Economic evaluation

Introduction

Evidence on the short- and long-term cost-effectiveness of pedometer-based interventions could support the development of policy and practice encouraging increased PA. Calls for evidence on ‘what works’ have been made, in both primary and secondary prevention policy documents, along with appeals to link interventions to clear health outcomes and ensure that resources are used most efficiently.94,95

To date, only one published estimate of the cost-effectiveness of pedometer programmes for the UK was based on primary evidence.58 A small (n = 79), highly selected (80% of women from one GP practice in Glasgow) sample was used. A ‘maximal’ pedometer-based walking programme, which included two 30-minute consultations (based on the transtheoretical model of behaviour change), was compared with a waiting list control group, which was asked to wait for 12 weeks, after which the group members received a ‘minimal’ walking programme, which included a pedometer and two 5-minute slots of brief advice.45,96 Compared with the 12-week waiting list control group, it cost an additional £92 per person to achieve an additional eight people meeting the target of 15,000 steps per day over a 12-week period. Comparing the maximal walking programme with the minimal walking programme, it cost a further £591 for one additional person to achieve the same target. No data were collected on QALYs, and long-term cost-effectiveness was not modelled.

Elsewhere, in Australia, New Zealand and the Netherlands, evidence on the benefits of pedometer-based interventions for primary prevention has been assessed for community-based adults with low PA levels.97–99 Interventions such as pedometer prescriptions and pedometer-based telephone coaching were compared with time-based activity prescriptions or usual practice. Evidence suggests that pedometers may be cost-effective in the long term, but estimates vary widely (from being cost-saving and having fewer disability-adjusted life-years in Australia in the long term to €11,110 per QALY gained). The generalisability of results to other contexts has also not been considered.100

This chapter examines the short- and long-term cost-effectiveness, from the NHS perspective, of alternative pedometer-based walking programmes to increase PA levels using PA outcomes for comparison with other PA programmes and QALYs to aid decision-making beyond PA programmes. The interventions are compared against usual practice, in inactive adults aged 45–75 years from south London, and are as described in Chapter 2:

1. provision, by post, of pedometers with written instructions
2. pedometers provided with tailored support from a practice nurse.

This chapter is structured into two sections: (1) a within-trial analysis with a time horizon of 1 year, and (2) beyond-trial modelling that takes a lifetime perspective. In each section, the methods and results are presented. This is followed by a discussion of the findings in the context of the strengths and weaknesses of the study, as well as current literature.

Within-trial cost-effectiveness analysis

The population, interventions and comparator are identical to Chapter 2. Harris et al.70 set out the protocol for methods, including for the economic evaluation. This section covers methods used to measure, value and aggregate costs and outcomes, the treatment of missing data and methods of assessing cost-effectiveness.

Methods

Identification, measurement, valuation and aggregation of cost

To identify NHS resource use, meetings with the trial team established ‘who did what, to whom and how often’,101 accounting for events that are likely to have high unit costs, be frequent or differ between trial arms, and excluding research-focused costs. NHS resources identified for collection included:

• Set-up of service (e.g. design, setting up the intervention in GP practices and staff training, but excluding trial set-up; see Appendix 3, Table 27).
• Delivery of service, including, for example, pedometers, post/telephone services, handbooks and staff time; all costs fell within months 0–3, except research assistant contacts with participants (4–12 months; see Appendix 3, Table 28).
• Health service use in primary care (GP and nurse consultations, excluding nurse PA consultations undertaken as part of the trial) and secondary care (hospital admissions, A&E, outpatients), as changes could occur from treating AEs and changes in lifestyle (see Appendix 3, Table 29); the health service data were available on the 1005 out of 1023 randomised participants who gave written informed consent for their primary care data to be downloaded. Of these, 956 participants (323 in the control group, 312 in the postal group and 321 in the nurse-support group) also had 12-month outcome data and, therefore, were included in the health economic analyses relating to health service use.

Resource use was measured using administrative/trial management records, electronic diaries and interviews of the trial manager and the principal investigator. Participant-level health service use (e.g. GP visits, referrals) was collected through a one-time download of GP records, for those who gave explicit consent for this, at the end of the trial (see Appendix 3, Tables 27–29).

