Cost–utility analysis of an intervention designed to reduce the critical handling error of insufficient inspiratory effort

A previous structured literature review of economic models in asthma identified 31 published cost-effectiveness models for asthma [11]. Many of the features of the models identified in the literature were utilised; including a Markov model structure and utility values from the Briggs et al. [12] model. However, full use of the existing models was limited as they did not include results specific to handling errors with asthma inhalers.

A new cost–utility model was therefore developed using Microsoft^® Excel 2010 and R-3.4.1, to estimate the economic and social impact of the critical handling error of ‘insufficient inspiratory effort’. The cost–utility model included patients defined as GINA Step 3 and above receiving ICS/LABA therapy via one of two DPIs, Diskus^® or Turbohaler^®, or an error-targeted intervention. These were the only DPIs assessed in the CRITIKAL study and were therefore deemed suitable for inclusion within this analysis [9].

The analysis takes the perspective of the National Health Service (NHS), the wider economy, and patients in the UK; therefore, the model included both direct (device costs and resource use) and indirect (productivity loss/reduction) costs as well as patients’ health-related quality-of-life in the form of utility values.

The patient population within this analysis matches that of the CRITIKAL study. The population of interest for this model is patients aged 16 years or over receiving Step 3 and above GINA-defined asthma therapy (ICS/LABA) in the UK, administered by DPIs. As this included patients of both sexes, as well as a range of BMIs, smoking status, age and other patient factors, the CRITIKAL study adjusted ORs based on these factors. Sub-group analyses were not feasible due to the absence of the reporting of the ‘non-handling error’ coefficients.

A two state Markov model was developed where patients could either be in a controlled asthma state or an uncontrolled asthma state (see Fig. 1). The decision to build a two state model was made to ensure consistency with the CRITIKAL study which reported the OR of moving to a less controlled health state conditional on experiencing a handling error [9]. Additionally, as the Demoly et al. [2] study looked at the health-care utilisation associated with only controlled or uncontrolled asthma, it was simpler to model a two health state Markov model, with any inputs that used controlled, partially controlled and uncontrolled states adjusted so that the weighted average of partially controlled and controlled was used. A one year time horizon was chosen to reflect the one year exacerbation data from the CRITIKAL study [13]. However, to reflect the time frame for budget holders’ decisions in the UK, the results are also given for when three and five year time horizons were applied.

No device switching and 100% compliance were assumed over the time horizon, due to the associations with change in asthma control and possible confounding effects [14]. As this model was designed to only investigate the effect of reducing a handling error through the introduction of an error-targeted intervention, the incorporation of device switching and varying compliance rates between devices may affect the results and mask the effect of the reduction of the handling error. The one year time horizon is also in place for consistency around the no-switching assumption.

Mortality was not included in the model because the risk of mortality associated with the different devices over a one year time horizon was considered minimal. For example, official UK guidance on asthma from the National Institute for Health and Care Excellence (NICE) states that ‘mortality from asthma is rare’; [15] the number of asthma-related deaths in the UK in 2012 was 1246, while the UK asthma population in the same year was more than 8 million [16]. Although NICE economic planning for asthma mentions incorporating mortality, it acknowledges that mortality data may not be available from studies due to lack of power arising from the rarity of asthma deaths and the size of the population [17]. The omission of mortality is thus a conservative simplifying assumption and by definition its inclusion would lead to even greater cost-effectiveness for the error-targeted intervention, as it has fewer exacerbations compared to DPIs.

A half cycle correction was implemented in the model to reduce bias of under-/overestimation of patients in each cycle, in line with standard practice for cost-effectiveness models [18]. Discounting of costs and QALYs was not included over the one year time horizon; however, a 3.5% discount rate for each was applied for the three and five year calculations.

An additional targeted literature search was conducted, using Embase and Medline, to identify published transitional probabilities for asthma patients using the GINA-defined health states of asthma to match the definitions utilised for the CRITIKAL study (see Appendix). Overall, the literature review identified 14 full text publications of relevance, of which Bateman et al. [19] was deemed the most appropriate for the baseline transition probabilities (see Table 1) as it was the most recent study which also explicitly used GINA classification for states of asthma. This source compared different drug therapies, rather than different inhaler devices specifically. To construct the two-by-two transition probability matrix, the model combined the probabilities of controlled and partially controlled health states to form a new ‘controlled’ state, and all transition probabilities were then normalised to ensure summation to unity, as shown in Table 2. All probabilities were fixed and therefore time independent, to be consistent with the transition probabilities in Bateman et al. [19] (transition probabilities for controlled status were for any week¹ and were based on the GINA-defined control status in the previous week²).

