Elective hospital admissions: secondary data analysis and modelling with an emphasis on policies to moderate growth

Article history

The research reported in this issue of the journal was funded by the HS&DR programme or one of its preceding programmes as project number 11/1022/19. The contractual start date was in January 2013. The final report began editorial review in January 2016 and was accepted for publication in August 2016. The authors have been wholly responsible for all data collection, analysis and interpretation, and for writing up their work. The HS&DR editors and production house have tried to ensure the accuracy of the authors’ report and would like to thank the reviewers for their constructive comments on the final report document. However, they do not accept liability for damages or losses arising from material published in this report.

Declared competing interests of authors

Maria Jose Aragon reports grants from the National Institute for Health Research (NIHR) during the conduct of the study, grants from the Department of Health, grants from the Department of Health via Economics of Social and Health Care Research Unit (ESCHCRU), grants from the Wellcome Trust via the Centre for Chronic Diseases and Disorders (C2D2) and grants from NHS England outside the submitted work. Martin Chalkley reports grants from NIHR during the conduct of the study, grants from the Department of Health, grants from ESCHCRU, grants from C2D2 and grants from NHS England outside the submitted work.

Copyright statement

© Queen’s Printer and Controller of HMSO 2017. This work was produced by Chalkley et al. under the terms of a commissioning contract issued by the Secretary of State for Health. This issue may be freely reproduced for the purposes of private research and study and extracts (or indeed, the full report) may be included in professional journals provided that suitable acknowledgement is made and the reproduction is not associated with any form of advertising. Applications for commercial reproduction should be addressed to: NIHR Journals Library, National Institute for Health Research, Evaluation, Trials and Studies Coordinating Centre, Alpha House, University of Southampton Science Park, Southampton SO16 7NS, UK.

Summary

Background

Hospital and Community Health Services (HCHS) constitute around 65% of NHS expenditure in England (2010 data), admitted patient care (inpatient and day cases) represents approximately one-third of total HCHS expenditure and elective care accounts for almost half of the admitted care expenditure. The starting point for our investigation is the observed pattern of growth in elective care in England over the past 15 years. There are a number of ways of measuring the extent of elective treatments, and Figure 1 shows one such measure – the number of continuous inpatient stays (CIPSs) – as well as the population-adjusted change in activity from 2000 to 2009. In Figure 1, as elsewhere in this report, we scale the measure of CIPSs per 1000 population. We consider the definition of different measures of activity in Data and empirical methods.

FIGURE 1.

Continuous inpatient stays per 1000 population in England.

Figure 1 appears to show a sharp increase in activity commencing in the middle of the decade and continuing until the end. The middle years of the decade correspond to a period of substantial reform in the organisation and, in particular, the financing of hospital care in England; one potential inference from the figure is that it is system reform that led to the increase in activity.

The main component of the system reform introduced in England was an activity-based payment system, known as PbR. 12 Such a payment system has been argued to provide incentives to reduce costs, by, for example, reducing length of stay (LOS), and to increase in the levels of activity, therefore reducing waiting times. 13 How strongly these incentives will affect the outcomes of hospitals will depend on how much of their revenue relates directly to their activity levels; in England it accounts for around 60% of hospitals’ revenue. 14 In terms of elective activity, there is evidence that PbR has increased activity levels and the proportion of activity performed as day cases and reduced LOS. 2,12 However, there have been concerns regarding the lack of efficiency-improving incentives in block contracts not linked to activity or its quality and the proportion of hospitals’ revenue being negotiated locally rather than by using the PbR system. 15 It has been argued that, given the incentives to increase activity and its efficiency, PbR is more appropriate for elective activity. 16 For more results on the effects of PbR, see reports by The King’s Fund16 and the Nuffield Trust. 17 PbR constitutes what is termed a form of prospective payment system, whereby the price of treatment is determined before treatment actually occurs and according to the patient’s medical condition or treatment requirements. The grouping of patients under PbR is by Healthcare Resource Groups (HRGs), which share a number of features with what are termed Diagnosis Related Groups (DRGs) in other jurisdictions; this terminology was established by the reform of the US Medicare Program in 1983. 18 Systems based on DRGs have proliferated around the world, especially in Europe. 19 Their introduction has been accompanied by extensive investigation of their impact, which has usually focused on hospital efficiency and quality of care,20 but there has been little focus in these other contexts on the impact on activity, which is our focus. The findings we present in this chapter form a contribution to the evidence concerning the impact of PbR (a ‘DRG-like’ prospective payment system) on hospital performance, but we do not attempt to separate out this effect from the general package of coinciding reforms that occurred with its introduction.

Other changes taking place in the NHS in England during the 2000s were patient choice21,22 and increases in the budget. 23 The key focus of investigations of the impact of patient choice has been on the impact for quality of care, usually proxied by hospital mortality (see Cooper et al. 24 and the references therein) and, as far as we can determine, there has been no investigation of the impact of patient choice on observed hospital activity. Previous analysis of the relationship between NHS expenditure and activity has focused on describing trends,25 whereas we utilise regression analysis to decompose those trends into their component drivers.

The current financial climate and pressures on public funding indicate a need to manage growth in activity going forwards and this task is one of the key responsibilities of the recently established purchasers of health care, namely CCGs. In this context, knowing more about what drives elective activity, whether or not the apparent trend in activity can be explained by factors that are controllable or endemic in the system and whether or not any impetus to growth given by system reform is likely to continue are questions of concern to CCGs.

Given that a CCG has responsibility for its constituent population, it will be especially concerned with the local position, whereas Figure 1 illustrates the overall picture for England. Hence, in addition to the broad questions outlined in this section, we anticipate that CCGs will wish to understand growth in elective care in their own regional context.

Therefore, the purpose of our investigation is to inform some answers to these questions with a view to offering CCGs a better understanding of the environment in which they operate and the constraints that they are likely to face in seeking to manage elective activity.

Modelling growth in elective activity: the role of another jurisdiction

The challenge to isolating the effect of any one influence on the growth of elective activity requires the identification of appropriate controls against which to assess the impact of a given treatment factor. Thus, for example, if we wish to isolate the effect of an ageing population it is most helpful to observe what happens to elective care in an area where there is population ageing and to compare that with (ideally) an otherwise identical area where there is no such ageing. Such ideal comparisons do not exist in practice and, as is standard, our approach is to look for variations across locations and across time that will reflect many influences and to isolate the effect of any one influence using multiple regression methods.

A key focus of our study is to understand whether or not (and if so, how) system reform in England might have impacted on growth in elective care, because, as noted above, such reform sets an important part of the context in which CCGs operate. Given that reform was enacted across all of England, we are limited in regard to variation that can be used to isolate its effect. There is some useful variation in the timing of the introduction of reform across different HRGs, but this is limited. We therefore make use of another jurisdiction of the UK, namely Scotland, as a comparator and control, because the NHS in Scotland is similar to, but independent of, the NHS in England and policy in regard to it is a reserved matter for the Scottish Government. In our analysis we control for variations in the number and composition of the populations of Scotland and England but note that they have evolved in a similar manner over the period under investigation. The respective populations of England and Scotland are shown in Figure 2 and the age composition of these populations is depicted in Figure 3 for Scotland and Figure 4 for England.

FIGURE 2.

Population 1998–2011 (thousands) in England and Scotland.

FIGURE 3.

Age composition in Scotland 1998–2011.

FIGURE 4.

Age composition in England 1998–2011.

Importantly for our study, Scotland and England have had a similar expansion in the funding of their health services, as illustrated in Figure 5. However, the NHS in Scotland did not undergo the reorganisation of financing of hospital care that was undertaken in England.

FIGURE 5.

Expenditure per capita in 2011/12 prices.

Figure 6 illustrates how having such a comparator may be potentially useful in assessing the growth in elective activity. The measure of activity is once again CIPSs per 1000 population, and the scale has been extended to cover the period 1997 to 2011. As can be seen, compared with what appears to be substantial growth in elective activity in England, activity in Scotland seems flat, which particularly applies to the period 2002–5, the period of most active system reform in England. Our empirical strategy is thus to make use of Scotland as a comparator for assessing the impact of changes that have taken place in England.

FIGURE 6.

Continuous inpatient stays per 1000 population in England and Scotland.

The information presented in Figure 6 can also usefully be summarised using regression methods and, because these are the methods that we use the most, a regression corresponding to Figure 6 is helpful. This is provided in Table 1.

TABLE 1 - Elective CIPSs per 1000 population dependent variable: log(CISt)

**, p < 0.05; ***, p < 0.01; log(CIS_t), natural logarithm of continuous inpatient stays per 1000 population in year t.

	Regression model
	(1)	(2)
Time trend, t (standard error)	0.007** (0.002)	0.005** (0.002)
t × England (standard error)	0.022*** (0.002)	0.016*** (0.004)
System reform	No	Yes
n	30	30
R 2	0.8593	0.8826

Table 1 reports the results of a regression of the natural logarithm of elective CIPSs per 1000 population against a linear time trend and that trend interacted with a dummy variable for England. It is included here simply to aid subsequent interpretation of regression estimates. Focusing first on regression model 1, the regression indicates that the annual growth in elective activity in Scotland is 0.7%, because the logarithm scale enables the interpretation of regression estimates in percentage terms. The annual growth rate for England is 2.2% higher than that (i.e. it is 2.9% per annum). Hence, the regression confirms and quantifies what is shown in Figure 6, namely that growth in elective activity has been substantially lower in Scotland than in England. Regression model 2 adds a variable capturing the occurrence of system reform in England to the regression. This potentially addresses the question of whether or not the differential trend that we see is coincident with the implementation of system reform. The details of our methods and the extension of this approach to account for numerous potential drivers of elective care are set out in Data and empirical methods and Results below, but here we simply note that the inclusion of system reform does not appear to change the underlying view that elective care grows faster in England than in Scotland. In regression model 2 we see that the growth rate in elective care in Scotland is estimated to be 0.5% and in England is 2.1%. We can tentatively conclude that system reform is important – its inclusion changes our estimates – but is a long way from being the complete story.

Different perspectives on activity

The current system of financing hospital activity in accordance with the National Tariff Payment System (NTPS), which sets a fixed price for each patient receiving an elective treatment within a defined HRG, naturally suggests measuring activity in terms of the number of treatments. The expenditure of a CCG depends on the number of inpatient spells that it has to fund and spells are closely related to CIPSs. Thus, from a CCG perspective, the measure of activity that must be managed is CIPS. More details of the definitions of spells and CIPSs and the reasons why we use the latter are set out in Data and empirical methods.

Although the number of CIPSs (weighted by the HRGs in which they occur) is a good proxy for the financial resources devoted to elective care by CCGs, from a broader societal perspective it fails to account for the real resources that are consumed in that activity. Over time it can be expected that the price per CIPS set under the national tariff will adjust to reflect changes in resource use within CIPSs and, thus, it is useful to have a perspective on at least some aspects of changing resource use in CIPSs.

The total resources devoted to elective activity comprise the combination of hospital capital, staffing and medical equipment and supplies. One possible proxy for these resources is the total number of hospital days that are devoted to elective care. Figure 7 shows the evolution of bed-days in England and Scotland from 1997 to 2011. There is a substantial decline in total elective bed-days in England, despite the rapid increase in elective CIPSs, but a less marked reduction in Scotland. Thus, even at an aggregate level, care must be taken before concluding that elective activity has increased substantially in England relative to Scotland; on one measure it has but on another it has not. Figure 7 can be further understood by noting that the average stay in hospital associated with an elective admission – the LOS – has declined rapidly, more so in England than in Scotland. Figure 8 shows the LOS in both countries over time.

FIGURE 7.

Total bed-days per 1000 population England and Scotland.

FIGURE 8.

Average LOS in England and Scotland.

The trend reduction in LOS has previously been noted, although as far as we can establish it has not been documented over such a long period as depicted in Figure 8. There have been previous studies concerned with the question of whether or not a shorter-term reduction can be attributed to system reform. 2 The relationship between growth in CIPSs, falling LOS and system reform is therefore of direct relevance to our main topic of enquiry.

A regression model again confirms and quantifies the changes depicted in Figures 7 and 8. The regression estimates presented in Table 2 follow the same pattern as those presented in Table 1 but focus on the LOS measure. Omitting the control for system reform, the rate of decline is estimated to be 2.8% per year in Scotland and 4.5% per year in England. Including system reform, the rates are 2.7% and 3.8%, respectively.

TABLE 2 - Elective LOS dependent variable: log(AvBDt)

***, p < 0.01; AvBDt, average bed-days in year t.

	Regression model
	(1)	(2)
Time trend, t (standard error)	–0.028*** (0.002)	–0.027*** (0.002)
t × England (standard error)	–0.017*** (0.002)	–0.011*** (0.003)
System reform	No	Yes
n	30	30
R 2	0.9444	0.9537

Data and empirical methods

Results

Regression estimates of our generic model set out in Equation 1 are central to our results. As noted previously, we consider two measures of activity. Tables 5 and 6 show the regression estimates for the number of CIPSs and average LOS, respectively.

In the first columns of Tables 5 and 6 the regression includes the common time trend, the differential time trend for England, controls for population characteristics that are expected to influence demand for elective care and the fixed effects for PCT clusters and HRGs. In the second columns the regressions additionally include a variable capturing system reform. The full details of the construction and rationale for this variable are set out in Modelling system reform. The results reported in this section adopt the most straightforward formulation, with results for alternative formulations being set out with commentary below (see Checks on the sensitivity of main results). All of the results we report here are robust to the choice of formulation.

There are a large number of coefficient estimates underlying the tables, most of which, for reasons of parsimony, we do not report. Thus, for example, there are 17 age category dummy variables, four deprivation category variables, 49 dummy variables associated with PCT clusters, 49 variables arising from allowing for cluster-specific growth effects and > 1200 implicit dummy variables capturing case-mix variation using HRGs. Under the computation method, which is described in detail in Empirical methods, these last estimates are not directly available but could be recovered if desired. We note simply that the majority of the dummy variables that are not reported here are significant at the 1% level and where there may be value or interest in establishing and reporting their values, we are able to do so. As evidenced by the F-test results reported in Tables 5 and 6 the different levels of fixed effects are jointly and separately significant at the 1% level. Overall, therefore, our model as specified in Equation 1 incorporates variables that are important in accounting for variation in elective activity.

Examining the effects across primary care trust clusters

Our results have hitherto been reported in terms of a comparison between Scotland and England, both for all HRGs and for subsets of HRGs. Our regressions allow for regional effects, which are captured as dummy variables defined according for each PCT cluster. Hence, PCT cluster might hitherto be interpreted as a potential confounding factor, which, if omitted, might bias the estimated growth in activity or decline in bed-days.