NHS resources were valued using national costs102,103 (see Appendix 3, Tables 27–33) to increase generalisability. Where national unit costs were not available, local unit costs from St George’s Hospital, London, were used. All costs were expressed in 2013–14 pounds sterling, inflated to the same base year when appropriate, using the Hospital and Community Health Service inflation index.103 As the study covers 1 year, costs and outcomes were not discounted. All resources and costs were collected at, or apportioned to, the trial participant level. The total cost per participant was the sum of each resource use multiplied by the relevant unit cost over 0–3 and 4–12 months, with costs censored at 12 months.

To support a sensitivity analysis of an alternative wider viewpoint that included participants, as participation can be affected by economic barriers, three types of costs borne by participants were collected: participation in the intervention (e.g. time and money spent accessing the intervention for months 0–3; see Appendix 3, Table 30), money shown to contribute to the costs of ‘walking’ and other PA (e.g. membership/event fees, shoes/clothing, food/drink)104 and money spent as a result of falls/fractures/sprains/injuries. These data were collected for months 1–3 using participant-completed questionnaires at 3 months and for months 10–12 using the participant-completed questionnaires at 12 months (see Appendix 1). As the data for months 10–12 were follow-up data beyond the intervention, costs from months 10–12 were multiplied by three (to approximate annual costs when added to the cost from months 1–3) and added to the costs from months 1–3 for an annual participant cost.

Measurement, valuation and aggregation of outcomes

The economic analysis uses indicators of PA outcomes. The use of cost per additional step aids comparison of inputs with directly intended and objectively measured outputs, which the trial was specifically powered to detect, and which therefore relates the economics to the main trial outcome. As objectively measured weekly minutes of MVPA in bouts of ≥ 10 minutes were statistically significantly different, have important health impacts, provide a second point of comparison to other studies and link directly to the longer-term model, these were also included as an additional outcome measure for assessing cost-effectiveness.

To facilitate the comparison of the PACE-UP trial with other health interventions using the quality-of-life measure recommended for evaluating health interventions in England, participants also completed the EQ-5D-5L questionnaire at baseline, 3 and 12 months. The EQ-5D-5L, rather than the EQ-5D-3L (EuroQol-5 Dimensions, three-level version), was selected as the EQ-5D-3L is known to suffer from ‘ceiling effects’ and the five-level version was expected to be more sensitive to differences in health among healthier people and, therefore, less subject to suffer from ceiling effects. This has been subsequently shown to be the case in England, especially for older populations, and, therefore, the EQ-5D-5L has been recommended for use in general population surveys (despite the ceiling effects of dimensions ranging from 58–90%).105 Utility weights were assigned at each time point, based on an interim scoring ‘cross-walk’ function106 linked to the standard UK-based weights.107 EuroQol-5 Dimensions (EQ-5D) utilities were converted to QALYs over the trial period using the ‘area under the curve’ method.108

Methods of analysis

Missing data

Data were investigated for patterns of missingness.109 Mean imputation was used when the proportion of missing data was ≤ 5%.110,111 Missing EQ-5D-5L data were replaced using an index, rather than domain imputation, as the sample size was > 500.112 Multiple imputation by chained equations was fitted to replace item non-response. In line with Rubin’s rules113 and other recommendations,114,115 the point estimate for imputations was derived by averaging estimates of the imputed data based on results from five imputations. The point estimate for categorical data was rounded up to the nearest decimal point. The imputation model included variables used in the main model for the analysis, while also including the predictors of missingness. The dependent variable was included in the imputation model to ensure that the imputed values have the same relationship to the dependent variable as the observed values.116

Incremental cost-effectiveness analyses

Incremental cost-effectiveness analyses were based on multiple regression models to adjust for variations not accounted for by randomisation, and to provide more robust estimates. Generalised linear models (GLMs) were fitted separately for costs and QALYs,117 accounting for the cluster effect (identified as household identifier) via clustered standard errors.118 Models used for step count and MVPA have been described in Chapter 2. The cost models used the Poisson distribution and the QALY models used the binomial 1 family, equivalent to beta regression.119 Although the generalised linear models do not account for the correlation between costs and QALYs, the efficiency loss (i.e. higher standard errors) will be minimal, as the inclusion of the cluster effect provides robust standard errors and mitigates the effects of potential inaccuracies in the family distribution used. The choice of distributional family rested on the modified Park test120 and the comparison of observed and predicted values. Covariates included the baseline level (for QALY-based models), as recommended,116 practice and variables found to be correlates of PA-related outcomes – that is, demography (age, sex, ethnicity, marital status, education, employment, socioeconomic status, cohabitation), health (number of disease conditions) and other lifestyle behaviours (smoking and alcohol intake).121 Reduced models were generated using Wald tests to examine the joint significance of variables found to be insignificant in the base model. Significance levels were set at 5%.

