Toward the end of life a significant proportion of cancer patients experience severe pain (Reference Haumann, Joosten and Everdingen1;Reference Breivik, Cherny and Collett2). A European survey of 5,000 cancer patients found that 72 percent experienced pain, 90 percent of which was of moderate-to-severe intensity (Reference Breivik, Cherny and Collett2). Inadequate management of pain at the end of life is likely to have a significant detrimental impact on quality of life (Reference Shi, Smith, Michonski, Stein, Kaw and Cleeland3) and may lead to increased healthcare costs through unplanned hospital visits and admissions (Reference Fortner, Okon and Portenoy4). Indeed, one study indicated that poor pain control is the most frequent reason that cancer patients at home need emergency medical help (Reference Adam, Wassell and Murchie5). Clearly, achieving good pain management at the end of life is a priority but service provision in this regard often falls short and interventions that are known to be effective are poorly implemented.
There is growing evidence that self-management strategies facilitated by better communication, pain assessment and patient education can lead to improved pain outcomes (Reference Adam, Bond and Murchie6;Reference Martinez, Aslakson and Wilson7). However, economic evidence for these types of interventions is less abundant. Despite the acknowledgement that economic evidence is key to improving access to effective palliative care (Reference Bickel and Ozanne8;9) evaluations are still relatively rare in this context (9;Reference McCaffrey, Cassel and Coast10) and evaluations in the more specific context of pain at the end of life are rarer still (Reference Bennett, Mulvey and Campling11).
We sought to generate evidence on the cost-effectiveness of pain self-management strategies at the end of life by conducting evaluations of case study interventions. Improving the Management of Pain from Advanced Cancer in the Community (IMPACCT) is a U.K. research program aimed at the development and testing of interventions for patients at home to facilitate improved pain outcomes through self-management when delivered in addition to routine community palliative care services. The Tackling Cancer Pain Toolkit (TCPT) is a small booklet and DVD containing information on pain and medicines, alongside self-directed learning activities and sources of further information. PainCheck is an internet based pain monitoring system that enables patients to communicate pain data to health professionals routinely. The system alerts professionals when pain scores are above specific thresholds, and allows them to provide feedback through the system or contact patients directly for further assessment.
Primary research evaluating the effectiveness of the IMPACCT interventions is on-going. The aim of the current research was to conduct economic evaluations of PainCheck and TCPT interventions when added to community palliative care delivery to estimate their value for money compared with usual care. The evaluation adopted a decision-analytic modelling approach incorporating published estimates of effectiveness from similar interventions and was designed to inform implementation strategies.
Methods
The economic evaluation followed the NICE reference case (12) and hence was a cost-utility analysis with the primary outcome cost per incremental quality-adjusted life-year (QALY) presented from the perspective of the U.K. healthcare and personal social services provider. The evaluation compared the PainCheck and TCPT pain management strategies in cancer patients at the end of life with usual care in this context. We defined usual care as routine care received by the patients at home from their local community palliative care team as determined by local policies and practices. As part of usual care in the United Kingdom, patients at the end of life with specialist needs (for example, poorly controlled pain) should be referred to community palliative care specialists and receive pain treatment and advice as part of that care. While this practice is increasing, services are highly variable across the U.K. and pain is often managed solely by means of GPs and community pharmacists. There is no set visit routine as patients access services as and when required but those on strong opiates will typically be reviewed 2–3 times per month, depending on response. During these contacts with health care professionals, response is assessed with pain rating items (such as those in the Numeric Pain Rating Scale) and with less formal questions about pain.
Decision Model
We developed a decision-analytic model to facilitate the economic evaluation, an earlier version of which has been previously described (Reference Bennett, Mulvey and Campling11). The decision model was a Markov cohort model the structure of which (Figure 1) was informed by patients, clinicians and other published decision models relating to pain. The model is structured around pain health states which are based on accepted 0–10 pain scale cut-offs: 0–4 = No/Mild Pain; 5–6 = Moderate Pain; 7–10 = Severe Pain (Reference Serlin, Mendoza, Nakamura, Edwards and Cleeland13). More severe pain states are associated with higher costs due to emergency and unplanned healthcare resource use and poorer quality of life. The hypothetical cohort of patients (mean age = 72.4 years based on the average age of IMPACCT survey respondents) transited through the model in weekly cycles until dying or until the end of the time horizon (one year). The model parameter values are presented in Table 1 and described below.