The empirical methodology that we have adopted allows us to focus greater attention on the potential disparities that arise across PCT clusters and we next turn to reporting on this level of analysis. The regressions reported above assume that a PCT cluster can have a higher or lower level of activity or average number of bed-days. This captures the idea that there are differences across different areas of the country that we do not observe but that nevertheless impact on elective activity. The PCT cluster-level effect is estimated alongside other parameters in the regression, and it would be possible to report this as an indicator of whether or not a particular PCT cluster starts the period of analysis with a higher- or lower-than-average level of activity.

A generalisation of this approach is to allow a PCT cluster to affect not only the level but also the growth rate of activity. We implement this approach empirically by adding variables that are formed by interacting the PCT cluster in which a treatment occurs with a time trend variable. To use this approach we must omit England as a separate time trend, because the complete set of PCT clusters span England. Thus, we replace the single England time-trend interaction with 49 PCT cluster time-trend interactions. Each estimated coefficient represents the differential growth in elective activity in that PCT cluster, over and above the benchmark trend growth which we express as the average growth across PCT clusters. This average is 5.6%, which is different from the 4.4% reported across England, because the average is taken across PCT clusters and some smaller PCT clusters have high growth rates. The same method can be applied to LOS, thus giving a PCT cluster-specific rate of decline for hospital LOSs, and can be applied to subgroups of HRGs, although we do not present subgroup results in this report.

The proliferation of regression coefficient estimates places constraints on reporting. Although we can estimate a growth rate for each of two outcomes (CIPSs and average bed-days) for each of 49 PCT clusters, we present results only in terms of ranking or graphically.

The results at PCT cluster level for CIPSs are reported in Figure 9 as a simple ranking illustrating the mean rate of growth for each PCT cluster. Recall that zero constitutes the average growth of CIPSs across PCT clusters. The lowest ranked PCT cluster, closest to the origin in the figure has a growth rate in CIPSs of approximately 8% below the average, which implies that it experienced decline in elective activity over the period. This one PCT cluster (and thus its associated CCGs) is an outlier and may therefore need further investigation. As we move to the right in the diagram we observe PCT clusters with higher growth rates. Towards the right hand limit of the figure, there are PCT Clusters with growth more than 5% above the average and hence they experience double-digit growth rates. CCGs in these clusters therefore are operating in local environments where growth in CIPSs is doubling in less than 8 years. This growth is per capita and after adjustment for changes in population age structure or deprivation. The figure immediately suggests that some CCGs face a very substantially greater challenge than others in regard to managing growth in elective activity. Referring back to the discussion above (see Growth in continuous inpatient stays) growth is not readily explained in terms of expansion in resources and so it is not clear that it is readily controlled by budgetary constraints.

FIGURE 9.

Deviation from average growth (%) in CIPSs per 1000 population.

For reasons of legibility Figure 9 is presented without labels for the names of PCT clusters. These are set out in the order in which the clusters appear on the horizontal axis, in Box 1.

BOX 1 - Primary Care Trust cluster: continuous inpatient stay growth ranking

Growth in elective CIPSs per 1000 population (in increasing order)

Cumbria.

Somerset.

North Yorkshire and York.

Cheshire.

South of Tyne and Wear.

Suffolk.

North of Tyne.

Lincolnshire.

West Mercia.

Norfolk and Waveney.

Derbyshire.

County Durham and Darlington.

Calderdale, Kirklees and Wakefield.

Leicestershire.

Gloucestershire, Swindon.

Cambridgeshire and Peterborough.

Merseyside.

Greater Manchester.

Cornwall and Isles of Scilly.

Bedfordshire and Luton.

Humber.

South Yorkshire and Bassetlaw.

Pan Lancashire.

Staffordshire.

Southampton, Hampshire, Isle of Wight and Portsmouth.

North Essex.

Bristol, North Somerset, South Gloucestershire.

Devon, Plymouth, Torbay.

Bath and North East Somerset, Wiltshire.

Kent and Medway.

South Essex.

Arden.

Surrey.

Hertfordshire.

Tees.

Bournemouth and Poole, Dorset.

Sussex.

Airedale, Bradford and Leeds.

Northamptonshire & Milton Keynes.

Black Country.

Nottinghamshire.

North Central London.

North West London.

Birmingham and Solihull.

Oxfordshire and Buckinghamshire.

South West London.

Berkshire West and Berkshire East.

South East London.

North East London and The City.

The disparities in growth depicted in Figure 9 may be a consequence of initial disparities in elective activity, whereby PCT clusters that have very low initial activity may have grown fast in a process of catching up. Our model permits us to examine this hypothesis by recovering the PCT cluster fixed effects and comparing these with growth rates. This is shown in Figure 10, in which a cluster’s growth rate (on the y-axis) is plotted against its initial (corrected for other variables) level of elective activity. If the catching-up hypothesis is correct, we would expect to see a well-determined negative trend in Figure 10, with high growth associated with a low initial level. Figure 10 also includes a regression trend line between the two variables that does exhibit negative slope. However, the relationship is weak with many clusters located well away from the trend and the trend line explaining < 1% of the variation depicted.

FIGURE 10.

Continuous inpatient stays per 1000 population.

We therefore conclude that there are substantial differences between PCT clusters in respect of growth of elective CIPSs and that these differences cannot be accounted for by initial discrepancies in activity.

The results set out in Figures 11 and 12 and in the associated labels (Box 2) show the same variation across PCT clusters with regard to average LOS decline. A low ranking is associated with large negative ‘growth’ rates of average LOS and the highest rankings correspond to the small number of PCT clusters to have experienced increasing average LOS. We again see that variations in the decline in LOSs have little correspondence to the initial circumstances of the PCT cluster.

FIGURE 11.

Deviation from average growth (%) in average LOSs.

FIGURE 12.

Average LOSs.

BOX 2 - Primary Care Trust cluster: length of stay growth ranking

Growth in lengths of stay (in increasing order)

Black Country.

Suffolk.

Norfolk and Waveney.

Arden.

North West London.

Cornwall and Isles of Scilly.

Bristol, North Somerset, South Gloucestershire.

Southampton, Hampshire, Isle of Wight and Portsmouth.

South Yorkshire and Bassetlaw.

Pan Lancashire.

North Essex.

Kent and Medway.

Bath and North East Somerset, Wiltshire.

Surrey.

South West London.

North Central London.

Oxfordshire and Buckinghamshire.

Cambridgeshire and Peterborough.

Birmingham and Solihull.

Somerset.

Cheshire.

Devon, Plymouth, Torbay.

Airedale, Bradford and Leeds.

Gloucestershire, Swindon.

Berkshire West and Berkshire East.

Greater Manchester.

Derbyshire.

Lincolnshire.

West Mercia.

South of Tyne and Wear.

South Essex.

Merseyside.

Bournemouth and Poole, Dorset.

Bedfordshire and Luton.

Cumbria.

Calderdale, Kirklees and Wakefield.

Nottinghamshire.

North Yorkshire and York.

Hertfordshire.

Leicestershire.

North of Tyne.

North East London and The City.

County Durham and Darlington.

Sussex.

Humber.

Staffordshire.

Northamptonshire and Milton Keynes.

Tees.

South East London.

Overall, these results give a different perspective on the organisation of hospital services in an area. CCGs in PCT clusters in the higher rankings operate in areas where bed capacity is more likely to be a constraint on growing activity.

Information for commissioning

In this section, we indicate a method by which the model we have developed might be used to provide commissioners with specific information regarding the growth of elective care in their area, defined in terms of the PCT cluster to which they belong. This information is intended to assist the process of planning services and alerting commissioners to the challenges they face with regard to managing elective activity.

Our model contains a number of dimensions by which growth may be decomposed. It is neither possible nor useful to consider every possible permutation of PCT cluster, time frame, HRG grouping and elective activity. However, the model can be deployed in responsive mode and provide relevant estimates for any given desired configuration in these dimensions.

One way of conceptualising this process is through the generation of commissioner-specific reports. We indicate here two possibilities, each of which is illustrated with a report template.

Introduction

It has long been known that changes to the way hospital payments are made can influence the levels of activity and quality of treatment. Coulam and Gaumer40 and Hodgkin and McGuire41 consider the consequences of the prospective payment system introduced in US medical care during the 1980s and find that it led to generally shorter lengths of stay and fewer admissions, whereas McClellan42 explains how quality of health care can be affected by payment systems. There is also extensive research examining how prospective payment schemes affect LOS by, among others, Farrar et al. ,13 Theurl and Winner43 and Newhouse and Byrne. 44 In addition, Appleby et al. 16 consider the effect of payment schemes on quality in English hospitals, and it is one factor analysed in Chapter 2 of this research as a possible driver for growth of elective care.

This report focuses on the effects of payment reform, and the introduction of tariffs in particular, on the variation of care, specifically in terms of hospital admission rates and LOSs across PCTs. We first examine whether or not there were reductions in the variation of NHS hospital bed-use, rather than its level, during the period of system reform (2003–9). Second, we explore the hypothesis that the introduction of a prospective payment tariff in England (PbR) may have reduced certain variation in patient treatment and resource use across English hospitals from 2002/3 to 2008/9, thereby reducing the ‘post-code lottery’ in the provision of treatments.

Many policies are assumed in statistical analysis to bear evenly on the affected hospitals, patients and clinicians, increasing volume or, perhaps, quality uniformly across the relevant population. However, some policies may increase or reduce the variation in the affected variables. This has been of particular interest for patient selection and quality of patient care. For example, Shen45 tests the hypothesis that performance-based contracting may encourage providers to treat patients with less severe symptoms at the expense of more severely affected patients. Dranove et al. 46 showed that prospective payment might encourage ‘cream-skimming’ of more profitable patients. These unintended effects of policy and more familiar unidentified ‘post-code lottery’ effects are examples of unwarranted variation.

As Skinner47 explains, some variation in health-care use can justifiably be explained as a consequence of different ‘prices, illness or income’. However, the residual variation not described by these factors is undesirable and suggests that some providers may be operating inefficiently. As such, policies that reduce the inequality of patient treatments and outcomes may increase overall efficiency and are recognised to be of potentially considerable social value. For example, a policy of regulating minimum quality thresholds is likely to have impacted most on the quality of lesser performing hospitals and their patient outcomes, and more so than on those that were performing well before policy innovation. In this case, the increase in patient value is a direct consequence of policy to reduce variance.

During the 6 years from 2002/3 to 2008/9, major changes to the NHS were introduced. In addition to considerable technological progress, changing the patterns of health-care demand across diseases and patient groups was actively encouraged. Reform, such as the introduction of a prospective payment model and waiting time targets, introduced a common set of economic incentives for managers that rewarded improved performance in a way that was likely to increase quality of care levels, as shown by Miraldo et al. 48 and Propper et al. 49

Prior to the tariff, health-care providers in England were reimbursed on a block grant basis, which gave physicians and managers considerable individual influence over the allocation of hospital resources, without sanction for low rates of treatment. The introduction of the tariff created different incentives for medical professionals, which may have altered behaviour towards that required to sustain hospital revenues and thereby jobs. The tariff may also have made internal political hospital problems of reallocating resources more tractable. Tariff policies are likely to have placed greater financial pressure on hospitals that offered long LOSs and low admission rates, and to have contributed to a narrowing of variation in activity levels and LOSs across England as scope for locally developed policies reduced.

This report examines whether or not variation has decreased by studying the system-wide pattern of hospital bed use, by investigating whether or not the frequency and lengths of admissions became more similar across PCTs, and, in particular, investigating whether or not the introduction of a tariff-based payment system may have shaped these changes.

In order to separate the consequences of tariff reform from other developments, this work exploits the fact that reform was not introduced uniformly across specialties. Whereas a reduction in the variation of LOS between 2002/3 and 2008/9 is found in emergency and, to a lesser extent, elective care, this pattern was not reproduced in mental health, a specialty that has yet to see a full introduction of the tariff. The next section studies variation across PCTs, with particular emphasis on LOS, which is likely to have been most affected by the introduction of the tariff. The tariff and variation in hospital bed-use applies similar methods to mental health data. An alternative explanation: the changing geography of resource allocation and bed-use considers the impact of funding patterns on variation, in particular whether the changing model of resource allocation to local areas, to capture need, provides an alternative explanation to system reform for the reductions in heterogeneity between 2002/3 and 2008/9. The relationship between efficiency gain and initial mean length of stay explores how LOSs have converged across the distribution, showing that areas with higher LOSs in 2002/3 have experienced greater decreases in LOS.

The changes in variation of hospital bed-use across primary care trusts, 2002/3–2008/9

This section discusses the evidence concerning particular bed use across English PCTs and how far patients’ usage became more similar.

The data come from HES, a database of ‘all admissions, appointments and attendances at NHS hospitals in England’. 50 Two financial years (2002/3 and 2008/9) are considered to see how resource use and variation across space altered during this time. These years are selected to capture the time period across which the tariff was introduced in the NHS. The measures considered are total bed-days, total admissions (finished consultant episodes) and average LOS, which is calculated by dividing the total bed-days by the total number of admissions.

Episodes are aggregated by the PCT of patient residence, so that there are 152 observations for each measure for each year. This is not the same as investigating treatment at hospitals. If hospitals were the unit of investigation, then this research would be a direct study of policy on hospital activity, but, by considering these alternative measures, it becomes more about how geographic groups of patients are affected by policy and its impact on hospital incentives.

The distributions are presented graphically as Kernel density plots in Figure 13.

FIGURE 13.

The distributions of emergency and elective admissions and LOS. (a) Emergency bed-days per 1000 population; (b) emergency admissions per 1000 population; (c) emergency LOS; (d) elective bed-days per 1000 population; (e) elective admissions per 1000 population; and (f) elective LOS.

The distribution of emergency bed-days has shifted slightly to the left between these 2 years and become more concentrated. This shows that average bed-days per PCT has fallen slightly, with a significant reduction in the variance across PCTs.

The distribution of admissions has seen different changes to the bed-days distribution, with the 2008/9 plot flatter than the 2002/3 plot and shifted to the right. This indicates an increase in the average rate of admissions and a higher variation in the admission rate across English PCTs.

Owing to the fall in emergency bed-days and an increase in the number of admissions, the average LOS has decreased significantly. Variation has also decreased markedly.

Elective data in Figure 13d–f show that bed-days have declined and that there is less variation. There has been a marginal decrease in total admissions and the variance has also fallen, but these changes are small and the distribution has a similar shape in both years. As a consequence of the changes in bed-days and admissions, the average LOS and its variance decreased between 2002/3 and 2008/9. Tables 16–18 show some descriptive statistics.

