Within-trial economic evaluation

Introduction

Background

Prior to analysis, a Health Economic Analysis Plan (HEAP) was developed (see: www.rescueasdh.org/trial-documents), though here, for ease of reading, we present the Methods and Results in sufficient detail that readers should not need to refer to the HEAP. Since the development of the HEAP (in 2021), new methods guide for NICE health technology evaluations has been published;6 we, however, refer to the previous methods guide, as that informed our HEAP, and the new methods guide was not available when we began our analysis.

Study rationale

There is potential for there to be differences in costs and benefits between craniotomy and DC. For example, DC involves an extra planned procedure (cranioplasty, with associated costs), but such costs could be partly/wholly offset if DC were found to have associated benefits, which in turn resulted in fewer AEs. Given this uncertainty, the high costs associated with such procedures, and the variation in practice as to whether craniotomy or DC is currently undertaken, a comparison as to the cost-effectiveness of craniotomy and DC was undertaken.

Objective

To estimate the cost-effectiveness of craniotomy, compared with DC, based on UK data from a multicentre, pragmatic, parallel-group randomised trial of DC versus craniotomy for patients undergoing evacuation of ASDH (RESCUE-ASDH).8

Methods

Trial design

As previously described, the RESCUE-ASDH trial was a multicentre, international, pragmatic, parallel-group randomised trial. For the purposes of this economic evaluation, the analysis was based on UK participants who were involved in the randomised controlled trial (i.e. excluding participants outside the UK and those who took part in the observation study). The trial compared DC, in which the bone flap is removed prior to closing the skin, with craniotomy, in which the bone flap is restored before skin closure. DC is undertaken on the basis that the missing part of the skull will be restored later (cranioplasty), in those patients who survive. Patients were eligible for enrolment if they were 16 years of age and older, had an ASDH on CT scan and the admitting surgeon felt that the haematoma needed evacuating either by craniotomy or DC.

Intervention

When an ASDH is evacuated, the bone flap can be left out (DC) or replaced (craniotomy) prior to closing the skin. Participants were randomly assigned in a 1 : 1 ratio to receive DC or craniotomy.

Additionally, there was an observation (non-randomised) cohort of patients who were also followed up and fulfilled the following criteria: (1) eligible patients who could not be randomised after evacuation of haematoma due to significant swelling of the brain, (2) patients who had a traumatic ASDH removed as part of their standard care and were enrolled postoperatively. These patients were not included in the economic evaluation.

Measuring costs

In line with the NICE methods guide,9 in the base-case analysis, costs were estimated from the viewpoint of the UK NHS and Personal Social Services (PSS). Subsequently described resource use data were collected, to which unit costs [estimated in Great British pounds (£) for the 2018–9 financial year] were assigned (see Table 35). Where unit costs were taken from previous years, they were inflated using the NHS Cost Inflation Index (NHSCII).10

There were two main sources for the resource use data: hospital-recorded data and a patient self-report (1-year follow-up) questionnaire. Both were specifically developed for the study (in consultation with hospital staff/non-trial patients) and informed by the guidance that one should focus on the large cost drivers and those resource items that might differ between study arms.11 Costs were not assigned to resource items that were undertaken for research purposes.

Hospital-recorded data

The hospital-recorded data included (if applicable) details of the operation to which participants were randomised (craniotomy or DC), including length of operation and graft details. Time spent in the ICU and NSU and any further neurosurgical procedures received during their initial (index) admission were also recorded, along with cranioplasties and shunt placements received after discharge from the NSU.

Unit costs were assigned to hospital-recorded data resource use items (see Table 23 for further details). In terms of the intervention costs, several procedure details were recorded, but costs were not assigned to all items, if they were considered relatively small or unlikely to differ between arms. For all procedures, a common unit cost of £142212 per hour (mean theatre direct costs per hour for neurosurgery) was applied. For the index procedures (DC or craniotomy), the length of operation was recorded, and the hourly rate was therefore applied accordingly. Furthermore, for participants who received DC, the method of bone flap storage was recorded. It was, however, assumed that the cost of storage inside the participant’s body was captured by the length of operation and that the marginal cost of storage outside the body was negligible, as tissue banks already exist in hospitals. In many cases, however, the bone flap was discarded, and synthetic materials used for replacement during a later cranioplasty procedure. Any cranioplasties were recorded (whether during index admission or post discharge from NSU), including the type of graft. If this indicated the use of a synthetic graft or, if this information was missing, the bone flap was recorded as ‘disposed’, ‘sent for implant modelling’ or ‘other’ postoperatively, an estimated synthetic material cost of £2500 (design and parts) was used. Additionally, as the length of time associated with cranioplasty was not recorded, it was assumed that this was 104 minutes.13 In addition, if the cranioplasty took place after discharge from NSU, an associated LOS of 4 days was assumed and costed. These costs and timings were informed by expert opinion.

