Methods for the Estimation of the NICE Cost Effectiveness Threshold

The threshold as a measure of opportunity cost

The National Institute for Health and Care Excellence uses cost-effectiveness analysis (CEA) to inform the decisions underlying most types of guidance that it publishes. The use of CEA is most prominent in appraisals relating to new medicines,3 but is also a key input into diagnostics appraisals as well as clinical guidelines and public health guidance. 3,5 For those interventions and programmes which impose additional costs on the NHS budget, their ICERs indicate the incremental cost per additional QALY achieved relative to appropriate comparators. Although the ICER is one of a number of evidential inputs into NICE committees’ decisions, is has been shown to be the most important, at least for technology appraisals. 6

Interpreting whether or not a given ICER is acceptable requires the use of a cost-effectiveness threshold. Given that NICE has no influence on the level of the NHS budget, its decisions need to consider that budget as a fixed constraint. 1 Therefore, the threshold should reflect the opportunity costs, in terms of health forgone, resulting from the imposition of additional costs on the NHS. When NICE issues positive guidance for a new intervention which imposes additional costs on the system, the resources required to deliver it must be found by disinvesting from other interventions and services elsewhere. 7 This displacement of existing services will result in health decrements for other types of individual. 8 Thus, the threshold represents the additional cost that has to be imposed on the system to forgo 1 QALY’s worth of health through displacement.

Resource allocation decisions based on comparing an ICER with a cost-effectiveness threshold uses some simplifying assumptions, including those of constant returns to scale and perfect divisibility of programmes. 9 Some have suggested that this makes these methods unreliable,10 although it has also been argued that they provide useful approximations to guide decisions. 11 This report takes NICE’s use of these methods as a starting point, and does not review the literature relating to this debate in any depth.

As Figure 1 illustrates, CEA effectively becomes an analysis of net health benefits (NHBs): does the health gain from the new intervention outweigh the health decrements associated with the displacement of existing services necessary to fund it? Figure 1 shows the incremental costs and QALYs associated with a new intervention relative to a comparator (the latter being shown at the origin). The new intervention generates 2 additional QALYs per patient and, at price P1, imposes an additional £20,000 per patient; the ICER is, therefore, £10,000 per QALY gained. At a threshold of £20,000 per QALY, the additional cost of £20,000 per patient translates into a decrement of 1 QALY (the distance between the y-axis and the threshold). This is because the threshold indicates the additional cost that needs to be imposed on the NHS budget in order to displace services that result in 1 QALY being forgone. Therefore, at that price, there is a net health gain of 1 QALY per patient (2 gained from the new intervention and 1 forgone through displacement). At a price of P2, the additional cost per patient of the new intervention is £40,000 and the net health gain is 0: the 2 additional QALYs from the new intervention are the same as the QALYs forgone through displacement. At the highest price of P3, the adoption of the new intervention would actually result in a net health decrement of 1 QALY as it generates fewer QALYs (2) than are forgone (3).

FIGURE 1.

Graph showing illustration of the NICE threshold as a basis for assessing NHB. Reproduced from Value based pricing for NHS drugs: an opportunity not to be missed?, Claxton K, et al. , vol. 336, pp. 251–4, 2008 with permission from BMJ Publishing Group Ltd.

The use of the threshold to facilitate this NHB analysis can be expressed as in Equation 1:

where Δh is the change in health generated by the new intervention, ΔC_h is the additional health-care cost imposed on the NHS, and k is the cost-effectiveness threshold. The net health gain from adopting the new intervention is therefore the health gained, Δh, minus the health forgone, ΔChk.

Understanding the NICE cost-effectiveness threshold as representing opportunity costs in terms of health is explicit in NICE documentation (e.g. the Guide to the Methods for Technology Appraisal3). It is also clear in reports published by the Department of Health, such as the consultation report on value-based pricing. 4,12,13 This conceptualisation of the principles of the NICE threshold is also described in the broader literature. 1,2 Formally, the threshold can be seen as the shadow price of the budget constraint. 1,2,8,14–16 Although this project focused on the use and estimation of a cost-effectiveness threshold for NICE decisions, the methods and estimates relate to any resource allocation decision within the NHS where the opportunity cost could fall anywhere in the system. Hence it could apply, for example, to Department of Health targets or to NHS England directives, as well as NICE guidance.

The threshold as the consumption value of health

Another view of what the threshold used in CEA should represent exists in the literature; however, in general terms this is based on the rate at which individuals are willing to forgo other forms of consumption to achieve health improvement [sometimes referred to as ‘willingness to pay’ (WTP)]. 17–36 Although this consumption value of health can provide information on the value of health improvement and may guide decisions such as the level of the overall NHS budget, it does not inform decisions regarding how to allocate a fixed budget within the health-care system.

The reason for this is that the consumption value of health applies equally to health gained as well as to health forgone. This is shown in Equation 2 where the consumption value of health, v, is added to the definition of NHB in Equation 1. This simply involves valuing both health gained and health forgone by the same consumption value of a unit of health, v. Therefore, the use of the consumption value is irrelevant: a treatment considered cost-effective in Equation 1 (i.e. to have a positive NHB) will inevitably be considered cost-effective in Equation 2, and an intervention with negative NHB (i.e. not cost-effective) will remain as such in Equation 2. a Therefore, the magnitude of the threshold, k, is not a value judgment but an empirical question which can, in principle, be estimated.

The National Institute for Health and Care Excellence’s threshold range

The National Institute for Health and Care Excellence has been reluctant to specify a single cost-effectiveness threshold used in its decision-making. 6 It has also consistently emphasised that factors other than CEA are taken into consideration by the various advisory committees. 3,5,6,38–40 Therefore, it has preferred to indicate the range within which its threshold value lies (i.e. £20,000–30,000 per QALY gained). 3,5 Alongside this, it has provided an indication of the role other factors play in determining which point of threshold range is relevant. The latest guide3 suggests that an ICER < £20,000 is likely to lead to recommendation unless the evidence is considered highly uncertain; an ICER between £20,000 and £30,000 will lead to recommendation if the committee is also happy with the levels of uncertainty in the evidence and/or the QALY does not capture all aspects of benefit; and an ICER > £30,000 would be recommended only if issues related to levels of evidential uncertainty and a failure to capture all benefits in the QALY are particularly compelling.

In 2009, NICE issued further supplementary guidance relating to the appraisal of interventions for patients with short life expectancy (LE), although this can be considered to relate more to the measure of benefit than factors to be considered outside of cost-effectiveness. 41 In 2012 NICE issued a draft update of its methods guide which added that, if a new technology has an ICER > £20,000 per QALY, the committee’s deliberations would also consider ‘aspects that relate to non-health objectives of the NHS’ (e.g. wider social considerations and/or costs that fall outside the NHS budget). 42

Although NICE has carefully argued the case for why its decisions are not driven entirely by a comparison of the ICER with its threshold range, it has not provided any empirical evidence for why the threshold range takes the value it does. Indeed it has been widely argued than an empirical basis for these values should be generated. 4,43–47 For example, the House of Commons Health Select Committee in 2008 argued:

The affordability of NICE guidance and the threshold it uses to decide whether a treatment is cost-effective is of serious concern. The threshold is not based on empirical research and is not directly related to the budget, it seems to be higher than the threshold used by [primary care trusts] PCTs for treatments not assessed by NICE. Some witnesses, including patient organisations and pharmaceutical companies, thought that NICE should be more generous in the cost per QALY threshold it uses, and should approve more products. On the other hand, some PCTs struggle to implement NICE guidance at the current threshold and other witnesses argued that a lower threshold should be used. We recommend that the threshold used by NICE in its full assessments be reviewed; further research comparing thresholds used by PCTs and those used by NICE should be undertaken . . .

p. 64

The basis for empirical work

Although there is acceptance of the need for empirical work on the NICE cost-effectiveness threshold, a set of issues exists regarding the starting point for such analysis. One aspect of this is the view that the nature of the services that are displaced in response to additional costs being imposed by NICE guidance, and hence the magnitude of the health forgone for other patients, will depend on the productivity of the NHS and its overall (inflation-adjusted) budget, both of which have increased since NICE initially defined its threshold range. 48,49 In principle an increase in the (real) NHS budget would allow it to introduce interventions which were previously not cost-effective which might be expected to increase the threshold if these interventions were the marginal ones displaced in response to the budget impacts of NICE recommendations. However, any increase in the NHS budget may be allocated to non-discretionary expenditure. This would include, for example, expenditure relating to national initiatives such as new contracts for consultants and activities to meet waiting list targets as well as, of course, the implementation of NICE guidance. The non-discretionary nature of such expenditure means that these types of activities cannot easily be disinvested from given a need to release resources to fund NICE guidance. Therefore, if an increase in the NHS budget is largely devoted to these types of non-discretionary expenditure, there will be a limited impact on the threshold.

Gains in productivity may come through doing worthwhile activities more cost-effectively, including for those marginal interventions displaced by NICE recommendations, suggesting a reduction in the threshold. Alternatively, productivity gains might come through discontinuing activities which are not worth doing (i.e. that produce no health improvement), freeing resources for additional cost-effective interventions which may be the marginal services displaced by NICE guidance – this can have the result of increasing the threshold.

The net effect of these changes on the threshold could not be determined a priori and would depend on how any additional (real) budgets were allocated and how the gains in productivity where achieved. This does emphasise the fact that the threshold may change over time in response to these and other broader developments, and this would have to be considered as part of any regular updating of the empirical analysis of the threshold.

A second issue to be considered relates to how decisions are taken locally about any displacement following NICE guidance. The principles of CEA suggest that such displacement should relate to interventions which are the least cost-effective of those currently covered by the budget. 14 The basis for how local commissioners and providers make their disinvestment decisions is not clear, however, and there have been calls for greater transparency and guidance in this area. 48 It would be entirely unrealistic to assume that displacement only takes place in those existing services which are the least cost-effective. The reality is that numerous criteria are likely to be used by commissioners in implementing disinvestment, and that significant variation will exist between local decision-makers. 8 Such criteria might include, for example, equity concerns about a particular disadvantaged group locally or capacity constraints regarding particular services. Therefore NICE needs to know what is likely to happen on average across the NHS given the reality of local decisions. If local decision-making changes over time – for example, if local commissioners become more focussed on displacing services which are the least cost-effective, in terms of population health – this may affect the estimate of the threshold.