In 2002/3, 30.1 million emergency bed-days were used in England to support approximately 3.8 million admissions (excluding day cases), indicating an average LOS of 7.8 days. Elective treatment was on a smaller scale, with 8.5 million bed-days and 1.8 million admissions at an average of 4.8 bed-days. Over the following 6 years, emergency bed-days fell by 6%, even though admissions increased by 30%. Consequently, the mean length of emergency stay fell by 28% to 5.6 days. There were 152 PCTs in England in 2008/9 but early analysis suggested that North Staffordshire and Stoke on Trent were outliers and were omitted from work carried out on the elective care data set. Elective admissions declined by 1.7%. There was a 20% reduction in bed-days and a similar decrease in average LOS, from 4.8 days to 3.9 days. This was at a time when a significant amount of hospital beds closed, with the number of available beds falling from 183,826 in 2002/3 to 160,254 in 2008/9.

Table 18 shows standard deviations (SDs) of rates per 1000 population for the PCTs in England. The SD for emergency bed-days fell by 4.7%. This was due to a large reduction in average LOS variation, which fell by 45%, whereas admission variation increased. Variation in elective admissions and average LOS remained fairly constant, but total bed-day variation decreased by 25%.

It is worth noting the increasing proportion of patients who are treated as day cases rather than overnight admissions. Methods such as key-hole surgery allow treatment to be less invasive and patients are regularly sent home after procedures that might have previously required significant recovery time in a hospital bed. The data used in this research include all patients who were expected to have an overnight stay and do not include day cases. By omitting day cases – a growing proportion of hospital treatment for which there is, by definition, minimal variation in treatment length – the results presented here are likely to underestimate the reduction in variation that has occurred.

These results suggest that variation in hospital resource use has declined, mainly as a consequence of shorter and less varied LOSs. Admissions have either remained fairly constant (in elective care) or increased (in emergency care), but patients are being discharged much more quickly on average and this appears to be the main reason for decreased variation. This fits with the patterns predicted above (see Introduction) in that clinicians appear to have become more similar in their behaviour.

The next section attempts to determine whether or not this decline in variation was due to the change from a block grant funding system to a prospective payment system.

The tariff and variation in hospital bed-use

Beginning in 2003/4 with 15 HRGs and a select group of providers, the block grant system of funding was replaced by a prospective payment tariff model that reimburses hospitals for treatments and procedures on a fixed unit-price basis. 51 This process continued until 2008/9, by which time all acute trusts were paid according to a tariff measure for 546 HRGs in elective, non-elective accident and emergency (A&E) and outpatient care, with mental health one of the few specialties still paid for with block grants. In this new framework there are strong incentives for hospitals to operate more efficiently to ensure that costs do not exceed tariff revenues. One way of achieving this is to reduce the LOS for patients in order both to reduce costs and to release capacity that might support additional admissions. Hospitals providing activity at costs above tariff have been increasingly pressed to reduce costs in order to avoid deficits. Given that bed use is a major source of cost, more homogeneous bed use appears to be a probable consequence of the tariff. In addition, it is predicted that admissions would increase in PCTs that had long lengths of stay before the introduction of the tariff, causing LOS variation to decrease. The impact on bed-days variation is less clear-cut, as this is likely to depend on demographic factors. The process of introducing the tariff had started sooner in elective than in emergency care, so it is expected that this had already had some impact on variation and that the consequences would be felt more strongly in emergency care.

Although there has been significant reform in most parts of the NHS, one of the least affected areas has been mental health, which has retained a block grant system instead of adopting the prospective payment policy. If mental health experienced similar patterns of change to the hospital system as a whole, it would cast doubt on the theory that system reform, at least the introduction of the tariff system, is responsible for reducing variation.

Distributions of emergency and elective resource use for mental health patients are presented in Figure 14.

FIGURE 14.

The distributions of emergency and elective admissions and LOS in mental health. (a) Emergency bed-days per 1000 population; (b) emergency admissions per 1000 population; (c) emergency LOS; (d) elective bed-days per 1000 population; (e) elective admissions per 1000 population; and (f) elective LOS.

A comparison of Figures 13 and 14, which uses all admission data, shows that the mental health sector has experienced different changes to hospitals in general. The biggest changes in all-admission data were a significant decreases in LOS variation for both elective and emergency admissions, but this did not occur in mental health. Looking solely at mental health, the emergency LOS distributions are similar in the 2 years and elective LOSs have become more dispersed across English PCTs. This suggests that incentives and clinician behaviour were affected in different ways by the policies adopted in different specialties of health care and highlights the role of the tariff introduction.

Tables 19–22 show descriptive statistics for mental health, where a mental health admission was defined as one in which the HES ‘tretspef’ variable was one of the following: 700 – Learning Disability; 710 – Adult Mental Illness; 711 – Child and Adolescent Psychiatry; 712 – Forensic Psychiatry; 713 – Psychotherapy; 715 – Old age Psychiatry; 720 – Eating Disorders; 721 – Addiction Services; 722 – Liaison Psychiatry; 723 – Psychiatric Intensive Care; or 724 – Perinatal Psychiatry.

Table 20 shows overall levels of resource use in the mental health sector. Between 2002/3 and 2008/9, emergency admissions fell from 103,977 to 68,737 per annum, and bed-days decreased by 36% to 3.4 million. The average LOS went down from 50.91 to 49.29, but this is not a statistically significant change. These values show that hospital stays have remained similar in terms of lengths for patients who are admitted but that fewer people receive care in this setting.

Figure 14d–f suggests that West Kent PCT is an outlier. This PCT is therefore dropped from the calculations of sample means and SDs for the elective data series. For the remaining PCTs, bed-days and admissions in mental health elective care have experienced declines, but with admissions falling much more than bed-days (33.5% vs. 1.3%, respectively), average LOSs have increased significantly. These changes do not reflect those discovered in the complete data set. Using aggregate values or average rates of sample PCT means (see Table 21) makes no difference to the conclusion that mental health has not changed in the same way as overall health care.

Table 22 presents SDs to show how variation in mental health admissions and bed-days across England has changed. In emergency mental health, bed-day and admissions variation decreased by 31% and 21%, respectively, and there has been no significant change in average LOS variation, with a small increase of only 3.5%. In mental health elective care, bed-day variation has fallen by an insignificant 0.7%, whereas the other two series have seen much larger changes in variation; admission variation fell by nearly 55% to 0.84, whereas average LOSs have become much more variable across PCTs, with the SD increasing by 48% to nearly 79 days.

New central policies are perhaps the most likely cause of changing patterns of variation but there are other potential contributing factors. Distinguishing between the effects of all relevant factors would be a complex task and is beyond the scope of this report. However, the evidence here suggests that policy had a major role.

An alternative explanation: the changing geography of resource allocation and bed-use

The analysis thus far has uncovered significant changes in the geographic distributions of resource use in both elective and emergency care settings, with a tendency for variation to decrease. One explanation for these changes, beyond policy or other endogenous factors, is a changing pattern of need and the associated flows of finance: PCTs could have become more similar if funding levels incentivised them to do so. For example, a narrowing in the distribution of funding owing to a compression in centrally assessed need could reduce the variance of behaviour, as hospital managers will be working with increasingly similar budgets.

To see if this is a likely explanation, this section looks at the relationship between the data series and a variable that captures different levels of need across PCTs, namely a ‘need index’, which is taken from the Department of Health ‘revenue resource limits’ publications. 51 This is a ratio of the actual population to a need-corrected population base and is used in the resource allocation framework.

There has been significant convergence in the need distribution. The average level has declined slightly, but a more noticeable difference between the two series is that the range has decreased and the variance has fallen.

Thus far, the investigation has used raw data on emergency and elective admissions. To accommodate changes in the distribution of need, it is necessary to needs-adjust these raw data. It is then possible to examine whether or not the distribution of data adjusted for need declines over the time period on a scale as great as the decrease noted in the distribution of unadjusted data.

Need has a large role in explaining emergency care, as the regressions in Tables 23 and 24 show. The coefficients all increase in absolute terms over the 6 years, with a statistically significant increase in the effect of need on admission rates and bed-days. This has been tested by running a stacked regression of both years data, with the t-statistic derived from the coefficient on a variable, which is zero if a 2002/3 observation and the need value if a 2008/9 observation.

The relationship between elective care and need is not well defined, with R-squared values never exceeding 0.07. The coefficients on need in equations explaining admissions and bed-day rates are positive, whereas those in the LOS regressions are negative, but none changes significantly between 2002/3 and 2008/9 and it is hard to draw strong conclusions given the lack of explanatory power in these regressions.

Need does not appear to make a significant difference to the patterns discovered earlier. The standard errors in Table 25, which take need into account, are all lower in absolute terms than the SDs for the emergency series, as would be expected, but the SD of bed-days fell by a similar magnitude to the standard error over these years and it was similar to LOS. The standard errors and SDs in the admission series both experienced slight increases.

Given that there is minimal correlation between need and the elective series, it is not surprising that the standard errors in the need-controlled distributions are very close to the SDs calculated using the unadjusted data.

These results appear to confirm the view that a changing pattern of need is not the main driver for the reduction in variation and that structural change, whether as a result of changing policies or other factors that cannot be separately identified, are responsible.

The relationship between efficiency gain and initial mean length of stay

The pattern by which LOSs have converged across England deserves further attention. It is likely that those PCTs starting with longest mean LOS had the greatest scope to gain from reducing LOS without damaging patient care. Regressions were performed to test this prediction. English PCTs were collected into five approximately equal groups, ranked from those with the longest LOS in 2002/3 (quintile 1) to the shortest (quintile 5). Dummy variables for these groups (omitting quintile 5) were then used, in addition to change in need, to predict the change in LOS from 2002/3 to 2008/9.

The regression equation estimated is:

In this specification, the coefficient on the constant term shows the change in average LOS for the 20% of PCTs that had the lowest average levels in 2002/3, with the coefficient on each quintile’s dummy variable showing the difference between the average change in the lowest 20% and the average change in the corresponding quintile. Therefore, a coefficient would be statistically significant when the given quintile has a change in average LOS that is significantly different from that in the 20% of PCTs that had the lowest LOSs in 2002/3. Results for elective and emergency care, in hospitals as a whole and using just mental health data, are presented in Table 26.

In all four data sets, the PCTs with the longest 2002/3 LOSs experienced greater declines than the PCTs with the shortest LOSs. This pattern was most apparent in the all-admission emergency data, which had significantly larger falls in quintiles 1–4 than in quintile 5, with increasingly greater coefficients from quintile 4 to quintile 1. These results suggest that the 20% of PCTs with the longest LOSs in 2002/3 reduced their LOSs by nearly 3 days more, on average, than those with short emergency stays. There was a similar pattern in the all-admission elective data (omitting North Staffordshire and Stoke on Trent, which previous analysis suggested were outliers), although only the top two groups of PCTs had significantly larger decreases than the bottom 20%. The coefficients for each quintile were smaller than the corresponding values from the emergency data. This supports earlier results, which suggested that the introduction of the tariff affected emergency care more than elective care.

As discussed above (see The tariff and variation in hospital bed-use), the tariff was not introduced in mental health care during this time period, so data from this sector can function as a useful control. The 20% of PCTs with the longest average LOSs in 2002/3 for both emergency and elective mental health admissions had significantly greater reductions, as was also found in the all-admissions data. However, the coefficient on the quintile 1 variable is the only quintile variable coefficient that is statistically different from zero. The lack of a consistent pattern in the size of the coefficients across the quintiles – in contrast to the steadily falling series of coefficients in the all admissions regressions – shows that care in a specialty that was excluded from the tariff changed differently to care across the whole system. The fact that data from PCTs in quintile 1 would be most at risk from the effects of ‘outlier’ observations, whether as a result of incorrect recording or patients who remained in hospital for a particularly long time, is a possible non-policy reason to expect these PCTs to tend to the mean over time.

Conclusions

The evidence here shows that variation in emergency admissions LOSs across English PCTs declined sharply between 2002/3 and 2008/9, and variation in LOSs for elective treatment also reduced. The variation in admission rates per thousand heads of population declines only slightly.

An analysis of the ways in which the system has changed suggests that the introduction of the tariff may have been responsible. The various influences on hospital behaviour in this period make it difficult to be certain, but evidence from the mental health sector and adjusting for changing need produces findings that are consistent with those expected if the introduction of the tariff was a major cause.

Nevertheless, it is apparent that PCTs have responded to a changing environment in different ways, with those that had the longest LOSs in 2002/3 reducing their stay lengths by more than others. Health-care managers across England are facing challenges going forwards and when reviewing policies, analysts need to ensure that they are using models that are sufficiently rich to assess the full impact.

Introduction

The total number of elective admissions rose by 50% between 1999/2000 and 2014/15, from 5.50 million in 1999/2000 to 8.26 million 2014/15, an average rise of 2.7% per year. The age-standardised admission rate per 1000 population rose from 112.2 in 1999/2000 to 146.3 in 2014/15, an average increase of 1.8% per year. If admission rates by age and sex had remained constant, the number of elective admissions would have risen by only around 13% rather than by 50% over this 15-year period.

An understanding of the factors driving this substantial rise in elective admissions is important in the context of this study. We need to consider whether or not the pressures responsible for the upwards trend are likely to persist and what they mean for the prospects of controlling future growth in numbers of elective admissions.

In an earlier study, we conducted APC analyses of trends in elective and emergency admissions and bed-days. 52 The aim was to examine how far the trends in hospital admissions can be explained by the effects of the age distribution of the population, together with rising numbers of older people, by cohort effects, that is, effects attributable to differing admission rates of people born in different years (different birth cohorts) and by period effects, that is, admissions effects relating to a specific year, which cannot be explained by either age or cohort effects.

Age–period–cohort analysis is a form of multivariate regression analysis. The dependent variable is emergency admission rates by age and year, using single years and single year age bands (but with a top band of ≥ 85 years). The independent variables are dummies for each age, year of birth and year of admission.

The findings of our APC analysis in respect of total elective admissions (including day cases) are reported in Wittenberg et al. 52 Annex A of that report explains the methodology. We found that a favourable cohort effect (the more recent the cohort, the lower the admission rate given the age) broadly offset the age effect (the older the age group, the higher the admission rate, given the cohort) and the impact of rising numbers of older people. The rate of elective admissions, averaged across all ages, fell for successive cohorts, after controlling for period effects. We also found a substantial positive period effect.

Although the age and cohort effects are likely to reflect demographic and epidemiological change, the period effects capture admissions that could not have been anticipated given evidence concerning the changing age structure and years of birth of the population: these period effects would reflect both changes in supply-side willingness to admit, perhaps as a result of policy or technological innovations, and any increases to admission demand unconnected to age and cohort.