Shunt placements were also recorded during index admission and post discharge from NSU. In line with expert opinion, an assumed length of procedure of 90 minutes was applied and costed, along with an estimated materials cost of £500. In addition, if the shunt took place after discharge from NSU, an associated LOS of 2 days was assumed and costed.

Revisions to cranioplasties and shunts were also recorded, and these were assumed to have the same procedure duration and associated LOS (post discharge only) as that described for the relevant index procedure, but without the materials costs.

Where a date was missing for cranioplasties or shunts, whether primary or revision, the operation was assumed to have occurred post discharge from NSU, if recorded on the post-discharge form, but during index admission if recorded on the neurosurgical intervention form used at NSU discharge.

As cranioplasties and shunts were recorded on several different forms during the study, there was potential for duplication of recording. To avoid this, the total number of cranioplasties was summed, having discounted any with duplicate dates. It was then assumed that one of these cranioplasties represented the primary cranioplasty operation and a materials cost (assuming synthetic graft was used and/or bone flap was recorded as ‘discarded’, ‘sent for implant modelling’ or ‘other’) was applied to this operation (as this was the most common type of graft). Any further cranioplasties, whether recorded as revisions or not, were assumed to be revisions, with no materials cost applied. The same principle was applied to shunt operations. If a cranioplasty and shunt (primary or revision) were recorded as having occurred on the same day, it was assumed that this was the same operation, and a duration of 150 minutes (expert opinion) was applied, along with a LOS of 4 days and relevant materials costs. These costs for cranioplasties and shunts are summarised in Table 35.

Details of further neurosurgical interventions were also requested at NSU discharge (i.e. during index admission), along with a date of procedure. Further craniectomies (DC), haematoma evacuations (HEs) (of any description) and wound revisions that were confirmed, by the procedure date, to have occurred during index admission (i.e. before NSU discharge/death date), were costed. Any DCs that were recorded as having occurred on the same date as the index DC procedure were excluded from the analysis, on the assumption this represented duplicate recording. As the length of HEs and wound revisions were not recorded, following expert opinion, these were assumed to be 90 minutes. External ventricular drain (EVD) placement and ICP monitoring insertions were also recorded but not costed, as these were deemed to be relatively inexpensive and unlikely to differ between arms (expert opinion). Where ‘other’ procedure was recorded on the form, these were recoded, following expert opinion, using the free-text responses that accompanied this response. Those that did not fit one of the pre-specified descriptions (DC, HE and wound revision), but were deemed to have resource use implications, were coded as ‘other’ and assumed to have an associated length of procedure of 90 minutes also. The remainder were not costed. Only those procedures that were costed were analysed and reported.

Details of LOS in NSU and ICU were also recorded for the index admission; this was assigned an appropriate cost per bed-day, as described in Table 35. The length of NSU stay was calculated from baseline (date of index surgery) to NSU discharge/death and mutually exclusive of any ICU stay recorded:

If ICU stay ≥ 0 and ICU admission date ≥ baseline date:

ICU stay = [(date of discharge from ICU or death) – (date of admission to ICU) + 1]

If ICU stay ≥ 0 and ICU admission date ≤ baseline date:

ICU stay = [(date of discharge from ICU or death) – (baseline date) + 1]

NSU stay

NSU stay = {[(date of final discharge/death) – (baseline date) + 1] – (ICU stay)}

‘Date of final discharge’ refers to the final NSU discharge date, excluding any previous discharges from NSU to ICU. For participants who had multiple stays in ICU, their total ICU stay was calculated as above, and all ICU stays summed.

For each participant, the component costs associated with the index procedure (craniotomy or DC), cranioplasty and shunt procedures (for index admission and post discharge), any other neurosurgical interventions during index admission and LOS in ICU and/or NSU (for the index admission) could thereby be estimated and were summed to estimate the total hospital-recorded costs.