Utility values were obtained from Briggs et al. [12] who mapped utility values from Asthma Quality-of-Life Questionnaire scores. Both the utility values [12] and transition probabilities above [19] were derived from the same study, the GOAL study [20], which provided consistency. As the GOAL study included three health states rather than two, the model took the weighted average of the controlled and partially controlled health state utility values based on the proportion of patients in each health state in the CRITIKAL study, as shown in Table 3 [9]. Utility values were applied to the two health states included in the model (controlled and uncontrolled) as shown in Table 4.

The model included the costs of the individual inhaler devices, and the costs associated with the degree of control [2], namely resource use and indirect costs, all of which were based on 2016/17 GBP.

Data for the average annual rate of direct resource use (physician visits, emergency room visits and hospitalisations) were identified in a systematic literature review in terms of the impact of critical handling errors on health outcomes and resource use [7]. The latest and most applicable data which provided resource use to match the two health states in our model were taken from Demoly et al. [2], which reported resource use by controlled and not-controlled status over a six month period. These data were then doubled to estimate the annual rates of resource use (shown in Table 5). These rates were then multiplied by the resource unit costs in Table 6 to provide estimates of resource use costs.

Device unit costs were estimated based on the assumption that patients would use a whole inhaler every 30 days with unit costs taken from the latest British National Formulary (BNF) 2017 as shown in Table 6 [21]. To compare the benefits of the removal of a single critical handling error, the cost of the error-targeted intervention was set to match either the cost of Diskus^® DPI (scenario 1) or Turbohaler^® DPI (scenario 2). Then, the economically justifiable cost for the error-targeted intervention was calculated based on the £20,000 per QALY NICE threshold (scenario 3).

As noted previously, the model objective was to focus on the impact of overcoming a critical handling error; therefore, it was assumed that there were no differences in drug effectiveness that would impact on control status. This is in line with NICE guidance on ICS devices for asthma in adults, which states it is reasonable to assume that there are no differences in clinical effectiveness for ICS devices at low doses. Their assessment found that there were no consistent differences between the clinical effectiveness of the different ICS/LABA combinations, and clinical and patient experts agreed with this finding [15].

To model indirect costs, 74.6% of patients in the model were assumed to be economically active, based on the estimated UK employment rate in April 2017 [25]. Therefore, only working patients were assumed to incur costs associated with absenteeism and presenteeism. The population working wage, which is used to calculate the indirect costs of absenteeism and presenteeism, was calculated by multiplying the estimated employment rate by the average UK salary in 2016 [24, 25].

Exacerbations were only considered as an event in the model based on rates from the CRITIKAL study (calculated from the number of exacerbations per health state) [13] with their impact on utilities only occurring in the uncontrolled health state [9]. The associated costs were not modelled directly as the resource use consequences of exacerbations were indirectly modelled via health state resource use which implicitly included differential exacerbation rates [26]. The number of exacerbations over the one year time horizon was estimated for each device.

The expected number of exacerbations in one year were calculated by multiplying the number of patients experiencing X exacerbations over one year, by X exacerbations, for each health state [13]. For the small number of patients in the CRITIKAL study falling into the 5+ exacerbations per year category (0.49%, 1.21% and 5.26% controlled, partially controlled and uncontrolled, respectively), it was assumed that they experienced 5 exacerbations per year. The expected number of exacerbations per year for the combined controlled health state was found by applying the weightings in Table 3 to the calculated expected number of exacerbations per year for controlled and partially controlled states. The resulting value for controlled, and the expected number of exacerbations in a year for a patient in the uncontrolled state were converted to probabilities of exacerbation in a week for each state, using the formula:where r is the expected number of exacerbations per patient in the state in a year and t = 52 (converting from one year to one week). This gave the weekly exacerbation rate 0.69% for the controlled state and 1.93% for the uncontrolled state, allowing the number of exacerbations to be found for each comparator across the time horizon. Due to data paucity on the cost per exacerbation for the asthma patient population in the UK, exacerbation costs themselves were not included in the total costs. Additionally, it was assumed that the resource use costs associated with each health state would contain the cost associated with an exacerbation (avoiding overestimation of cost).