Studying displacement locally

A reasonable conclusion from a consideration of these issues is, therefore, that local decisions about disinvestment are likely to be an important determinant of the NICE threshold. 50–55 Appleby et al. 56 sought to assess whether or not it was possible to study local decisions about service investment and disinvestment to infer the cost-effectiveness thresholds being used (implicitly) locally and to draw conclusions about the appropriate level of the NICE threshold. They identified six PCTs and undertook structured interviews with each of the directors of public health. They also administrated questionnaires to an opportunistic sample of finance directors from NHS trusts. On this basis they developed a list of new services as well as those that had been deferred or discontinued. An attempt was made to estimate the implicit local ICER relating to these decisions by using any cost-effectiveness evidence used to inform the decisions together with relevant evidence on cost-effectiveness from the published literature.

The study found it quite straightforward to identify specific services that had been introduced, discontinued or deferred, but concluded that these decisions were typically based on clinical and other non-economic factors. A number of ‘decisions at the margin’ were identified but none of these were based on CEA. Instead, the basis for changes in services was a ‘business case’, or overall cost impact. It was possible to impute cost-effectiveness for most of the services affected, but the study concluded that, even with a larger sample of commissioners and providers, it would be very difficult to estimate an implied cost-effectiveness threshold locally. This would be because, first, most PCT decisions were service reconfigurations including demand management and waiting list initiatives. By their nature, teasing out the incremental cost and health effects, potentially across numerous types of patients, would be an enormous challenge. Second, there would be difficulty in identifying all local decisions as many options for investment, deferment or discontinuation are rejected before they are made more explicit in documentation. A third problem would be the finding that a range of criteria is used to make local decisions, with relatively little concern for cost-effectiveness, making a local threshold estimated in this way hard to interpret. A final challenge would be that it would be very difficult to establish a causal link between a change in local NHS budget and specific local investment and disinvestment decisions. The Appleby et al. 56 study highlights the problems that exist in deriving a cost-effectiveness threshold from a bespoke study of specific local resource allocation decisions.

What evidence is needed?

Given the challenges of studying local decisions as a means of establishing the NICE threshold, and keeping in mind NICE’s remit, it is possible to suggest a series of important characteristics that estimation methods should have from the perspective of principle and practice:

• They should reflect the effect of NICE guidance on the average of the displacement decisions taken across the NHS, with less consideration on which types of patients and interventions are affected and why the decisions are taken. NICE cannot be expected to reflect what is likely to be marked variation between local commissioners and providers in how they react to an effective reduction in their budget as a result of positive guidance. Given NICE’s remit, it is the expected health effects (in terms of length and QoL) of the average displacement within the current NHS (given existing budgets, productivity and the quality of local decisions) that is relevant to the estimate of the threshold.

• The methods used should not be a ‘once and for all’ effort but should facilitate regular updates to reflect changes in the broader NHS context such as changes in the overall real budget and productivity. This requires the use of data sources that are currently routinely available, are expected to become so in the future or could be made available at reasonable cost. It may be possible to glean some idea of how the threshold may change in the future by studying how it has changed in the past, which would require routine data sources to extend back over a period of time. Periodic updating using explicit scientific methods would encourage accountability through scrutiny of estimates by relevant stakeholders. It would also provide more predictability in likely changes to the threshold for the investment decisions of technology manufacturers.

• The nature of the displacement of existing services (and hence the magnitude of the health forgone) will depend on the scale of the budget impact coming through NICE guidance. Therefore, the methods used to estimate the threshold should ideally be able to reflect this budget impact.

• The methods should recognise the inevitable uncertainty relating to the evidence currently available for threshold estimation and translate this into an expression of the uncertainty in the estimate of the threshold. As well as providing information with which NICE can determine the appropriate implications for its choice of a threshold value, this consideration of uncertainty can help to prioritise further research or the collection of routine data.

Past work

The study was able to build on some key existing research relating to the relationship between NHS expenditure and mortality. 57–59 Since 2003 data on expenditure on health care across 23 programme budget categories (PBCs) of care have been available for each PCT in the NHS in England. These PB data seek to allocate, to broad areas of illness according to the primary diagnosis [using International Classification of Diseases, Tenth Edition (ICD-10) codes], all items of NHS expenditure, including expenditure on inpatient care, outpatient care, community care, primary care, and pharmaceuticals and devices.

For the purposes of this study, the merit of these data is that they open up the possibility of examining the relationship between differences in local spending and associated disease-specific mortality outcomes routinely available from the National Centre for Health Outcomes Development. In each programme, the elasticity of outcome with respect to changes in expenditure was estimated controlling for differences between PCTs in need. Changes in mortality were then transformed into YLG using assumptions regarding LE without the change in expenditure. This provides estimates of the marginal cost per YLG on average across the NHS by PBC.

This work focused largely on spending and outcomes in two of the largest programmes: circulatory disease and cancer,60 but has also informed the link across other programme categories. 58,61 Estimates of the cost per YLG for 2006/7 were £15,387 for cancer, £9974 for circulation problems, £5425 for respiratory problems, £21,538 for gastrointestinal problems and £26,428 for diabetes. These estimates were based on a straightforward, though carefully constructed, theoretical model of health production which informs the specification and estimation of a set of equations. These dealt with the challenge of there being alternative plausible directions of causation (e.g. between expenditure and health outcomes within a programme). This problem of endogeneity was addressed by identifying and testing suitable instrumental variables (IVs). In doing so, they accounted for variation in the clinical needs of the local population relevant to each programme together with broader local environmental factors relevant to the costs of care and outcomes.

This earlier work provides a strong foundation for the current study through its consideration of the average marginal elasticity of outcome with respect to programme expenditure. However, to estimate the threshold suitable for NICE decision-making, a number of further elements of research are necessary, and these are described below.

Further econometric analysis

This further econometric research is covered in Chapter 3, with full details in Appendix 2. The earlier work estimated the cost per YLG for the major programme areas. The NICE threshold needs to relate to the whole NHS and will, therefore, depend on all the programmes of care where disinvestment takes place. Given that each programme of care has been estimated separately, it is not clear how expenditure on particular programmes changes with the overall budget. For example, does disinvestment tend to fall on respiratory care or diabetes following a budget impact from NICE guidance? Therefore, the current study has further developed the econometric analysis to reflect the need for PCTs to operate within a fixed overall budget. This provides an estimate of the ‘budget elasticity of expenditure’ in each PBC, and facilities estimates of the impact of marginal increases (or decreases) in overall PCT budgets on spending in each PBC.

As well as indicating budgetary influences on programme spending, these elasticities have then been linked to changes in mortality outcomes by programme. These changes are used to estimate years of life lost (YLL) taking account of the fact that some of the observed deaths would have occurred anyway (had the same population not been at risk in the particular PBC); that is, taking account of unobserved counterfactual deaths. This takes into account how such budgetary changes (such as those imposed by NICE guidance) translate through local decisions into changes in expenditure on programmes of care and then to health outcomes.

Changes in budgets are in practice incremental rather than marginal, and it may be the case that the outcome elasticities of programme expenditure in times of budgetary increase (when new initiatives are introduced) are not the same as in times of budgetary decrease (when the focus is on disinvestment). The possible effects of non-marginal changes have therefore been explored. The project has also sought to explore how both expenditure and outcome elasticities, and hence the threshold, vary over time, and this has been assessed by generating relevant estimates for three sets of data.

A development from earlier work has been to relate expenditure in period t to mortality in periods t, t + 1 and t + 2. Although the data used are largely cross-sectional, mortality data are linked so as to follow expenditures. Given the inevitable uncertainty relating to assumptions in the analysis, extensive sensitivity analysis is undertaken to consider the implications for the estimates.

Moving from life-years to quality-adjusted life-years gained

A key element of the research has been to take the results of the econometric work linking NHS spending and mortality, and to translate this into effects on life-years and QALYs. The methods planned for the study included a consideration of local data, collected routinely by PCTs, on the types of interventions in which local decision-makers were investing and disinvesting. The aim was to inform the link between the effects of expenditure changes on mortality and impacts on broader health in terms of QALYs. These data may have indicated the types of interventions and services, within a given PBC, on which investment and disinvestment were taking place. Using targeted literature reviews, estimates of QoL for those activities may have been identified. However, it was established that there were limited data available at a local level to facilitate this type of analysis, so other data sources were used for this purpose (see Addendum 2: the role of data on local NHS decisions in Appendix 3).

It has therefore been necessary to consider alternative data and approaches. This is tackled using three sequential steps:

1. translate the estimated effects on mortality from the econometrics work into life-years by exploring the limitations of the mortality data available at PCT level and the published YLL figures used in the econometric analysis, and by considering how to improve the estimates using additional data and analysis

2. consider how estimates of life-year effects can be adjusted for the QoL in which they are lived, taking account of the gender and the age at which life-years are gained or lost as well as the QoL implications of particular diseases

3. explore ways to take account of those effects on health not directly associated with mortality and life-year affects (i.e. the ‘pure’ QoL effects) to estimate an overall cost per QALY threshold.

This aspect of the analysis is described in Chapter 4 with further details provided in Appendix 3.

The central or ‘best’ estimate is based on two assumptions relating to the health effects associated with expenditure, one conservative and the other more optimistic. The first assumption is that the health effects of changes in 1 year of expenditure are restricted to 1 year. This is implicit in the estimates of outcome elasticities estimated in the econometric analysis. This is likely to underestimate effects on mortality as expenditure that reduces mortality risk for an individual in 1 year may well also reduce their risk over subsequent years, and expenditure may also prevent disease in future patient populations. Therefore, total health effects will be underestimated and the cost per life-year or QALY threshold will be overestimated. Although undoubtedly conservative, it may be offset to some extent by the more optimistic assumption. It is assumed that any death averted by expenditure in 1 year will return the individual to the mortality risk of the general population, that is the years of life gained (YLG) associated with each death averted are based on what would have been their LE taking account of their of age and gender (using life tables for the general population).

The extreme upper and lower bounds for cost per life-year and cost per QALY thresholds are based on making both of these assumptions either optimistic (providing the lower bound for the threshold) or conservative (an upper bound for the threshold). The lower bound is based on assuming that health effects are not restricted to 1 year but apply to the remaining disease duration for the population at risk during the expenditure year. The upper bound is based on the combination of assuming that health effects are restricted to 1 year and that any death averted is only averted for the minimum duration consistent with the mortality data used to estimate the outcome. It is very important to note that the lower and upper bounds are very much extreme values with limited plausibility.

Programme budgeting in England

Prior to October 2006, there were 303 PCTs in England with an average population of about 160,000 people. In October 2006 the 303 PCTs became 152 PCTs. Some PCT boundaries remained unchanged while other PCTs were merged with one or more neighbours to form a new, larger, PCT. In a few cases the geographic area covered by an existing PCT was split between two or more new PCTs. These 152 PCTs have an average population of about 330,000 people. PCTs are allocated fixed annual budgets within which they are expected to meet expenditure on most aspects of health care, including inpatient, outpatient and community care, primary care and pharmaceutical prescriptions.