The trends in elective admissions were not due to a substantial shift in the balance between elective and emergency admissions. The number of emergency admissions also rose rapidly, by 2.1% per year between 1997/8 and 2014/15. The age and cohort effects did not differ significantly between the two types of admissions, but the period effect was not quite as high for emergency as for elective admissions. 52

We have investigated trends in elective admissions for four selected groups of procedures to examine variation between procedures in the pattern of APC effects. The purpose of looking at the four groups of procedures is to check whether or not the forces identified for electives as a whole also apply at a disaggregated level. For example, evidence of a weaker upwards age effect, or a stronger negative cohort effect, or a weaker or negative period effect for these procedures may suggest that the growth that we have seen in the overall number of elective admissions may not inevitably continue.

The four sets of procedures that we have examined are:

1. diagnostic procedures, the numbers of which rose somewhat more slowly than total elective admissions

2. hip and knee replacements, the numbers of which rose especially rapidly

3. coronary circulation procedures, the numbers of which fell

4. menorrhagia procedures, the numbers of which also fell.

These four sets of procedures account for 26% of total elective admissions in 1999/2000 and 21% of total elective admissions in 2014/15.

Our analyses use HES for England for each year from 1999/2000 to 2014/15. Although our earlier analysis of total elective admissions related to the 17-year period from 1997/98 to 2014/15, in our analyses of these four groups of procedures we considered the 15-year period 1999/2000 to 2014/15. The reason is concern about the accuracy of the coding of procedures in the early years of the HES data for 1997/98 and 1998/99.

The following sections of this chapter set out the findings of our APC analyses for the four sets of procedures. We present the findings in two ways for each of the APC effects. We first describe how the coefficients on the APC variables in the regression vary with changes in age, period and cohort, respectively. We then present an example of the ratio between the admission rates at different ages holding the cohort and period constant (and similarly for the other two effects). This form of presentation is intended to illustrate the varying effects of APC in a helpful way but it should be considered illustrative. This is mainly because the ratios vary with the values of the two factors that are held constant (e.g. with the cohort and period in the case of the ratio of admission rates between different ages).

Diagnostic procedures admissions

The total number of elective admissions for a range of diagnostic procedures (Box 3) rose from 1.08 million in 1999/2000 to 1.49 million in 2014/15 (Figure 15). This is an increase of 38.5% over the full 15-year period and an average annual increase of 2.2%. This compares with an annual average increase of 2.7% for all elective admissions (including day cases) over this period.

FIGURE 15.

Diagnostic procedures: admissions by broad age band, England, 1999/2000–2014/15.

The numbers of elective admissions for diagnostic procedures rose for all age groups, but, in the case of the 20–34 years, 35–64 years and 65–74 years age groups, the rise followed a decline in the years 1999/2000 to 2004/5.

The overall admission rate for these procedures per 1000 population first fell slightly from 22 in 1999/2000 to 20 in 2005/6 and then rose to 27 in 2014/15, which amounts to a rise of 25% between 1999/2000 and 2014/15. The age-standardised rate rose by 19%, from 22 in 1999/2000 to 26 in 2014/15, an annual average increase of around 1.2% (Figure 16). This compares with an annual average increase, on the same basis, of 1.8% for total elective admissions over this period.

FIGURE 16.

Diagnostic procedures indexed age-standardised admission rates by age band in England, 1999/2000–2014/15.

The age-standardised admission rate for diagnostic procedures rose most rapidly for children and young people aged < 20 years (by 35%) and least rapidly for the 20–34 years age group (by 14%).

If the admission rates for these procedures by age band had remained constant at their 1999/2000 levels, the number of admissions would have reached 1.25 million in 2014/15 rather than 1.49 million, an increase since 1999/2000 of 16% rather than of 39%.

The age effect findings show that, after controlling for cohort and period effects, the elective admission rate for diagnostic procedures rises with age from 9 years to 74 years. It falls with age up to 9 years and from ≥ 74 years. The admission rate at 75 years of age is over twice the rate at 25 years of age and around 35% higher than the rate at 50 years of age. This is on the basis of constant 2010 period effect and 1970 cohort effect (Figure 17).

FIGURE 17.

Diagnostic procedures: predicted admissions by age, for fixed cohort and period.

The cohort effect findings are that each successive cohort from the cohort born since 1915 onward has experienced a lower elective admission rate for diagnostic procedures at a given age than the preceding cohorts, after controlling for period as well as age effects. This is subject to minimal change for cohorts born between 1985 and 1995. The admission rate for those born in 1960 is around one-third lower than for those born in 1940 and half that of those born in 1920. This is at 50 years of age and a constant 2010 period effect (Figure 18).

FIGURE 18.

Diagnostic procedures: predicted admissions by cohort, for fixed age and period.

The period effect fell slightly between 1999/2000 and 2004/5 and then rose sharply almost every year to 2014/15, after controlling for age and cohort effects. The elective admission rate in 2014/15 is around 60% higher than the rate in 1999/2000, for the 1970 cohort and age 50 (Figure 19).

FIGURE 19.

Diagnostic procedures: predicted admissions by period, for fixed age and cohort.

If there had been no period effect, the annual number of elective admissions for diagnostic procedures would have fallen between 1999/2000 and 2014/15 to around 925,000 in 2014/15 (Table 27). This indicates that the downwards cohort effect has more than offset the age effect over the period 1999/2000 to 2014/15.

Hip and knee replacement procedures

The total number of elective admissions for the prosthetic replacement of hip or knee joint (procedures codes W37 to W42) rose from 76.99 thousand in 1999/2000 to 157.46 thousand in 2014/15 (Figure 20). This is an increase of 104.5% over the full 15-year period and an average annual increase of 4.9%. This compares with an annual average increase of 2.7% for all elective admissions (including day cases) over this period.

FIGURE 20.

Hip and knee replacements: admissions by broad age band, England, 1999/2000–2014/15.

The numbers of elective admissions for these procedures rose in each year except 2009/10, with the fastest rise in the period 2000/1 to 2003/4. The numbers rose for each of three age groups (40–64 years, 65–74 years and ≥ 75 years). Younger age groups are not included in the analysis because joint replacements are rare below the age of 40 years.

The overall admission rate for these procedures per 1000 population rose from around 3.4 in 1999/2000 to 5.8 in 2014/15, which amounts to a rise of 71%. The rate rose most rapidly for the 40–64 years age group and least rapidly for the ≥ 75 years age group. The age-standardised rate rose slightly more slowly, by 70%, over the period 1999/2000 to 2014/15, an annual average increase of around 3.6% (Figure 21). This compares with an annual average increase, on the same basis, of 1.8% for total elective admissions over this period.

FIGURE 21.

Hip and knee replacements: indexed age-standardised admission rates by broad age band, England, 1999/2000–2014/15.

If the admission rates for these procedures by age band had remained constant at their 1999/2000 levels, the number of admissions would have reached 93.5 thousand in 2014/15 rather than 157.5 thousand, an increase since 1999/2000 of 22% rather than of 105%.

The age effect findings show that, after controlling for cohort and period effects, the elective admission rate for hip and knee joint replacement procedures rises with age from 40 to 73 years and then falls with age. This pattern would be expected because need is lower at lower ages and rises with age, but in late old age the risks of joint replacement are higher and the benefits may be lower. The rate at 75 years of age is around 65% higher than the rate at 60 years, 65% higher than the rate at 83 years and around 3.5 times higher than the rate at 50 years of age. This is on the basis of constant 2010 period effect and 1970 cohort effect (Figure 22).

FIGURE 22.

Hip and knee replacements: predicted admissions by age, for fixed cohort and period.

The cohort effect findings are that each successive cohort from the cohort born in 1915 onwards has experienced a lower elective admission rate for hip and knee joint replacement procedures at a given age than the preceding cohorts, after controlling for period, as well as age, effects. This is subject to a few single year rises and to minimal change for cohorts born between 1950 and 1958. The rate for those born in 1960 is around two-thirds that of the rate for those born in 1940 and slightly over one-third that of the rate for those born in 1920. This is at 50 years of age and constant 2010 period effect (Figure 23).

FIGURE 23.

Hip and knee replacements: predicted admissions by cohort, for fixed age and period.

The period effect rose each year between 1999/2000 and 2014/15 except 2009/10. The elective admission rate for hip and knee joint replacement procedures in 2014/15 is around three times the rate in 1999/2000, for the 1970 cohort and at 50 years of age (Figure 24).

FIGURE 24.

Hip and knee replacements: predicted admissions by period, for fixed age and cohort.

If there had been no period effect, the annual number of elective admissions for hip and knee joint replacement procedures would have fallen between 1999/2000 and 2014/15 to around 54,100 in 2014/15 (Table 28). This indicates that the downwards cohort effect has more than offset the age effect over the period 1999/2000 to 2014/15.

Coronary circulation admissions

The total number of elective admissions for coronary circulation procedures (K40–K46 – replacement or bypass of coronary artery, and K49–K51, K75 – transluminal operations on coronary artery) among people aged ≥ 35 years first rose from 37,580 in 1999/2000 to peak at around 64,750 in 2005/6 and then fell back to 51,210 in 2014/15 (Figure 25). This is an increase of 38.1% over the full 15-year period and an average annual increase of 2.2%. This compares with an annual average increase of 2.7% for all elective admissions (including day cases) over this period.

FIGURE 25.

Coronary circulation admissions: admissions by broad age band, England, 1999/2000–2014/15.

The numbers of elective admissions for these procedures for those aged 35–64 years and 65–74 years peaked in 2005/6 and then declined, with levels remaining fairly static since 2009. The ≥ 75 years age group experienced more steady growth, with the number of admissions increasing from 3718 in 1999/2000 to 12,854 in 2014/15.

The overall admission rate for these procedures per 1000 population first rose from 1.43 in 1999/2000 to 2.30 in 2005/6 and fell to 1.71 in 2014/15, which amounts to an increase of 20% between 1999/2000 and 2014/15. The age-standardised rate increased by 14%, from 1.43 in 1999/2000 to 1.63 in 2014/15, an annual average increase of 0.9% (Figure 26). This compares with an annual average increase, on the same basis, of 1.8% for total elective admissions over the period 1999/2000 to 2014/15.

FIGURE 26.

Coronary circulation admissions: indexed age-standardised admission rates by broad age band, England, 1999/2000–2014/15.

The age-standardised admission rate for these procedures fell (by 13%) for the 35–64 years age group, but it rose considerably (by 213%) for the ≥ 75 years age group (see Figure 26). Although the rate for the ≥ 75 years age group was three times the rate for the 65–74 years age group in 1999/2000, the rates for both groups are now similar.

If the admission rates for these procedures by age band had remained constant at their 1999/2000 levels, the number of admissions would have reached 44,140 in 2014/15 rather than 51,910, an increase since 1999/2000 of 17% rather than of 38%.

The age effect findings show that, after controlling for cohort and period effects, the elective admission rate for coronary circulation procedures rises with age to 62 years and then falls, especially after the age of 73 years. This is for constant 2010 period effect and 1970 cohort effect (Figure 27).

FIGURE 27.

Coronary circulation admissions: predicted admissions by age, for fixed cohort and period.

The cohort effect findings are that each successive cohort from the cohort born in 1932 onward has experienced a lower elective admission rate for coronary circulation procedures at a given age, after controlling for period as well as age effects, subject only to a rise for the 1946 cohort and some minor deviations in the late 1970s. Prior to 1932, the rate rose for cohorts born between 1915 and 1932. The admission rate for those born in 1950 is around 51% lower than for those born in 1932 but similar to the admission rate for those born in 1922. This is at 50 years of age and constant 2010 period effect (Figure 28).

FIGURE 28.

Coronary circulation admissions: predicted admissions by cohort, for fixed age and period.

The period effect rose every year from 1999/2000 to 2005/6, fell rapidly in 2006/7 and was then fairly constant with a small upwards trend between 2007/8 and 2014/15, after controlling for age and cohort effects (Figure 29).

FIGURE 29.

Coronary circulation admissions: predicted admissions by period, for fixed age and cohort.

If there had been no period effect, the annual number of elective admissions for coronary circulation procedures would have remained constant from 1999/2000 to 2005/6 and then fallen slightly to around 33,070 in 2014/15 (see Figure 29). Table 29 illustrates further evidence of the importance of the period effect, which has caused a rise of 57% in the levels of admissions.

Menorrhagia procedure admissions

The total number of elective admissions for menorrhagia procedures (codes Q07–Q10 and Q16) fell from 106,620 in 1999/2000 to 66,740 in 2014/15 (Figure 30 and Table 30). This is a decrease of 37.4% over the full 15-year period and an average annual decrease of 3.1%. The most likely reason for this decrease is increased use of the Mirena^® coil (Bayer, Whippany, NJ, USA), which is very effective for the treatment of menorrhagia. The decrease of 3.1% per year contrasts with an annual average increase of 2.7% for all elective admissions (including day cases), over this period.

FIGURE 30.

Menorrhagia procedures: admissions by broad age band, England, 1999/2000–2014/15.

The number of elective admissions for these procedures fell sharply between 1999/2000 and 2004/5, rose slightly between 2004/5 and 2007/8 and then fell slowly to 2014/15. More than 70% of these admissions are of women aged 35–64 years.

The overall admission rate for these procedures per 1000 women fell from 5.19 in 1999/2000 to 2.91 in 2014/15, which amounts to a fall of 44% between 1999/2000 and 2014/15. The rate fell most rapidly for the 15–34 years age group and least rapidly for the 65–74 years age group. The age-standardised rate also fell by 44% over the period 1999/2000 to 2014/15, an annual average decrease of around 5.3% (Figure 31). This contrasts with an annual average increase, on the same basis, of 1.8% for total elective admissions over this period.

FIGURE 31.

Menorrhagia procedures: age-standardised indexed admission rates by broad age band, England, 1999/2000–2014/15.

If admission rates for these procedures by age band had remained constant at their 1999/2000 level, the number of admissions would have reached 119,700 in 2014/15 rather than fallen to 66,700, an increase since 1999/2000 of 12% rather than a decline of 37%.

The age effect findings show that, after controlling for cohort and period effects, the elective admission rate for menorrhagia procedures rises with age from 15 years to around 45 years, then falls with age to around 60 years and then remains constant from 60 years to around 75 years of age. The rate at 45 years of age is more than twice the rate at 35 years of age. This is on the basis of constant 2010 period effect and 1970 cohort effect (Figure 32).

FIGURE 32.

Menorrhagia procedures: predicted admissions by age, for fixed cohort and period.