Patient self-reported resource use

At the 12-month follow-up point, UK participants were sent a questionnaire that requested the following resource use data (where a participant was unable to fill in the questionnaire, proxy responses by a relative/friend/carer were requested). The questionnaire only requested information in relation to the time since discharge from the NSU, as the resource use prior to NSU discharge was assumed to be captured by the aforementioned hospital-recorded data. Participants who died during their index admission, that is, were reported as having a discharge destination of ‘died’ at any point, were given a value of ‘0’ in response to all questions in the patient-reported questionnaire. Those who died after discharge from NSU were reported as ‘missing’ in response to all questions (this allows these data to be imputed for the base-case analysis, while adjusting for time to death – see Missing data).

Participants were asked to report any overnight stays in a hospital or other healthcare facility (for any reason), where information was also requested (if applicable) in relation to the number of nights, type of unit, and whether there was an associated operation on the skull/brain. Participants were also asked to report any healthcare professional (HCP) visits (for any reason), where (if applicable) information was requested in relation to the type of HCP seen, frequency of visit and where (most) visits took place (home, community or hospital). Additionally, participants were asked to report any scans of the head/brain, where (if applicable) information as to the type [magnetic resonance imaging (MRI), CT or ‘other’] and number of investigations was requested.

Participants were also asked to report if they had spent any time in a care home, where (if applicable) the associated time in weeks/months was requested. Participants were similarly asked to report if they had received any help from a family member/friend or other carer. If a respondent reported ‘yes’, then (for each carer) they were asked to report the average number of hours of care received (per week since discharge from NSU).

Finally, participants were asked several questions in relation to work, including: (1) whether they were currently working (paid/unpaid), and if yes: (a) the number of hours per week they work, and (b) their return-to-work date. Within the HEAP, it was specified that we would estimate the ‘mean lost work time over the 12-month follow-up period’. However, implementation of this was difficult, as (in error) those who reported they were not currently working at the 12-month follow-up were not asked (1) whether they were working before their brain injury and, if they were, (2) the number of hours they worked. Consequently, we did not seek to estimate/value lost work time. We did, however, seek to estimate the number of hours participants had worked since their brain injury (in the 12-month follow-up period) in the following manner. Participants were asked how many hours they currently work and the date they returned to work (if applicable). Assuming the reported current working hours were applicable to the time between the date they returned to work and the 12-month follow-up point, it was possible to estimate (for responding participants) the number of hours they have worked since their brain injury (in the 12-month follow-up period).

Unit costs were assigned to each item of resource use data reported by participants (see Table 35). This enabled the component costs associated with any overnight stays (post NSU discharge), HCP visits, scans, care home and carer costs to be estimated for participants from UK sites. However, to avoid the potential for cost duplication, pre-analysis we stated that patient self-reported overnight stays with an associated skull/brain operation would only be included in the base-case analysis if the total number reported was greater than the total reported number of post-discharge cranioplasties/revisions and shunt placements/revisions, as reported in the hospital-recorded data. This was seen as possible as hospital-recorded data would only record operations that took place at the same hospital as their index admission. If this was not the case, only costs associated with self-reported overnight stays, where an operation to the skull/brain was not noted, would be included in the base case, along with the costs associated with post-discharge cranioplasties and shunt placements (based on hospital-recorded data). Component costs were summed in order to estimate the total post-discharge costs, though in order to align with the NHS and PSS cost perspective recommended in the NICE methods guide,4 care home and carer costs (the same hourly cost was assigned to all reported hours of care regardless of whether, for example, a payment was made/the carer lived with the participant) were not included in the base-case analysis.

Measuring outcomes

To estimate HRQoL, and in line with the NICE guidance,18,20 the EQ-5D-5L was used, and respondents were asked to report the level of problems (none to extreme/unable) they had on five dimensions (mobility, self-care, usual activities, pain/discomfort and anxiety/depression).7 Participants were asked to complete the EQ-5D-5L at discharge from NSU, and again at 6- and 12-month follow-up (though this was not applicable if they were yet to be discharged from NSU at these time points). Where a participant was unable to fill in the questionnaire, proxy responses by a relative/friend/carer were requested. As recommended in the NICE position statement,21 the crosswalk mapping function22 was then used to convert responses into utility scores, where a score of zero corresponds to death and one to full health.22 NICE18 has recently updated the preferred mapping function to be used in reference-case analyses for the EQ-5D-5L from that published by Van Hout et al.22 to that of Hernández Alava et al.23,24 We have adhered to the earlier guidelines9 as the analysis started prior to the change in guidelines. Participants who died were assigned a utility score of zero on their date of death (assuming this occurred within the 12-month follow-up period). The quality-adjusted life-year (QALY) score that was accrued over the trial period was then estimated for each individual for whom the EQ-5D had been completed at the three aforementioned time points. This was based on the total area under the curve method and the assumption of linear interpolation.25 In the base-case analysis, the score at the date of discharge was assumed to be the baseline score (12 months prior to the final follow-up point).