Fig. 1. Decision model structure.
Table 1. Model Parameter Values
a Distributions used in the probabilistic sensitivity analysis.
b Derived using generalized linear model (Gamma family, Log link).
c Derived using ordinary least squares regression.
dChange in pain rating scale vs. usual care.
e Weekly probability applied only to usual care arm, all other transition probabilities assumed to be 0.0.
f Weekly mortality following referral to palliative care.
gPain reductions are first converted to probabilities based on likelihood of pain category change in IMPACCT survey respondents.
hUses estimates based on the regression covariance matrix.
Costs
A costing of the IMPACCT interventions is provided in supplementary Table 1. Costs for TCPT related only to printing of the material (one off cost of GBP14.34 [EUR16.33]) and brief telephone contact with a nurse (weekly cost of GBP9.10 [EUR10.36]). Resources required for the provision of PainCheck were larger due mainly to programming requirements. PainCheck incurred a one-off cost (covering maintenance programming, leaflet printing and nurse time to introduce the tool) of GBP49.01 [EUR55.81] and a weekly cost (covering nurse time for monitoring and patient response) of GBP24.47 [EUR27.86]. Only intervention implementation costs are included and those relating to intervention development are excluded. The health state costs (Table 1) are derived from the IMPACCT patient survey where respondents completed a questionnaire capturing primary and community (e.g., GP visits, nurse contact, etc.) and secondary or hospital (e.g., visits to A&E, hospice stays, etc.) care use in the previous month. Unit costs were obtained from national sources including the PSSRU Costs of Health and Social Care (Reference Curtis14), NHS Reference cost database and the British National Formulary (BNF). All costs are U.K. pounds and 2017 prices with equivalent Euro values provided in brackets using a historical currency converter (for June 30, 2017).
Health State Utility
The utility values for the health states (Table 1) were derived from the IMPACCT patient survey in which respondents completed the EQ-5D-3L measure scored using the U.K. tariff (Reference Dolan, Gudex, Kind and Williams15). We also conducted a sensitivity analysis where utilities were based on the EORTC-8D measure (Reference Rowen, Brazier and Young16). In view of the fact that the QALY framework has been criticized in palliative care (Reference Wichmann, Adang and Stalmeier17), we conducted another analyses based on the ICECAP measure (Reference Al-Janabi, Flynn and Coast18) which adopts the capability framework but which may enable cost-utility analyses.
Transition Probabilities
Meta-review of Educational and Monitoring Interventions
We conducted a meta-review of educational and feedback/monitoring interventions for improving cancer pain at the end of life. We searched for randomized controlled trials (RCTs) of nonpharmacological interventions in advanced cancer pain and for reviews in this area. We searched Embase Classic and Embase 1947–2017; Ovid MEDLINE(R) 1946–2017; MEDLINE(R) In-Process & Other Non-Indexed Citations. Searches were conducted in February 2017. We short-listed and reviewed those studies reporting a systematic review of either educational or feedback/monitoring interventions for pain. Non-English language publications were excluded. Two researchers reviewed abstracts and differences were resolved by consensus meeting. The shortlisted publications were discussed with a clinician to identify which best reflected the properties of the IMPACCT interventions and hence were suitable proxy estimates of effectiveness. In particular, we were interested in studies of interventions that encouraged self-management (as opposed to having significant levels of health care professional input). In identifying a suitable review study, we assessed their inclusion criteria to ensure this aligned with the two IMPACCT interventions.