To provide more precise estimates of uncertainty, the ‘margins method’ was used to generate sample means for trial arms and incremental point estimates for costs and QALYs.116,122 A different standard error and calculated CIs123 accounted for the cluster design.

Sensitivity analyses

To reflect the stochastic uncertainty surrounding the mean incremental cost-effectiveness, cost-effectiveness planes and cost-effectiveness acceptability curves (CEACs) were constructed using 2000 non-parametric samples from the base-case estimates. Bootstrapping used a new unique identifier for the clusters in addition to the original cluster identifier (household ID).116

Deterministic sensitivity analyses performed included:

• all randomised participants (not only those who provided accelerometry data)
• varying exclusion of costs of health service use beyond the immediate intervention
• a method of accounting for AEs
• only spontaneously self-reported SAEs
• only GP data relating to AEs; these were predefined by GP investigators and collaborators (TH, SI, SDW, JI) as possibly being related to increased walking, and included musculoskeletal events (falls requiring medical attention, fractures, sprains or injuries, pain in back or lower limb) and cardiovascular events [myocardial infarction, coronary artery bypass graft, angioplasty, stroke, transient ischaemic attack or new-onset ischaemic heart disease (angina)]
• a perspective of analysis (i.e. NHS with or without users)
• variation in the length of the life of a pedometer (between 1 and 4 years)
• scenario combinations (excluding all health service use costs, and including participant costs related to participation in PA and interventions, except the health service use cost borne by participants), to ensure that the long-run model could use evidence of ‘worst-case’ findings.

Results

Table 9 summarises the data on costs, EQ-5D-5L utility scores and quality of life (see Appendix 3, Tables 31–35). At 3 months, the average cost per participant was highest in the nurse group (£249), followed by the postal group (£122) and the control group (£107). The mean cost and distribution is affected considerably by the inclusion of health service use, which resulted in the control group costing £36 more per participant than the postal group and £12 more than the nurse-support group. QALYs varied marginally, with the gap being the greatest between the control and postal groups at +0.002. At 12 months, the average cost per participant was highest in the nurse group (£603), followed by the control group (£461) and the postal group (£375). The inclusion of health service use resulted in the control group costing £86 more per participant than the postal group, but £142 less than the nurse-support group. QALYs were marginally higher in the postal group (0.843 QALYs) than in the control group (0.837 QALYs) and the nurse-support group (0.836 QALYs) group.

The main results (Table 10), which are adjusted for baseline differences, show that, at 3 months, the cost of the nurse-support group was statistically significantly higher than that of the control group (£135, 95% CI £99 to £171), whereas this was not the case for the postal group (£15, 95% CI –£15 to £45). Table 10 also shows that there was a statistically significant increase in daily steps and minutes of MVPA in ≥ 10-minute bouts for the intervention groups compared with the control group. The ICER per additional minute of MVPA in ≥ 10-minute bouts was £0.35 for the postal group and £2.21 for the nurse-support group, compared with the control group. However, both intervention groups accrued slightly fewer QALYs than the control group (postal group: –0.0005 QALYs; nurse-support group –0.0004 QALYs), although this difference was not statistically significant. Nevertheless, it contributed to the dominance (lower costs and more QALYs) of the control group compared with both intervention groups at 3 months.

Comparing the two interventions at 3 months (see Table 10) shows that the nurse group achieved 481 more steps (95% CI 153 to 809 steps) and 18 more minutes of MVPA (95% CI 1 to 35 minutes) per person than the postal group, at a statistically significant additional cost of £120 (95% CI £95 to £146). The estimated cost per additional step and additional MVPA minute (in bouts of ≥ 10 minutes) was £0.25 and £6.67, respectively. The nurse-support group had slightly fewer QALYs, although this was not significantly different (0.0004 QALYs, 95% CI –0.0026 to 0.0018 QALYs); however, it contributed to the nurse-support group being dominated (higher costs and fewer QALYs) by the postal group.

