 are smaller than that for the EQ-5D-5L, the probability of the trabeculectomy intervention being cost-effective is still very small, so the conclusions remain unchanged. A similar result was also seen for the GUI.

The inclusion of the participants’ time and travel costs in Appendix 5, Table 46, and Figures 19 and 20, demonstrate an increased difference in costs between arms and further supports the conclusion that medical management is likely to be considered more cost-effective than surgery over a 24-month period.

Background and rationale

Although the within-trial analysis is useful for informing decisions about cost-effectiveness in the short to medium term, it provides limited evidence to inform decisions about the cost-effectiveness of interventions in the longer term. 104 This is a very important limitation owing to the chronic nature of glaucoma,105 which can progress over time,106 and because the effects of treatment on costs and outcomes may persist into the future, with patients often living for many years after their glaucoma diagnosis. Therefore, a further analysis of the longer-term effects based on a Markov model107,108 was undertaken to extrapolate results of the trial beyond the 2-year follow-up and over the expected lifetime of patients. The results of this model can be used to inform decisions about the longer-term cost-effectiveness of trabeculectomy and medical management for people with newly diagnosed advanced glaucoma. In simple words, we tried to anticipate (extrapolate) the long-term impact of medication compared with surgery based on trial data and other related data from the literature. The logic of the modelling for this research is that the surgical procedure is expensive and, therefore, surgery is associated with higher initial costs than medical management. A longer-term time horizon would allow more time for any additional benefits of surgery to offset these initial higher costs. Essentially, a short time horizon could act as a bias against surgery.

Objective

The purpose of this model-based economic evaluation is to compare primary medical treatment with primary augmented trabeculectomy (glaucoma surgery) for patients presenting with advanced glaucoma (HPA classification severe64) in terms of costs and QALYs.

Methods

This section describes the methodology underpinning the model-based economic evaluation results. Detail is provided on the estimations of parameters used within the model and on how these values are used within the model to inform how patients move through the model. This section also describes how the model was validated to ensure that it estimates the costs of real-world glaucoma care service in the UK. Furthermore, the section describes the sensitivity analyses performed. These were carried out to analyse how uncertainty in the model’s parameters affect the final outcomes.

Model overview

A Markov model was constructed to model both treatments (trabeculectomy and primary medication)1 and disease progression that reflects the timing and allowed modelling of the logical and temporal sequence of events following the initial treatment strategies. The primary source of data for the model was the trial data set, supplemented with data from the literature where necessary. The uncertainty surrounding the model findings was assessed using probabilistic sensitivity analysis (PSA) and deterministic analysis. 107 The model was developed using TreeAge Pro 2019 (Williamstown, MA, USA).

To design the structure of the model and to classify the stages of glaucoma, we used the Enhanced Glaucoma Staging System (GSS2). 109 An external independent reading centre (CARF) reviewed and categorised TAGS visual fields into stages of VF loss based on the GSS2 system of grading. The consequences considered were the costs (of treatment and of subsequent management of patients with severe glaucoma) to the NHS and the effects on quality of life (QALYs). QALYs were estimated by assigning utility weights based on data on health state utility values derived from responses to the EQ-5D-5L (and GUI or HUI). 1 Combining these data with information on the probabilities of events occurring over time enabled cost, patient outcomes and QALYs to be estimated for a hypothetical cohort of patients undergoing each treatment strategy.

Model structure

The Markov model was developed using in-house experience of previous evaluations of glaucoma treatment,110–113 as well as the literature and advice from clinical colleagues. The model is used to simulate patient pathways from initial treatment until the end of life using data from the TAGS trial along with the best available UK relevant data to define transition probabilities for the model. All other model parameter estimates beyond 2 years were informed by the data from TAGS, other existing data sources (routine databases and the literature) and expert opinion.

Typically, Markov models have states (Markov states) in which individuals stay for a period of time. Each state reflects the subject’s level of well-being at that time. In each Markov state, the model will assign a cost and utility weight to each individual depending on the different interventions received by that individual while in that state and/or how long the individual remains a member (time spent) of that state. 107,114 The model is run for a length of time (e.g. 25 years), known as the ‘model time horizon’. This time horizon is broken down into equal parts, denoted as Markov cycles (e.g. 1-year intervals).

The cycle must be a period relevant to the condition considered (e.g. 6 months, 1 year). At the end of each cycle, individuals can remain in the state in which they started the cycle or move to a different state. Movement is dependent on the transition probability, which is defined as the probability of moving from one state to another. 115 Such models can be useful in capturing the costs and benefits of treating chronic medical conditions, such as glaucoma, over time. 116 In addition, all Markov models have at least one absorbing state, typically death, which all individuals will eventually enter if the model is given a sufficiently long time horizon. The sum of the cost in each year and the product of the utilities in each year were summed over the time horizon of the simulated patient cohort to compute total cost and quality-adjusted life-years (QALYs) for that cohort. 115

Within the model, and mirroring the trial inclusion and exclusion criteria, on entry into the model patients could have disease in one or both eyes. Again mirroring the trial inclusion criteria, one eye is defined as the index eye in terms of initial treatment and, in terms of disease severity, has advanced glaucoma, as this was also an inclusion criterion for the trial. Based on the (modified) disease classification (GSS2), we defined the stages for the economic model shown in Box 1. Other possible options for model structure are provided in Appendix 7, Tables 48–52.

We adopted this model structure because it allows us to model both eyes independently and estimate the chance of unilateral and bilateral progression more precisely than modelling a single eye, even with a cohort Markov model. This is especially important in the case of eye disease because HRQoL and, therefore, QALYs are thought to be determined by the quality of vision in the better eye. 117 The model allows progression of the disease in both eyes over time. For the purposes of modelling, the time step over which progression could occur is 1 year because glaucoma is generally only slowly progressive. 116

Within the model, the mean age and sex distribution of the modelled cohort matches that of the trial participants at baseline. For those with binocular disease, the risk of visual loss in each eye is modelled independently. The model structure is presented in Figure 9.

FIGURE 9.

Model structure.