To undertake the cost-effectiveness analysis (CEA), the primary outcome measure, the GOSE26–29 was used, with the GOSE collapsed onto a binary scale (i.e. favourable vs. unfavourable). Favourable outcomes were defined as upper severe disability or better, while unfavourable outcomes included death, vegetative state and lower severe disability. The favourable versus unfavourable scoring was based on the participants’ 12-month responses.

A further binary (favourable vs. unfavourable) categorisation was also used: if GCS at randomisation was between 3 and 8 (patient comatosed),4 a favourable outcome was defined as upper severe disability or better, but if GCS at randomisation was between 9 and 15 (responsive patient), a favourable outcome was defined as lower moderate disability or better.

Analyses

Missing data

Missing data is common in randomised trials and can lead to bias and lack of precision.30 As recommended for within-trial analysis of cost-effectiveness, in line with previously described methods30 in the base-case analysis multiple imputation (MI) was thereby undertaken.

Based on the pattern of missing data, costs were aggregated and imputed at the level of ‘index admission costs’ (from hospital-recorded data), ‘cranioplasty and shunt costs’, including both pre- and post-discharge procedures (also from hospital-recorded data) and ‘post-discharge costs’ (from patient self-report questionnaire). Costs were imputed at this level as the hospital-recorded and patient self-report data had different levels of missing data, whereas individual questions within each of these data types had similar levels of completion. Cranioplasty and shunt forms were completed separately to other hospital-recorded data and therefore had different levels of completion to those previously mentioned. Similarly, where missing, overall EQ-5D utility scores were imputed at NSU discharge and at 6- and 12-month follow-up, as response rates differed over time, though if EQ-5D data were missing, in over 50% of cases it was missing in all dimensions. Missing data for GCS score at baseline and GOSE at 12 months were also imputed, though both had a low level of missing data (< 5%).

Based on a descriptive analysis demonstrating the association between costs and outcomes and baseline variables and between missing costs and outcomes and baseline variables, the data were assumed missing at random (MAR). MI via chained equations under MAR was therefore used and the missing data imputed, by treatment group, using the ‘mi impute chained’ command [Stata 17.0 (StataCorp LP, College Station, TX)] to create 30 data sets (it is recommended the number of imputed data sets should be similar to the percentage of incomplete cases30), which were then pooled using Rubin’s rules.31

In addition to the costs (index admission, cranioplasty and shunt, and post discharge), EQ-5D scores (baseline and 6 and 12 months), GCS (baseline) and GOSE scores (12 months), the MI model included baseline covariates associated (p < 0.05) with missing data, costs or outcomes: age (years) and sex. Finally, the MI model was validated by comparing the distributions of the observed and imputed data to check that the distributions were similar.30

Cost–utility

The NICE methods guide recommends that costs be estimated from the viewpoint of the NHS and PSS.18 Therefore, in the base-case analysis, the previously defined total hospital-recorded costs and total patient-questionnaire reported costs (excluding care home and carer costs) were summed to estimate overall costs.

A 12-month within-trial analysis was undertaken, where (after excluding any participants randomised in error) an ITT approach was adopted, where participants were analysed within the group to which they were allocated, regardless of whether they received craniotomy or DC. A within-trial analysis was deemed appropriate, as we are not aware of any previous economic evaluations that have compared craniotomy with DC for patients with ASDH. No discounting was undertaken, as the analysis period matched that of the trial duration (1 year).

To estimate the mean incremental cost and mean incremental effect (QALY gain) associated with DC compared with craniotomy, seemingly unrelated regression analysis was undertaken, which is generally robust for skewed data and allows for any correlation between costs and effects.32 All regressions included those baseline variables expected to be predictive of total costs and outcomes: age (years), sex and baseline utility score.