The main results at 12 months (see Table 10) were somewhat different from those at 3 months. Although the mean costs were lower for the postal group (–£91, 95% CI –£213 to £33) and higher for the nurse-support group (£126, 95% CI –£37 to £290) than the control group, neither was statistically significantly different. However, the increase in the cost of moving from postal delivery to nurse-support delivery was statistically significantly higher (£217, 95% CI £81 to £354). Although both interventions were associated with a statistically significant increase in both step count and weekly minutes of MVPA in ≥ 10-minute MVPA bouts compared with the control group, the difference between intervention groups was not statistically different at 12 months. The postal group took more steps on average (+ 642 steps) and cost less on average (–£91) than the control group, and dominated the control group in terms of the PA outcomes. Compared with the control group, the nurse-support group cost an additional £0.19 per step and £3.61 per additional minute of MVPA in bouts of ≥ 10 minutes. None of the small decrements in QALYs at each incremental comparison (–0.0009 QALYs for the postal group vs. the control group, –0.0007 QALYs for the nurse-support group vs. the control group and –0.0004 QALYs for the nurse-support group vs. the postal group) was statistically significantly different. Compared with the control group, the postal group had fewer QALYs (although this was not statistically significantly different) and lower costs (also not statistically significantly different). However, the magnitude of the cost-savings is such that they outweigh the forgone QALYs at a threshold of £20,000 per QALY, and would be considered cost-effective. Using QALYs, the nurse-support group is dominated by both the control group and the postal group.

Comparing the two interventions at 12 months (see Table 10) shows that the estimated cost per additional step and additional MVPA minute was £6 and £109, respectively, and that, in terms of QALYs, the nurse-support group was still dominated by the postal delivery group.

The cost-effectiveness planes (Figures 5 and 6; see also Appendix 3, Figure 19) broadly confirm the findings that the postal delivery group had a strong likelihood of lower costs, but also fewer QALYs, and that the nurse-support group tended to have fewer QALYs and higher costs than the control group. However, different levels of uncertainty surround these mean estimates, as reflected in the CEACs (Figure 7; see also Appendix 3, Figure 20). At £20,000 per QALY, the postal delivery group has a 50% chance of being cost-effective compared with the control group, which falls to 42% at £30,000 per QALY. This is because, as the willingness-to-pay threshold increases (and, therefore, the higher the value that is placed on forgone QALYs), the value of QALYs lost begins to outweigh the cost-savings (see Figure 9). This is reflected in the CEAC, on which the probability moves towards zero. The nurse-support group had only a 5.5% chance of being cost-effective compared with the control group at a willingness-to-pay threshold to gain, or give up, a QALY of £20,000, and this fell to 4.9% when compared with the postal delivery group.

FIGURE 5 Cost-effectiveness plane for the postal delivery group vs. the control group at 12 months. (Figure not reproduced here; see the original article.)

FIGURE 6 Cost-effectiveness plane for the nurse-support group vs. the control group at 12 months. (Figure not reproduced here; see the original article.)

FIGURE 7 Cost-effectiveness acceptability curve for the postal delivery and nurse-support groups (vs. the control group) at different willingness-to-pay-per-QALY thresholds. (Figure not reproduced here; see the original article.)

The deterministic sensitivity analyses (see Appendix 3, Table 36) show that, with the exceptions of (1) using health service use including only self-reported serious adverse effects, (2) excluding all health service costs, (3) changing perspective (including all participant costs) and (4) the worst-case ‘combined scenario’, the sensitivity analyses produced results that were consistent with the base-case findings. For the exceptions, the postal delivery group was dominated by the usual-care control group at 12 months.

Beyond-trial modelling

Systematic reviews have indicated the positive influence that PA has on primary prevention for a range of conditions, including coronary heart disease (CHD), stroke, type 2 diabetes mellitus and cancers, 124–128 and, more recently, for improving cognition in older adults.129 Carlson et al.5 have separately shown that increasing amounts of leisure time PA is associated with decreasing health expenditure. Therefore, by reducing the risk of disease, increased PA can increase future QALYs, as well as lower future costs. The next section provides the methods used to estimate the long-term cost-effectiveness of the PACE-UP trial.