The CRITIKAL study, as mentioned previously, calculated the OR of a patient being in an uncontrolled asthma state for each inhaler, conditional on having made the critical handling error of ‘insufficient inspiratory effort’ [9]. This was the only critical handling error associated with asthma control after adjusting for patient factors including age, smoking status, and adherence. These adjusted ORs are shown in Table 7.

The CRITIKAL study assessed a variety of handling errors, using iHARP review data (a multicentre cross-sectional study of adults with asthma) for patients aged 16 years or over, receiving Step 3 and above GINA-defined asthma therapy of ICS/LABA in the UK. This resulted in 4276 patients in the iHARP database after inclusion and exclusion criteria (such as receiving antibiotics for lower respiratory tract infection in the 2 weeks prior to iHARP review—to minimise potential confounding by other medicines) were applied. Error frequency, asthma symptom control, and exacerbation rate were analysed to identify critical handling errors. The study found the association between critical handling errors and asthma control and exacerbations for a range of inhalers including DPIs and MDIs. The critical handling error of ‘insufficient inspiratory effort’ was made by 32%–38% of DPI users and was associated with uncontrolled asthma (adjusted ORs [95% CI], 1.30 [1.08, 1.57] and 1.55 [1.15, 2.08] in those using Turbohaler^® DPI and Diskus^® DPI, respectively) and increased exacerbation rates [9].

To apply an RR of a critical handling error to the model’s baseline transition probabilities, the ORs were transformed to RRs using a common statistical approach reported in Zhang et al. [27]. This approach required the ORs to be between 0.5 and 2.5, so this approach of converting an OR to an RR (based on the incidence of the outcome of interest, P₀) was able to be applied. Additionally, this approach required the incidence of the event to be less than 10% which was not the case in our study; therefore, this assumption was explored in the sensitivity analysis by assuming that OR is equal to RR. The approach reported in Zhang et al. [27] was as follows:The RR, adjusted for the probability of a patient experiencing a handling error, was applied to the transition probability of patients going from controlled to uncontrolled and patients remaining uncontrolled. The inverse of the RR was applied to the remainder of the transition probabilities. The transition probability matrix was then normalised to ensure that the probabilities added up to one.

As the objective of the model was to evaluate the economic and societal impact of the critical handling error of ‘insufficient inspiratory effort’, through the introduction into the market of an intervention designed to avoid this error, the model made the assumption that this error-targeted intervention had an RR of one (as shown in Table 7), and so the transition probabilities remained unchanged. To explore the complete or partial eradication of the ‘insufficient inspiratory effort’ error, the model varied the error-targeted intervention OR by an increase in 25%, 50%, 75% and 100% with a probability of 35.25% of experiencing a handling error, which is the average probability of each comparator. As seen from the associations found in the CRITIKAL study and Table 8 [9], the critical handling error ‘insufficient inspiratory effort’ increased the likelihood of a patient having uncontrolled asthma and therefore increased the likelihood of an exacerbation.

Cost and QALY calculations were used to estimate the net monetary benefit (NMB) for the error-targeted intervention compared to either Diskus^® DPI or Turbohaler^® DPI in both the deterministic and probabilistic analyses, and also simultaneously in the probabilistic analyses using a cost-effectiveness acceptability curve (CEAC) which effectively estimates the probability of each of the devices included, of being the most cost-effective, i.e. having the highest NMB. The formula for NMB is:

Each comparator had an NMB value for three incremental cost-effectiveness ratio (ICER) thresholds: £10,000, £20,000 and £30,000 per QALY as recommended by NICE for approval [28]. The £10,000 per QALY threshold was included because drugs or devices meeting this threshold can be fast-tracked for reimbursement approval to provide faster access to innovative and cost-effective treatments [29]. The £20,000 to £30,000 per QALY threshold is the recommended threshold by NICE [28].

Both deterministic sensitivity analysis (DSA) and probabilistic sensitivity analysis (PSA) were conducted to evaluate the robustness of the model values.

A simple DSA was conducted to determine which parameters had the greatest influence on the incremental NMB (INMB) value, which compares the NMB of the main results with the sensitivity analysis NMB value, at a threshold of £20,000 per QALY by increasing/decreasing each of the input values by 20%. The parameters included in the DSA were health-care unit costs, utility values, the proportion of patients starting in each health state, indirect cost inputs including UK average salary and probability of experiencing the handling error for each comparator (related to efficacy).

Each device was analysed separately against the error-targeted intervention. A separate analysis was conducted to calculate the economically justifiable price to identify a ceiling price for an error-targeted intervention to be cost-effective at £20,000 per QALY which is the lowest cost-effectiveness threshold for NICE.