Where appropriate, the incremental cost-effectiveness ratio (ICER), defined as the mean incremental cost/mean incremental effect,18 or DC compared with craniotomy, was estimated. If one intervention was both less costly and more effective, estimating the ICER was not necessary, and instead, that intervention could be categorised as ‘dominant’.33 The ICER can be used to assess whether the extra cost of an intervention (in this case, DC) constitutes value for money. In the UK, NICE refers to a cost-effectiveness threshold value of £20,000–30,000 per QALY.18 As such, if DC had an ICER (incremental cost per QALY) lower than this level, it could be argued to be cost-effective.

Cost-effectiveness

With the following exceptions, the analysis replicated that of the previously described cost–utility analysis (CUA). For the first CEA, the aforementioned GOSE favourable versus unfavourable categorisation (see Measuring outcomes) was used for the measure of effect, and in the absence of dominance, the ICER was the cost per additional favourable outcome. In a second CEA, this was repeated but with the favourable versus unfavourable categorisation based on GCS score at baseline, as described previously. As the outcome is binary (favourable/unfavourable), logistic (logit) regression was undertaken to estimate the OR (95% CI) of a favourable outcome for DC compared with craniotomy. Mean incremental costs associated with DC compared with craniotomy were estimated using linear regression. All regressions included variables age (years) and sex, which were expected to be predictive of total costs and GOSE outcomes. Given the measure of effect in the second CEA took account of the baseline GCS score, baseline GCS was only included as a covariate in the linear regression to estimate mean incremental costs.

Decision uncertainty

To estimate the level of uncertainty associated with the decision regarding cost-effectiveness, Fieller’s theorem was used to calculate30 the probability of being cost-effective at different thresholds of cost-effectiveness: the cost-effectiveness acceptability curve (CEAC).34 Here we report the estimated probability that craniotomy is cost-effective compared with DC at the recommended18 threshold of £20,000 per QALY.

Sensitivity analyses

The following sensitivity analyses (SAs)34 were undertaken to assess the robustness of conclusions to changes in key assumptions. Unless otherwise stated, all other assumptions remain as in the aforementioned base-case analysis. To analyse the data from a wider societal perspective, the care home and carer costs (which were excluded from the base-case analysis) were included in the first SA (SA1) (it should be noted that this is in contrast to the HEAP, as lost productivity costs were not included for reasons given above). A second SA (SA2) excluded patient self-reported resource use data altogether (there was expected to be more missing data from this source) and used only total hospital-recorded costs. Additionally, as there may be differences in the reported levels of post-discharge operations from different sources, a third SA (SA3) was conducted, which included only patient-reported post-discharge skull/brain operations, along with their associated LOS data, instead of the hospital-reported data on post-discharge cranioplasties and shunts (where LOS data were not recorded and instead assumptions about LOS had to be made). The fourth SA (SA4) tested the use of the EQ-5D score at discharge from NSU as the baseline point for QALY. As any benefits associated with the interventions could already have been partially/wholly achieved by discharge, the QALY scores were therefore re-estimated in SA4 with the assumption that, given the grave nature of the condition and following expert advice, participants had the lowest possible EQ-5D score at baseline (date of index surgery), that is, −0.594 was assumed to be the baseline score. On this basis, the total QALY scores were re-estimated for each patient. A fifth SA (SA5) re-analysed the data on a per-protocol basis, as defined in the SAP, that is, all patients included in the ITT population, except patients where the primary treatment was not as allocated, for example, patients allocated to DC who had craniotomy and vice versa. A complete-case analysis based on the base case was also undertaken in a sixth SA (SA6), where participants were only included if they have complete hospital records, participant self-report and QALY data, with no imputation undertaken. Aside from SA4, the above SAs were also applied to the GOSE.

Results

Participants

Between September 2014 and April 2019, 452 patients were recruited and randomly assigned to a trial group, though two patients were withdrawn owing to ineligibility [see the Consolidated Standards of Reporting Trials (CONSORT) diagram, Report Supplementary Material 3]. Thus, 450 patients were enrolled in the trial, 228 in the craniotomy arm and 222 in the DC arm. These patients constitute the full analysis population and their baseline characteristics have been outlined previously (see Table 18) or the purposes of the economic evaluation, as the perspective is from the viewpoint of the UK NHS and PSS, only participants from UK sites were evaluated and their baseline characteristics are outlined in Table 24. The UK economic evaluation base-case population included 248 patients, 126 in the craniotomy arm and 122 in the DC arm. Unless specified otherwise, this constitutes the population used in the base-case and other analyses in this economic evaluation. Compared with the full analysis population, these UK patients are slightly older (3.5 years on average) and more likely to be on antithrombotic medication. They are also more likely to have an extracranial injury requiring admission, combined with a lower GCS score.