Methods

An existing Markov model,130 designed to assess the cost-effectiveness of PA interventions, was used. This model has evolved from assessing the cost-effectiveness analysis of exercise referral schemes131–133 to brief interventions129 and beyond.134 It has been used by NICE to update national guidance (PH44 on brief interventions, PH2 on exercise referral schemes guidance)36,135 and is the basis for the NICE Return on Investment Tool136 used by local authorities.

The model is driven by evidence of the impact that PA interventions have on the proportion of people meeting the recommended PA levels, short-term quality-of-life gains (associated with meeting the recommended PA levels) and the impact of PA on future rates of CHD, stroke and type 2 diabetes mellitus. Figure 8 shows the pathways within the model. In the original Markov model,129 a cohort of 100,000 33-year-old people were followed in annual cycles over their lifetime. At the end of the first year of the model, the cohort would be either ‘active’ (doing 150 minutes of MVPA per week) or ‘inactive’, and could have had one of three events (non-fatal CHD, non-fatal stroke, type 2 diabetes mellitus), remain event free (i.e. without CHD, stroke or diabetes mellitus) or die, either from cardiovascular disease (CVD) or from non-CVD-related causes. Active individuals have a better life expectancy and quality of life, attributable to lower risks of developing CHD, stroke and type 2 diabetes mellitus. People who become active in the first year (irrespective of the trial arm) accrue a one-off utility gain associated with achieving the recommended level of PA. QALYs reflect the health outcomes from a reduced risk of disease. A discount rate of 3.5% per annum is used for costs and QALYs, as recommended by NICE, and the analysis is conducted from a NHS perspective. Full details of the model are provided in Appendix 3 and elsewhere.130

FIGURE 8 Illustration of pathways within the long-term cost-effectiveness model.130 (Figure not reproduced here; see the original article.)

The model by Anokye et al.130 was adapted for use in five ways:

1. The population begins with the mean age of the trial population (i.e. 59 years rather than 33 years) and was followed to 88 years (average life expectancy at 59 years in the UK).137 This change was also reflected in the age-specific estimates used.
2. The intervention was either pedometer plus nurse support or a posted pedometer.
3. Within-trial costs were used, with a second year of annuitised values added for the intervention arms (£5.03 per person for the postal delivery group and £4.14 per person for the nurse-support group – the number of pedometers to the postal delivery group was relatively higher as a result of more replacements), as the trial analysis had assumed pedometers lasted 2 years.
4. Effectiveness estimates from the PACE-UP trial were used as follows: the probability of moving from an inactive to an active state was based on estimated relative risks (RRs) for achieving 150 minutes of MVPA in ≥ 10-minute bouts over 7 days at 12 months. RRs of achieving ≥ 150 minutes of MVPA in ≥ 10-minute bouts at 12 months were estimated from ORs using the formula OR/[(1 – P_ref) + (P_ref × OR)], in which P_ref is the proportion of all subjects achieving 150 minutes of MVPA in ≥ 10-minute bouts at baseline (i.e. 218/1023 = 0.21). The OR was derived from a logistic regression model in which achieving 150 minutes of MVPA in bouts of ≥ 10 minutes at 12 months was regressed on baseline minutes of MVPA in bouts of ≥ 10 minutes, month of baseline accelerometry, age, sex, general practice and treatment group, with household as a cluster.
5. The short-term psychological benefits associated with achieving 150 minutes of MVPA per week used trial data; incremental EQ-5D-5L scores (at 12 months) for active people were regressed, via beta regression, on EQ-5D-5L scores at baseline, ethnicity, education, employment status, intervention group, practice and disability.

Other parameters were informed by estimates from the original model. Cost and utility estimates for disease conditions were originally sourced from literature reviews of economic evaluations conducted for NICE on CVD and diabetes mellitus.5 Estimates for the health impacts of PA were taken from national/international evidence-based guidance on PA and health, such as the US Physical Activity Guidelines Advisory Committee Report, 2008.138 Appendix 3, Table 37 details all parameter values.

The probabilistic sensitivity analysis (PSA) addressed the uncertainties around all parameters in the model (except for baseline mortality, as the mortality census data have little uncertainty). A total of 10,000 Monte Carlo simulations were used to generate stable estimates. Given that the within-trial sensitivity analysis showed the decisional influence of some assumptions, a deterministic sensitivity analysis using the lifetime model explored two alternative, conservative scenarios:

• Scenario 1 – combined exclusion of all health service use costs during the trial period (year 1 of the model), with no short-term QALY gain associated with achieving the recommended level of PA. This was considered as a result of the uncertainty around short-term changes to health service use, and because previous studies found the exclusion of short-term QALY gain associated with being physically active to the recommended level to be decisional.5,129
• Scenario 2 – scenario 1 plus user costs related to participation in PA and the interventions. This combination represented a ‘worst-case’ scenario in the trial. Although this perspective is not one adopted by NICE, it represents the most conservative scenario based on this evidence.