An additional sensitivity analysis was conducted to determine the effect of assuming the RRs were equal to the ORs in the model. This was included as a sensitivity analysis because there was some uncertainty around the base case transformation from OR to RR utilising the approach described above, given that Zhang et al. [27] reported that their method may not be applicable in all situations based on its technical requirements. Given that the RRs are conditional on making a critical handling error, the magnitude of this error needs to be controlled for in a Bayesian conditional probability context for the OR = RR approach. To make the ORs probabilistic for the OR = RR approach, they were adjusted based on a Bayesian approach with the formula below:

In line with standard practice for cost–utility models, PSA was performed. Where possible, the estimated standard errors (SEs) of the mean for the parameters used in the PSA were based on published data. In the absence of published data or any other data, the estimated SEs of the mean of the costs distribution were assumed to be the same as the mean cost. These values have been highlighted in the data tables above. Resource use rates followed a Gamma distribution to reflect their non-negativity, while utilities and transition probabilities followed a Beta distribution as they ranged from zero to one.

The error-targeted intervention dominates the two DPIs individually when setting its price to match each DPI separately over one year (scenario 1 at £35 and scenario 2 at £38), as the error-targeted intervention had a lower total cost and higher number of QALYs in each scenario. Additionally, it resulted in only 636 exacerbations per 1000 patients over a one year time horizon, compared to 677 for Turbohaler^® DPI and 713 for Diskus^® DPI. For a three year time horizon, there were 1864, 1987 and 2097 exacerbations per 1000 patients for the error-targeted intervention, Turbohaler^® DPI and Diskus^® DPI, respectively, while for five years, there were 3010, 3209 and 3388 exacerbations per 1000 patients for the error-targeted intervention, Turbohaler^® DPI and Diskus^® DPI, respectively.

Positive NMB for the error-targeted intervention indicated that it was cost-effective compared to both DPI devices. For one year, at a threshold of £20,000 per QALY, and incorporating both direct and indirect costs, the error-targeted intervention priced at £38 had an NMB of £303.73 over Turbohaler^® DPI, and an NMB of £541.94 over Diskus^® DPI (see Table 9). Similar results were found when the error-targeted intervention was priced at £35 with an NMB of £340.13 over Turbohaler^® DPI, and an NMB of £578.34 over Diskus^® DPI (see Table 9). The economically justifiable cost based on direct costs only for the error-targeted intervention in the one year time horizon, based on a cost-effectiveness threshold of £20,000 per QALY, was calculated at £50.52 and £58.83 per error-targeted intervention (which approximates to a monthly price) compared to Turbohaler^® DPI and Diskus^® DPI, respectively. For the three year time horizon and the five year time horizon (see Tables 10 and 11), the results were approximately triple and quintuple the NMB for the one year time horizon.

Multiple one-way deterministic sensitivity analyses based on an increase/decrease of 20% in the value of each parameter showed that the weekly utility value for each state and the RR of moving to an uncontrolled state for the comparator were the key factors driving variance from the base case NMB result for error-targeted intervention priced at £38, as shown in Fig. 2. Similarly, the same key factors were identified when the error-targeted intervention was priced at £35 and can be found in the Appendix (Fig. 5).

The overall results of the two price scenarios did not change greatly when RR was equal to OR from the CRITIKAL study. The incremental direct costs savings compared to Diskus^® DPI and Turbohaler^® DPI increased by £26 and £8, respectively. The incremental QALYs decreased by 0.0031 and 0.001 for Diskus^® DPI and Turbohaler^® DPI, respectively.

The PSA demonstrated that the two DPIs had lower number of QALYs and were more expensive compared to the error-targeted intervention priced at £38 (see Fig. 3). Additionally, the error-targeted intervention priced at £38 had a probability > 99% of being cost-effective for all ICER thresholds £14,000 or above (see Fig. 4). This was expected when the no handling errors assumption associated with the intervention was made. When the error-targeted intervention was priced at £35, it had a 100% probability of being cost-effective for all ICER thresholds. These results can be found in the Appendix (Figs. 6 and 7).

The error-targeted intervention was cost-effective when the OR of moving to an uncontrolled health state after making an ‘insufficient inspiratory effort’ error increased to 1.25 (RR = 1.15) compared to both comparators in both scenarios. However, when the OR was 1.50, Turbohaler^® DPI became more cost-effective in both scenarios. When priced at £38, with OR 1.75, Turbohaler^® DPI became more cost-effective (see Table 12).