Some patients did not receive their allocated treatment and were dropped from the per-protocol analysis (SA5), leaving 113 patients who were allocated to and received craniotomy and 114 patients who were allocated to and received DC.

Costs

At NSU discharge 211/248 (85%) of participants had complete hospital-recorded resource use data (index admission) (Table 26).

Table 28 presents the reported levels of resource use at discharge, based on available hospital-recorded data for both the craniotomy and DC groups. This includes reported levels of resource use associated with the mean duration of index surgery, mean LOS in NSU and ICU, and details of further neurosurgical interventions, including additional DCs, HEs, wound revisions and ‘other’ neurosurgical interventions. Duration of index surgery was comparable across groups, as was LOS in ICU. There was a slightly higher non-significant mean LOS in NSU reported for the craniotomy group compared with the DC group (16.75 vs. 21.30 days, p = 0.210). With regard to further neurosurgical interventions received during index admission, there were more HEs (9 vs. 2) and DCs (15 vs. 4) in the craniotomy group compared with the DC group (see Table 28). This is not unexpected as it represents the fact some patients who initially had a craniotomy, went on to have a DC. There were, however, less wound revisions (one vs. six) in the craniotomy group compared with the DC group (see Table 28). When combined with associated unit costs (see Table 35) this enabled costs to be estimated for the same resource components and, in turn, the total index admission costs (excluding cranioplasties and shunts) (see Table 29), which were slightly lower in the craniotomy group compared with the DC group, though not significantly (£30,790 vs. £34,759, p = 0.195).

With regard to cranioplasty and shunt procedures, 242/248 (98%) had complete resource use data (Table 27). Table 27 presents the levels of resource use relating to these cranioplasties and shunts (index and revision procedures), as reported in the hospital-recorded data. These occurred during index admission to hospital and/or post discharge, up to 12 months. As cranioplasty is only conducted after a craniectomy has taken place, as would be expected the mean number of primary cranioplasties per patient was significantly lower in the craniotomy group than in the DC group (0.169 vs. 0.512, p < 0.0001). Of note, however, there were 30/126 patients who were randomised to craniotomy that went on to have a DC at some point; 13 received DC instead of their randomised treatment, while the remainder received the treatment to which they were randomised (craniotomy) but went on to have a DC at a later date. Twenty-two of these (1 missing) were recorded as having a cranioplasty in the 12-month follow-up period. The majority of patients that received a cranioplasty did so using a synthetic material plate (81% in the craniotomy group, 74% in the DC group), as opposed to replacement of their own stored skull. As expected, the mean costs for cranioplasties were significantly lower in the craniotomy group than in the DC group (p < 0.0001, for all related costs, see Table 29).

With regard to shunt procedures, Table 27 shows that the numbers were broadly similar between the groups, and this was also true for shunt-related costs (see Table 27). When cranioplasty and shunt costs were combined, due to the high costs associated with cranioplasty procedures, they were significantly lower in the craniotomy group compared with the DC group (£1258 vs. £3,228, p < 0.0001) (see Table 29).

Eighty-two post-discharge cranioplasty and shunt (index or revision) procedures were reported in the hospital-recorded data. This exceeds the 45 overnight stays with associated skull/brain operations reported in the patient self-report questionnaire. Therefore, as described in the methods, we have not included patient-reported overnight stays with an associated skull/brain operation in the base-case analysis, to avoid the potential for double-counting and cost duplication.

In terms of patient self-reported resource use data, at 12 months, 180/248 (73%) had complete resource use data (see Table 27) based on available data, Table 28 presents the associated levels of post-discharge resource use, including overnight stays not associated with a skull/brain operation, HCP visits and number of scans. Levels were broadly similar in both groups, with no significant differences between any of the resource use items or costs (see Tables 28 and 29).

Results between arms were also similar when the index admission, cranioplasty and shunt, and post-discharge costs (excluding care home and carer costs) were summed to estimate per patient overall NHS and PSS costs (mean per patient cost, craniotomy vs. DC group: £48,509 vs. £53,573, p = 0.510, see Table 29).

Outcomes

There was 132/248 (53%) complete EQ-5D data (see total QALYs in Table 27) at all time points (NSU discharge, 6 and 12 months post randomisation), 72% at baseline, 79% at 6 months and 86% at 12 months (those that died were given an EQ-5D score of ‘0’ and would therefore not be recorded as ‘missing’).

Table 30 shows the mean EQ-5D utility scores at each time point on an available case basis, along with the change in score. The EQ-5D scores at follow-up were higher (non-significant at 6 months, p = 0.016 at 12 months) in the craniotomy group compared with the DC group. Furthermore, the change in EQ-5D score from baseline to 12 months was significantly higher in the craniotomy group compared with the DC group (difference: 0.146, p = 0.013). There was no significant difference between groups for the total QALY score, based on available data.

Table 31 shows the percentage of favourable and unfavourable outcomes, as defined previously in the methods, for the GCS score at baseline, the GOSE score at 12 months and the GOSE score at 12 months dependent on the GCS score at baseline (where a lower GCS score at baseline means a lower GOSE score at 12 months is required to be defined as ‘favourable’). The percentage of favourable GCS scores at baseline were lower in the craniotomy arm compared with the DC arm, though not significantly (29% vs. 40%, p = 0.093). At 12 months, the percentage of favourable GOSE scores was higher, but not significantly so, in the craniotomy arm compared with the DC arm, both in the GCS-independent measure (48% vs. 37%, p = 0.102) and GCS-dependent measures (44% vs. 32%, p = 0.7).

Analyses

Material with this chapter has been reproduced with permission from Pyne et al.19 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) licence, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: https://creativecommons.org/licenses/by/4.0/. The text below includes minor additions and formatting changes to the original text.

Cost–utility analysis

Table 32 presents estimates of mean incremental costs and incremental QALYs, generated from the regression analysis (both costs and QALYs adjusted for baseline EQ-5D, age and sex), along with the ICER and probability of being cost-effective at a threshold (λ) of £20,000 per QALY.

For the base case (based on ITT/MI), the mean difference in cost for the craniotomy group compared with the DC group was −£5520 (95% CI −£18,060 to £7020), with a mean QALY difference of 0.093 (95% CI 0.029 to 0.156). That is to say that the mean total cost was lower (non-significantly) and the mean QALY scores were higher (significantly) in the craniotomy group compared with the DC group, which means that craniotomy is dominant over DC; it is estimated to be less costly and more effective.

Cost-effectiveness analysis

Table 33 presents estimates of the mean incremental costs and OR of a favourable outcome, generated from linear regression and logistic (logit) regression, respectively. For the craniotomy group compared with the DC group, the mean difference in cost was −£4536 (95% CI −£17,374 to £8301), with an OR of favourable outcome on the GOSE score of 1.682 (95% CI 0.995 to 2.842, p = 0.052). That is to say that the mean total cost was lower (non-significantly) and the odds of a favourable score was higher (non-significantly) in the craniotomy group compared with the DC group, which means that craniotomy again dominated DC. Similar results were found for the GCS-adjusted binary GOSE score (see Table 33).

Decision uncertainty

In terms of uncertainty, the base-case probability that craniotomy is cost-effective compared with DC at a threshold of £20,000 was 87% according to the CEAC (see Table 32). This indicates that there is a high degree of certainty associated with the decision that craniotomy compared with DC is cost-effective at a threshold of £20,000 per QALY.

Sensitivity analyses

The results of the SAs that were undertaken for the CUA are shown in Table 32, and for the CEA in Table 33. Of note, when, carer and care home costs were included in the wider perspective analysis (SA1) the cost difference increased between craniotomy and DC. In all these analyses craniotomy dominates DC, it is less costly and more effective. In every cost–utility SA, other than the complete-case analysis (SA6), the probability that craniotomy is cost-effective compared with DC at a threshold of £20,000 per QALY is ≥ 87%, which reinforces the high degree of certainty associated with the decision. For the complete-case analysis (SA6) the mean difference (craniotomy vs. DC) in cost was −£1917 (95% CI −£15,564 to £11,729) with a mean QALY difference of 0.071 (95% CI −0.011 to 0.153). Craniotomy therefore still dominates DC, though there was a slightly lower probability of cost-effectiveness (68%) at a threshold of £20,000 per QALY. This may reflect the nature of the missing data.

Discussion

Main findings

Based on the results of the cost–utility and cost-effectiveness analyses craniotomy dominates DC, it is less costly and more effective (see Tables 32 and 33). There is a high degree of certainty associated with the conclusion of the CUA, as choosing DC would be estimated to be associated with an 87% probability (at a willingness-to-pay threshold of £20,000 per QALY) of making the wrong decision. Furthermore, this is supported by the results of the SAs, all of which also found craniotomy to be less costly and more effective, with a 68–99% probability (at a willingness-to-pay threshold of £20,000 per QALY) of making the wrong decision by using DC rather than craniotomy in these patients.

Comparisons with other studies

We are not aware of any previous economic evaluations that have compared craniotomy with DC for patients with ASDH, although the general literature on this topic shows that there is much clinical uncertainty as to which operation to undertake.19,32,35

Study limitations

In line with good practice recommendations for cost-effectiveness analyses,5 we concentrated on the large cost drivers and excluded resources that were not expected to differ between the two treatment arms (e.g. routine monitoring scans or tests). That said, the mean costs could be heavily influenced by outliers, due to the high unit costs associated with resource use, such as LOS in ICU and operation costs.

An additional potential limitation is that we specified, within the HEAP, that lost productivity costs would be estimated, but this was not undertaken for the aforementioned reasons (details of working patterns before their brain injury were not requested). We did, however, estimate the number of hours participants had worked since their brain injury (in the 12-month follow-up period), where the results in Table 28 (based on available data) show that the estimated mean number of hours worked was higher in the craniotomy arm compared to DC. Assuming that, due to randomisation, the number of hours worked was broadly similar between groups at baseline, then this would suggest that had lost productivity costs been estimated, this would not have changed the base-case or SA1 results that craniotomy had a lower mean cost compared to DC. Indeed, as the estimated mean number of hours worked was lower in the DC arm, this might suggest lost productivity costs were higher in the DC arm, which would only reinforce the base-case and SA1 results that craniotomy was associated with lower costs.

Regarding HRQoL, QALY scores (EQ-5D recorded at all time points) were available for 53% of participants only, and the amount of missing data were greater at discharge than at 6 (79%) and 12 months (86%) (see Table 27). Some of the missing EQ-5D baseline (NSU discharge) data may be due to participants being discharged at short notice or at the weekend when a research nurse was not available. As some patients had not yet been discharged from hospital by 6 months, this may also explain why there was a greater degree of missing EQ-5D data at this time point compared with 12 months. Such missing data is a limitation, but it should be noted that the same conclusion can be drawn from both the results of the complete-case and imputed (base-case) analyses.

A further potential limitation is that, for ethical reasons, baseline HRQoL (EQ-5D) scores were taken at discharge from NSU rather than at the point of randomisation. The date of discharge from NSU is post intervention, and there is, therefore, the potential for any benefits to be underestimated by assuming the score at the date of discharge to be the baseline score. To test the potential impact of this, a SA (SA4) was done, where the baseline EQ-5D score was assumed to be that of worst possible health (−0.594). The results of this SA differed little from the base case (see Table 32), with craniotomy still dominant over DC; less costly and more effective.

A further potential limitation is that the conclusions might differ if results were estimated over a longer follow-up period. However, if the treatment effect found in this study was maintained beyond 12 months, the conclusions would be unchanged, because extrapolation would result in further QALY gain for the craniotomy group compared with the DC group, while the costs would remain similar between the two groups.

The main strength of this economic evaluation is that it is based on a large, multicentre, randomised study. Previous evidence was inconclusive and highlighted the uncertainty in how to treat patients with ASDH.32,35,36 Furthermore, until now, there were no randomised studies investigating this topic.

Implications

It has been argued that decisions should be based on the mean estimates (of cost and effect), regardless of, for example, the associated level of the p-value.37 This is advocated on the basis that decisions (about implementation or not) need to be made, regardless of the associated quality of the evidence, and that the mean value is the best estimate to use, for example, to predict total cost. As such, economists tend to focus on the mean estimates of cost and effect when making recommendations as to whether treatments should be provided by the NHS or not.

Conclusion

Our economic evaluation has shown that in a UK population of patients with traumatic ASDH, craniotomy had lower mean costs and a higher mean effect (dominating DC). There was an associated high degree of certainty that craniotomy compared with DC was cost-effective at a threshold of £20,000 per QALY. Additionally, when several SAs were conducted, the main conclusion remained unchanged. Consequently, the health economic analysis supports the aforementioned recommendation (based on the primary outcome) that a craniotomy should be undertaken, rather than a DC, if it is operatively feasible to replace the bone flap.