Candidate studies had to report synthesized estimates of effectiveness (and standard error/deviation). Effectiveness had to be based on the Numeric Pain Rating Scale (NPRS) (Reference Jensen and McFarland19). The NPRS is a reliable and valid, self-complete measure of pain. Responses are captured on an 11-point numeric scale ranging from “0” (i.e., “no pain”) to “10” (i.e., “pain as bad as you can imagine”). NPRS questions cover worst intensity, least intensity and average intensity (over the last 24 hr). Our review specifically attempted to identify reports of average pain intensity as this item was also included in the IMPACCT patient survey and was the basis of health state parameters. We also considered RCTs that were published after reviews had been completed.
Effectiveness, Pain Progression, and Survival
The initial distribution between the pain severity groups was determined by the IMPACCT patient survey (Table 1). Effectiveness translated into health state transition probabilities by observing the pain category change in the IMPACCT survey respondents following the relevant pain reductions. Intervention effects (rather, a sixth of) were assumed to occur on an incremental basis for 6 weeks after which pain levels were maintained. In the usual care arm pain was assumed to progress over time. Progression was based on a multinomial regression model predicting change in EQ-5D-3L pain/discomfort item response over time and after controlling for survival in a recent trial including cancer patients at the end of life (Reference Meads, Marshall, Hulme, Dunn and Ford20).
The survival of the cohort was estimated using parametric regression which was fitted to other IMPACCT data. The data (n = 4,638; 84 percent with a cancer diagnosis) was retrospectively collected on all patient referrals to specialist palliative care services in the city of Leeds, United Kingdom over 2 years (2012–14). The sample had a mean survival of 80.77 days (SD = 117.81). Several models were applied to the data including exponential, Weibull and Gompertz. Based on best visual fit with the observed Kaplan-Meier curve and lowest Akaike information criterion (AIC) value, Weibull was selected. As the gamma factor was significant the use of the Weibull model is justified as this indicates a nonconstant (and declining) hazard function. The same risk estimates from this analysis were applied to all health states. The survival model estimates were permitted to vary in the probabilistic sensitivity analysis following Cholesky decomposition for correlated regression parameters. During the 52-week model time horizon, 97 percent of the cohort were expected to have died. Although there is some evidence that pain can independently explain survival, the evidence is mixed (Reference Zylla, Steele and Gupta21) and thus here we assumed they are unrelated.
Analysis
Cost-effectiveness was assessed based on incremental cost-effectiveness ratio (ICER) and Incremental net monetary benefit (INMB) values. These were generated separately for PainCheck and TCPT in pairwise comparisons versus usual care. A fully incremental analysis was not conducted because these interventions might be used in combination. To test this, a further analysis evaluated receipt of the combined active interventions. The NICE willingness to pay per incremental QALY threshold ([λ] = GBP20,000 [EUR22,774]) was adopted with ICERs below this figure indicating cost-effectiveness. The INMB is a transformation of the ICER where positive INMB values indicate cost-effectiveness; INMB was calculated thus (where ΔE are incremental effects and ΔC are incremental costs):
$$INMB = \lpar {\lambda {\ast}\Delta E} \rpar {\rm -}\Delta C$$A range of one-way, deterministic sensitivity analyses were conducted where alternative model parameter values or assumptions were applied and their impact on cost-effectiveness observed. The deterministic sensitivity analyses tested: basing effectiveness on the upper confidence interval of pain reduction values from meta-analyses (i.e., assuming reduced effect); assuming an additional cost (10 percent of the cohort receiving a nurse visit) in the PainCheck and TCPT arms; assuming that 50 percent of those in the PainCheck and TCPT arms experienced pain progression (at usual care rate) after 6 weeks rather than maintaining pain levels; assuming the starting cohort all had either moderate or severe pain (no mild cases); basing QALY calculations on the EORTC-8D and ICE-CAP; and removing the half cycle correction. We conducted additional analyses exploring the impact of using costs from individual studies identified by the systematic reviews. For PainCheck, a U.S. study was used as an alternative source of intervention costs (Reference Kroenke, Theobald and Wu22) and for TCPT we used a Dutch study (Reference de Wit and van Dam23); these were chosen from the review as they had relatively large sample sizes, levels of effectiveness similar to the overall mean and as they reported the resources required to deliver the interventions (see Supplementary Table 2).