The cohort effect findings are that each successive cohort from the cohort born in 1968 onwards has experienced a lower elective admission rate for menorrhagia procedures at a given age than the preceding cohorts, after controlling for period as well as age effects. Prior to the 1968 cohort, each successive cohort from 1920 to around 1945 experienced a higher admission rate for these procedures. The rate then fell for cohorts born between 1945 and 1953 before rising to reach a peak for the cohort born in 1968. The rate for those born in 1965 is around 20% higher than the rate for the 1955 cohort and 20% higher than for the 1975 cohort. This is at 50 years of age and constant 2010 period effect (Figure 33).

FIGURE 33.

Menorrhagia procedures: predicted admissions by cohort, for fixed age and period.

The period effect fell each year between 1999/2000 and 2014/15. The elective admission rate for menorrhagia procedures in 2014/15 is only around 60% of the rate in 1999/2000, for the 1970 cohort and 50 years of age (Figure 34). As indicated above, the most likely reason for declining period effect is increased use of the Mirena coil.

FIGURE 34.

Menorrhagia procedures: predicted admissions by period, for fixed age and cohort.

If there had been no period effect, the annual number of elective admissions for menorrhagia procedures would have risen between 1999/2000 and 2014/15 to around 120.4 thousand. This indicates that the cohort effect has marginally more than offset the age effect over the period 1999/2000 to 2014/15.

Conclusions

We have explored in this chapter how far the considerable increase in the number of elective hospital admissions over the 15-year period from 1999/2000 to 2014/15 is due to the ageing population, how far to cohort effects and how far to period effects. We have considered in our APC analyses the trends in total elective admissions (including day cases) and trends in elective admissions for four sets of procedures. These four sets of procedures accounted for 26% of all elective admissions in 1999/2000 and 21% in 2014/15.

Total elective admissions rose by 2.7% per year over this 15-year period. Elective admissions for diagnostic procedures rose somewhat more slowly (2.2% per year) and admissions for hip or knee replacements rather more rapidly (4.9% per year). In contrast, coronary circulation procedures and procedures for menorrhagia both fell (by 4.4% and 3.1% per year, respectively) over this period.

The age effects are fairly similar across all the procedures we considered (Table 31). They are, not surprisingly, a little higher than average for joint replacements, which are concentrated on older ages, and a little lower for menorrhagia procedures, which are concentrated on younger women.

The cohort effects are more strongly negative for three of the sets of procedures than for electives in general (see Table 31). There has been an especially large downwards cohort effect for hip and knee replacements and substantial downwards cohort effects for coronary circulation procedures and diagnostic procedures. A potential explanation is that health at a given age has improved more than average for coronary artery disease and for arthritis of the hips and knees. For menorrhagia, however, there has been no significant cohort effect.

The variation in period effects for the different procedures is of considerable interest (see Table 31). Period effects relate to factors that cannot be explained by age or cohort effects: they may include demand effects such as shifts in demand at a particular time period, as well as supply factors such as uptake of new technologies.

For menorrhagia procedures, the large downwards period effect is likely to be due to the introduction of the Mirena coil, which has proved to be an effective treatment for the condition. For coronary artery disease, new technology may also be an explanation for the downwards period effect. The upwards period effect for diagnostic procedures is higher, albeit not much higher, than the period effect for all elective admissions. It may be that the scope for diagnostic procedures has risen greatly, even faster than the scope for other valuable elective procedures.

For three of the groups of procedures the negative cohort effect outweighs the positive age effect, and for one of these three (hip and knee replacements), the cohort effect is strongly negative. We would therefore conclude that demographic/epidemiological effects (in so far as the balance of age and cohort effects captures them) do not inevitably generate increases in the number of elective admissions.

The period effects are positive for the two sets of procedures for which new technologies have offered improved prospects of benefit (diagnostic procedures and hip and knee replacements) and negative for the two sets of procedures for which new technologies have provided alternatives to surgery (coronary circulation procedures and menorrhagia procedures). A positive period effect is not inevitable. To reduce elective admissions, the best approach may be to promote research to find less invasive alternatives to surgery.

Introduction

In the decade to 2011/12, hospital admissions in England increased by 35.4%. 53 Policy reforms, including a new GP contract, and increased numbers of GPs, were introduced to reduce admissions growth by strengthening less costly and more accessible primary care. Several explanations have been proposed for the growth in hospital admissions. They include an increase in illness and frailty linked to the ageing population (see, e.g., Blatchford and Capewell54 and Gillam55); increased ability to detect and treat illness (e.g. Hobbs56); the effects of incentives in the tariff model of paying hospitals (e.g. Farrar et al. 13 and Health and Social Care Information Centre57); and ‘targets’ to reduce waiting for both elective and emergency care. The working practice of GPs has also changed, notably the new ‘out-of-hours’ (OOH) services (Coast et al. ;58 Bunn and Kendall;59 Sibley et al. 60), and this may also have contributed to admissions. For example, Dusheiko et al. 61 find that since GP contract changes, primary care performance has been associated with lower hospital costs in stroke care only. However, there is a substantial body of literature that argues that primary care physicians provide high-quality care at a lower cost than specialists (Chernew et al. ,62 Franks and Fiscella63 and Macinko et al. 64). This may be the case particularly in situations in which primary care physicians operate as gatekeepers (Forrest and Starfield65). The primary intention of the new contract was to incentivise several patient management interventions to improve patient access and increase cost-effectiveness by moving treatment for chronic care out of hospitals. Austerity funding has heightened interest in understanding how increasing GP services might reduce both referrals and elective admissions, as well as emergency care, with the two main UK political parties in the 2015 general election both proposing an increase in GP supply. This report addresses these issues in two ways: (1) a model of GP referrals is developed, which examines the interaction between GP and specialist, but does not assume that gatekeeping arises because GP treatments are cheaper, as do, for example, Mariñoso and Jelovac,4 in the first of the few theoretical studies of the relative benefits of gatekeeping. Instead, gatekeeping (i.e. limiting access) enables greater efficiency in supplying diagnostic information and, hence, in maximising health gain net of cost. (2) The chapter reports estimates of the influence on referrals and admissions of increasing (1) GP supply and (2) mean GP practice size. The changing supply of NHS GPs to areas across England in the period 2004–11 is used to construct panel data estimates of their effects on referrals and admissions, which allow for the endogeneity of GP location choice.

Theoretical studies of gatekeeping by GPs have been particularly concerned with the implications of alternative payment schemes on GPs’ effort to diagnose accurately, and for the incentive to refer or treat efficiently (see, e.g., Malcomson,66 Mariñoso and Jelovac4, Gonzáles67 and Allard et al. 68). The schemes most often analysed are fee for service, fundholding and capitation, which is the UK arrangement during the period we analyse and commonplace in Health Maintenance Organisations (HMOs). We study a model in which GPs are paid by capitation and there is a competitive specialist sector. We use a characterisation of disease drawn from these models to study the consequences for referrals and admissions of an increased market supply of GPs.

Given that NHS GPs are paid by capitation on registered patients, and that their use of time is not observed, it is necessary to decide how to model how GPs are constrained from pursuing their own utility rather than patient health. There are three approaches in the literature: (1) a physician’s objective function reflects his or her imperfect role as agent to the patient (principal) and includes both his or her own utility and, for altruistic or professional reasons, her patients’ health. This approach is widely adopted, for example, in Iversen and Ma69 and Gaynor et al. ,70 and was studied by Ellis and McGuire71 and, recently, by Chandra et al. ;72 (2) clinical decisions reflect a bargain between the treating clinician and patient, as discussed by Ellis and McGuire73 and Ma and Riordan;74 and (3) GP referrals are constrained towards efficient levels by patient access to a second opinion, an idea that is analysed by Malcomson. 66

A difficulty with using the familiar ‘altruistic clinician’ model to explain NHS GP referral behaviour is that, because GPs are paid by capitation, a GP offered the choice either to refer or to increase effort and treat a sick patient will refer to a specialist because it serves the interests of both the patient and the GP; hence, as Malcomson66 argues, gatekeeping cannot be explained. The puzzle of why GPs paid by capitation may nevertheless limit specialist access and provide treatments is resolved here by exploring the nature of patient ‘agency’, and by introducing a GP objective suited to the aims of GPs employed in single-payer insurance systems and insurer organisations including HMOs that employ GPs. In single-payer and managed-care contexts, the insurer influences the GP, and the principal may be regarded as the insurer and not the patient as in the ‘altruistic clinician’ model. However, the insurer has imperfect information about a GP’s decisions, and the GP can be an imperfect agent of the insurer. Unless the GP is constrained to efficient behaviour by ethics or a second opinion, moral hazard problems arise, as they do in a different form in the ‘altruistic clinician’ model.

In the model the first best captures the intuition that if specialists can treat at no higher cost than GPs and can provide low-cost patient insight before costly diagnostics, an efficient solution would be to refer all patients and for a specialist to decide diagnostic expenditure and treatment. We show, however, that limiting patient access to specialists can increase efficency in a health system with information failures and insurance. We discuss why a single-payer insurer is more likely to address these failures by encouraging gatekeeping. In this system we study the effects of more GPs, and a more elastic demand for GP registration, on referrals and specialist treatment.

A single information constraint underpins the argument in the model that a health authority (HA) will regard gatekeeping as in the patient’s interest: the HA imperfectly observes the evidence used by specialists to choose diagnostics. This creates a moral hazard, with a higher than optimal level of diagnostic expenditure by consultants who compete to attract patients by maximising health gain. Efficient equilibria in which GP patient information is used by a specialist to refuse an exact diagnostic test are not selected by the specialist, and such equilibria can be obtained only if a GP limits access to specialists. However, a competitive GP is not incentivised to limit access. If a GP is trained to refer in the interests of the insured, then a perfect agent can fully correct the distortion of excess diagnostics, whereas an imperfect agent can reduce the distortionary costs.

This report discusses why GPs in single-payer systems and those employed by insurance companies may behave differently from GPs in a devolved primary care system, as is found across much of the USA, and why a single-payer insurer is more incentivised than a competitive insurer to invest in GP training to represent the interests of the insured. Nevertheless, GPs who value their leisure and income are likely to be imperfect agents of the HA. Their incentive is to set higher referral rates than insurers prefer to secure more patient revenue and personal leisure. If a specialist can order diagnostics without HA oversight then competition to maximise health gain leads to using the best diagnostics.

We model a GP whose intrinsic concerns are his/her own utility and the objective of the insurer, which wishes to maximise the social welfare of those insured, as in Malcomson. 66 This approach may be viewed as being professionally underpinned by the ‘population health’ dimension of ethical medical practice. GPs learn to be aware of the opportunity cost of their decisions for other insured patients by following ‘value for money’ guidelines. The NHS reinforces this by requiring GPs to lead budget-holding organisations and monitor referrals, which holds GPs decisions to account and subject to peer pressure (see Department of Health75). We discuss how this supports GPs voluntarily limiting access to specialists, even if this is not in the interests of their patient, or themselves, and why this is efficient. Unlike a competitive insurer, this investment is incentivised because the HA is a single payer and receives all the benefits from the training investment, which it may view as ‘firm-specific’ rather than ‘general’. Even if GPs in competitive primary care systems wish to maximise the expected utility of insurees, information on the patient value of alternative treatments is costly, and clinicians have little incentive to invest in cost-effectiveness information if their decisions are imperfectly observed by the insurer. Instead, to compete for patients, the clinician selects the treatment offering the greatest benefit.

Individuals may become mildly or severely ill. A ‘single-payer’ HA funds treatments by GPs and specialists. However, the evidence on which a specialist decides whether or not to use a costly diagnostic test, and GP effort to treat, is unobserved by the HA. In view of the specialist tendency to purchase excess diagnostics in order to compete in providing patient gross health gain, the HA trains the GP to refer to maximise the expected utility of registered patients. The GP gathers patient information to estimate severity and uses it to decide on referrals, recognising also the distortion in specialist behaviour. If, for some patients, GP information is valuable to economise on diagnostic spending, this is possible only if a patient is not referred. We discuss assumptions under which there exists an interior solution, in which a GP acting as an imperfect agent of the insurer treats some ill patients and refers the others. We discuss the consequence of increasing the supply of GPs, and reducing their local monopoly power, for patient welfare, referrals and costs/taxes.

Increasing the market supply of GPs results in fewer patients per GP. This reduces the loss of utility from complying to a welfare-maximising professional norm that requires that GPs treat more patients to refer a smaller proportion. It thereby increases the expected health gain of insurees, while reducing the demand for specialists. In contrast, GPs who are ‘altruistic clinicians’ will, given the absence of fees for most NHS GP services, refer all patients to specialists regardless of GP supply. In this model, GP referrals are unaffected by offering patients a second GP opinion, if the second opinion is obtained from a new GP provider and not an adjudicator without the responsibilty to treat.

‘Gatekeeping’ is a feature of primary care systems in various countries, including Scandinavian countries and the UK, and, as Glied76 explains, in US HMOs, there is little evidence of the effect of the impact of gatekeeping or increasing the supply of GPs. Gulliford,77 in a cross-sectional study of 99 UK HAs, finds that an increased supply of GPs is associated with lower hospital admissions for both chronic and acute conditions. Harris et al. ,78 in a cross-sectional study of 68 English practices, find little evidence that GP supply influences emergency admissions. Baicker and Chandra79 and Chernew et al. 62 find that US Medicare reimbursement is lower in markets with a higher percentage of primary care physicians. Wright and Ricketts80 show that inpatient admissions are lower in those US areas with a higher density of GPs. There is evidence that financial incentives may influence GP referrals and admissions, as with GP fundholding, for example, in Dusheiko et al. 81 The recent literature (e.g. de Bruin et al. ,82 Carroll and Dowling83) focuses on the impact of primary care disease management programmes on health-care expenditure, finding that the cost saving is small. Improved primary care may not necessarily result in fewer admissions; indeed, more GPs may improve patient screening or access and thereby increase elective care (see Blustein et al. 84). Iversen and Ma69 present a model of how the degree of competition for patients may influence GP rate sof referral and, using Norwegian data, find no evidence that more GPs and competition reduces referrals to specialists. Fang and Rizzo85 show that the empirical influence of an increased supply of GPs on referrals is sensitive to whether or not GPs are self-employed or working in a HMO, in which an increased supply of GPs is found to reduce referrals.

Using panel data from the NHS in 2004/11, and instrumental variable (IV) methods to correct for the possible endogenous choice of location by GPs, we are able to estimate the effect of GP supply and the density of GP practices on hospital admissions. Using a method to address the problem of endogenous variables, proposed by Chernozhukov and Hansen,86 the effects of these variables at different points along distribution of hospital admissions is estimated by quantile regression. Panel data estimates suggest that increases in the local supply of GPs helps to explain the pattern of referrals and admissions in a way consistent with the hypothesis developed below (see A model of general practitioner gatekeeping, diagnosis and referrals).