Programme budgeting data collection was initiated by the Department of Health in April 2003 when each PCT was required to prepare expenditure data disaggregated according to 23 programmes of health care. These programmes are defined by reference to ICD-10 codes at the four digit level, and most PBCs reflect ICD-10 chapter headings (e.g. cancer and tumours, circulation problems, renal problems, neonates, problems associated with the skin, problems associated with vision, problems associated with hearing, etc.). In some cases the 23 categories are broken down into further subareas to achieve a closer match with the various National Service Frameworks (NSFs); for example, the large mental health category is broken down into ‘substance abuse’, ‘dementia’, and ‘other’.

Programme budgeting seeks to allocate all types of PCT expenditure to the various PBCs, including secondary care, community care and prescribing. However, the system acknowledges that a medical model of care may not always be appropriate, and two specific non-clinical groups – ‘healthy individuals’ and ‘social care needs’ – have been created. These are intended to capture the costs of disease-prevention programmes and the costs of services that support individuals with social rather than health-care needs. In addition, in some cases it is not possible to assign activity by medical condition, preventative activity, or social care need and, in these cases, expenditure is assigned to a residual category (PBC 23) entitled ‘other’. The most important element of this residual programme is expenditure on general practitioner (GP) services (PBC 23a). In principle it should be possible to allocate each GP consultation to a particular care programme. However, at the moment the available data information systems do not permit such an allocation and so all primary care expenditure is allocated to this residual programme. The use of this residual category ensures that all expenditure is assigned to a programme of care. 73

The aim of the programme budget classifications is to identify the entire volume of health-care resources assigned to broad areas of illness according to the primary diagnosis associated with an intervention. It serves a number of purposes, most notably to assist in the local planning of health care. However, for this study its crucial merit is that it opens up the possibility of examining the statistical relationship between local programme spending and the associated disease-specific outcome. Various forms of data collection and analysis are required to map PCT expenditure on acute, community and other services to the 23 PBCs. From the PCT perspective, however, the construction of each PCT’s return largely involves collating information provided by other bodies and drawing on other information already in the PCT’s own annual accounts. Details of how expenditure is assigned to programmes of care can be found in Appendix 2, The collection of programme budgeting data.

Table 1 shows the expenditure per head and the growth in this expenditure for each PBC for 2003/4–2008/9. d Year-on-year comparisons of expenditure in each group are complicated by the fact that the algorithms used to allocate activity to PBCs are regularly revised. e However, by 2008/9 total PCT expenditure per person had increased to £1531 (up 28% from 2004/5). The residual ‘other’ category (PBC 23) still accounted for the largest share of expenditure (14.9%) with per capita expenditure of almost £228, of which £145 was accounted for by primary care expenditure. Mental health (PBC 5) accounted for just over 12% of expenditure, but the expenditure share recorded by circulation problems (PBC 10) had fallen from 10.2% to 8.5%. Other categories recording a fall in budget share of more than a half of 1 percentage point included: the gastrointestinal system (down from 6.1% to 5.1%), the musculoskeletal system (down from 6.0% to 5.2%), trauma and injuries (down from 6.0% to 4.2%), and maternity (down from 4.6% to 3.9%). Categories recording an increase in budget share of more than a half of 1 percentage point included neurological problems (up from 2.9% to 4.4%) and dental problems (up from 1.1% to 4.1%).

Some of these changes will partly reflect revisions to the algorithms used to allocate expenditure to particular PBCs. For example in 2006/7 expenditure per person on musculoskeletal problems fell by 11% and expenditure on trauma and injuries fell by 25%. In the same year, expenditure on neurological problems increased by 35%. This suggests that some types of activity, which were previously allocated to musculoskeletal problems and/or trauma and injuries, were reallocated to neurological problems.

Similarly, up to and including 2006/7 expenditure that was not directly attributable to a particular programme category was apportioned using admitted patient care percentages. f In other words, if x% of total admitted patient care expenditure was allocated to PBC 1, then x% of all expenditure that was not directly attributable to a particular programme category was also allocated to PBC 1. With effect from 2007/8, however, NHS organisations were asked to select an appropriate basis for the apportionment of this non-programme-specific expenditure and that, where no reasonable basis existed, such expenditure was to be allocated to the ‘other – miscellaneous’ (PBC 23X) category. These two changes to the algorithm used to allocate expenditure to particular PBCs illustrate that year-on-year comparisons of expenditure need to be interpreted with care.

Expenditure per head on any given programme varies from one PCT to another and Table 2 presents some statistics that indicate the degree of variation in expenditure levels across PCTs by PBC. Columns 3–6 of Table 2 present descriptive statistics for PCT expenditure per person. These reveal that, for example, PCT per capita expenditure in the cancer programme averaged £96.30 across all PCTs, with the minimum spend being £62.90 and the maximum being £155.70.

Some PCTs will be spending more than other PCTs simply because they face higher input costs. Columns 7–10 in Table 2 present descriptive statistics for PCT per capita expenditure that has been adjusted for the unavoidable geographical variation in costs (input prices) faced by PCTs. g However, if anything, this adjustment appears to increase the variation in expenditure across PCTs; for example, the range of per capita expenditure on cancer increases from between £62.90 and £155.70 (unadjusted) to between £59.10 and £163.10 (adjusted for local health-care input prices).

Another cause of the variation in expenditure levels is the fact that the need for health care varies from one PCT to another. For example, areas with a relatively large proportion of elderly residents, or PCTs operating in relatively deprived locations, can be expected to experience relatively high levels of spending. The Department of Health has a well-developed methodology for estimating the relative health-care needs, which it uses as the basis for allocating health-care funds to PCTs. 75

Columns 11–14 in Table 2 present descriptive statistics for PCT per capita expenditure that has been adjusted for both the unavoidable geographical variation in costs and the local need for health care faced by PCTs. h For virtually every PBC, this adjustment reduces the variation in expenditure across PCTs; for example, the standard deviation of PCT per capita expenditure falls from £19.70 to £15.30 for the cancer programme. Although this adjustment reduces the variation in expenditure levels across PCTs, this decline is quite modest and there are still substantial differences in expenditure even after allowing for differences in local cost and need. For example, expenditure per head in the circulation problems category varies between £78 and £328 using cost-adjusted expenditure data, but falls between £76 and £327 using cost- and need-adjusted population data.

The variation in expenditure across PCTs has led some commentators to question the reliability of the PB data. The National Audit Office (NAO)77 undertook a survey of trusts, PCTs and Strategic Health Authorities (SHAs) to assess the quality of the data. They concluded that although the processes for collecting the budgeting data were well defined in most areas, there remained scope for improvements to the robustness of some of the data (e.g. non-admitted patient care). Appleby et al. 78 also considered the issue of data reliability in variations in spending on cancer services and noted some large year-on-year changes. However, the authors point out that it is difficult to define what might be either an implausible level of expenditure or an implausibly large change in expenditure. This is complicated by the fact that the Department of Health makes regular improvements to the way in which activity is matched to programme categories.

As with most data sets, there are likely to be recording and other errors associated with the PB data. However, although we note that the allocation of PB data might not be perfect there is no systematic evidence of this. Accordingly, for each disease category, we observe that PCT expenditure per person varies considerably and this variation – holding constant input prices and the need for health care – offers the opportunity to examine whether or not PCTs that spend more on health care achieve a better outcome and, if so, at what cost. Empirical estimates of the strength of this relationship for several programmes of care are presented in this report.

Health outcome data

Most studies of the relationship between expenditure and outcome have used some measure of mortality as an indicator of the latter. We also employ mortality as an outcome measure. First, it is a relevant (albeit not comprehensive) measure of the outcome of health-care expenditure; and second, it is available for more disease areas than any other outcome measure at PCT level.

Although mortality is available (by PCT) for several disease areas, it is not available for just over a half of all programmes not least because it is simply not relevant for these programmes (e.g. for learning disabilities, vision problems, hearing problems, dental problems and skin problems). Moreover, even where a mortality measure is available, the ICD-10 coverage of the mortality data often falls short of the coverage of the expenditure data. For some programmes, therefore, we have combined the published mortality rates for two or more disease areas in an attempt to match the ICD-10 coverage of the mortality data with that of the expenditure data.

Table 37 (see Appendix 2) shows how we have attempted to marry the mortality data (column C) and the expenditure data (column A). ICD-10 coverage of the component mortality rates for some PBCs falls short of the expenditure data and the extent of this shortfall is illustrated by the ratio reported in the final column of Table 3. For example, the cancers and tumours programme covers all expenditure associated with ICD-10 codes C00–C97 and D00–D49 but the PCT-based mortality data only relates to ICD-10 codes C00–C97. At the national (all England) level, figures are available which show that, in 2008, there were 62,072 deaths of those aged < 75 years from codes C00–C97 and that there were 63,076 deaths from codes C00–C97 and D00–D49 combined. In other words, the PCT-level mortality data reflect 98.4% of all deaths associated with the expenditure codes. We adjust our cost of life (year) estimates for this mismatch.

We acknowledge that mortality is a more relevant outcome indicator for some programmes (e.g. for circulatory problems) than for others (e.g. for epilepsy) and, for this reason, we would expect better results in some programmes than others. We also acknowledge that this focus on mortality ignores the impact of expenditure aimed at chronic care and at palliative care. Nevertheless, our focus on mortality is purely practical: it is both a widely available measure and clearly a relevant outcome indicator. i

The mortality data provide us with a number of possible outcome indicators including the < 75 years of age standardised mortality rate (SMR) and the (< 75 years) standardised years of life lost rate (SYLLR). The SMR gives equal weight to all deaths irrespective of the age at which they occur but the SYLLR gives greater weight to deaths that occur at earlier ages. For our purposes we focus on a measure of the avoidable YLL. j This is calculated by summing over ages 1–74 years the number of deaths at each age multiplied by the number of years of life remaining up to age 75 years. The crude YLL rate is simply the number of YLL divided by the resident population aged < 75 years. Like conventional mortality rates, the crude YLL rate can be age standardised to eliminate the effects of differences in population age structures between areas, and this (age) standardised YLL rate is the health outcome variable generally employed in this study. 79

Other variables

We employ an IV estimation technique to our empirical models of the outcome and expenditure equations as described in the next section. This is due to (i) own programme expenditure is likely to be endogenous in the outcome equation and (ii) other programme need is likely to be endogenous in the own programme expenditure equation. Endogeneity of programme expenditure results from expenditure levels being responsive to levels of outcomes and/or unobserved need rendering expenditure correlated with the residuals in an ordinary least squared (OLS) regression of outcomes on expenditure. Due to limitations in the data available, need in the expenditure equation in the ‘other’ programmes is proxied by death rates (minus that due to the programme under investigation). This will be influenced by expenditure decisions, including expenditure in other programmes and is treated as endogenous in the expenditure model.

Instrumental variable estimation basically involves replacing the endogenous variable in the equation of interest with its predicted value from an OLS regression which regresses the endogenous variable on a set of IVs. These instruments should be good predictors of the endogenous variable (i.e. they should be relevant and strong predictors) but should be appropriately excluded from the equation of interest (i.e. they should be valid instruments).

We have a number of potential instruments available, mostly derived from the 2001 Population Census. 80 In our earlier studies we found that a small subset of these instruments proved sufficient to generate plausible results. These included the proportion of the population providing unpaid care; the proportion of households that are one pensioner households; index of multiple deprivation; and proportion of the population in the white ethnic group.

We also had available a further set of potential instruments and, where our more limited set of instruments failed to generate plausible results, we extended our instrument search to include this wider set of variables. This extended set of instruments is shown in Table 3. k

Our instruments reflect factors, such as socioeconomic deprivation and the availability of informal care in the community, which might indirectly impact on mortality rates and/or health-care expenditure levels. As we shall see, although our instruments ‘pass’ the appropriate statistical tests, some commentators claim that such tests may have ‘low power’ to detect the presence of invalid instruments. Consequently, in Appendix 2, The sensitivity of the outcome elasticity to the validity of the instrument exclusion restrictions we examine how sensitive our results are to the presence of invalid instruments.

Table 3 reports descriptive statistics for the socioeconomic and needs variables used in the study (these statistics are for the variables in absolute form). For example, on average, lone pensioner households comprise 14% of all households, the ‘white ethnic’ group accounts for 89% of the population and 10% of the population provide unpaid care.

In addition to the IVs, Table 3 also reports descriptive statistics for the Department of Health’s ‘need for health care’ index,l its need for human immunodeficiency virus (HIV) services index, and its need for maternity services index. The latter two indices are used to either supplement or replace the all service measure of need when estimating our models. The ‘need for health care’ index averages about 1 but varies substantially, with some PCTs having a needs index more than 25% below the national average and others facing a need for health care more than 30% above the national average. Table 3 also reports descriptive statistics for some disease prevalence rates (e.g. for diabetes and for epilepsy) and, again, these are used to either supplement or replace the all service measure of need when estimating our models. Finally, the MFF index shows that input prices in the most expensive PCT are almost 20% above those in the least expensive PCT.

Instrumental variable estimation

Other programme need, m_i, in the expenditure Equation 7 and expenditure, x_i, in the outcome Equation 8 are both likely to be endogenous rendering OLS both biased and inconsistent. Endogeneity of programme expenditure results from expenditure levels being responsive to levels of outcomes and/or unobserved need. Other programme need in the expenditure equation is proxied by death rates which is influenced by expenditure decisions and hence is treated as endogenous. To deal with this endogeneity we employ IVs estimation and implement two-stage least squares (2SLS). Unlike OLS, IV is a consistent estimator in the presence of an endogenous regressor and, although in finite samples the IV estimator will be biased, with the bias (providing certain assumptions are met) being less than that associated with OLS.

For the health outcome equation, IV estimation can be viewed as finding variables (instruments) that are good predictors of programme expenditure but which are appropriately excluded from the outcome equation of interest (that is, from Equation 8) because they are not predictive of outcome. The assumption is that these instruments impact on health outcome through their impact on expenditure only, and that they do not have a direct effect on the outcome. o

Similarly, for the expenditure equation, IV estimation can be viewed as finding variables (instruments) that are good predictors of the proxy for other programme need (m_i) but which do not belong in the expenditure equation of interest (that is, Equation 7). The assumption is that these predictors impact on own programme expenditure only through their impact on other programme need and that they do not have a direct effect on own programme expenditure.

The outcome and expenditure equations for any given programme may contain different IVs because these instruments are trying to predict different variables (own programme expenditure and other programme mortality respectively). In addition, the instrument set for, say, the expenditure equation may vary across programmes because the other programme need variable will reflect need in a different basket of programmes for each expenditure equation.

We have a number of potential instruments available, mostly derived from the 2001 Population Census. 80 In previous studies, we have often found that a small subset (four) of these instruments often proved sufficient to generate plausible results. However, if plausible results were not obtainable with some combination of these four instruments, we employed an extended instrument set. Further details of the identification of suitable instruments for each model can be found in Appendix 2, Re-estimation of poorly performing models with an extended instrument set.

The available instruments reflect factors, such as socioeconomic deprivation and the availability of informal care in the community, which might indirectly impact on mortality rates and/or health-care expenditure levels. The set of instruments associated with each estimated equation was selected on both technical and pragmatic grounds. From a pragmatic point of view, we require a parsimonious set of instruments that satisfy the necessary technical criteria. These are, first, that they have face validity, that is, that they are plausible determinants of the endogenous variable being instrumented, and second, that the instruments are both relevant and valid. The relevance of an instrument set refers to its ability to predict the endogenous variable of concern, whereas validity refers to the requirement that instruments should be uncorrelated with the error term in the equation of interest.

Should the instrument set be strong, relevant and valid, 2SLS will produce consistent estimates of the parameters of the reduced form models. We subject the instrument sets to tests for validity using the Sargan–Hansen test of over identifying restrictions. The joint null hypothesis is that the instruments are valid instruments (i.e. they are uncorrelated with the error term), and that the instruments are correctly excluded from the outcome equation of interest. A rejection of the null hypothesis casts doubt on the validity of the instruments. We test for instrument relevance using Shea’s81 partial R² measure; this reflects the correlation between the excluded instruments and the endogenous regressor. However, even where valid and relevant, a non-zero but small correlation between the set of instruments and the endogenous regressors can lead to the problem of weak instruments, again rendering IV estimation biased. We test for the presence of weak instruments using the procedures set out in Stock and Yogo82 and the Kleibergen–Paap Lagrange multiplier (LM) statistic. A general test of model specification is provided through the use of Ramsey’s83 reset test for OLS and an adapted version of the test for IVs. 84

Finally, we check that the presumed endogenous variable is in fact endogenous using the test proposed by Durbin. 85 If the null hypothesis of exogeneity cannot be rejected, then we revert to using OLS. Although, in general, our instruments ‘pass’ the appropriate statistical tests, some commentators claim that such tests may have ‘low power’ to, and hence may fail, to reject the validity of the instruments when this is false in small samples. Consequently, in Appendix 2, The sensitivity of the outcome elasticity to the validity of the instrument exclusion restrictions we examine how sensitive our results are to the relaxation of the assumption that the instruments are valid.

Further details of our approach to IV estimation are set out in Appendix 2.

2006/7 expenditure data and mortality data for 2006/7/8

This section presents results that relate expenditure in 2006 to mortality in the same year and in the 2 following years (i.e. in 2006, 2007 and 2008). Throughout our measure of the need for health care is derived from the Department of Health’s resource allocation model based on the CARAN needs formula. 74 This represents a more up-to-date needs adjustment than the AREA based model76 that has been applied in previous studies,60,63 and is directly applicable to the 152 PCTs in existence in the 2006/7 expenditure year. Expenditure data has been adjusted for differences in input prices using the MFFs for HCHS and prescribing. r The outcome and expenditure results for the big four programmes are shown in Table 4 with the relevant outcome and expenditure elasticities highlighted.

In all four outcome models expenditure has a significant negative effect on mortality and the all service measure of need has a significant positive effect. The squared value of the measure of need is also positive and significant in the cancer outcome equation. In the respiratory outcome model, there is an additional indicator of need – the proportion of the population that are permanently sick – and this is both positive and statistically significant. The diagnostic statistics suggest that, in all four cases, own programme expenditure is endogenous and that the instruments are valid. They also suggest that the instruments are relevant. There is no evidence that the instruments are weak in three of the four outcome results. The Pesaran–Taylor test suggests that there is no evidence of model misspecification.

However, the Kleibergen–Paap F-statistic for the respiratory disease outcome model is 7.022 and this is less than the ‘critical’ target of 10.0. This indicates that the instruments may be weak and not good predictors of the programme expenditure. However, if we re-estimate this model having dropped the least significant instrument, the coefficient on own programme expenditure becomes −2.622 and is significant at the 1% level. Moreover, there is now no evidence of weak instruments (the Kleibergen–Paap F-statistic is 11.025) and it is this coefficient that we use for the respiratory outcome model in the cost of a life-year calculations below.

In three of the four expenditure models both the need and budget variables have a positive and significant effect on own programme expenditure. In addition, the proxy for need in other programmes is negative and significant in all four cases. The diagnostic statistics suggest that, for all four expenditure models, expenditure is endogenous and the instruments are valid. They also suggest that the instruments are relevant and there is no evidence that the instruments are weak. The Pesaran–Taylor test suggests that there is no evidence of model misspecification.

2007/8 expenditure data and mortality data for 2007/8/9

Outcome and expenditure models were estimated using updated data for expenditure (from 2006/7 to 2007/8) and updated mortality data (from 2006/7/8 to 2007/8/9). Appendix 2, Analysis of programme budgeting expenditure for 2007/8 and mortality data for 2007/8/9 presents detailed discussion of the findings including tables of results.

2008/9 expenditure data and mortality data for 2008/9/10

Outcome and expenditure models were estimated using updated data for expenditure (from 2007/8 to 2008/9) and updated mortality data (from 2007/8/9 to 2008/9/10). Detailed results for the outcome model and expenditure model are shown in Appendix 2, Tables 85 and 86 respectively. First-stage regressions for these IV models can be found in Tables 99 and 100 in the Annex to Appendix 2.

Comparing the cost of life-year estimates associated with different data sets

Table 6 presents expenditure and outcome elasticities for the five combinations of expenditure and outcome data that have been used to estimate our model. It also reports the corresponding unadjusted cost of life-year estimates (i.e. estimates that are unadjusted for the mismatch in the ICD-10 coverage of the expenditure and mortality data). It is clear from this Table 6 (see row 13) that the (unadjusted) cost of a life-year for the 10 programmes with a mortality-based outcome indicator fluctuated around £22,000 for the first three sets of estimations (see columns M–O). However, using the two most recent sets of expenditure data (i.e. for 2007/8 and then for 2008/9), the figures in the table suggest that this cost has increased to about £38,000.

What are the proximate causes of this increase? Recall that the cost of a life-year is calculated as:

For 2006/7 (using mortality data for 2006/7/8) and for the 10 programmes with a mortality-based outcome indicator, the change in expenditure associated with a 1% budget increase is £184.53M and the change in the number of life-years lost associated with this increase is 7760 (see Appendix 2, Table 67 for the calculation of these figures). Thus the cost of a life-year is £23,780 (= £184.53M/7760).

For 2007/8 (using mortality data for 2007/8/9) and for the 10 programmes with a mortality-based outcome indicator, the change in expenditure associated with a 1% budget increase is £257.94M and the change in the number of life-years lost associated with this increase is 6768 (see Appendix 2, Table 83 for the calculation of these figures). Thus the cost of a life-year is £38,110 (= £257.94M/6768).

It is clear that the 60% increase in the cost of a life-year between 2006/7 and 2007/8 is largely attributable to (a) the 40% increase in the additional expenditure (up from £184.53M to £257.94M) directed towards these 10 programmes following a 1% budget increase; and (b) the 12% decline in the number of life-years saved by this increase in expenditure (down from 7760 to 6768 life-years).

The rise in the share of the budget increase directed towards these programmes can be attributed to the increase in the implied expenditure elasticity associated with these 10 programmes (up from 0.561 to 0.749). The decrease in the number of years of life saved appears to be due to (a) an overall reduction in the (absolute) size of the outcome elasticities; and (b) a shift in the additional expenditure towards those programmes with a relatively high cost of a life-year. For example, the cost of a life-year for the ‘small six’ programmes is much larger than for the ‘big four PBCs’. However, in 2007/8 the spend elasticity for the small six increases from 0.561 to 0.961 (71%) whereas the expenditure elasticity for the big four rises from 0.528 to 0.559 (6%). A similar pattern – of additional expenditure shifting away from the low cost PBCs – can be seen within the big four programmes. However, it is not clear why such rather dramatic changes should have taken place.

If we correct the cost of life-year estimates adjusting for the mismatch in the ICD-10 coverage of the expenditure and mortality data, these reveal similar increases in the cost of a life-year between 2006/7 on the one hand and 2007/8 and 2008/9 on the other. The cost of a life-year increased from £19,965 in 2006/7 to £28,983 in 2007/8 for the 10 programmes with mortality rate, an increase of 45%; and it increased from £22,565 to £31,846 for all programmes if we assume a zero health gain in PBC 23 and the same gain in the other 12 programmes as in the 10 with a mortality rate (an increase of 41%).

A potential reason for this apparent step change in the cost of a life-year is the adjustment that was made to the methodology for the collection of the 2007/8 PB data. In previous years expenditure that was not directly attributable to a particular programme category was apportioned using admitted patient care percentages. ac In other words, if x% of total admitted patient care expenditure was allocated to PBC 1, then x% of all expenditure that was not directly attributable to a particular programme category was also allocated to PBC 1. With effect from 2007/8, however, NHS organisations were asked to select an appropriate basis for the apportionment of this non-programme-specific expenditure and that, where no reasonable basis existed, such expenditure was to be allocated to the ‘other – miscellaneous’ (PBC 23X) category.

The Department of Health estimates that this allocation rule change increased the amount of expenditure attributed to PBC 23X by £700M. It will also, of course, have reduced expenditure across other programmes by the same amount in total. However, not all programmes will have been equally affected; PBCs that are more heavily inpatient based would have ‘lost’ expenditure whereas others, such as learning disabilities, social care and mental health, will have ‘lost’ considerably less. In addition, not all PCTs will have been equally affected because each will have employed different apportionment rules for the non-programme-specific expenditure (Bryn Shorney, Department of Health, 2012, personal communication).

Although this allocation rule change has considerably increased the estimated cost of a life-year, we believe that this rule change has led to a more accurate allocation of expenditure across PBCs, and that the more recent estimates of the cost of a life-year (for 2007/8 and 2008/9) are more accurate than those for the earlier years (for 2005/6 and 2006/7).

Adjusting the cost of a life-year estimates to constant prices

The estimates of the cost of a life-year presented above are all at current prices. To put them on a constant price basis, we need an index of pay and price inflation for the labour and goods/services purchased by the NHS. Curtis87 reports a pay and prices index for HCHS and this implies an inflation rate of 3.7% in 2006/7, 2.9% in 2007/8 and 3.9% in 2008/9. ad If we assume that similar inflation rates also apply to the purchase of pharmaceuticals and the provision of primary care (items that are excluded from the HCHS index), then we can use these figures to put the estimates of the cost of a life-year on a constant price basis.

For example, if we assume that PBC 23 generates a zero health gain and that the gain attributable to the 12 programmes without a mortality indicator is, on average, the same as that attributable to those with a mortality outcome measure, then the cost of a life-year across all programmes in 2008/9 is £33,333 at current (2008/9) prices. The cost for 2007/8 is £31,846 at current (2007/8) prices or £33,088 at constant (2008/9) prices, and the figure for 2006/7 is £22,565 at current (2006/7) prices or £24,125 at constant (2008/9) prices. The conversion of the costs from a current to constant price basis has relatively little impact because the inflation rate over the relevant period is quite small.

Mortality and years of life lost coverage

The mortality data that is available at PCT level does not offer full coverage of all deaths across all the ICD-10 codes that make up each PBC (see Table 37 in Appendix 2 for how three-digit ICD-10 codes are mapped to PBCs). However, national (English) data are available that covers all deaths associated with all the ICD-10 codes that make up each PBC. Therefore, it is possible to adjust the incomplete reporting of mortality at PCT level (see Chapter 3, Previous studies) before applying the estimated outcome elasticities to calculate the deaths averted due to expenditure. b Applying published estimates of YLL per death to all the deaths averted provides the estimate of the cost per life-year reported in Chapter 3.

The published estimates of YLL (National Health Service Information Centre; NHS IC) used in Chapter 3 only include deaths below the age of 75 years (but exclude deaths below age 1 year) and are based on the difference between age 75 years and the age of each death below 75 years. These estimates have the same limited coverage as PCT-level mortality data so are not available for all the ICD-10 codes that make up each PBC. Therefore, applying the available estimates of YLL per death to the estimated number of deaths averted requires an assumption that the YLL per death is similar for those groups of ICD-10 codes covered and not covered by the published YLL figures.

This can be examined by using national Office for National Statistics (ONS) data to calculate YLL in the same way as NHS IC, but with full coverage of all the ICD-10 codes that make up each PBC. c Although ONS data provides complete coverage and reports gender, age at death is only reported in 5-year ranges (these data are not available at PCT level so could not be used when estimating outcome elasticities in Chapter 3). Therefore, using ONS data to estimate YLL requires taking the mid-point of each range as the age of death, i.e. assuming reported deaths are equally likely over the range in which they are reported. For this reason it is not possible to precisely recover the published YLL figures using ONS data for those ICD groupings that can be precisely matched to the NHS IC coverage. However, the differences are small (see Appendix 3, Table 102), suggesting that taking the mid-point of each range as the age of death may be a reasonable approximation.

The differences between estimates of YLL based on ONS and NHS IC data are, however, much more significant and are reported in Table 8. These reflect differences in the distribution of ages at death between those groups of ICD-10 codes covered and not covered in the NHS IC figures. For example, NHS IC figures available at PCT level for PBC 7 (neurological problems) have low coverage of all deaths in this PBC [0.14 in column (1)]. The deaths that are reported in NHS IC are associated with epilepsy and the YLL [22,046 in column (2)] reflects the generally younger age at death in this group. When adjusted for full coverage [22,046/0.14 = 162,100 in column (3)] the estimated YLL is much greater than the YLL based directly on all deaths by age group reported in ONS. This difference in YLL reflects the fact that the deaths in PBC 7 which are not covered by NHS IC figures tend to be in older age groups so generate fewer YLL.

Using ONS data also allows deaths under the age of 1 year to be appropriately assigned to PBCs via the ICD-10 code in which they occurred (NHS IC YLL figures exclude deaths under 1 year), rather than assigning them all to PBC 18 + 19 as in the previous chapter. d This explains the large reduction in YLL for PBC 18 + 19 (maternity and neonates) as much of the mortality is reassigned to ICD-10 codes which contribute to other PBCs. As most of the deaths that are reassigned are allocated to PBC 1 (infectious diseases) the YLL for this PBC increases despite complete reporting of deaths at PCT level and full coverage by NHS IC figures (see also Table 104 in Appendix 3).

Life expectancy and years of life lost

As noted above, the NHS IC estimates of YLL only include deaths below age 75 years and are based on the difference between age 75 years and the age of each death below 75 years. Implicitly this treats 75 years as the appropriate normal LE for males and females for the population at risk in each PBC. However, with the exception of maternity and neonates most deaths in PBCs occur above the age of 75 years and LEs are significantly > 75 years. For example, based on 2006–8 data, LE for the general population is 80.7 years for males and 84.4 years for females (considering age distribution) and even LE at birth is > 75 years (77.74 years for males and 81.88 years for females). 90

Based on ONS data YLL can be recalculated using gender-specific LE for the general population. e When increasing LE two effects occur, both of which tend to increase estimates of YLL. First, more deaths are included in the YLL calculation (those that occur between age 75 years and LE) and second, each death previously counted below 75 years will generate 5.7 or 9.4 more YLL for males and females respectively. The effect on the number of deaths and the YLL for each PBC of using the LE of the general population is reported in Table 9 [see columns (1), (2) and (3)].

The number of deaths counted below LE increases for every PBC except for maternity and neonates because, as expected, all deaths are below age 75 years in PBC 18 + 19. However, YLL increases for all PBCs reflecting the additional years otherwise expected to be lived to an older LE. Of course including more of the deaths observed in each PBC and the greater YLL associated with them will generate more deaths averted and more YLGs when applying the same proportionate effects from the outcome elasticities estimated in Chapter 3. Therefore, the cost per death averted and cost per life-year threshold are lower using these figures than those reported in Chapter 3 (see Table 13 and Table 107 in Appendix 3 for a summary of the effects on the thresholds). However, there are good reasons why YLL figures calculated as the difference between age of death and LE are likely to be overestimated. This is dealt with in the next section (see Years of life lost and accounting for counterfactual deaths). In Inferring excess deaths we take account of the fact that some of the deaths observed in a PBC would have occurred anyway in a similar ‘normal’ population (i.e. the counterfactual population not at risk through membership of the PBC) so not all observed deaths are ‘excess’ and generate YLL.

Years of life lost and accounting for counterfactual deaths

The estimates of YLL based on ONS data overcome many of the limitations of the published NHS IC figures. However, the YLLs reported in Tables 8 and 9, are calculated in the same way as the NHS IC figures, by taking the difference between a fixed LE and the age at death of deaths observed below that LE. This will tend to overestimate the YLL for two reasons: (1) it does not account for the fact that not all deaths observed below LE are ‘excess’ deaths in the sense that some deaths would have occurred (at the same age) in a similar population not at risk in the PBC; and (2) some of the deaths observed above LE may be ‘excess’ deaths that would not otherwise have occurred at that age. The overall effect on YLL, and the cost per life-year, will depend on the number of deaths above and below LE that are excess. Therefore, estimates of YLL are required which take account of the ‘counterfactual’ deaths that would have occurred even if the population in the PBC was not at risk through membership of the ICD-10 codes that make it up, but faced the same mortality risks as the general population, accounting for the age and gender distribution of the PBC population.

Ideally, with reliable information about the size of the population at risk in each PBC and its age and gender distribution it would be possible to estimate the number of deaths that would be expected to occur had this population not been at risk, based on mortality data for the general population. The difference between deaths observed across all ages and the deaths expected to have occurred in this matched ‘normal’ population would provide the number of ‘excess’ deaths by age and gender. f The YLL associated with each of these excess deaths is the LE conditional on gender and on surviving to the age at which the excess death occurred. The total YLL for the at risk population is simply the sum of these YLLs over all excess deaths, which could occur at any age. This YLL is equivalent to the area between the survival curve for the population at risk in a PBC and the counterfactual survival curve for the same population but not at risk from membership of the PBC. The difficultly is that routinely available data do not provide any information about the size of the population at risk or its age and gender distribution. All that is routinely available are observed deaths (by age and gender). Therefore, it is not possible to directly estimate excess deaths or compare survival curves.

Even if the size of the at risk population is unknown we can still use information that might be available about its age and gender distribution (or make reasonable assumptions) to estimate a matched ‘normal’ LE using life tables for the general population (such a LE summarises the area under the counterfactual survival curve). Unfortunately, it is not possible to also calculate the LE for the population at risk in the PBC (or represent the survival curve) without information about the size of the at risk population (if it was possible the difference between these life expectancies would approximate the YLL per patient at risk in a PBC).

Fortunately, we can still recover a consistent estimate of YLL using observed deaths and a LE that represents the normal LE of a matched population that is not at risk. This requires all observed deaths, both those that occur below and those that occur above this LE to be taken into account. Those deaths occurring below LE generate YLL – compared with the average of a matched population not at risk. However, we must also account for those deaths that occur at ages above LE. These deaths generate YLGs compared with the average of a matched population not at risk. Therefore, the appropriate estimate is a net YLL (i.e. YLL – YLG). In effect, by subtracting YLG from YLL we take account of the fact that not all deaths below LE are excess deaths but some deaths above LE are (see Appendix 3 for more formal explanation of the equivalence of these ways of calculating YLL). g

Inferring excess deaths

We have been able to establish a measure of net YLL, which takes account of deaths that would have occurred anyway below a normal LE for the PBC population (i.e. not all deaths observed in a PBC are excess), and that some deaths observed above this LE would not otherwise have occurred at that age (i.e. some of these deaths are excess). As explained in Years of life lost and accounting for counterfactual deaths, net YLL calculated in this way is equivalent to first establishing the number of excess deaths at each age, then calculating YLL for each excess death (based on the LE conditional on the age at which each excess death occurred) and then summing these YLL across all excess deaths (i.e. across all ages). In other words, the estimates of net YLL imply a number of excess deaths required to generate them in each PBC. Therefore, it is possible to solve for the total number of excess deaths based on the net YLL and the average YLL per observed death. n The net YLL divided by the average YLL per death provides the number of excess deaths required, which on average will generate the estimated net YLL. o

The implied excess deaths associated with net YLL based on the LE of the PBCs [see Table 12, column (7)] are reported in Table 14. With the exception of PBC 18 + 19, excess deaths are some proportion of total observed deaths in each PBC. The proportion of excess deaths differs by PBC reflecting the distribution of deaths relative to the LE of the PBC. p For example, in those PBCs where a large proportion of deaths occur below LE [see Table 12, columns (3) and (4)], excess deaths tend to be a greater proportion of the total deaths (e.g. PBC 2, 13 and 10). Where most deaths occur above LE, excess deaths as a proportion of the total deaths tend to be lower (e.g. PBC 1, 11 and 17).

Estimates of net YLL and changes in life-years due to expenditure (see Tables 12 and 13) have already accounted for the fact that not all deaths are excess and don’t generate YLL. Nevertheless, solving for the number of implied excess deaths associated with these net YLL estimates allows a comparison of the cost per excess and observed PBC death avoided and an examination of the interpretation that can be placed on the life-years expected to be gained from an excess or observed death averted. As only deaths observed in the PBC can be used to estimate the effects of expenditure (excess deaths are not directly observed as they rely on an unobserved counterfactual population and would occur outside the PBC), the outcome elasticities can be interpreted as the proportionate change in observed PBC mortality due to a proportionate change in PBC expenditure. Equally, however, they can also be interpreted as the proportionate effect on excess death due to a proportionate change in expenditure so can be applied to either total observed or total excess deaths. q

The cost per excess death and the cost per PBC death averted are reported in Table 15 (see Table 119 in Appendix 3 for a detailed breakdown of changes in spend and excess or PBC deaths across PBCs). The cost per PBC death averted is, of course, significantly lower than the cost per excess death as excess deaths are only a proportion of total deaths (see Table 14). Also the cost per PBC death averted is substantially lower than those reported in Chapter 3 (see Tables 68 and 69 in Appendix 2), as these estimates do not restrict the effects of expenditure to PBC deaths under the age of 75 years. r The cost per PBC or excess death averted (or life saved) should not be over interpreted as they are of little direct policy interest because lives are never saved (death is only delayed), and the significance of a death averted depends critically on how long it is averted for (the YLGs, see Table 13) and the QoL in which additional years are lived (see Adjusting life-years for quality-of-life).

However, establishing the number of excess and PBC deaths averted which are associated with net YLL is useful because it enables an assessment of the number of YLGs associated with each death averted. On average, across all 11 PBCs, each excess death averted is associated with 11.4 YLGs. These are reported for each PBC in Table 121 in Appendix 3, and range from 74.3 years per excess death for PBC 18 + 19 maternity and neonates to 8.3 years for PBC 17 genitourinary. However, clinicians or the evaluative literature cannot distinguish whether or not an observed death is excess. What can be observed is whether or not groups of similar patients with and without access to a treatment survive and for how long. Therefore, it is the life-years associated with each observed death that provide a context that can be interpreted based on experience and evidence of how effective those interventions that could be invested or disinvested intend to be. The average life-years expected to be gained associated with each observed PBC death averted takes account of the fact that some deaths that are avoided in the PBC are not delayed for very long but quickly occurs elsewhere and do not generate YLGs (i.e. they were not excess deaths). These are also reported for each PBC in Table 121 in Appendix 3 and range from 74.3 years per observed death for PBC 18 + 19 maternity and neonatest to 1.0 year for PBC 11 respiratory problems (i.e. the YLL per PBC death are much lower for those PBCs where a small proportion of observed deaths are excess). Each PBC death averted is associated with 4.1 YLGs on average across all 11 PBCs.

Summary of cost per life-year estimates

The sequence of analysis set out above has enabled an examination of the impact of the limitations associated with the incomplete reporting of mortality data at PCT level and incomplete coverage of published YLL estimates. We have also been able to consider effects above the age of 75 years while taking account of that fact that many deaths would have occurred anyway, despite the limited information available about the population at risk within a PBC. The GBD study does provide some information about the age and gender distribution of the population at risk in a PBC so offers some improvement over the other assumptions that would otherwise be required (i.e. that the distribution of age and gender is the same as the general population or follows the distribution of observed deaths). For this reason the cost per life-year threshold in column (4) of Table 13 and repeated in lines (1)–(4) in Table 16 are regarded as the central or best estimates given the evidence available and the credibility of alternative assumptions that could be made. As explained in the Introduction of this chapter, these are based on the conservative assumption that any health effects of changes in expenditure are restricted to 1 year, which, to some extent, may be offset by the more optimistic assumption any death averted returns the individual to the mortality risk faced by the general population, matched for age and gender.

It does not seem credible to imagine that NHS expenditure has no health effects in the 12 PBCs which do not have sufficient mortality reported at PCT level to estimate outcome elasticities – what is implied by the estimate reported in Table 16, line (3). Therefore, it is the estimates reported in Table 16, lines (2) and (4) that are of policy interest. The estimate of £15,628 per life-year [see Table 16, line (2)] is restricted to the effects of changes in expenditure in the 11 PBCs where outcome elasticities can be estimated. The threshold of £17,663 per life-year uses the estimated health effects of expenditure in these PBC as a surrogate for health effects in the others, that is assuming that the effects that can be observed will be similar to those that cannot. However, no health effects are assigned to PBC 23 (GMS) on the basis that any health effects of this expenditure would be recorded in the other PBCs. u

The extreme upper and lower bounds for the cost per life-year thresholds in Table 16 are based on making the necessary assumptions about duration of health effects and how long a death might be averted optimistic (providing the lower bound for the threshold) or conservative (an upper bound for the threshold). The lower bound [see Table 16, lines (5)–(8)] is based on assuming that health effects are not restricted to 1 year but apply to the whole of the remaining disease duration of the population at risk in PBCs during the expenditure year. v Although this combines optimistic assumptions, it is possible, indeed likely, that at least some expenditure may have effects on the health outcomes of future patients who are not currently part of the population at risk in a PBC (e.g. investments or disinvestment in prevention will have an impact on populations that are incident to PBCs in the future). Such effects are not captured in any of the estimates presented in this chapter so all are conservative with respect to this type of health effect of changes in expenditure.

The upper bound [see Table 16, lines (9)–(12)] is based on the combination of assuming that health effects are restricted to 1 year for the population currently at risk and that any death averted is only averted for the minimum duration consistent with the mortality data. The econometrics work used the average of 3 years of mortality (2006–8), so the estimated outcome elasticities are based on differences in mortality that remain after averaging over 3 years. Therefore, the estimated effects are based on differences in observed PBC deaths that must have been sustained, on average, for more than a minimum of 2 years. w

Quality of life based on the general population

The most commonly used metric of health-related QoL in the UK is European Quality of Life-5 Dimensions (EQ-5D),91 which is specified in the NICE reference case for methods of technology appraisal. 3 This metric has five dimensions of quality each with three possible levels. Each of these 243 possible health states is valued relative to a score of one, which represents full or best imaginable health (the best score across all five dimensions), and a score of zero, which represents death, based on a representative sample of the UK population. 92 Therefore, insofar as the years of life expected to be gained (or lost) through changes in expenditure would be lived in this state of full health, the cost per life-year thresholds reported in Table 16 would also be the cost per QALY thresholds, albeit ones that only account for the health effects of mortality changes. However, unsurprisingly, there is good evidence that, on average, the general population is not in this state of full health. Therefore, the QoL score associated with the health states experienced by the general population are less than 1, decline with age and differ by gender. These QoL ‘norms’ for the general population by age and gender are illustrated in Figure 3 based on an analysis of data from the Health Survey for England (HSE). x

FIGURE 3.

Quality of life for the general population by age and gender. Pooled QoL estimates provided by personal communication with Dr Anju Keetharuth, University of Sheffield, 2013.

These QoL norms can be applied to the YLL associated with all observed deaths in each PBC, taking account of gender and age at death. The results are reported in columns (4)–(6) of Table 17. There are two effects of adjusting life-years for quality: (i) as QoL norms are always < 1 the adjusted YLL and YLG are always lower than the unadjusted values in columns (1) and (2) (previously reported in Table 12); and (ii) deaths above LE are necessarily at older ages with poorer QoL norms than those below, so the difference between adjusted and unadjusted values is greater for YLG than YLL. The overall effect of quality adjustment on net YLL is the balance of these two effects. The overall effect of quality adjustment is to reduce the net YLL [compare Table 17, columns (3) and (6)]. y

The quality adjusted net YLL figures in Table 17, column (6) suggests that the health effects of mortality are lower than when relying only on unadjusted life-years in From mortality to life-years. Therefore the health effects of changes in expenditure on this more complete measure of health are lower. The implications of these adjustments on a cost per QALY threshold that only accounts for the direct health effects of mortality are reported in Table 18. As expected the cost per QALY threshold based on adjusting the YLGs or lost [see Table 18, column (2)] is higher than a threshold based on unadjusted life-years [see Table 18, column (1) and previously reported in Table 16].

Adjusting age-related quality-of-life for disease decrements

Adjusting life-years for age- and gender-related QoL norms assumes that any YLG through a change in expenditure would be lived in a similar QoL to the general population. It is possible, however, that patients benefiting from reduced mortality may, nevertheless, continue to be affected by the type of diseases that make up each PBC and experience the QoL associated with the original disease.

The HODaR93 provides over 30,000 observations of EQ-5D measures of QoL by ICD-10 code and the age and gender of the patients in the sample (see Appendix 3, Addendum 1: data sources). Although this is a rich UK data set, there were a limited number of observations for some of the less common ICD-10 codes. For this reason HODaR was supplemented with information from the Medical Expenditure Panel Survey (MEPS)94 which also provides EQ-5D by ICD-10 and reports the average age of respondents (see Appendix 3, Addendum 1: data sources). These data provided a means of estimating the QoL associated with each ICD-10 code at the average age of respondents in the pooled sample. z The QoL associated with each PBC can be expressed as an average of the QoL associated with its component ICD-10 codes. aa The QoL effects of being in each PBC can then be expressed as a disease-related decrement compared with the population norms at the same age (see Table 129 in Appendix 3). This is illustrated for PBC 1 (infectious disease) in Figure 4, where the weighted average of QoL scores across the component ICD-10 codes was 0.667, at an average age average age of 54 years for male respondents. As the QoL norms for males age 54 years is 0.859 this suggests a decrement associated with membership of PBC 1 of 0.192, which can then be applied to QoL norms by age. ab

FIGURE 4.

Quality of life for males in PBC 1 (infectious disease) and the general population by age.

Quality-of-life norms adjusted for disease-related decrements can be applied to the YLL associated with observed deaths in each PBC, taking account of gender and age at death in the same way as in Quality of life based on the general population. ac The results are reported in columns (4)–(6) of Table 19. The overall effect of quality adjustment that also applies a disease-related decrement is to reduce the net YLL to a greater extent than adjustment with population norms alone [compare column (6) in Table 19 with column (6) in Table 17].

It should be noted that combining QoL adjustments for both population norms and disease-related decrements assumes that any YLGs due to a reduction in mortality will be lived in the diseased state until LE (i.e. that all diseases are not just chronic but disease duration is lifelong). Inevitably this assumption means that the health effects of changes in mortality will be reduced. Consequently, the cost per QALY threshold reported in Table 20, column (2) will be higher than adjusting YLGs for population norms in Table 18.

Summary of the cost per quality-adjusted life-year threshold based only on mortality effects

The analysis to this point is summarised in Table 21. The three estimates of a cost per QALY threshold are based on assuming that each YLG is either lived in full health [see Table 21, column (1), equal to the cost per life-year estimates in Table 16]; lived in a QoL that reflects age and gender norms of the general population [see Table 21, column (2)]; or lived in a QoL that reflects the original disease state [see Table 21, column (3)].

Assuming that YLGs are lived in full health is not credible and should be regarded as an underestimate of the threshold, given what is known about QoL norms for the general population (see Figure 3). Equally, assuming that all YLGs are lived in the QoL of the original disease state does not seem credible either and is likely to overestimate the threshold as it assumes that all disease is not only chronic but lifelong and all life-years would be lived in the diseased state until death. ad Although adjusting YLGs for the QoL of the general population, taking account of age and gender [see Table 21, column (2)], is likely to underestimate a cost per QALY threshold based only on mortality effects, it probably represents the ‘best’ of the three alternative estimates available at this stage of the analysis (see Using estimates of the QALY burden of disease for how analysis based on measures of QALY burden allows this assumption to be relaxed). ae The lower and upper bounds are based on combining optimistic and pessimistic assumptions about the duration of health effects and how long a death might be averted as described in Summary of cost per life-year estimates.

However, it should be noted that these cost per QALY thresholds only account for the direct health effects of changes in mortality due to changes in expenditure. Insofar as much, or at least some, of NHS activity and expenditure is intended to improve QoL, not just mortality, then these estimates will underestimate total health effects and overestimate a cost per QALY threshold based on a more complete measure of possible health effects. In Including quality-of-life effects during disease we explore the ways in which the likely effects of expenditure on QoL (other than through mortality) might also be taken into account.

Using ratios of quality-adjusted life-years to years of life lost

The ratio of the total QALYs to YLL due to a disease indicates the number of QALYs associated with each YLL. Therefore, any change in YLL is likely to generate a number of QALYs indicated by the ratio – if it is reasonable to interpret the estimated effects on mortality and life-years as a surrogate for a more complete measure of total health effects. For example, a disease with a ratio > 1 suggests that each YLL across the at risk population is associated with more than 1 QALY (i.e. where there are significant QoL effects while experiencing the disease). af Therefore, a change in expenditure that leads to 1 YLG in this type of disease may be expected to generate > 1 QALY and a greater QALY effect than the same life-year effects in a disease where this ratio is < 1 (i.e. where most of the effect of disease is on mortality rather than QoL). Therefore, information which allows these ratios to be estimated for the diseases that make up each PBC provides a means of accounting for the likely effect on QoL other than through effects on mortality.

To understand the differences between the three ratios presented below it is useful to regard the total QALY lost to YLL ratio (R) for a particular disease as the sum of two ratios: (i) the QALYs lost due to premature death to YLL ratio (R_death);ag and (ii) the QALYs lost during disease (while alive) to YLL ratio (R_alive) (see Using ratios of quality-adjusted life-year to years of life lost in Appendix 3 for more detailed explanation).

Using estimates of the quality-adjusted life-year burden of disease

In this section we use estimates of the QALY burden of disease to infer QALY effects in those PBCs where the mortality effects of changes in expenditure can be observed and then extrapolate the estimated proportionate effects to those PBCs where the health effects of changes in expenditure cannot be observed. The estimated proportionate effect of change in expenditure on the life-year burden of disease in the 11 PBCs where mortality-based outcome elasticities could be estimated are applied to measures of QALY burden in each of these PBCs (i.e. effects on the mortality burden of disease are used as a surrogate for effects on QALY burden). The proportionate effect on burden of disease due to the change in expenditure across these PBCs can then be applied to measures of QALY burden in the other 11 PBCs where mortality effects could not be estimated (i.e. the observed effects of changes in expenditure on burden of disease is extrapolated to the other PBCs where health effects cannot be observed). In this way we can use all the information available about the mortality and QoL effects of the different types of disease that make up each PBC, particularly those where mortality-based outcome elasticities are not available. Applying a proportionate effect to measures of QALY burden of disease is equivalent to assuming that any effects on life-years are lived at a QoL that reflects a proportionate improvement to the QoL with disease. ap It also allows QoL effects of changes in expenditure to be included; also based on proportionate improvement in the QoL with disease.

Previously in Chapter 3 and From mortality to life-years, Adjusting life-years for quality-of-life and Using ratios of quality-adjusted life-years to years of life lost, expenditure elasticities were not estimated for the other 11 PBCs where outcome elasticities could not be estimated because the same health effect of changes in expenditure was assumed (i.e. it did not matter how changes in expenditure was allocated between them). However, in this section it does matter how the remaining change in expenditure is allocated between the other 11 PBCs as they have different QALY burdens so different implied health effects of expenditure. Therefore, expenditure elasticities are estimated for all 23 PBCs (see column 2 of Table 108 in Appendix 3). However, it is not possible to estimate expenditure equations for all 23 PBCs simultaneously (see Chapter 5, Future research and improving estimates of the threshold), so the 23 independently-estimated expenditure elasticities do not account for all of the change in overall spend (i.e. the sum of changes in PBC expenditure based on the estimated PBC expenditure elasticities accounts for a < 1% change in total spend). This remaining change in total spend is allocated between all 23 PBCs reflecting their relative share of changes in expenditure based on their estimated expenditure elasticities [see column (4) of Table 108 in Appendix 3]. aq

The total QALY burden of disease for the population with disease in a particular year includes (i) the quality-adjusted YLL due to all the disease-related mortality that could occur in this population over their remaining duration of disease; and (ii) the reduction in QoL while alive also for their remaining disease duration. However, applying the estimated proportionate effects on mortality and life-years to such a measure of total burden would provide an estimate of the effects of a change in expenditure, not just in 1 year, but in all the remaining years of disease for the population at risk in that year. Recall from From mortality to life-years that we have adopted the conservative assumption that changes in expenditure will only have health effects in 1 year for the population with disease in that year. Therefore it is not a measure of total burden that is required, but a measure of the QALY burden of disease during 1 year for the population with disease (prevalent and incident) in that year. The estimated outcome elasticities can then be appropriately applied to this measure of burden. ar

The information from GBD used to derive QALY ratios in Using ratios of quality-adjusted life-years to years of life lost includes information about the YLL and duration of disease for those incident to a U-code [i.e. the measure of QALY burden from the information included in the ratios is a measure of the total burden of the disease, but only for the population that is incident (rather than the total population with disease) in 1 year]. Assuming that incidence is stable over the disease duration, this is also equivalent to the QALY burden of disease during 1 year for the population with disease (i.e. those that are incident and prevalent) in that year. as

However, in moving from ratios to absolute measures of burden it becomes more important to examine and then adjust for any inconsistency between information about YLL and size of the incident population from GBD (which is available by U-codes and can be mapped to ICD-10 codes), and the information about net YLL and observed deaths for each PBC based on ONS data as described in Inferring excess deaths (see Table 146 in Appendix 3). at

The implications for the cost per QALY threshold of using information about the QALY burden of disease for all PBCs rather than QALY ratios for those where an outcome elasticity can be estimated are reported in Table 27. The QALY effects of a change in PBC expenditure are a weighted average of the QALY effects within each of the ICD-10 codes that contribute to the PBC. The figures reported in Table 27, column (2) are based on weighing the effects at ICD-10 level by the proportion of the total PBC population within each contributing ICD-10 code, rather than the contribution to variance in HES costs. au

The cost per QALY threshold for the 11 PBCs with outcome elasticities is lower using a measure of QALY burden (£5128) rather than the QALY ratios (£10,297) described in Mortality and years of life lost. This is in part because GBD calculates YLL in the same way as published NHS IC figures and so will tend to overestimate a net YLL which accounts for counterfactual deaths (see Years of life lost and accounting for counterfactual deaths). This will make little difference to the first term in the QALY ratio (R_death) used in Mortality and years of life lost coverage as an overestimate of YLL affects both denominator and numerator of the ratio. However, the second term (R_alive) is likely to be underestimated. Therefore the ratios used in Using ratios of quality-adjusted life-years to years of life lost will tend to underestimate the QALY effects of expenditure and overestimate the cost per QALY threshold (see Table 27). We are able to adjust the GBD-based measure of QALY burden for this overestimation of net YLL in calculating the QALY threshold reported in Table 27, column (2). av

As the purpose of this section is to use the estimated mortality and life-year effects as a surrogate for a more complete measure of likely health effects, it is the cost per QALY threshold for all 23 PBCs that is of most relevance. The cost per QALY threshold for all 23 PBCs is based on applying the proportionate effects on the QALY burden of disease, based on the observed effects of changes in expenditure on mortality in the 11 PBCs with outcome elasticities,aw to the QALY burden of disease in the other PBCs. This generates a higher cost per QALY threshold (£10,187) than the one based only on the 11 PBCs with outcome elasticities (£5128). The reason is that the QALY burden of disease in the other PBCs is, in general, lower than the QALY burden of disease across those PBCs where outcome elasticities can be estimated (see Table 147 in Appendix 3). Therefore, applying the same proportionate effects to a lower QALY burden generates a smaller health effect of a change in expenditure. ax

In essence the difference between the estimates in Table 27 is that in column (1) the absolute effect on health associated with an absolute change in expenditure is extrapolated to the other PBCs, whereas in column (2) it is the relative effect on health of an absolute change in expenditure that is extrapolated. As we know that QALY burden differs between (and within) PBCs and especially between the groups of PBCs with and without estimated outcome elasticities (see Table 147 in Appendix 3),ay it is the values based on QALY burden in column (2) that are regarded as most credible and represent our central or best estimate.

A detailed breakdown of changes in expenditure and changes in QALYs across all PBCs is provided in Table 150 in Appendix 3, both when the analysis is based on QALY ratios and when it is based on QALY burden of disease. A comparison of these values suggests that QALY effects for the other PBCs are generally lower and therefore the cost per QALY for each of these PBCs are, in general, higher when based on a proportionate effect on QALY burden. Of course, we have not directly observed QoL effects in these PBCs but inferred them from the proportionate effects that we can observe. Insofar as investment and disinvestment opportunities in these PBCs might have been more valuable (offered greater improvement in QoL)az than suggested by the implied PBC thresholds, then overall QALY effects will tend to be underestimated and the cost per QALY threshold overestimated. For the reasons discussed in previous sections, we regard all the costs per QALY threshold reported in column (2) of Table 27 as on balance conservative with respect to overall health effects of a change in expenditure. However, the estimate of £10,187 may be conservative with respect to health effects (i.e. overestimated), as it is based on an extrapolation of the proportionate effects to measures of burden on these PBCs, rather than observations of the direct impact of changes in expenditure on QoL in these types of disease. This is especially important in PBC 5 (mental health disorders), which accounts for a large proportion of the change in overall expenditure (22%) and where a review of the evidence suggests that the investment and disinvestment opportunities in this PBC may have been more valuable than the implied PBC cost per QALY of £13,876 (see Table 155, Appendix 3 and Addendum 3: characterisation of the investment and disinvestment decisions in mental health – depression and schizophrenia). az The lower cost per QALY threshold for the 11 PBCs with outcome elasticities (£5128) might be regarded as more secure in this respect but they only account for a proportion (38%) of any change in overall expenditure (see Table 155 in Appendix 3).

Table 28 reports how the estimated QALY effects based on measures of QALY burden for each PBC can be decomposed into that part associated with life-year effects adjusted for quality and that part associated with ‘pure’ QoL effects. These results are similar to those reported in Table 26 which were based on QALY ratios for the 11 PBCs with an estimated outcome elasticity. Those PBCs for which mortality is the major concern have a much greater share of total QALY effects associated with avoidance of premature death (e.g. PBC 2 and PBC 10) compared with those where QoL is the major concern (e.g. PBC 7). The differences tend to favour QALYs gained though avoidance of disability, which reflects the underestimation of the effects on ‘pure’ QoL when using QALY ratios based on estimates of YLL from GBD (see the discussion above). aaa The QALY to YLL ratios that are implied by this analysis are reported in Appendix 3, Table 152. As expected the implied QALY ratio across all 11 PBCs with outcome elasticities is higher (3.05aab) than reported in Using ratios of quality-adjusted life-years to years of life lost because the previous bias against QoL effects by using QALY ratios based on unadjusted GBD information has been removed.

In Using ratios of quality-adjusted life-years to years of life lost the ratios of QALYs to YLL due to disease in those PBCs where outcome elasticities could not be estimated could not be used to inform estimates of the threshold or indicate how any total health effects in these other PBCs are likely to be ‘shared’ between life-year effects adjusted for quality and that part associated with ‘pure’ QoL effects (see Table 26). By applying the observed proportionate effects of changes in expenditure to measures of QALY burden of disease in these other PBCs, the likely share of any effects on QALYs between avoidance of premature mortality and avoidance of disability more closely reflect the nature of these types of diseases (see Table 28). As expected, a much greater proportion of QALY effects are associated with QoL during the disease compared with the 11 PBCs where mortality-based outcome elasticities could be estimated. The share of effects in particular PBCs are also much more credible. For example, in PBC 5 (mental health disorders) the overwhelming share of QALY effects are associated with QoL itself and for others, such as PBC 12 (dental problems), PBC 9 (problems of hearing) and PBC 8 (problems of vision); almost all effects are associated with QoL rather than mortality and life-years. For this, and the other reasons discussed above, the analysis based on measures of QALY burden are regarded as the best estimate of a cost per QALY ratio that reflects a more complete picture of the likely health effects of changes in overall expenditure.

Summary of the cost per quality-adjusted life-year threshold

The results of the three sequential steps of analysis described in this chapter are summarised in Table 29. In From mortality to life-years we explored ways in which the estimated effects on mortality from the econometrics work in Chapter 3 might be better translated in to life-year effects by overcoming some of the limitations of mortality data available at PCT level and taking account of counterfactual deaths. The results of this analysis were reported in Table 28 and are repeated in column (1) of Table 29. aac In Adjusting life-years for quality-of-life we considered how the estimated life-year effects might be adjusted for the QoL in which they are likely to be lived, taking account of the gender and the age at which life-years are gained or lost (see Table 21). The results of this analysis are repeated in column (2) of Table 29. Finally in Including quality-of-life effects during disease we explored ways to also take account of the likely effects of changes in expenditure on QoL during disease as well as the effects associated with mortality and life-years [see Table 29, column (3)]. These estimates provide our central estimate of a cost per QALY threshold, because they make best use of available information while the assumptions required, which on balance are likely to be conservative with respect to health effects, appear more reasonable than the other alternatives available. aad

The estimate of £5128 per QALY [see Table 29, line (2)] is restricted to the effects of changes in expenditure in the 11 PBCs where outcome elasticities can be estimated. Although this might be regarded as more secure, these PBCs only account for a proportion of a change in overall expenditure [approximately 38%, see column (6) in Table 108 in Appendix 3]. The threshold of £10,187 uses the estimated proportionate effects of expenditure on the QALY burden of disease in these PBCs as a surrogate for proportionate effects in the others (i.e. assuming that the effects that can be observed will be similar to those that cannot). As discussed in Using estimates of the quality-adjusted life-years burden of disease there are reasons to suspect that this may underestimate health effects in these PBCs which have most influence on the overall threshold. As in previous sections, no health effects are assigned to PBC 23 (GMS) on the basis that any health effects of this expenditure would be recorded in the other PBCs. aae Therefore, the best or central estimate of cost per QALY threshold is £10,187 [see Table 29, column (3), line (3)]. However, this estimate reflects changes in undiscounted QALYs associated with changes in expenditure. Although all the health effects of a change in expenditure are restricted to 1 year (so no discounting is necessary) some of the QALY effects of a change in mortality in that year will occur in future years, so in principle should be discounted. However, discounting these life-year effects, even at the higher rate of 3.5% recommended by NICE, only increases the cost per QALY threshold to £10,333 (see Table 154 in Appendix 3 for discounted values).

As in previous sections of this chapter, the upper and lower bounds for the cost per QALY thresholds in Table 29, column (3) are based on making the necessary assumptions about duration of health effects and how long a death might be averted optimistic (providing the lower bound for the threshold) or conservative (providing an upper bound for the threshold). The lower bound [see Table 29, lines (4)–(6)] is based on assuming that health effects are not restricted to 1 year but apply to the whole of the remaining disease duration of the population at risk in PBCs during 1 year. Although this combines optimistic assumptions, it is possible that at least some part of a change in expenditure may prevent disease so will have an impact on populations that are incident to PBCs in the future. Such effects are not captured in any of the estimates presented in this chapter so all are conservative with respect to this type of health effects of expenditure. The upper bound [see Table 29, lines (7)–(9)] is based on the combination of assuming that health effects are restricted to 1 year for the population currently at risk and that any death averted is only averted for 2 years (see Summary of cost per life-year estimates).