Results

Table 11 shows that the postal delivery group dominated both usual practice and the nurse-support group, as the lifetime costs were lower and the number of QALYs was greater. The stochastic uncertainty associated with the mean ICER (Appendix 3, Figure 21) indicates that these findings are robust, as there is a 100% likelihood, at a willingness-to-pay threshold of £20,000 per QALY, that the postal delivery group is cost-effective compared with the control group and with the nurse-support group (Figure 9). This is consistent with the net monetary benefit estimates, which show that, although we can be 95% confident that the postal delivery group is better than the control group and the nurse-support group, we cannot be 95% confident that the nurse-support group is better than the control group.

FIGURE 9 Cost-effectiveness acceptability curve showing the probability of lifetime cost-effectiveness for the postal delivery group and the nurse-support group (vs. the control group) at different willingness-to-pay threshold levels. (Figure not reproduced here; see the original article.)

The results for scenario 1 of the sensitivity analyses for the 100,000 cohort were as follows:

• postal delivery group versus control group – postal delivery moved from a dominant position to being a more expensive option (+£4M) with greater QALY gains (+609 QALYs) and an ICER of £6100
• nurse-support group versus control group – the ICER increased from £16,000 to £26,000 (+£14M, +538 QALYs)
• nurse-support versus postal delivery group – the nurse-support group remained dominated by the postal delivery group (+£10M, –87 QALYs).

For scenario 2, the sensitivity analyses for the 100,000 cohort showed the following:

• postal delivery group versus control group – the postal delivery group moved from a dominant position to being more expensive (+£16M) with greater QALY gains (+609 QALYs) and an ICER of £26,600
• nurse-support group versus control group – the ICER increased from £16,000 to £25,400 (+£13.7M; +538 QALYs)
• nurse-support group versus postal delivery group – the nurse-support group moved from a dominated position to a cost-effective position (–£2M, –87 QALYs).

Discussion

The within-trial analysis shows that, at 3 months, compared with the control group, both interventions cost something to achieve increases in PA. Compared with the control group, the postal delivery group cost an additional £0.02 per additional step gained to an average of + 692 steps per day, which was a cheaper buy than the nurse-support group (at £0.12 per additional step). However, the nurse-support group achieved more steps on average, and the additional 448 steps were achieved at an incremental cost of £0.25 per step. Although this pattern of results was replicated for additional minutes of MVPA (in bouts of 10 minutes), the results for QALYs were very different as both intervention groups were dominated by the control group (i.e. the control group cost less and had more QALYs).

The main results at 12 months were different, leaning more favourably towards the postal intervention. Compared with the control group, the postal delivery group achieved statistically significantly better PA outcomes, and did so at a lower cost. This was much better than the results for the nurse-support group, and the insignificant difference in PA outcomes between the nurse-support and postal delivery groups at 12 months implied very high costs per additional step (£6.00). The analysis of the costs per QALY confirmed that the nurse-support group was not a cost-effective alternative. It also showed that, although the postal delivery group has fewer QALYs (although not statistically different) and lower costs (with costs saved higher than £20,000 per QALY), the postal delivery group could be considered to be cost-effective. Assuming a value of £20,000 per QALY, there was a 50% probability that the postal delivery group was cost-effective compared with the control group, and a 5% probability that the nurse-support group was cost-effective compared with the control group or the postal delivery group at 1 year. The sensitivity analyses did not change the conclusions, except in three cases (using self-reported SAEs, excluding health service use, but including all participant costs), when the postal delivery group was dominated by the usual-care control group.

The lack of evidence on effectiveness in terms of quality-of-life outcomes is not necessarily evidence of no effect, as the trial was not powered to detect a change in quality of life. The results indicate a lot of variation around the change in QALYs (95% CI –0.017 to 0.009 QALYs) and we are aware of some ceiling effects at baseline (98% self-care, 83% usual activities, 79% mobility, 73% anxiety, 43% pain). Although this might contribute to raising questions about the relevance of general quality-of-life measures for capturing the quality-of-life impacts of public health interventions within the first year, it also serves to highlight the importance of capturing the QALY impact of public health interventions on disease avoidance in longer-term economic models. Cost-per-QALY results from short-term public health trials have the potential to mislead decision-makers on the efficiency of investments in the context of changes that lead to longer-term reductions in the risk of disease.

A lifetime cost-effectiveness model characterised the long-term impact of PA interventions on CHD, stroke and type 2 diabetes mellitus.129 This showed that the postal delivery group would dominate both the control group and the nurse-support groups, as quality-of-life gains (759 QALYs, 95%CI 400 to 1247 QALYs) add to increased cost-savings (–£11M, 95% CI –£12M to –£10M), resulting in an incremental net monetary benefit of £26M (95% CI £18M to £36M) for a 100,000 cohort. There was a 100% likelihood that the postal intervention was cost-effective compared with the control group and the nurse-support group. The conservatively framed deterministic scenario analyses showed that excluding both the short-term reduction in health service use and the utility gain seen in the trial would not alter the main conclusion that the postal delivery group would be an extremely cost-effective intervention (ICER of £6100 per QALY). Even taking the unusual step of including participant costs did not raise the ICER beyond a threshold value of £30,000 per QALY.

Strengths and weaknesses

To our knowledge, this is the first published study of short- and long-term costs and effects of a pedometer intervention. Its strengths are the use of detailed individual-level cost and effectiveness data from a well-designed population-based RCT, which had nearly complete (93.4%) follow-up data to 1 year, to estimate the cost per QALY at 1 year, and also to input into a lifetime model of cost-effectiveness. It also included both provider and user perspectives in costing service provision and participation, which allowed the impact of perspective to be investigated, and which provides a basis for further investigation of the role of cost in its association with participation in PA. The estimates of uncertainty extended commonly used techniques to account for increased precision in the context of clustered data. The sensitivity analysis pushed both the short- and long-term analyses to very conservative outcomes given the trial data and, therefore, provides a good indication of the robustness of findings based on the current evidence.

The weakness of the within-trial cost-effectiveness study connects to the cost of health service use over the period of the trial; no information was available on the severity or procedures used for hospital admissions, or cause of admission to the A&E department. We relied on the principal investigator’s ‘best guess’ or ‘nearest appropriate code’ (while blind to the treatment group) and on averaging across elective/non-elective admissions and we therefore explored alternative assumptions in our sensitivity analyses. There was considerable variation in costs in each trial arm, and the trial was not powered to detect a difference. Data were also not collected on costs to participants for months 4–9, and the last 3 months were multiplied to represent the missing data. These may have over- or underestimated participant costs and, if significantly underestimated, could be decisionally important. With respect to the long-term modelling, the model assumes that people would revert to PA patterns observed in long-term cohort studies. This estimate could be improved with longer-term trial data. All the challenges set out in previous work130 are also relevant here (e.g. some diseases, such as cancer and AEs, are not accounted for, which could lead to either over- or under-estimation of cost-effectiveness).

The study feeds into an area sparse of primary data139,140 populated only by small studies.95,96 Leung et al.97 showed a 95% likelihood that pedometers would be a cost-effective addition to green prescriptions (in New Zealand) at 12 months, which is much higher than the 50% likelihood that we found. Our study also provides long-term estimates based on the population-level primary data for comparison with the larger body of cost-effectiveness estimates134 from decision models.97,98,141,142 Some have identified cost-savings and an improved quality of life at a population level from pedometers in the long term.98,137 Others97,138 have indicated high probabilities that pedometers will be cost-effective in the long term, with Brennan et al.142 indicating that, even with long-term support at £25 per year (for monitoring and support), ICERs fell well below £10,000 per QALY gained. Our study provides further support to indicate that pedometer-based programmes are a cost-effective method of improving health.

Conclusion

A range of sensitivity analyses of both short- and long-term cost-effectiveness confirmed the view that postal delivery of pedometer interventions to people aged 45–75 years through primary care has a high chance of being cost-effective in the long term and has a 50% likelihood of being cost-effective, through resource-savings from changes in health service use, within 1 year. Further research is needed to ascertain the level of maintenance of PA beyond 1 year and the impact of PA on quality of life and general health service use in both the short term and long term.