The next section describes the model and hypotheses. The empirical strategy, data and results are then presented (see Economic strategy, Clinical data and Results). Robustness checks, including whether or not the findings are similar in less and more deprived areas, and quantile regression estimates of the effect of GP supply, are then discussed (see Robustness checks) and conclusions are then outlined.

Model of general practitoner gatekeeping, diagnosis and referrals

The economy has two sectors, a consumption good sector and a health-care sector. Labour is the only input. There is a population of n, all of whom are equally vulnerable to illness and equally productive in three jobs: hospital specialists, GPs and craftsmen. Hours of work are fixed except for GPs, who work the hours required to treat their patients, which they choose. Individual utility is an additive separable in income, leisure and health. Health care is produced using GPs and specialists, who produce an economic surplus in the form of a patient health gain net of cost. This has a value that depends on the match between the patient’s severity of illness and the treatment. The surplus from each illness–treatment pair is known with certainty, but the patient’s initial illness is uncertain. This uncertainty may be reduced by a diagnosis, which is exact and costs c if made by a specialist, and is inexact and free if made by a GP. A single-payer HA financed by a lump sum tax, t, funds services and chooses GP numbers and a GP capitation fee. Specialist and craftsmen employment is endogenous. Patients first contact a GP, who refers all or only a fraction of patients to a specialist.

Discussion

What is the intuitive reason that an increase in the supply of GPs, and hence fewer patients per GP, increases s*, and reduces referrals? Essentially, having fewer patients reduces the loss of utility that follows if s* is increased towards the referral rate that maximises health gain. Maximising health gain for the representative patient requires that a GP limits specialist access and treats more patients than if income and leisure were his or her only concern. This requires that he or she accepts less income and leisure than at maximum GP utility. A GP with fewer patients has a smaller loss of capita income and leisure when treating a larger fraction of his or her patients than is utility maximising; and, because the cost of more closely conforming to his or her professional ethic of a ‘health maximising GP’ is lower, he or she is willing to treat a higher proportion of patients. By increasing the referral threshold, s*, a GP reduces patients’ gross health gain, because more referrals always add to health gain, but increases health gain net of costs, and insuree utility. In this model it is optimal for GPs to have a minimal list of patients, and this may be constrained at a plausible level by a requirement for a minimum daily income or leisure hours, at a social optimum. The effect of reducing λ on GP income is ambiguous.

Econometric strategy

Our empirical strategy is to estimate a fixed-effects panel data model for three classes of hospital referrals/admissions (outpatients, elective and emergency) at LSOA level, controlling for a group of area-specific characteristics, and primary care variables. The model is the following:

where F_jt represents the number of outpatient referrals or hospital admissions (elective or emergency) per 1000 population resident at each LSOA (j) in each year. X_jt is a vector of socioeconomic characteristics that are time-varying at LSOA j in time t. It includes percentage of sex, age and ethnicity, and β is a vector of the slope effects of these variables.

The key explanatory variables that capture the supply of GP services in each year t, are a measure of both the number of GPs employed and FTE GP employment. Although the latter may be a more representative measure of local labour supply, unfortunately, FTE information is available only for GPs working in traditional practices. The term αGP_jt represents this effect of the density of GPs for each 1000 population of the LSOA j in the PCT (p) in time t, or the GP FTE supply in PCT p in time t. Instead, ρGR_jt represents the density of GP practices at time t in PCT p per 1000 of population. Both of these effects (αGP_jt and ρGR_jt) are estimated in models with LSOA fixed effects, so that each captures the influence of changes in the variable over time within a LSOA.

In addition, in some models we also control for certain other new models of delivering primary care services, namely walk in centres (WICs) and OOH services, that are labelled ‘other cost prescribing centres’. In each case we calculate density by dividing the number of each of type of provider by the local PCT population in 1000s. Because the geographic area of PCTs does not change over time the fixed LSOA effects control for large and small area PCTs, allowing the density variables to capture the average effects of within-LSOA changes in density of traditional practices, or WICs.

The terms z_jt and d_jt capture measures of the urban and deprivation local area dummy variables. Finally, we control for time (µ_t) with year effects, disease prevalence (ω_jp) and LSOA (σj) fixed effects. To avoid any possibility of the endogenous recording of conditions following hospital admission, we use the prevalence data for the year prior to that for the year of study for hospital admissions. The inclusion of these effects allows us to identify the impact of variations in primary care supply on the hospital admission at the LSOA area level. Our principal objective in this report is to measure how variations in the scale and concentration of GP services affects hospital admissions. ε_jt is a random error term, which is assumed to be normally distributed.

Data

Summary statistics

Table 33 provides a summary description of the variables used to study the period 2004/11. The three categories of hospital service (outpatient referrals, elective admissions and emergency admissions) all increase over time, with outpatient referrals increasing most rapidly. The mean proportion of the elderly in local populations increases slightly for both males and females over time. The mean PCT density of GPs per head increases by about 27% between 2004 and 2011; the mean PCT density of GP FTEs increases by about 11% between 2001 and 2009 but then declines by 3% in the period 2009/11. The mean number of traditional GP practices per 1000 population is unchanged at one practice per 5000 persons. Mean age and ethnic population proportions have also experienced increases. The WICs and OOH services are primary care services that were started around 2005/7 and, as reported by Exeter data, remain few in number.

TABLE 33 - Summary statistics

This includes conventional partnership practices.

This includes WICs, OOH centres and other prescribing cost centres, which include addiction services.

Variable	2004	2005	2006	2007	2008	2009	2010	2011	Total
Elective per 1000 population (SD)	114.92 (35.36)	120.23 (36.70)	124.02 (38.03)	130.64 (37.99)	138.78 (39.75)	142.34 (40.13)	145.19 (40.26)	147.93 (43.36)	133.01 (40.69)
Emergency per 1000 population (SD)	116.33 (44.20)	122.70 (45.66)	123.18 (46.92)	124.16 (48.17)	132.20 (49.41)	138.33 (51.59)	141.30 (52.73)	140.47 (52.04)	129.83 (49.73)
Outpatients per 1000 population (SD)	261.65 (79.27)	294.42 (86.01)	298.29 (93.87)	321.53 (101.38)	362.30 (132.89)	400.18 (182.78)	405.76 (239.71)	411.86 (312.76)	344.50 (181.26)
Female population, % (SD)	51.04 (2.30)	51.01 (2.31)	50.98 (2.36)	50.95 (2.48)	50.92 (2.67)	50.90 (2.85)	50.87 (3.11)	50.97 (2.28)	50.96 (2.56)
Female population aged ≥ 60 years, % (SD)	11.78 (4.59)	11.83 (4.61)	11.90 (4.65)	12.10 (4.73)	12.28 (4.82)	12.43 (4.88)	12.38 (4.82)	12.56 (4.94)	12.16 (4.77)
Male population aged ≥ 65 years, % (SD)	6.87 (2.71)	6.92 (2.75)	6.96 (2.80)	7.02 (2.86)	7.14 (2.94)	7.28 (3.04)	7.42 (3.15)	7.27 (3.06)	7.11 (2.92)
Black ethnicity, % (SD)	2.60 (4.40)	2.67 (4.26)	2.72 (4.12)	2.78 (3.99)	2.84 (3.87)	2.88 (3.75)	2.93 (3.64)	2.91 (3.29)	2.79 (3.93)
Asian ethnicity, % (SD)	5.65 (6.89)	5.88 (6.80)	6.11 (6.71)	6.37 (6.66)	6.58 (6.57)	6.78 (6.50)	7.01 (6.46)	7.12 (6.08)	6.44 (6.61)
Revenue per head (SD)	1.00 (0.69)	1.00 (0.69)	1.00 (0.68)	1.00 (0.68)	1.00 (0.13)	0.99 (0.13)	1.00 (0.14)	1.00 (0.14)	1.00 (0.49)
Headcount GPs density at PCT per 1000 population (SD)	0.83 (0.10)	0.87 (0.10)	0.89 (0.11)	0.93 (0.12)	0.96 (0.13)	1.00 (0.13)	1.04 (0.17)	1.06 (0.18)	0.95 (0.15)
FTE GPs density at PCT per 1000 population (SD)	0.61 (0.07)	0.62 (0.07)	0.65 (0.08)	0.64 (0.08)	0.66 (0.08)	0.68 (0.08)	0.68 (0.10)	0.67 (0.09)	0.65 (0.09)
General practices density at PCT per 1000 population (SD)a	0.17 (0.04)	0.17 (0.04)	0.16 (0.04)	0.16 (0.04)	0.16 (0.04)	0.16 (0.04)	0.16 (0.04)	0.16 (0.04)	0.16 (0.04)
WIC and OOH density at PCT per 1000 population (SD)b	0.00 (0.01)	0.01 (0.01)	0.01 (0.01)	0.01 (0.01)	0.01 (0.01)	0.01 (0.01)	0.01 (0.01)	0.01 (0.01)	0.01 (0.01)
n	32,482	32,482	32,482	32,482	32,482	32,482	32,482	32,482	259,856

Figure 35 describes the distribution of mean GP practice size by PCT in 2004 and 2011. This highlights both the wide variation of mean PCT practice size and the increasing mean PCT practice size over this period. The additional GPs in traditional practices have mostly joined existing general practices, which has increased their mean size. The number of single-handed GPs is falling as older GPs retire, and is down by 28% since 2004. 91

FIGURE 35.

Distribution of mean GP practice size by PCT in 2004 and 2011.

Figure 36 gives a regional breakdown of the change in admissions by PCT in the period 2004/11 per 1000 population with the south-west, East Anglia and the north-west having smaller increases.

FIGURE 36.

Change in hospital admissions between 2004 and 2011 by PCT per 1000 population. Contains OS data © Crown copyright and database right 2016.

Results

For each type of hospital admission (outpatient referral, elective and emergency), two sets of results are presented corresponding to the two GP supply variables per head of PCT population: (1) total GP headcount; and (2) FTE (hours) data. Each set of results also estimates the influence of the density of places of provision, which is measured using the number of GP traditional practices per head of population, and also the density of other ‘prescribing cost centres’, which includes the recent innovations of WICs and OOH service centres. Estimates for both ordinary least sqaures (OLS) and 2SLS are given, the latter of which allow GP density to be endogenous and estimated using IV methods, as described above (see Identification and instrumental variables).

Robustness checks

The aim of these robustness checks is to explore potential differences between LSOAs that may have been overlooked by the modelling thus far. One motivation is that the supply of health-care services may be different across LSOAs in a way that alters how demand is presented and how admissions are determined. Our particular concern is with potential differences between deprived and more prosperous areas of the country. To better understand how the influence of GP supply may differ across LSOAs with quite different levels of admissions, we first compare two groups of LSOAs: the 20% most deprived LSOAs compared with the 20% least deprived LSOAs, using the Index of Deprivation by decile.

Consider Figures 37 and 38. Figure 37a gives the density of practices and prescribing cost centres in the population for deprived and least deprived LSOA areas, 2004–11; Figure 37b gives the density of headcount and FTE GPs in the population for the least and most deprived LSOA areas, for the same years. The structure of primary care is both different and in an important sense changing in a more problematic way in the most deprived areas than in less deprived areas. The deprived areas have about twice the number of practices per head and on average smaller practices, but although GP headcount increases steadily in all areas, the GP FTE falls after 2007 in both the least and most deprived areas. This decline in FTE also steadily narrows the difference in FTE GP supply between the most and least deprived areas.

FIGURE 37.

Density of general practices and prescribing cost centres in the population for most deprived and least deprived areas, 2004–11. (a) Density of practices and prescribing cost centres in the population for most deprived and least deprived LSOA areas; and (b) density of headcount and FTE GPs in the population for the least and most deprived LSOA areas.

FIGURE 38.

Hospital referrals and emergency admissions for the least and most deprived areas, 2004–11. (a) Average emergency admission rates for the most/least deprived 10% of areas; (b) average elective admission rates for the most/least deprived 10% of areas; and (c) average outpatient rates for the most/least deprived 10% of areas.

Figure 38 shows how specialist referrals and emergency admissions are relatively high for the most deprived areas, although admissions for elective care in the least deprived areas have increased relative to those in the most deprived areas. Given that the levels of admissions are quite different across these deprived and non-deprived LSOAs, it appears important to examine whether the effects of GP supply on admissions should be robust to whether the LSOA has high or low expected levels of each type of admission/referral.

In Table 37 we report the 2SLS model for each of the two groups of LSOAs, applying the same estimation methods described above (see Identification and instrumental variables) to models of hospital admissions (outpatients, elective and emergency). These estimates suggest that increasing GP supply significantly decreases hospital admissions in deprived areas, but does not do so in the least deprived areas. The influence of more practices, holding GP supply constant, has a positive and significant effect on admissions in both types of area. The influence of age follows the same pattern for both areas: it has a positive and significant influence on admissions. Ethnicity has different effects in deprived and less deprived areas. Black ethnicity has a positive influence on referrals and both types of hospital admission in deprived areas but has no effect in the least deprived areas. The influence of Asian populations on hospital admissions has a positive influence in deprived areas but, in the non-deprived areas, has a negative effect. We may conclude that the influence on admissions of individual characteristics, and GP variables, are different in these two types of areas, as well as the levels of admission also often being quite different.

To address this heterogeneity across LSOAs, and, in particular, whether or not the impact of GPs is similar for both low and high admission areas, we estimate a quantile regression model accounting for the endogeneity of GP supply. Quantile regression is a widely used tool for estimating conditional quantile models (see, for example, Koenker and Bassett92 and Koenker. 93 Quantile regression modelling gives estimates of quantile-specific effects that describe the impact of covariates not only on average, but also on the tails of the conditional outcome distribution. Although central effects, such as the mean effect obtained through conditional mean regression, provide useful summary statistics of the impact of a covariate, they fail to describe the distribution of the impact. Standard linear regression techniques summarise the average relationship between a set of regressors and the outcome variable based on the conditional mean function E(y ∣ x). This provides only a partial view of the relationship, as we might be interested in describing the relationship at different points in the conditional distribution of y. Quantile regression provides that information. Analogous to the conditional mean function of linear regression, we may consider the relationship between the regressors and outcome using the conditional median function Qq(y ∣ x), where the median is the 50th percentile, or quantile q, of the empirical distribution. The quantile q∈(0,1) is that y, which splits the data into proportions q below and 1 – q above: F(yq) = q and yq = F – 1(q): for the median, q = 0.5.