Treatment pathway

A patient starts with one of the two randomised treatments (trabeculectomy or medical management). The medical treatment strategy was chosen as the comparator and trabeculectomy as the intervention in this trial. A patient is categorised in different stages of glaucoma (based on the GSS2 classification system; see Box 1) for both the trabeculectomy arm and the medical management arm. The initial and transition probabilities of patients in each stage are based on trial data. Disease progression may occur during the trial and some patients move to more severe levels of glaucoma. Disease progression in the model is determined by a set of transition probabilities, each of which is defined in the sections below. Patients may also experience some complications that may affect related cost/utility in both arms. Therefore, complication-related cost/utility has been included in the model as well.

In the model, a patient may progress through a number of different states, each represented by increasing severity of disease (Figure 9). The Markov model used in this study (see Figure 9) considers increasing severity of glaucoma for each eye separately, as has been undertaken by others. 118 The level of the patient‘s disease on entry into the model determines the Markov model state at which the patient starts. The progress of patients through the model reflects disease progression, but it possible that patients can stay in a state and not progress further. However, they cannot move back to a less severe state because glaucoma is an irreversible disease. 117

During each model cycle, a portion of the cohort progresses based on the probabilities of progression derived from the analysis of the trial and other relevant data. In addition, it is assumed that patients move sequentially between states and cannot skip states because of the relatively slow evolution of glaucoma. 106 Finally, death from all causes is included in the model as an absorbing state (a state that someone can enter but cannot leave). Transition probabilities to this state are assumed to be independent of severity and treatment history and are derived from age-/sex-specific UK life tables. 119 Patients who are blind (in one or both eyes) are taken to have a higher risk of death than the general population and, hence, standardised mortality ratios were used to adjust the risk of death for those who are blind in one or both eyes. 120 Surgical mortality was not included in the model. All of the programming for the model was implemented in TreeAge Pro 2019.

Model assumptions

• The model-based economic evaluation analysis took the perspective of the NHS and Personal Social Services (PSS).

• For extrapolation, the cost and utility values (by clinical severity state and treatment allocation) beyond 24 months (beyond the duration of trial follow-up) in the model were assumed to be the same as those incurred in the second year.

• Independent variables (e.g. progression state, have trabeculectomy treatment) and constant value of year 1 were assumed as a proxy to estimate the cost and utility for the second year.

• All model input parameters are defined as statistical distributions in the model, allowing probabilistic sensitivity analysis to be conducted. 107

• Ranges and distributional assumptions for input parameters were based on the trial data. We assigned gamma distributions for costs and beta distributions for utility data. 121 We have described these in more detail in Model parameters.

• All future costs and QALYs were discounted at 3.5% per annum, the UK recommended rate, as the duration of follow-up (time horizon) was > 1 year. 122

Model parameters

Utilities specified in the model

Utilities associated with each health state were attached to the modelled severity states, allowing cumulative model-based QALYs to be estimated in both (trabeculectomy/medical management) arms. We assigned a zero utility weight for death. The mean QALYs for each intervention were calculated by multiplying the amount of time that patients spent in each health state by the associated health state utility values. The estimated utility used in the model was based on data retrieved from the trial using instruments such as EQ-5D-5L;123 data from the trial were also available for the HUI-3124 and the GUI,55 and these data have been used in the sensitivity analysis.

As with costs, regression techniques were carried out to derive the drivers of the difference in QALYs for a 12-month period between the arms of the trial (trabeculectomy vs. medical management) after controlling for the key predictors of QALYs. The purpose of this regression was to determine the consequences of a glaucoma state (based on GSS2 measure) and its effect on a participant’s overall QALYs:

The variables considered were:

• dependent variable – QALYs (total QALY score across the two arms of the trial controlling for the independent factors)

• independent variables – progression rate (P) and trabeculectomy (T).

Here, the dummy variable for the treatment arms estimates the difference in QALYs between the trabeculectomy and the medical management arms, controlling for all other factors in the model. For example, if the dummy is specified as trabeculectomy arm = 1 and medical management arm = 0 and the beta value of the coefficient is 0.50, this indicates that the trabeculectomy (coded as 1 in the dummy) provided 0.50 more QALYs over 12 months than current practice, on average, after controlling for all other factors. If the coefficient was –0.50, the QALY gain resulting from the intervention would, on average, be –0.50 lower than that achieved by medical management. As with costs, we also used Cholesky decomposition and assigned multinormal distributions to these parameters to enable the model for sensitivity analysis. 121

Model validation

The model was validated by checking the model structure, calculations and data inputs for technical correctness. 115 The structure was reviewed by clinical experts from our TAGS team to establish that it was appropriate for the disease and its treatment. The robustness of the model to change input values was tested using sensitivity analyses to ensure that any changes to the input values produced changes to the results of the expected direction and magnitude. To establish its external consistency, the model results were compared with outcomes reported in other trials and other economic evaluations. Moreover, we assigned the same time horizon of 24 months of for the trial-based economic evaluation (reported in the first half of this chapter) to the model-based analysis (2 years) and compared the model-based with the within trial-based incremental cost-effectiveness analysis to check the accuracy of the model compared with the within trial analysis.

Base-case analysis

The base-case analysis considered a cohort of 67 year olds (65% were male) as the start of the model. Each person had severe glaucoma in at least one eye. The costs and outcomes were modelled over their estimated lifetimes (i.e. a lifetime horizon adopted). As described above, the analysis adopted a UK NHS and PSS perspective, the cycle length was set at 1 year and a discount rate of 3.5% for costs and benefits was used. 122

The results of the model are presented in ICERs. The ICER is a ratio of the difference in costs divided by the difference in the effectiveness between two alternative strategies. These data can be interpreted as how much society would have to pay for an extra unit of effectiveness. 108 Central to the assessment of cost-effectiveness is the value that society would put on gaining an additional QALY. NICE states that ‘Below a most plausible ICER of £20,000 per QALY, judgements about the acceptability of a technology as an efficient use of NHS resources are based primarily on the cost-effectiveness estimate’122 (© NICE 2018 Guide to the Methods of Technology Appraisal 2013. Available from www.nice.org.uk/process/pmg9/resources/guide-to-the-methods-of-technology-appraisal-2013-pdf-2007975843781. All rights reserved. Subject to Notice of rights. NICE guidance is prepared for the National Health Service in England. All NICE guidance is subject to regular review and may be updated or withdrawn. NICE accepts no responsibility for the use of its content in this product/publication). Between £20,000 per QALY and £30,000 per QALY, judgements about the acceptability of the technology should take into account factors such as the degree of uncertainty surrounding the calculation of ICERs, the innovative nature of the technology, the particular features of the condition and population receiving the technology, and, where appropriate, the wider societal costs and benefits. Above an ICER of £30,000 per QALY, the case for supporting the technology on these factors has to be increasingly strong. In the absence of a more definitive statement, this report focuses on a willingness-to-pay threshold of £30,000 for a QALY.