We also conducted threshold analyses to establish the costs and effectiveness required to achieve cost-effectiveness. We conducted a probabilistic sensitivity analysis (PSA) to characterize overall parameter uncertainty in the model by assigning probability distributions to each of the input parameters, and randomly drawing from these probabilities over 10,000 Monte Carlo simulations, yielding 10,000 estimates of ICERs and INMB. PSA results were plotted on the cost-effectiveness plane and INMB estimates used to generate the cost-effectiveness acceptability curve (CEAC) (Reference Fenwick, O'Brien and Briggs24) The CEAC illustrates the probability that each intervention would be cost-effective given a range of willingness-to-pay thresholds. Discounting was not required as all costs and benefits were experienced within one year. A half-cycle correction was applied but a sensitivity analysis was also conducted where no half-cycle correction was applied
Finally, we explored the value of further research by conducting a value of information (VoI) analysis which attaches a formal cost to the uncertainty in the cost-effectiveness results (Reference Claxton and Sculpher25). We estimated population level expected value of perfect information (EVPI) which required information on the number of patients who could benefit from the interventions (incidence and number of years the decision is relevant for). According to national data, 149,152 people die each year of cancer in England (26). Based on European survey results (Reference Breivik, Cherny and Collett2) indicating that 72 percent have pain which in 90 percent of cases is moderate/severe, yields a relevant annual population of 96,650. We assumed that the decision problem was relevant for 10 years and we discounted values beyond one year based on a 3.5 percent discount rate. We used the Sheffield Accelerated Value of Information (SAVI) tool (Reference Strong, Oakley and Brennan27) to estimate the Expected Value of Perfect Parameter Information (EVPPI) for single parameters. Using the incidence figures we also estimated the budget impact of implementing the interventions over the same period. All analyses were conducted in Stata IC software (version 14; StataCorp) and Excel (Microsoft).
Results
The PRISMA flow diagram for the literature review is included in the Supplementary Materials (Supplementary Figure 1). Seventeen reviews (Supplementary Table 3) were identified and discussed with the lead clinician. Reviews were not deemed suitable for several reasons including: not synthesizing study results (e.g., Adam et al. (Reference Adam, Bond and Murchie6) and Ling et al. (Reference Ling, Lui and So28); including noncancer studies (e.g., Bennett et al. (Reference Bennett, Bagnall and Raine29); having a restricted population (Reference Santos Salas, Fuentes Contreras and Armijo-Olivo30); being superseded by more recent reviews (Reference Sellick and Zaza31); or synthesizing outcomes from studies investigating interventions comprising a significant element of face-to-face healthcare professional delivery (Reference Jho, Myung, Chang, Kim and Ko32–Reference Marie, Luckett, Davidson, Lovell and Lal34).
Results from a review and meta-analysis of pain assessment and feedback interventions was thought a suitable proxy for the effectiveness of PainCheck (Reference Adam, Burton, Bond, De Bruin and Murchie35). The meta-analysis estimated mean reductions in pain ratings of -0.59 (lower CI = −0.87; upper CI = −0.3). These figures led to a 0.46 and a 0.39 probability of transiting from Moderate to No/Mild pain and Severe to Moderate pain categories, respectively, over 6 weeks. The meta-analysis selected as a proxy for TCPT intervention (Reference Bennett, Bagnall and José Closs36) estimated mean reductions in pain ratings of -1.1 (lower CI = −1.8; upper CI = −0.41). These figures led to a 0.56 and 0.51 probability of transiting from Moderate to No/Mild pain and Severe to Moderate pain categories, respectively, over 6 weeks. We could not identify a study that would provide a reasonable approximation of the joint effectiveness of PainCheck and TCPT thus we made an assumption that this was: -1.1 + (-0.59 × 0.5) or -1.395. We tested this value in the model in a comparison with usual care but included the full costs for both active interventions.
The results from the base case and sensitivity analyses are included in Table 2. PainCheck and TCPT were both cost-saving and more effective than (i.e., they dominated) usual care. However, TCPT led to greater NMB. In general, the incremental costs and benefits were modest. Sensitivity analyses exploring alternative utility values, costs, and maintenance of effect after 6 weeks in most cases did not change the conclusion for either intervention. Only changes to the levels of effectiveness substantively altered the results. Using the upper confidence interval from the respective meta-analyses led to PainCheck no longer being cost-effective while TCPT remained cost-effective. Adopting alternative intervention costs (and effects) from individual studies from within the systematic reviews had the same effect.
Table 2. Cost-effectiveness results
Threshold analyses indicated, all else being equal and at current costs, PainCheck requires a pain reduction of at least 0.50 to remain cost-effective; at current levels of effectiveness, PainCheck is cost-effective up to one-off intervention cost of GBP401.86 [EUR457.59] or weekly intervention costs of GBP58.25 [EUR66.33] being incurred. At current costs, TCPT requires a pain reduction of at least 0.40 to remain cost-effective; at current levels of effectiveness TCPT is cost-effective up to one-off intervention costs of GBP744.61 [EUR847.88] or weekly intervention costs of GBP43.44 [EUR49.46] being incurred. The combined intervention was cost-effective but represented less value for money than TCPT alone.
The probabilistic results yielded similar results to those from the deterministic analyses although INMB is reduced. A higher number of simulations in the TCPT arm were cost-saving versus usual care than is the case with PainCheck. In both comparisons, all simulations showed positive QALY gain over usual care. The cost-effectiveness planes (Figure 2) indicates the spread of ICERs from the Monte Carlo simulations, representing parameter uncertainty. The cost-effectiveness acceptability curves (Supplementary Figure 2) indicate that PainCheck has a 65 percent chance of being the cost-effective option and TCPT a 99.5 percent chance of being cost-effective at a willingness to pay threshold of GBP20,000 [EUR22,774] per QALY.
Fig. 2. Cost-effectiveness planes.
For PainCheck, the population EVPI per year was GBP2,945,026 [EUR3,353,460] and GBP24,592,210 [EUR28,002,805] over 10 years, indicating a significant cost of uncertainty in the decision. Given that TCPT was highly likely to be cost-effective, the population EVPI was low (GBP10,464 [EUR11,915] per year) indicating further research on the topic may be unnecessary. The EVPPI figures for TCPT were effectively zero given the lack of uncertainty in the decision. They were also zero for PainCheck parameters except for the severe and moderate pain health state costs. The cost of decisional uncertainty surrounding the moderate health state cost was GBP971,100 [EUR1,105,778] over 10 years. This was much higher for the severe health state cost parameter (GBP15,590,000 [EUR17,752,115]) and the magnitude suggests additional research into the cost of cancer pain may be warranted. The budget impact estimates indicate that PainCheck would lead to savings of GBP663,831 [EUR755,895] per year or GBP5,543,272 [EUR6,312,046] over 10 years. The estimated cost savings for TCPT were estimated to be GBP23,369,253 [EUR26,610,241] per year and GBP195,143,121 [EUR222,206,740] over 10 years.
Discussion
These analyses represent the most comprehensive assessment to date of the value for money of pain management interventions at the end of life. Using a decision modelling approach, we compared two types of intervention (educational and pain monitoring/feedback) against usual care from the perspective of the health and social care provider. We relied on estimates of effectiveness from published meta-analyses. PainCheck and TCPT are relatively inexpensive and the evaluations suggest that both have the potential to be cost-effective. Indeed, in the base case analyses, both interventions were cheaper and more effective than usual care. The conclusions were relatively robust to several sensitivity analyses. The effectiveness parameter appeared relatively more influential in determining cost-effectiveness than intervention costs or utility values. Assuming all patients were either in moderate or severe pain improved the benefit of the IMPACCT interventions noticeably. A scenario where both interventions were received would be a worse strategy than implementing TCPT alone, mainly due to the additional costs of PainCheck. However, assumptions were made here on the level of combined effect.
Although the use of the generic QALY framework (based on EQ-5D) to evaluate palliative care interventions has been questioned (Reference Wichmann, Adang and Stalmeier17), here the use of condition specific QALY (EORTC-8D) and capability-based approaches (ICE-CAP) yielded reduced incremental benefits for the active interventions. It is possible that the EQ-5D fails to capture additional benefits this patient group may experience following improved pain management such as a greater feeling of control and the emotional positives that come with being able to stay at home. However, the EQ-5D appears adept at discriminating between people based on pain level and this may explain the relative performance of the utility measures as the decision-model is predicated on pain categories.
The probabilistic analyses suggest that both PainCheck and TCPT are highly likely to deliver QALY gains over usual care. However, in both cases, the interventions were less likely to lead to cost-savings. Incorporating parameter uncertainty in the model suggests that PainCheck and TCPT have 65 percent and 99.5 percent chance of being cost-effective, respectively. The greater uncertainty surrounding PainCheck relates to the higher cost and lower assumed effect. The Value of Information analysis suggest that additional research on PainCheck is warranted and the EVPPI values indicate that reducing uncertainty surrounding cost estimates should be a focus.
The costs predicted here are similar although lower than those presented in a recent publication (Reference Round, Jones and Morris37). However, we used a different definition of end of life and much lower survival periods. Although the cost savings associated with each intervention were modest, values for the estimated population are potentially substantial (GBP5,543,272 [EUR6,312,046] and GBP195,143,121 [EUR222,206,740] over 10 years for PainCheck and TCPT, respectively).
Limitations
We did not have direct estimates of the effectiveness of either active case study intervention evaluated here and relied on synthesized estimates from meta-analyses. While the studies used in the analyses as the basis of effectiveness estimates were selected following a meta-review and due consideration, it is possible that these reviews incorporate studies that are not accurate reflections of the PainCheck and TCPT interventions.
It is possible that these reviews and synthesized outcomes derive from disparate study interventions or designs which may have biased results. Adam and colleagues (Reference Adam, Burton, Bond, De Bruin and Murchie35), reviewing patient feedback/monitoring studies, found most were prone to some element of bias and concluded that two studies contributing to the synthesis should be treated with caution. As the effectiveness estimates in those two studies were above the mean, their exclusion would reduce the assumed overall effectiveness (albeit slightly given study weightings) for PainCheck.
While there were very few reviews relating to patient feedback/monitoring, there were several targeting educational interventions. The review by Bennett and colleagues was selected based on appropriateness of their study inclusion criteria. There is limited information in the review of the quality of studies included and potential for bias. Examination of study outcomes indicates the presence of significant heterogeneity with one outlier study reporting a very large intervention effect and this may have biased results. However, it is worth pointing out that the uncertainty in outcome should be captured in the probabilistic sensitivity analyses presented here. Furthermore, excluding that study from the weighted mean still yielded an effect greater than that required (−0.40) for TCPT to be cost-effective. Although, of course, this is based on the use of costs estimated for the TCPT intervention. It is also possible that there may be individual randomized controlled trials that better reflect the potential effect of either active intervention.
Although it has been suggested that informal care costs are an important consideration in palliative care economic evaluations (Reference Gardiner, Ingleton, Ryan, Ward and Gott38), we did not include these in the current analyses. We wished to adhere to the NICE reference case which excludes these costs but, more importantly, we did not have health state data relating to informal care costs. It is likely that carers of patients in higher pain categories incur higher costs; thus, adopting a broader perspective and including informal care costs would likely increase the estimates of value for money for PainCheck and TCPT. However, increased self-management may also increase informal care requirements and further research is needed to explore this.
Further Research and Implications
The decision model generated and tested here is robust and may be a tool that, following adaptations, has other useful applications in palliative care. It may also be useful for local decision makers considering commissioning alternative pain management strategies. The active interventions evaluated here have the potential to be cost-effective and additional research, for example, in the form of randomized trials or observational data collection and analysis, may be warranted to add to the evidence base.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0266462319000114.
Conflicts of interests
The authors declare that there are no conflicts of interest.