In many studies, the variables of interest (education, health or prices) are endogenous. Endogeneity of covariates renders the conventional quantile regression inconsistent for estimating the causal effects of covariates on the quantiles of economic outcomes, just as with the conventional linear model. To study the relation between GP supply across the conditional distribution of hospital admissions (outpatient, elective and emergency) we use a new quantile instrumental variable (QIV) panel estimator, developed by Chernozhukov and Hansen. 86 The principal identifying assumption of the model is the imposition of conditions that restrict how rank variables (structural errors) may vary across treatment states. These conditions allow the use of IVs to overcome the endogeneity problem and hence recover the true quantile treatment effects (QTEs). This framework also ties naturally to simultaneous equations models, corresponding to a structural simultaneous equation model with non-additive errors. Estimation and inference procedures for linear quantile models have been developed by Chernozhukov and Hansen,86 Chernozhukov et al. 94 and Chernozhukov et al. 95

The QIV estimator allows us to obtain estimates of the influence of GP supply that vary across the conditional LSOAs hospital admissions distribution. Unlike estimators of the conditional mean, which can be very sensitive to values in the tail of the distribution, conditional quantile estimators are inherently more robust to extreme values, which is particularly advantageous. In addition, the QIV estimator allows us to address the endogeneity described in the previous section with an IV identification strategy. To estimate the quantile regression we use the same variables as used above to estimate the 2SLS of the hospital admissions, except that Strategic Health Authority (SHA) fixed effects are included instead of LSOA fixed effects owing to the problem of the incidental parameter giving inconsistent estimates using individual fixed effects in a quantile regression (see Kato et al. 96).

Table 38 reports the quantile regression accounting for the endogeneity of GP FTE density, and the intervals of confidence for outpatient appointments, elective admissions and emergency admissions. The major variation across the distribution is observed in the outpatient referral equation. This implies that an increase in FTE GPs of the same amount will have a larger effect on outpatient referrals in a LSOA with low rate of referrals per head than one with a high rate of referrals. Figures 39 and 40 show the quantile estimation for the influence of four independent variables: the number per head of population of (1) FTE GPs, (2) general practices, (3) Asian populations and (4) black populations.

FIGURE 39.

Quantile regression for hospital admissions: estimated coefficients for independent variables. (a) FTE GPs density per 1000 population; and (b) GPs density per 1000 population.

FIGURE 40.

Quantile regression for hospital admissions: estimated coefficients for independent variables. (a) Percentage black ethnicity; and (b) percentage Asian ethnicity.

We find that the influence of GP supply is unchanged across the distribution for elective and emergency admissions, whereas its influence on referrals is positive and larger in the LSOAs with high levels of referrals. In contrast, the influence of the density of general practice is positive and larger in the middle of outpatients distribution, whereas along the tails it decreases until it becomes negative at the top of the distribution. Again, the influence of practice density is constant along the distributions of elective and emergency. The same flat pattern is found for the impact of ethnicity (black and Asian population shares) for the elective, outpatients and emergency distributions. The sole exception is that the coefficient of Asian ethnicity is negative in LSOAs with few outpatients admissions and positive in the LSOAs with more outpatients admissions. This is consistent with the estimates discussed above in this section for deprived and non-deprived areas. This estimate gives insight into how the influence of GP supply is greater for outpatient referrals to hospital in areas with high referrals, which are the more deprived areas. This is consistent with the findings concerning the relative responsiveness to increases in GP supply given for deprived areas in Table 38. In contrast, the levels of elective and emergency admissions respond in a similar way at all levels of admission rates to changes in GP supply.

Conclusions

This report discusses a model of GP referrals in which a GP, in order to increase efficiency, may not refer certain patients despite this not being in the interests of either the patient or the GP. It explores in this model the hypothesis that adding GPs to a given patient population will reduce the demand for elective hospital services. It gives assumptions under which a shorter list of registered patients will lead a GP to reduce specialist referrals and to treat a larger share of his or her patients in primary care. These assumptions include that a GP acts as an imperfect agent of the insurer, maximising an objective function of his or her own utility and the expected health gain of a representative patient in the health system. Although the GP and his or her patients both benefit from referring all ill patients, the GP’s professional ethic prompts GP treatments for patients that he or she believes to be only mildly ill. Increasing the number of GPs is found to lead each GP to treat a larger share of their shorter patient list and to make fewer specialist referrals. This increase brings referrals closer to the level that maximises expected health gain, because, as long as GPs consider their own welfare, as well as that of patients, referrals are below that level.

We find that both the local supply of GPs and whether or not GPs are concentrated into larger practices help to explain hospital admissions, but only in deprived areas; an increased density of GPs tends to reduce both outpatient referrals and elective, but not emergency, admissions. In more prosperous areas these effects are either not present or weaker. This may be because in the more prosperous areas patient choice predominates and the gatekeeping role of the GP is de-emphasised. PCTs with a large number of small practices, but the same density of GPs, tend to have more outpatient referrals and elective and emergency admissions. Various robustness checks also suggest that the structural relationships between GP and practice density and admissions differ between deprived and non-deprived areas, with additional GPs in the more deprived areas reducing both referrals and elective admissions. However, even in deprived areas, the value of these savings does not by itself meet the cost of additional GPs.

Introduction

In 20011/12, first referrals by GPs to specialists reached a level of 11.2 million, with referrals to orthopaedic and general surgeons, ophthalmologists and gynaecologists being particularly significant, and in the preceding eight years, both first referrals and total elective care grew by about one-third. Referrals are usually intended to initiate treatment by specialists (see Donohoe et al. 97) and thus cause a large part of NHS expenditures, even though GPs do not refer all patients seeking a specialist consultation. This amelioration of referrals is usually thought of as reflecting NHS and GP concerns to limit the impact of insurance on patient demand for care, so that care is consumed to the point at which the patient’s marginal benefit equals the marginal social, and not private, cost. With austerity-driven cost-saving programmes currently in place across government, it is unsurprising that policy to reduce the growth of elective admissions has included the consideration of whether or not GP ‘gatekeepers’ might support this constraint on unaffordable growth by reducing referrals. For example, Lambeth CCG has incentivised GPs to refer at mean practice rates, whereas Lincolnshire CCG has incentivised GPs to refer at a rate similar to the low referral practices. Such policies implicitly assume that fewer referrals might play a decisive part in reducing admissions. However, this overlooks the role of specialists, who determine thresholds of illness below which a patient is not admitted. Following changes in referral policy, the mix of patients eligible for a specialist’s threshold might be increased if patients with lower priority for treatment are no longer referred. The literature, usefully summarised by O’Donnell,98 discusses the socioeconomic determinants of GP referrals and variation in referral rates between practices. The samples used are usually quite small. In contrast, partly owing to obstacles in obtaining appropriate data and the difficulty of linking/measuring the hospital admissions of the patients of specific practices, there are very few studies about the variation of treatment rates between practices, and none has linked the analysis of these two critical building blocks of the relationship between GP and specialist.

The aims of this chapter are, first, to contribute to a small body of literature on understanding the determinants of large variations in ‘first’ referral rates across practices and, second, to examine the determinants of treatment rates by practice following first referrals and how far a reduction in GP referrals at practice level will reduce specialist treatments. We wish to answer the following questions and, in part, to illuminate the benefits of alternative policies to increase GP supply: what are the GP and patient characteristics that explain high rates of GP referrals? Are certain groups of GPs more likely to refer patients who induce particularly productive work by the specialist, insofar as they refer a higher proportion of patients who are either subsequently treated or not discharged without follow-up by the specialist? How should we think of the relationship across NHS GP practices between a practice’s rate of patient referral and the rate of elective treatment following first referral of patients at that practice? An intuitive supposition would be that practice variation in both referral and treatment is driven by patient health, and that practices with less healthy patient populations will have both higher referrals and admissions. This suggests that practice admissions per registered patient will broadly increase with practice referrals. We shall explore this relationship, and also explore hypotheses regarding referrals that shed light on the appropriateness of referrals for various socioeconomic groups, by age for both very young and very old patients, and by the level of prosperity in the patients’ local area. This may also indicate whether or not savings can be made to the health economy by caring for patients in primary care rather than by referring them to specialists (see Foot et al. 99). The absence of experimental data leads us to use observational data comparing the level of treatments received by patients of practices with observationally similar practice, local area and patient characteristics, but with differing referral rates. This requires that we estimate the joint roles of (1) practice referral rates, (2) patient and GP practice characteristics and (3) local socioeconomic and health system characteristics to explain the likelihood that a patient is referred and that a referred patient is admitted for treatment. To do this, we have created a rich data set that combines practice-, hospital- and area-level data for all English GP practices, for the period 2004–12, together with the characteristics of their clinicians and the morbidities and socioeconomic characteristics of their patients. The sample size of practices declines from about 8400 practices in 2004 to 7950 in 2012. One feature of the data is that we have created an index of deprivation for each practice based on the socioeconomic groups of the LSOA in which their patients reside.

We shall examine the implications of the recent changing structure of GP practices for first referrals. There is a growing dependence on salaried GPs and locums within GP practices; there has also been an increase in part-time working by GPs and a growth of average practice size. Another GP characteristic that may influence the referral rate is the sex of the GP, as other studies have found that female GPs are more likely to refer, all else being equal. 100

O’Donnell98 summarises the literature concerning variations in the referral rate among GPs, discussing how the variation of referral rate among GPs remains largely unexplained in the literature. In an early contribution, Coulter et al. 101 examined the variation in rates of admission to hospital among general practices to determine the relationship between referral rates and admission rates. Analysing 19 general practices in three districts in the Oxford Regional Health Authority, the authors found that patients referred to surgical specialties were more likely than those referred to medical specialties to be admitted after an outpatient referral, and they report only modest increases of patient treatment rates at practices with higher levels of referral rates – a ‘flattening’ of the treatment/referral relationship at higher referral levels. Hippisley-Cox and Jumbu102 describe the trend increase in consultation in general practices, from 3.9 per patient per year in 1995 to 5.4 per patient per year in 2007. Consultation rates vary markedly by age, with the highest rates for the elderly. Consultation rates for females are generally higher than for males, although the consultation rates at the extremes of age (i.e. the very young and the very elderly) were similar for males and females.

There has been considerable interest in whether or not use of NHS services differs by socioeconomic group. Recent evidence from studies showing the distribution of use of different levels of service find that use of GPs is broadly equitable but that specialist treatment favours the more affluent. Recent microstudies of cardiac surgery, elective surgery, cancer care, preventative care and chronic care support the findings of an earlier review that use of services was higher relative to need among higher socioeconomic groups (Dixon et al. 103). We explore these issues below.

In cities, health services are more widely available than in the countryside, where GPs are often the only providers of care. With highly mobile populations and a plentiful supply of doctors in cities, the prevailing regulations for access and use of services are more difficult to maintain. It is also more difficult to control access and, thus, opportunities for inappropriate use are greater, as reported by Boerma et al. 104 The implications of this for referrals are ambiguous: first referrals may fall if patients approach emergency care more often in cities, but GPs in cities may feel under more pressure to refer if their patients can more readily seek emergency care or move to another practice. Hospital characteristics, such as unoccupied hospital beds in a particular area and consultant-to-consultant referrals restrictions, are shown to register an increase in GPs referrals to hospitals in some areas. 105

Ringberg et al. 100 show that in Norway, male GPs and specialists in family medicine refer significantly less frequently to secondary care, but that the latter refer more frequently to radiological examination. They conclude that high referral rates to secondary care by GPs are associated with GPs’ sex and specialist qualifications in family medicine. Evans et al. 106 examine how peer review with consultant engagement may influence GPs and improve the quality of referrals. van Dijk et al. 107 looked at the variation in the referral rate in primary care and if more provision of services can influence referral rate. They found that practice-level services were not associated with the referral rate. Noone et al. 108 studied the variation in referral rates among practices after a new hospital opened in the Oxford region in order to analyse how referral rates may increase after a positive shock to the supply.

Madeley et al. 109 found no difference in referral rates between single-handed GPs and those in partnerships in Lincolnshire. However, in a multivariate study, Christensen et al. 110 found a significant association between single-handed practices and high referral rates in Nottinghamshire. Srirangalingam et al. 111 surveyed GPs to determine whether or not implementation of the General Medical Services (GMS) contract has led to changes in referrals to a secondary care diabetes mellitus clinic and found no significant change in numbers of referrals received pre and post contract. We shall return to both of these issues below.

There is a broad literature on variation in referral rates between practices, although the samples used are usually quite small. In contrast, there are very few studies about the variation of treatment rates between practices on a national basis, and none has linked the analysis of these critical building blocks of the relationship between GP and specialist. This is partly due to obstacles in obtaining appropriate data and also to the difficulty of linking/measuring the hospital admissions of the patients of specific practices.

The remainder of the chapter is structured as follows. The next section describes the data. The empirical strategy and results are presented in Empirical strategy and hypotheses and Results, respectively. A conclusion then follows.

Data

We use three blocks of data that come from different sources referring to practices, patients and hospital admissions, for the period 2004–12. The GP data are primarily from the HSCIC, under its previous guise as ‘NHS Connecting for Health’, and relate to current GPs, including the practice at which they work, and the start and end dates of employment from 2004 to 2011. These data refer to all GPs in work, not only in a general practice, but also employees of a PCT, SHA or an OOH service/WIC. Unfortunately, these data do not provide information beyond headcount about the hours of work for each GP, and, specifically, the FTE labour supply. However, data from the NHS Annual Work Force Census of NHS employees provide practice-level information about work hours.

Additionally collected data provide a rich information base for each practice: the size of the patient list, the total FTE GP hours, the total number of GPs, the sex of GPs, if they have studied outside the UK, their years of qualification and birth, and the their type of contract [salaried, provider (partner) or in training]. There is further information regarding the practice, including the practice location, whether or not it is single handed and the type of contract [GMS or Personal Medical Services (PMS)] agreed with the PCT. For each practice, we also link individual-level patient information. In the UK, each individual must be registered with only one general practice. Patients’ information at each practice includes their age structure (divided in three groups: 0–34 years, 35–59 years, ≥ 60 years), sex mix and the number of patients resident in each LSOA.

As we know the residential location of the patients, we can calculate indices of deprivation and rurality for each practice. Socioeconomic status level is measured using the deprivation domains of the English Indices of Deprivation. 90 The general practice IMD is estimated by taking a weighted average of the IMD scores for each LSOA in which a given practice has registrations. We have done the same to calculate the index of rurality for each practice.