To further this aim, we have also adopted the net monetary benefit (NMB) statistic. NMB represents the value of an intervention in monetary terms when a willingness-to-pay threshold for a unit of benefit (e.g. QALY) is known. 125 A net benefit in terms of NHS costs and benefits, expressed in commensurate units, was calculated for each participant in the model according to Equation 3:

where i = individual and WTP = maximum WTP threshold for a QALY.

Evaluation of uncertainty

Both probabilistic and deterministic sensitivity analysis were used to explore parameters and other forms of uncertainty surrounding model-based estimates of cost-effectiveness. 115 The deterministic sensitivity analyses were conducted to investigate the impact of varying key assumptions and/or parameter values used in the base-case analysis. A deterministic sensitivity analysis with the results presented in the form of tornado diagrams was also conducted. For the deterministic sensitivity analysis, we explored the impact of the results of changes in costs of medical management in year 1 (Cost_Year1_MD). For this analysis, the costs of medical management in year 1 were varied between £1500 and £1700. We likewise varied the cost of trabeculectomy in year 1 (Cost_Year1_TB). The cost of this was varied between £2000 and £4000. The cost of year 2 (Cost_Year2) was also varied in the range £700–10,000. Similar deterministic sensitivity analyses were conducted for utilities. For the utility of year 1 for medical management (Util_Year1_MD) and the utility of year 1 for trabeculectomy (Util_Year1_TB), utility values were varied between 0 and 1. Other deterministic sensitivity analyses included changing the time horizon for the analysis from 0 to 30 years; changing the age at which severe glaucoma started (start age) by varying the age on entry into the model from 30 to 80 years; and changing the percentage of males and females (sex ratio), with the percentage of females varying between 40% and 60%.

Alternative scenarios have been explored using PSA. PSA is a technique used in economic modelling that allows the modeller to quantify the level of confidence in the output of the analysis, in relation to uncertainty in the model inputs. 125 For the PSA, all parameter values, including costs and utilities, were defined as statistical distributions using parameters, such as a mean and SD, for a normal distribution, for example. Ranges and distributional assumptions for input parameters were based on the trial data. In a PSA, a set of input parameter values is drawn by random sampling from each distribution, and the model is ‘run’ to provide the results as probability of being cost-effective for different thresholds. This is repeated many times (typically 1000 to 10,000), resulting in a distribution of outputs that can be graphed on the cost-effectiveness plane and analysed. Different distributions are generally appropriate for different types of variable. 125 We assigned gamma distributions for costs and beta distributions for utility data. 121 We also calculated correlations between the coefficients of cost and utility for the variables included logistic regression analyses using Cholesky decomposition and assigned multinormal distributions to these parameters in the model to account uncertainty in the estimated transition probabilities.

A key output of a PSA is the proportion of results that fall favourably (i.e. considered cost-effective) in relation to a given cost-effectiveness threshold. This may be represented using a plot of costs and QALYs and a CEAC. The CEAC is a graph summarising the impact of uncertainty on the results of an economic evaluation, estimating the NMB for each strategy for each iteration of the Monte Carlo simulation so that a CEAC can be generated. The uncertainty surrounding the cost-effectiveness of the treatment is represented on a CEAC according to the probability that the intervention will be cost-effectiveness at a particular willingness-to-pay threshold.

Results

This section reports the proportion of patients in the cohort models by the level of severity of glaucoma in both eyes over time. It also reports the model-based cost-effectiveness results for primary medical treatment with primary augmented trabeculectomy (glaucoma surgery) for patients presenting with advanced glaucoma in terms of costs and QALYs. Results in terms of costs and QALYs are presented for each arm, with costs and benefits discounted at 3.5% per annum. 122

Progression from one heath state to another

In the analysis, all patients start with severe glaucoma in at least one eye (see Figure 9). Patients were assumed to progress from less severe disease states to more severe disease states. Once patients were in a more severe state, they could either remain in that state or continue to progress to the next, more severe, disease state.

For each cohort of patients with advanced glaucoma, we determined disease progression from lower stages of glaucoma to higher stages based on the TAGS trial data. We used the GSS2 progression rates to capture the progression for both the trabeculectomy arm and the medication treatment arm who progressed from less severe to more severe disease states. We generated plots of progression through disease states on a year-by-year basis for both arms for a 30-year time horizon (Figure 10).

FIGURE 10.

Markov tracings of disease states over time for (a) the trabeculectomy arm; and (b) the medical management arm.

Incremental costs, quality-adjusted life-years and cost-effectiveness

Table 21 shows the QALYs (for EQ-5D-5L, HUI-3 and GUI, presented separately), total cost, incremental cost per QALY and NMB for three time horizons (and utility measures). In the base-case analysis (lifetime horizon and EQ-5D-5L-based QALYs), trabeculectomy had an additional cost of £2687, an additional 0.28 QALYs and an incremental cost per QALY gained of £9679 compared with primary medical treatment. Furthermore, the results of the PSA show that, should society be willing to pay £20,000 per QALY, the likelihood of trabeculectomy being cost-effective compared with medical treatment would be 73%.

The model-based estimates of mean costs and QALYs at 2 years indicate that trabeculectomy for the treatment of glaucoma is expected to cost an additional £2106, on average, and to provide an average QALY gain of 0.04 compared with medical management. The corresponding incremental cost per QALY for surgery compared with medical management would be £47,663 (see Table 21). At a 10-year time horizon, the corresponding figures are £2362 and 0.17 QALYs (ICER £13,911). When running the model over 10 years, the results of the PSA show that, should society be willing to pay £20,000 per QALY, the likelihood of trabeculectomy being cost-effective compared with medical treatment would be 59% (see Table 21). Table 21 also reports the incremental cost, incremental QALY and incremental cost per QALY when QALYs are based on HUI-3 or GUI utility values. Figure 11 presents a plot of costs and QALYs in the base-case analysis. Similar results for the analyses using HUI-3- and GUI-based QALYs are presented in Appendix 7, Figures 21 and 22.

FIGURE 11.

Cost-effectiveness analysis in the base-case analysis.

Sensitivity analysis

One-way deterministic sensitivity analysis

Deterministic sensitivity analysis shows that the model-based findings are generally robust to the changes examined. Main model inputs were varied in the sensitivity analyses to determine whether or not they may affect the ICER of trabeculectomy versus medical management (Table 22). The results of these sensitivity analyses are depicted in the form of a tornado diagram (Figure 12). The tornado diagram helped to identify main inputs that could be altered to make trabeculectomy more or less cost-effective relative to medical management. Similar analyses are shown in Appendix 7, Figures 23 and 24, for the HUI-3 and GUI, respectively.

TABLE 22 - Sensitivity analyses of main parameters in the model

Parameter	Value	ICER (£ per QALY)
		EQ-5D-5L	HUI-3	GUI
Cost of year 1 (trabeculectomy) (£)	2000	5594	4055	9713
	3000	9196	6666	15,967
	4000	12,798	9277	22,222
Cost of year 1 (medical management) (£)	700	10,986	7963	19,076
	1100	9546	6919	16,574
	1500	8105	5875	14,072
Utility of year 1 (trabeculectomy)	0.2	Dominated	Dominated	Dominated
	0.6	70,594	14,483	Dominated
	1	6134	4589	8625
Utility of year 1 (medical management)	0.2	3012	2819	3331
	0.6	5461	4856	6606
	1	29,197	17,528	395,757
Cost of year 2 (£)	700	7726	5600	13,415
	1100	7878	5711	13,679
	1500	8031	5821	13,944
Utility of year 2	0.2	52,124	50,146	96,391
	0.5	34,466	33,590	49,497
	1	20,546	20,231	25,087
Time horizon (years)	2	47,663	39,724	147,247
	10	13,910	10,506	24,179
	30	9679	7016	16,805
Age at (severe) glaucoma (years)	30	7924	5614	13,914
	42	8037	5703	14,099
	55	8406	5997	14,701
	67	9776	7093	16,972
	80	14,438	10,769	25,930

FIGURE 12.

Tornado diagram for the main parameters (EQ-5D-5L measure). EV, expected value; WTP, willingness to pay.

Impact of missing data

The base-case analysis made use of the EQ-5D-5L measure and lifetime horizon duration. Running multiple imputation shows a slight difference in the incremental cost value (£2707 vs. £2687) and no difference in the incremental QALYs value (0.28) compared with the base-case analysis. The input parameters of the model were based on a data set where multiple imputations (MI) were performed for missing data. These data were used to parameterise uncertainty surrounding the joint incremental costs and effects. This is presented graphically as confidence ellipses on the incremental cost-effectiveness plane (Figure 13). Moreover, the model-based CEACs based on the lifetime horizon are presented in Figure 14. The related results of all other scenario analyses (HUI-3 and GUI QALYs measures) are presented in Appendix 7; for the incremental cost-effectiveness plane these are shown in Figures 25 and 26, respectively. The CEACs are shown in Appendix 7, Figures 27 and 28, for the GUI and HUI-3, respectively.

FIGURE 13.

Incremental cost-effectiveness scatterplot: trabeculectomy vs. medical treatment in the base-case analysis. WTP, willingness to pay.

FIGURE 14.

Model-based CEAC for the base-case analysis (lifetime horizon; EQ-5D-5L measure).

Discussion

The focus of the present model-based analysis was to perform a comparison of the cost-effectiveness of trabeculectomy and medication for treatment in patients with severe glaucoma. This analysis was completed alongside the trial-based economic evaluation and was based on TAGS data, but sought to extrapolate findings over the expected lifetime of patients treated with either trabeculectomy or medical therapy.

Quality of life

The outcome of management of glaucoma that is most important to patients is their ability to continue to live an independent life and maintain their QoL. 126,127

Quality of life

Intraocular pressure

Both of the interventions employed in TAGS are recognised to effectively lower IOP and are recognised by NICE. 65

In TAGS, the investigators were advised to use the guidance provided by the Canadian perspectives in glaucoma management: setting target IOP range to set the IOP level for individual patients; this suggests an IOP for advanced disease of < 15 mmHg, with reduction of ≥ 30% from the baseline IOP. 137

There was a clear reduction in IOP in both arms for the duration of the study from similar baseline values (holding medical management implemented at diagnosis). In the trabeculectomy arm, IOP fell to 12.4 mmHg (SD 5.73 mmHg) at 4 months and remained at around 12 mmHg for the remainder of the study. In the medical management arm, the IOP reduction was initially 4 mmHg lower than that achieved in the trabeculectomy arm [16.4 mmHg (SD 4.12 mmHg) at 4 months and 16.12 mmHg (SD 4.54 mmHg) at 12 months]. There was a further reduction in the medical management arm to 15.07 mmHg (SD 4.8 mmHg) at 24 months, but this reduction was still 3 mmHg lower than that achieved in the trabeculectomy arm. Therefore, for the duration of the study an additional 3–4 mmHg reduction in IOP was achieved in the trabeculectomy arm compared with the medical management arm.

At diagnosis, the IOP was 26.9 mmHg (SD 9.1 mmHg) in the trabeculectomy arm and 25.9 mmHg (SD 8.4 mmHg) in the medical management arm; therefore, the 24-month IOPs represent a reduction of 14.9 mmHg (55%) in the trabeculectomy arm and 10.83 mmHg (42%) in the medical management arm.

The value of medical management in prevention of VF or optic disc progression in patients with ocular hypertension, primary OAG and normal tension glaucoma is well established. 62 Two recent RCTs, both using prostaglandin analogues as primary treatment, as per NICE guidelines,65 have been undertaken. In the UKGTS,58 the baseline untreated IOP was 19.6 mmHg (SD 4.6 mmHg) in the medical management arm; by 24 months, an average reduction of 4.0 mmHg (SD 3.4 mmHg) had been achieved, resulting in an IOP of 15 mmHg, mostly on a single medication. The use of prostaglandin monotherapy had a clear benefit in terms of lowering IOP compared with no treatment, and this translated into a statistically and clinically significant reduction in VF progression rate at 24 months. 58 There was a 50% difference in VF progression in the medical management arm compared with the no treatment arm, further establishing the relationship between IOP lowering and VF progression.

The LiGHT trial59 compared selective laser trabeculoplasty with medical management in the treatment of patients presenting with ocular hypertension and mainly early primary OAG. In the primary OAG group, the baseline IOP was 23 mmHg (IQR 20–27 mmHg). In the medical management arm, IOP fell to a mean of 16.6 mmHg, and only 35% of patients required more than one type of eye drops to achieve this. Despite these results, however, 11 patients required trabeculectomy to prevent further progression.

In the medical management arm of the TAGS trial, a much greater reduction in IOP was achieved in terms of both the absolute IOP reduction and the percentage drop in IOP. This probably reflects the guidance provided for IOP lowering and the recognition that vision is more likely to be preserved in patients with advanced disease if IOP can be reduced.

The IOP lowering achieved in the trabeculectomy arm of the TAGS trial is consistent with current results61,138,139 reported from the NHS. Kirwan et al. ’s61 multicentre service evaluation of augmented trabeculectomies found that the mean IOP achieved at 24 months following augmented trabeculectomy was 12.4 mmHg (SD 4.0 mmHg). 61 Stead and King138 looked at a group of eyes with advanced glaucoma (mean deviation < –20 dB) and found that, at 24 months, the mean IOP in the cohort was 13.1 mmHg (SD 5.4 mmHg) and the mean number of different types of eye drops required was 0.6. In both of these studies, longer-term follow-up indicated a continued sustained low IOP.

Since the most recently undertaken trial of primary trabeculectomy compared with medical management, CIGTS,31 which recruited in the early 1990s, there have been significant advances in the technique of trabeculectomy surgery140 in terms of both anti-metabolite use and post-operative management, which now includes a series of routinely undertaken adjustments to improve IOP control. 141 Therefore, both the safety and the efficacy of trabeculectomy have been improved in this time. 142,143 Similarly, in CIGTS, the primary medical management was initial treatment with a beta-blocker followed by escalating medical therapy, as required. Since CIGTS recruited, several further agents have been introduced, and one type in particular, the prostaglandin analogues, has superior IOP lowering, can be applied once per day and has a low side effect profile and, therefore, is well tolerated, improving both persistence and adherence to the medication. 63 These improvements may reflect that in CIGTS the mean IOP was 15 mmHg in the trabeculectomy arm and 18 mmHg in the medical management arm post treatment; in both cases this was considerably higher than our results. One caveat, however, is that TAGS patients have advanced glaucoma and it is possible that clinicians may have undertaken a more aggressive treatment approach as indicated by consensus guidance to pursue a lower IOP and treat more aggressively, thus aiming for and achieving a lower IOP.

A sustained reduction in IOP is recognised to be the most effective method of preventing further VF loss in glaucoma. 33,34,58,144 In AGIS, long-term follow-up over 8 years indicated that lower IOP in both the predictive analysis (early IOP level behaviour) and the associate analysis (long-term IOP level behaviour) resulted in less VF loss. In the associate analysis, the cohort with the lowest sustained IOP, < 18 mmHg at all follow-up visits, had virtually no VF loss progression. For this arm, this equated to a mean IOP of 12.3 mmHg, which is virtually identical to that achieved for the trabeculectomy arm of TAGS. A difference of 3 mmHg between treatment arms is a clinically important difference and achieving a sustained IOP of 12 mmHg is a clinically important level of IOP reduction for minimising further VF progression.

Glaucoma drop usage

This profile of IOP control is also reflected in the need for glaucoma eye drops. At 4 months, only 61 participants (28%) in the trabeculectomy arm were using eye drops, equivalent to a mean of 0.43 (0.79) different types of eye drops and reflecting the fact that the majority of patients had undergone trabeculectomy and, consequently, were drop free. At the same time point, 210 (96%) participants in the medical management arm were using a mean of 1.77 (0.92) different types of eye drops. At 12 months, eye drop usage in the medical management arm remained unchanged, whereas there was further reduction in the trabeculectomy arm, probably because some patients underwent trabeculectomy as part of treatment allocation after only 4 months. It is also important to remember that 16 patients declined surgery following randomisation to the trabeculectomy arm and, therefore, remained on eye drops for IOP control. By 24 months, there was a small increase in the number of eye drops in the trabeculectomy arm, suggesting some additional medications required to maintain IOP at a sufficiently low level, and this is consistent with observations from other studies reporting trabeculectomy outcomes. 138,145 By contrast, in the medical management arm, there was a reduction in the mean number of eye drops used and the number of participants using them, reflecting the fact that, during this period, 39 patients (17.3%) required trabeculectomy to control their IOP adequately, following failure of medical management, and subsequently stopped eye drops.

Visual acuity

There is no evidence of a difference in logMAR VA at baseline between the arms. However, at 4 months the VA in the trabeculectomy arm dropped from 0.15 (SD 0.25) to 0.25 (SD 0.31), whereas it remains unchanged in the medical management arm (p < 0.001). This is likely to correspond to some visual reduction experienced by patients undergoing trabeculectomy during the postoperative recovery phase and has resolved by 12 months. At 24 months there has been some further deterioration, VA falling to 0.21 (SD 0.28) in the trabeculectomy arm compared with 0.16 (SD 0.26) in the medical management arm (p < 0.006). It is likely that this deterioration is because of the development of cataract in the trabeculectomy arm that has not yet been removed. The development of cataract is well recognised following trabeculectomy146 and in the CIGTS study cataract extraction was significantly higher in the trabeculectomy arm as opposed to the medical management arm in the first 5 years following randomisation. The hazard ratio for cataract extraction in the trabeculectomy arm compared with the medical management arm was 3.76 (95% CI 2.16 to 6.54; p < 0.001) for CIGTS. 147 This difference was clearly visible by 24 months, at which those in the trabeculectomy arm were nearly six times more likely than those in the medical management arm to have had cataract surgery.

Safety

In the UK, consultants did not consider medical management to be better or equivalent to primary augmented trabeculectomy when surveyed; however, there was a concern about exposing patients to the risks associated with trabeculectomy and this was the major barrier to carrying out primary surgery. However, clinicians indicated that if robust evidence existed that supported primary surgery they would change their practise accordingly. 35 The perceived ‘high risk’ associated with trabeculectomy may stem from the National Trabeculectomy Study performed in the late 1990s, which indicated particularly high rates of early complications following trabeculectomy. 151 Improvement in surgical techniques has greatly reduced these perceived risks. 142,143 In addition, virtually all surgery is now carried out by fellowship-trained glaucoma specialists instead of primarily general ophthalmologists, as was the case at the time of the National Trabeculectomy Survey.

One of the main concerns of clinicians is the unexplained irreversible loss of central vision – ‘wipe-out’ – believed to occur in patients with advanced glaucoma undergoing trabeculectomy. 152 There is no suggestion of unexplained vision loss occurring immediately after surgery in TAGS, which indicates no episodes of wipe-out. In two prospective studies specifically exploring the development of ‘wipe-out’, no cases of severe, irreversible, unexplained central vision loss were identified, which suggests that ‘wipe-out’, if it exists, is a very infrequent occurrence. 153,154

Although visual loss is not uncommon after trabeculectomy, in the majority of cases this is reversible and explained;153–156 this is the case in TAGS, for which all episodes of visual loss have been identified.

Treatments received

In the trabeculectomy arm, 201 participants (88.5%) received surgery in their index eye and 34 participants (15.0%) received a trabeculectomy in their non-index eye (see Appendix 3, Table 27). Of these, four participants did not have a trabeculectomy in their index eye. In the medical management arm, 30 participants (17.3%) received trabeculectomy (see Table 4). The reason for participants receiving trabeculectomy in the medical management arm was mainly because of uncontrolled IOP.

Sixteen participants in the trabeculectomy arm declined to have trabeculectomy once allocated to the trabeculectomy arm. It is well recognised that some patients who agree to participate in a trial refuse to have the allocated treatment once randomisation has occurred. 162 In addition, several participants with bilateral advanced glaucoma who were allocated to the trabeculectomy arm did not have surgery performed on their index eye. For these patients, there was a reluctance on the part of patients and patients’ clinicians to perform trabeculectomy on the better eye (eye with greater mean deviation value) first and, therefore, in such cases, a decision was made to operate on the worse eye (non-index eye) first and then operate on the index eye when this had settled; however, this did not always occur.

Economic evaluation

The EQ-5D-5L has been a recommended QoL measure used to estimate QALYs in economic evaluations in the UK. However, a limitation of this generic QoL measure is that it may not be sensitive to small but clinically important changes. 111,112 The GPI and the GUI were developed by Burr et al. 55 as a health outcome measure that could be used to estimate preference-based weights in glaucoma. 104 These weights can be used as an alternative to the EQ-5D-5L to estimate QALYs to identify whether or not preference weights from a disease-specific QoL measure are more sensitive to changes in QoL than those from the EQ-5D-5L in individuals with glaucoma. The GUI QoL weight values currently available were estimated based on the responses to a DCE from individuals with self-reported mild or moderate glaucoma, with few responses from those with advanced glaucoma. 104 The aim of the DCE study was to estimate the QoL weights for the GPI in a population with advanced glaucoma and incorporate these values into the economic evaluation.

As expected, participants had a preference for reduced difficulty in the attributes central and near vision, lighting and glare, mobility, and activities of daily living compared with the base level (i.e. severe difficulty). Although some attribute values were estimated to be, on average, negative, there was no evidence that these attribute levels were any different compared with severe difficulty.

The QoL weights were estimated using the results from the logistic regression under the assumption that the health state 111111 (no difficulty with any of the attributes) was equivalent to 1 (perfect health) on the QoL scale. The values for the other attribute levels were anchored on this value to maintain the ratio between the attributes and the levels. Levels with a negative coefficient were assumed to be zero in the estimation of the QoL weights because these levels were close to zero and not statistically significant.

A sensitivity analysis was undertaken incorporating the QoL weights estimated from the DCE and participants’ responses to the GUI questionnaire administered at baseline and at 1, 3, 6, 12, 18 and 24 months post randomisation to estimate QALYs and subsequently the cost-effectiveness of surgical intervention in the management of glaucoma. On average, the GUI produced higher average total QALYs for both interventions than the EQ-5D-5L and HUI-3. There was, however, a smaller difference between treatment arms.

The results of the DCE, support the use of the EQ-5D-5L in individuals with advanced glaucoma. Out of the six attributes valued in the DCE, we found a strong preference for three attributes (central and near vision, mobility, and activities of daily living) with very little utility to be gained from reduced difficulty in the other three attributes. Two of the attributes preferred by individuals with advanced glaucoma are also included in the EQ-5D-5L (mobility and usual activities); therefore, any benefit from an intervention that targets these attributes can arguably be captured by the EQ-5D-5L.

This trial sought to address a paucity of economic evidence for medical compared with surgical interventions in the management of advanced glaucoma. 30 In the economic analysis, the trabeculectomy arm was, on average, more costly and provided more QALYs at 24 months. At 24 months, it was unlikely that trabeculectomy would be considered cost-effective compared with the medical therapy over the range of values for society’s willingness to pay for a QALY considered. The driving cost was the initial costs of the trabeculectomy procedure incurred in the in the first year of the trial in the trabeculectomy arm.

The economic evaluation model suggests that trabeculectomy is likely to be a cost-effective approach compared with medication in patients with severe glaucoma over a lifetime time horizon. The reason for this is that the initial costs (cost of year 1 treatment) are higher for patients assigned to trabeculectomy and these are partly offset in the longer term because these patients receive fewer subsequent procedures and have lower medication use.

It should be noted that there is a small difference in the estimated mean costs and QALYs when comparing the results of the within-trial analysis with those of a 2-year time horizon model-based analysis. These differences are explained partly by the way that we used the trial data to populate our model and the impact that ‘random noise’ has on the Monte Carlo simulation. A further difference between the trial and the model-based analysis is that the model used age-specific and sex-specific probabilities of death derived from UK life tables, rather than the observed data from obtained from the trial data set.

Discrete choice experiment

The aim of the DCE was to identify the preferences of individuals with advanced glaucoma and update the QoL weights estimated by Burr et al. 55 Burr et al. 55 estimated their QoL weights from a population with self-reported mild or moderate glaucoma, with only 15% of the sample reporting severe glaucoma. Participants were asked to complete 36 choice-set questions, four of which were rationality tests. The response rate (70% vs. 62%) and number of participants who attempted the choice-set questions (n = 308 vs. n = 289) were similar. Our results were comparable to Burr et al. 55 in that the strongest preferences were the attributes central and near vision, mobility, and activities of daily living. The ranking of the attribute levels was also similar. Similar to Burr et al. ,55 the remaining three attributes (i.e. lighting and glare, eye discomfort, and other effects of glaucoma and its treatment) were not statistically significant in our model but, on average, there were negative preferences. In a sensitivity analysis, we replicated their analysis by combining the levels of these attributes, however, our conclusions did not change and these attributes were not statistically significant. Hence, we can conclude that participants with advanced glaucoma do not have a strong preference for improvements in these attributes (lighting and glare, eye discomfort, and the effects of glaucoma and its treatment). A limitation of combining the attribute levels is that the ratio of preference for that attribute level compared with severe difficulty is lost in the estimation of utility scores. Second, in Burr et al. 55 all three levels of lighting and glare (no difficulty, some difficulty and quite a lot of difficulty) were combined but the level no difficulty has a different QoL weight to the other two levels, which were assumed to be zero. This is inconsistent with how they created QoL weights for the other attribute levels that were combined.

Cost-effectiveness

Within the 24-month time horizon of the trial, medical management was likely to be cost-effective over the range of values for society’s willingness to pay for a QALY considered. This recommendation changes when the benefits were extrapolated beyond the time horizon of the trial and are considered across the participant’s lifetime. This is important because the median age of glaucoma diagnosis in TAGS is 69 years. Their current average projected life expectancy is 15 years, according to published figures regarding average life expectancy. 119

For the model-based economic evaluation when a lifetime horizon is taken, trabeculectomy is likely to be considered as cost-effective compared with medical management over the range of values for society’s willingness to pay for a QALY considered compared with medication.

There are other studies that have estimated the cost-effectiveness of surgery compared with medical therapy for glaucoma treatment in the UK population. These studies were conducted in different populations of glaucoma patients and gave conflicting results. 112,163–165 Stein et al. 112 compared medications with laser trabeculoplasty for the treatment of patients with newly diagnosed mild OAG. They used a Markov model with a 25-year time horizon; the results of this suggested that medication was superior to laser trabeculoplasty, although their analysis assumed optimal medication adherence. 112 Choi et al. 113 conducted a cost-effectiveness analysis comparing medication, laser trabeculectomy and trabeculectomy in a South Korean population with mild OAG. Their analysis reported that medication was cost-effective compared with laser trabeculectomy and trabeculectomy. 113 These two studies have limited relevance to our study because they focused on treatments for patients with less severe glaucoma. Only one study was identified that also considered more severe glaucoma, as we have done in TAGS. Guedes et al. ’s166 analysis was based on Markov models and aimed to identify the most cost-effective treatment strategy for each severity of glaucoma. 166 The results of this study were similar because this study found that the surgery is cost-effective in participants who are aged < 70 years. Similarly, this study finds that patients who will live for at least 10 years after surgery are more likely to have the benefits outweigh the investment. 166

Strengths and limitations

This pragmatic study is representative of what currently happens to patients presenting with advanced glaucoma being managed in the NHS. Consequently, the concluding findings on disease progression, reduction of IOP, vision retention, safety profile of interventions and cost are highly relevant to normal clinical practice.

This trial was unmasked, which allowed the trial to capture any treatment effects on patients’ perception, which is clinical reality and reflects the patients experience of the intervention. It answers specific concerns of clinicians regarding the safety of the interventions, which was a major barrier, particularly for consideration of trabeculectomy as a primary procedure.

A common limitation of DCEs is that they can be difficult for participants to understand. However, the majority of those who responded to the DCE questionnaire (n = 283, 92%) completed all 15 choice-set questions, which suggests that they had a good understanding of the DCE exercise.

Unlike Burr et al. ,55 we found negative preferences for 5 of the 18 levels included in the regression model; however, these preferences were not statistically significant. Thus, we have no evidence that the coefficient value is different from the worst state, severe difficulty. We assigned these coefficients a QoL weight of zero where we had no evidence of any difference between these attribute levels and the most severe level (e.g. no difficulty compared with severe difficulty). This assumption potentially created a bias in our QoL weight values as positive coefficients (which were not statistically significant) were assigned a positive QoL weight, albeit close to zero, and it reduced the ratio of preference between the attributes as the negative values were given a slightly higher weighting by assuming that they were zero. However, it is unlikely that this assumption had any implications on our overall results because when the attribute levels, which were not statistically significant, were combined in a sensitivity analysis the results remained consistent, suggesting that there was no evidence of individuals having a preference for reduced difficulty in these attributes. The second possible bias is that our assumption of assigning a zero value (i.e. levels are equivalent) makes the assumption that evidence of no difference is the same as the there being evidence of no difference in levels. An alternative would be to incorporate the imprecision in estimates into the estimation of QALYs. Although this may be possible, we are unaware of any work that has undertaken this.

There was also an issue with theoretical validity in our model because we found a stronger preference for some difficulty with activities of daily living than for no difficulty with activities associated with daily living, and we had hypothesised that the coefficients would increase with reduced levels of difficulty. A reason for this could be that using a d-efficient design meant that there was no attribute level balance that may have produced this unusual finding. However, a d-efficient design is still recommended because it optimises efficiency as it reduces the standard errors (uncertainty) in the model estimation. 85 Finally, asking participants which alternative; they consider worse is more cognitively challenging than choosing the best alternative; this may explain some of the negative preferences estimated from our model. Further estimation work will be undertaken to explore these negative preferences.

For the economic evaluation, one of the key limitations of the within-trial analysis is the limited time horizon. Primary OAG is a chronic lifelong condition, so the capture of full benefit of any interventions is unlikely to be shown within this time frame. It is also possible that there have been issues with recall for individual participants when completing the resource and QoL questionnaires. A recent study that examined the use of a laser-based intervention and medical management strategy for patients presenting with earlier stage glaucoma was the LiGHT trial. 59 The LiGHT trial assessed selective laser trabeculoplasty versus medical management for glaucoma and ocular hypertension. 59 This study concludes that laser is very likely to be cost-effective within the 3-year trial outcomes. This, potentially, is for a number of reasons. First, a selective laser trabeculoplasty could be a smaller initial intervention cost (the economic results of the trial are yet to be published in full) and that the benefits were examined for an additional year. This would allow the benefits of the initial investments to accrue over a longer time period. It is for this reason that the results of the within-trial analysis should be considered alongside the decision model. Extrapolating the results beyond the initial treatment period could accrue benefits into the future and change the decision regarding which intervention is likely to be cost-effective. It is possible that trabeculectomy may be the most efficient use of resources in a patient with a greater life expectancy who will likely accrue the benefit.

Another consideration is the inclusion of the costs of the non-index eye. When considering the calculation of total costs, the costs for both the index eye and the non-index eye were included. This was the case for a number of reasons. First, the outcome measures of principle interest were not specific to the index eye and when considering vision and QoL the eyes do not work independently of one another. Second, the prognosis of one eye may affect the decisions regarding the management of the other eye. This can be a difficult consideration when carrying out work with vision-related QoL because there is a potential to under value or over value any potential costs. In this study, as shown in Table 20 within the context of the model, the spread of the severity of glaucoma in the non-index eye is equally spread between the two arms. This means that costs of including the non-index eye should be spread evenly between the two study arms and not affect the marginal difference between the costs.

The main strength of this study is that this model-based economic evaluation was carried out in parallel with the trial-based economic evaluation and was populated with data derived from the trial. All defined parameters in this model-based economic evaluation analysis were based on data collected prospectively as part of TAGS. Moreover, although other economic evaluations exist, the TAGS-based analyses (both the within trial and the model-based analyses) included patients at severe stages of glaucoma only. The results should, therefore, be generalisable across the UK NHS for patients with severe glaucoma.

Using GSS2 for classifying patients (baseline, 12 months and 24 months) into different stages of glaucoma, is another strength because it allowed our study to detect these differences in both cost and utility outcomes between trabeculectomy and medical treatment. Within the model, the severity of glaucoma has been defined based on the state of each eye separately and data on the progression of glaucoma is based on data for both eyes, allowing the modelling of the index eye and non-index eye within the model. Other studies defined health states within their model based on the condition of both eyes together and did not consider each eye separately in the model as we have. 110,111 Although HRQoL and, hence, QALYs are thought to be determined by the quality of vision in the better eye in the case of eye disease,117 the model structure estimated the chance of unilateral and bilateral progression more precisely over time.

Our evaluation has also repeated the analysis using QALYs based on different measures of utility (EQ-5D-5L, HUI-3 and GUI). The advantage of this is that the EQ-5D-5L is widely used to estimate QALYs in the UK and is the recommended approach by NICE. However, it does not include a specific vision component. The HUI-3, however, is also a generic measure of health status but unlike the EQ-5D-5L has a specific vision component. It should be noted, however, that health state utilities values are derived from a Canadian population rather than a UK population. The GUI is a condition-specific measure for which utilities have been derived directly from TAGs trial conditions. The results of our study showed that using the HUI-3 measure over a patient’s lifetime provides the largest difference in QALYs between treatment methods (0.38 additional QALYs of trabeculectomy compared with medical management); QALYs based on the EQ-5D-5L measure are ranked second (with 0.28 additional QALYs of trabeculectomy compared with medical management) whereas the GUI provides the smallest difference in QALYs (0.16 additional QALYs of trabeculectomy compared with medical management).

For the model-based economic evaluation, our analysis has some limitations that need to be considered when interpreting the results. The first limitation is the limited data to extrapolate beyond 2 years; for example, needing further surgery (cataract surgery) to manage complications was not considered in the model because the trial results provide no evidence of a difference between the two arms. Longer-term follow-up, as is planned for TAGS, may further inform the model. Medical therapy is preferred by many patients because it is non-invasive and, from the perspective of the health service, it avoids the initial cost of surgery but may be associated with higher VF loss in the longer term. In part, this may be caused because adherence to medication in the long term may be challenging for patients. 167 Finally, to address uncertainty caused by missing data, we have used multiple imputation data for the model. The results of the multiple imputations suggest that the complete-case analysis may underestimate the difference in effects between the alternatives as a result of those with poorer health outcomes being more likely to withdraw the follow-up.

Further research planned

Further funding has been secured to evaluate clinical and patient-reported outcomes at 5 years to further explore the lifetime experience, patient-reported outcomes and visual loss (VA and VF survival) in this group of participants. These data will be incorporated into an updated economic model once they become available.