Patient demand for admissions is likely to be shaped by patient illnesses. Clinical information concerning the prevalence of specific diseases for the patients registered with the practice, and quality care of practice, is provided by the QOF database. The data cover almost all GP practices in England and are extracted from disease registers submitted to the national QMAS. The prevalence of all illnesses except epilepsy and asthma increased during the study period. By the end of the period, 90% of the population was registered with a GP, and the observed illness prevalence is quite representative at the England level. However, there could be some under-reporting bias, especially in the early years. In addition, for each practice there is a score from 100 to 1000 about the quality of care. The QOF, which was introduced in 2004, requires that some patients with conditions such as chronic kidney disease and angina are routinely referred for specialist review, which will inevitably generate a higher number of referrals from primary care. However, these will be appropriate referrals that reflect adherence to agreed protocols and/or best practice. For example, one study has shown an increase in referrals for poor glycaemic control after implementation of the new GMS contract to a secondary care diabetes mellitus clinic; see Srirangalingam et al. 111

To estimate the models of GP referrals and specialist treatments, we have used data from the HES database. These data provide information concerning all inpatients and outpatients admitted to the English NHS, as well as A&E attendances. It includes private patients treated in NHS hospitals, patients resident outside England and care delivered by TCs (including those in the independent sector) funded by the NHS. Admitted patient care data are available for every financial year from 1989–90 onwards. Each HES record contains information about individuals admitted to NHS hospitals, including clinical information about diagnoses and operations; patient characteristic information, such as age group, sex, ethnicity, registered practice; administrative information, such as time waited, dates and methods of admission; and geographical discharge information, such as where patients are treated and area of residence. Because the focus is on GP influence upon admissions, our analysis concerns only the ‘first admission’ to the hospital, which the GP is most likely to influence, rather than subsequent admissions for continuing treatments. Moreover, we focus on only those referrals made by GPs and we link these referrals with treatment data (inpatients).

In 2011/12 there were 91.0 million outpatient appointments, of which 72.6 million (79.8%) were attended. Although the total number of appointments has increased year on year, the percentage that are attended (79.8%) has remained relatively stable since 2007–8, decreasing by 1.8 percentage points. First appointments accounted for 21.8 million attendances (30.1% of all attendances), of which 11.3 million (51.5%) were referred from a GP. London SHA had the greatest number of appointments, at 18.5 million appointments, compared with the North East SHA which had the lowest number of appointments (4.9 million). London SHA also had the highest rate of appointments (by population), with an average of 2.25 per person, and South Central SHA had the lowest at 1.34 per person. The main consultant specialties with the most attendances were Trauma & Orthopaedics (7.1 million; 9.8%), Allied Health Professional Episodes (6.5 million; 9.0%), Ophthalmology (6.3 million; 8.6%) and General Surgery (4.1 million; 5.7%), which together made up one-third (33.1%) of all attendances.

Three stylised facts

A description of the distributions of the rates of (1) first referrals and (2) planned hospital admission rates following first GP referrals for English GP practices in 2011 is given in Figure 41. The referrals rate has a mean of 201 per 1000 registered patients, and the planned admission rate from these referrals is 72 per 1000 registered patients. There are also planned admissions following referrals by hospital specialists, in the case of emergency attendance, and nurses at medical centres, which are excluded from this study.

FIGURE 41.

Admission vs. referral rates (per 1000 patients) across practices, 2011.

Fact 1: we confirm, for the first time in a national sample, that the variation in referral rates is considerable, and that using the SD of referral rates as the metric of variation suggests that variation has risen since 2004. The practice referral rate had a SD of 40.4 in 2004 and 54.4 in 2012. However, the mean referral rate also rose during this period so that the coefficient of variation remained unchanged at 0.27.

Fact 2: although referral rates vary considerably by practice, following a GP first referral at practice level, the number of planned hospital admissions per 1000 patients exhibits considerably more variation. Moreover, the practice admission rate shows increased variation since 2004. The coefficient of variation in 2012 was 0.27 for referrals, and 0.59 for elective admissions following GP first referral, having increased from 0.45 in 2004.

Fact 3: the joint distribution of practice first referral rates and subsequent hospital admission rates has a bimodal structure. There is a concentration of practices around 200 referrals per 1000 patients and 50 admissions per 1000 patients, and a weaker second concentration around a slightly lower referral rate, but with a considerably higher admission rate that is slightly > 100 per 1000 patients. The differences in the health of patient populations between practices readily explains why both referrals and treatments should be higher at some practices, and we should expect random variation in both variables. However, it is far from clear why at some practices, admissions are so much lower, given their referral levels, or, put differently, why referral rates are similar when the practice treatment rates are so different. If it is correct to assume that the practices with less healthy patient populations have higher treatment rates, we can ask why many practices with these unhealthy populations have low patient referral rates.

There are two contrasting explanations for the bimodal relationship found in Figure 41. First, the demand for admissions may present itself differently across practices. GP referral behaviour may follow guidelines/protocols which result in a narrow referral range despite quite different patient populations, or patients in different areas may seek GP advice at different stages of disease. This may inhibit referrals in some areas until patients’ illnesses are more advanced and treatment is more likely. Second, it may be that supply differs between areas and that providers vary in treatment thresholds despite referral patterns being otherwise quite similar.

Figure 41 also suggests that a careful model of factors explaining elective treatment rates across practices, other than GP first referrals, is crucial if the relationship between referral and treatment rates is to be estimated accurately. Figure 42 gives the same cross-sectional portrait as Figure 41 but without the indication of point density and is a reminder of the difficulty of using scatterplots with very large samples.

FIGURE 42.

Admission vs. referral rates (per 1000 patients) across practices, 2011.

Figure 43 gives the distributions of practice referral rates, distinguishing between referrals for which the patient is given a second appointment and those that are ended after one appointment. About two-thirds of referrals are followed up by specialists with further appointments. Although this is not to argue that only those followed up are of high value, it is probable that those followed up are, on average, more likely to be the referrals that lead to larger patient health gains net of cost being provided by the secondary sector. The distribution of referrals that are discharged is less variable, but there is a long tail of practices with non-followed up referral rates in excess of twice the national mean rate of about 50 per 1000 patients.

FIGURE 43.

Outcome of referrals (1000 per registered patients) by practice: discharge rates vs. follow-up rates.

In Table 39 we report the descriptive statistics by year of the variables that we include in our estimated models. The time trend of the first referrals made by GP practices shows steady growth, and the resulting treatments from first referrals increased until 2008 and subsequently declined until flattening out after 2010. There is considerable variation across general practices in referrals and treatments, as we can see from the SD and Figure 41. However, although the coefficient of variation scarcely changes for referrals, the coefficient of variation for treatments increased from 0.45 to 0.59 in only 8 years and most of this happened after 2006 when system reforms, including the new GP contract, initiated change.

TABLE 39 - Descriptive statistics of variables used in the study

Ln, natural log.

Variables description	2004	2005	2006	2007	2008	2009	2010	2011	2012
First referrals per 1000 patients by GPS (SD)	153.61 (40.8)	162.4 (41.15)	156.89 (39.35)	164.07 (41.33)	183.31 (45.52)	196.74 (50.61)	200.7 (51.86)	200.88 (51.39)	205.05 (55.76)
Related treatments per 1000 patients (SD)	75.16 (33.57)	75.82 (34.58)	75.13 (34.67)	72 (34.27)	75.04 (38.87)	74.42 (39.64)	71.98 (40.53)	71.92 (41.94)	71.93 (42.38)
Proportion of beds occupied at PCT	0.8	0.81	0.82	0.81	0.81	0.82	0.82	0.79	0.8
Proportion of patients aged 0–34 years	0.44	0.44	0.44	0.44	0.44	0.44	0.44	0.44	0.44
Proportion of patients aged 35–59 years	0.35	0.35	0.35	0.35	0.35	0.35	0.35	0.35	0.35
Proportion of patients aged ≥ 60 years	0.21	0.21	0.21	0.21	0.21	0.21	0.21	0.21	0.21
Proportion of female patients	0.50	0.50	0.50	0.50	0.50	0.50	0.50	0.50	0.50
Index deprivation by practice	23.47	23.4	23.42	23.33	23.35	23.37	23.2	23.25	23.23
Rural indicator per practice	5.33	5.34	5.33	5.32	5.33	5.32	5.34	5.33	5.35
Ln QOF points	6.88	6.92	6.87	6.88	6.86	6.85	6.86	6.88	6.87
Mean patient list per practice	6664	6720	6794	6871	6918	6992	7042	7095	7193
Total FTE hours per practice	3.29	3.38	3.76	3.88	3.91	4.02	3.91	3.92	3.91
Proportion of single-handed practice	0.20	0.20	0.18	0.16	0.14	0.13	0.12	0.12	0.12
Mean age of GPs (years)	48.36	48.84	48.75	48.79	48.77	48.88	49.01	49.07	49.22
Proportion of FTE female GPs	0.3	0.31	0.34	0.35	0.36	0.37	0.37	0.38	0.39
Proportion of FTE immigrant GPs	0.29	0.3	0.31	0.3	0.31	0.31	0.31	0.31	0.31
Proportion of GPs working FTE ≥ 1	0.82	0.82	0.88	0.84	0.8	0.78	0.7	0.67	0.67
Proportion of GPs working 0.6 ≤ FTE < 1	0.18	0.18	0.06	0.08	0.09	0.12	0.16	0.16	0.18
Proportion of GPs working FTE< 0.6	0.00	0.00	0.06	0.08	0.11	0.11	0.14	0.16	0.15
Years of education of GPs	18.86	18.86	18.86	18.85	18.83	18.82	18.81	18.8	18.79
Proportion of GPs provider	0.92	0.93	0.91	0.87	0.84	0.83	0.82	0.83	0.83
GMS contract	0.57	0.57	0.57	0.57	0.57	0.57	0.58	0.58	0.58
Total observations	8364	8355	8259	8216	8167	8125	8241	8174	7951

We also observe that registered patients per practice increased by about 8% over the period 2004–12 and the share of single-handed practices decreased from 20% in 2004 to 12% in 2012. The percentage of female GPs has increased from 30% to 39% and this has consequences for the average hours worked: the percentage of full-time GPs had declined from 82% to 67%. The percentage of partner GPs has decreased from 92% to 83% in the period 2004–12, whereas the mean age of GPs has increased by almost 1 year during this period.

Empirical strategy and hypotheses

This work aims to improve understanding of how policy might influence the level of hospital admissions by studying two central and inter-related aspects of the relationship between GPs and specialists: first, we study the determinants of the rate of practice first referrals; and, second, we study the determinants of the probability that a patient given a first GP referral is admitted for hospital treatment, which can be affected by exogenous influences that are linked to policy as well as the total rate of first referrals from the referring practice. By studying these two decisions separately, we are able to decompose the effect of a variable of interest on the likelihood that a registered patient is admitted for hospital treatment into a part arising from its influence on the probability of a referral, which is decided by the GP, and a second part influencing the probability that the specialist decides to admit the referred patient. Thus, for example, we give further insight into why patients from deprived areas may receive too little secondary care relative to need.

Our empirical strategy consists of two steps: first, we estimate a model of how the characteristics of general practices, GPs and their patients influence the rate of referrals. This is the first paper to provide this evidence for a comprehensive set of controls, for almost all of the practices in England and for a panel data set covering a long period.

We estimate a model of the form:

where lnR_it represents the logarithm of the number of first referrals at each practice i in each year t per 1000 patients. Referrals made by others (specialists, nurses, etc.) are excluded to ensure that we capture the effects of various primary care characteristics on GP referral decisions and the subsequent hospital admissions. GPX_it is a vector of GP and practice characteristics that are time-varying at practice i in time t. The GP characteristics include the proportion of female GPs, mean practice GP age, ethnicity of GPs, type of GP contract (provider/partner vs. salaried) and practice characteristics (single-handed practice, QOF score and GMS/PMS contract), and θ is a vector of the slope effects of these variables.

The term Pat_it represents a vector of patients’ characteristics at practice i in year t (the share of patients in each of several age categories, sex share, rural practice and the deprivation index). The terms z_it capture the proportions of patients registered at each practice with each of several chronic diseases indexed i (asthma, cancer, chronic obstructive pulmonary disease, coronary heart disease, diabetes mellitus, epilepsy, hypertension, hypothyroidism, left ventricular failure, mental health and stroke). Finally, we control for time (µ_t) with year effects, HA (λ_i) and practices (σ_i) with fixed effects. We use panel data fixed effects to estimate our model. In panel data, individuals (cities, firms, etc.) are observed at several time points (e.g. days, years, etc.). Panel data are most useful when we suspect that the outcome variable depends on explanatory variables which are not observable but correlated with the observed explanatory variables. If such omitted variables are constant over time, panel data estimators allow the consistent estimation of the effect of observed explanatory variables. The main advantage of working with panel data is that we are able to control for practice-specific, time-invariant, unobserved heterogeneity, the presence of which could lead to bias in standard estimators such as OLS. The fixed-effect component, σ_i, captures unobserved heterogeneity across practices, which is fixed over time. In addition, we correct the standard errors in order to control for heteroscedasticity. In practice, the idiosyncratic errors, ξ_it, are often serially correlated when T > 2. Bertrand et al. 112 show that the usual standard errors of the fixed-effects estimator are drastically understated in the presence of serial correlation. It is therefore advisable to always use cluster-robust standard errors for the fixed-effects estimator. However, cluster-robust standard errors have no impact on the estimated coefficients but do have an effect on standard errors. Therefore, potential bias in the estimated coefficients remains.

We also estimate a practice-level model of the probability of elective admissions of first referral patients. To specify this model, we begin by assuming that the probability of a referred patient from practice i being admitted is determined by the various (patient, practice and GP) characteristics associated with the practice, as well as by the practice rate of referrals. An increase in the practice referral rate is hypothesised to reduce the probability of admissions if the practices with higher referral rates refer at the margin those patients whose expected health value of treatment is lower. In this case, if the proportion of referred patients who are admitted from practice i in year t is g_it, and the practice referral rate per registered patient is R_it, then the treatment rate per registered patient T_it may be estimated in the following model:

where z_it, P_it and GPX_it are exogenous influences on the proportion of referrals being admitted. If the proportion being treated, g_it, declines with increases in the practice referral rate, in a multiplicative separable relationship, then we have:

We assume a semi-log specification for γ, so that a unit change in R gives β% reduction in γ, the probability of admission. Thus, γ = R^-β and hence in Equation 29,

If we cannot reject the hypothesis that β = 0, the likelihood of treatment for patients from practices with higher rates of referrals, all else being equal, is not significantly different from patients referred from smaller practices. The greater β is, the greater the fall in the probability of treatment for patients from practices with higher levels of referrals. If we assume that g is an exponential function of its determinants, then in Equation 30, the expression for admissions takes the form:

To examine whether, and to what extent, the treatment rate is declining with the referral rate, we test whether or not β is significantly greater than zero and interpret the size of β. Using this formulation, we estimate a model of the elective admissions that follow a GP referral, using IV panel data estimation at practice level, and controlling for the level of referrals, and a group of GP- and patient-specific characteristics as described above, which may influence g, the probability that a referral is treated. The estimated model is: