Equity of utilisation of cardiovascular care and mental health services in England: a cohort-based cross-sectional study using small-area estimation

Inverse care law and the NHS: a ‘prevailing paradigm’

Literature that yields insights into differential access to NHS services has been conceptualised and empirically studied by researchers from a wide range of different disciplines, giving rise to a conflicting body of research. Indeed, ‘access’ (which may refer to need, provision or utilisation of health services in empirical research) is not always the primary outcome of relevant studies. As a result, systematic reviews can overlook evidence, the existence of which is not immediately clear from the title or abstract. 2

There is, moreover, evidence that researchers have made key assumptions (explicit and implicit) about the nature of access to and use of NHS services, which, in turn, affect research design. According to Dixon-Woods et al.,3 the inverse care law has come to operate as:

a meta-narrative: a distinct research tradition invoking specific sets of ‘normal science’ assumptions that has developed as a storyline over time. Within this tradition, there has been ongoing concern to demonstrate that socioeconomically disadvantaged people are disadvantaged in their access to health care [our emphasis].

If inverse care has come to function as a ‘taken-for-granted’ assumption, it is important to recognise that this may have shaped researchers’ choices about the most important questions to raise, the study designs and methods best suited to answering those questions, and even the way in which they interpret their results. 4 It is also important to accept that, where a set of assumptions guides commonplace views about a particular topic, results that challenge those assumptions can be hard to publish. Against this background, the task of interpreting previous evidence on variations in access to and use of NHS care is not straightforward. The strategy we have used to review existing evidence is described in Chapter 2 (see Review strategy).

Health-care equity: is it still a normative goal of the NHS?

To further complicate matters, not everybody even agrees that the principle of ‘health-care equity’ (whereby health-care resources should be geographically distributed to ensure ‘equal opportunity of access to health care for people at equal risk’) is the most important principle of the NHS. Since 1999, the NHS has also been charged with contributing to the reduction of avoidable inequalities in health, or the promotion of ‘health equity’ (a ‘vertical’ definition of equity).

It was only in the 1980s that a sustained critique developed in the UK of the ‘equal access for equal need’ principle, first because such ‘horizontal’ definitions of equity were considered to be mutually incompatible – equality of expenditure for equal need not necessarily translating into equity of access or equity of treatment5,6 – and, second, for failing to consider need in terms of capacity to benefit, an approach that (a) promoted ‘efficiency’ and (b) lent itself to the more ‘ethical’ objective of achieving health equity or equality in terms of health outcomes. 7–10

In 1997, Mooney and Jan11 observed that ‘vertical’ equity considerations had tended to be overlooked in the health policy literature. Reviewing both literature and policy developments since, the reverse would appear to be true. Compared with a substantial body of literature exploring the case for distributing health care so as to secure a more equal distribution of health,6–28 it is today quite difficult to find any literature that asserts the moral and philosophical case for health-care equity. 29,30 Indeed, the concept has been roundly attacked: ‘Slavish devotion to time-honoured principles of allocation in proportion to need and equal universal access should be avoided. They are misleading principles’;8 ‘[T]he equal access for equal need objective can be seen as a piece of grand or flamboyant rhetoric of symbolic politics, representing a misreading, or at least an oversimplification, of history’. 31

This raises the question of whether or not health-care equity is still a normative goal of the NHS. If it is not, then the overall purpose of this project – to examine equity in the utilisation of cardiovascular and mental health services in England – might be considered to be of little policy relevance. Yet we concur with Sen32 that accepting the importance of health equity does not mean that the relevance of other claims such as non-discrimination in the delivery of health care should be denied. We also propose that, although the principle of universalism has been criticised in recent years,33 there remains a public understanding that access to key services such as the NHS are a ‘right’ in the ‘just’ society. 29,34 This suggests that, although they receive relatively little attention in contemporary academic literature, traditional conceptions of social justice around notions of equality, social rights, non-discrimination and universalism35,36 have not been entirely abandoned.

There are, however, debates about the extent to which public perspectives on ‘fairness’ reflect Fabian notions of equality and universalism or targeted approaches to welfare. Although several studies suggest that the public supports the understanding that the NHS should provide care from cradle to grave and respond to the ‘rule of rescue’ (i.e. giving priority to those in immediate need),28,37–40 others note that public preferences suggest a willingness to give higher priority to some categories of people than others, including those who are socioeconomically disadvantaged. 41–54 Interestingly, persistent media storylines about inequalities in health-care equity may have played a role in this. The narrative is as follows: if the ‘inverse care law’ characterises access to NHS services, then people living in deprived areas are disadvantaged not only with respect to their health outcomes but also in their access to health care. According to compensatory principles of social justice, this makes a strong case for positively targeting NHS resources at the poor.

A shifting policy context

Policy-makers also appear to have accepted that English NHS resources should be positively targeted at deprived areas in order to address the dual aims of promoting health equity and addressing unmet need. Responding to academic interest in the potential for incorporating equity weights into cost-effectiveness analysis,55–57 NICE recently set out proposals to incorporate social value judgements in its methods for appraising health technologies. These included an attempt to capture the difference between people’s relative production and consumption of resources, through, for example, a treatment enabling a patient to return to work and thereby pay more tax or contribute more to family finances. 58,59

The NICE was responding to early criticisms that measures of cost-effectiveness such as cost per quality-adjusted life-year (QALY) do not always account for what matters to patients, their families and society. In a major review of oral and written evidence from patients, patient organisations, pharmaceutical companies and academics, the House of Commons Health Committee had noted that ‘Many witnesses argued that the exclusion of societal gains compromised the validity of QALY-based cost-effectiveness calculations’ (contains Parliamentary information licensed under the Open Parliament Licence v3.0). 60 NICE was undoubtedly placed in a difficult position. 61 That said, its willingness to so explicitly favour the young, using a rationale of ‘productivity ageism’,62 does raise questions about the extent to which the principle of non-discrimination continues to have resonance among senior policy-makers.

The approach taken to NHS resource allocation has strongly responded to the policy requirement to reduce inequalities in health,63 recent research concluding that increasing the proportion of resources allocated to deprived areas between 2001 and 2011 was associated with a reduction in absolute health inequalities from causes amenable to health care. 64 In fact, this is open to interpretation. As the authors themselves acknowledge, one cannot rule out the possibility that the associations they observed were attributable to other factors that pose health risks. More problematically, although absolute differences in health between areas reduced, relative inequalities remained constant. As absolute declines will always be greater where the starting point is higher (put simply, a 10% decline in a mortality rate of 40 per 100,000 will be greater than a 10% decline in a mortality rate of 20 per 100,000), it is not at all clear that additional investment led to a reduction of health inequalities. Finally, the conclusions conflict with those of a recent study that found no relationship between Quality and Outcomes Framework (QOF) achievement and reduced mortality, which appeared to be driven by factors outside primary care such as declining rates of smoking. 65 Thus, the jury is still out on whether or not health care can address inequalities that are fundamentally embedded within the unequal structures of society. 32,66–68

In 2013, the assumption of unmet need became a primary justification for continuing to positively target NHS funding at deprived areas. The stimulus for this was the introduction of separate budget streams for NHS England and Public Health England (PHE), which threatened to reduce the funding of more deprived Clinical Commissioning Groups (CCGs), who stood to lose previous health inequality adjustments to local authorities (LAs). To prevent this from happening, the finance director of the NHS made assurances that areas with worse outcomes would not receive less NHS funding, as the new formula would adjust for a health economy’s unmet need, where life expectancy suggests that people are not accessing health services. 69 Against this background, it is important to explore whether or not evidence of socioeconomic bias in access to care supports this assumption.

If evidence suggests that there is a pro-rich bias in the access to and utilisation of NHS care, it is interesting to ask why this is the case and what policy-makers can further do to address inequity. This is because highly deprived areas have already benefited from very high allocations (Table 1). Although the budgets given to inner-London primary care trusts (PCTs) distort the national picture, the figures in Table 1 have been symptomatic of a pattern in which young, deprived populations with lower crude but high standardised rates of illness and death have received and spent significantly higher NHS allocations than their older, more affluent counterparts (which have higher crude but lower standardised illness rates).

The profound redistribution towards deprived areas that occurred after 2002 was largely due to a technical flaw in the way in which the ‘AREA’ (Allocation of Resources to English Areas) formula76 was implemented. Rather than proportionately reflecting the effects of age and deprivation in influencing health service need, the formula effectively cancelled out the effect of age through the sequential inclusion of deprivation indicators. 77–79 Although subsequent formulae attempted to address this,80,81 the overall distribution of NHS resources has remained largely unchanged, in part owing to the inherent circularity of utilisation-based formula,82–85 and in part because, for mainstream policy-makers, allocations made with respect to deprivation appear to have looked ‘about right’. 86 This very positive targeting of funding does raise questions as to whether or not and why Tudor Hart’s ‘law’ should continue to be a defining feature of the NHS.

Shifting patterns of morbidity: implications for working at the ‘deep end’

There are, however, a number of factors that may account for the durability of inverse care. The first relates to shifting patterns of morbidity. The reasons why deprivation has been strongly correlated with standardised but not crude morbidity and mortality are that (a) for most conditions (mental health being a notable exception), demographic gradients in health have been steeper than socioeconomic gradients87 – in other words, as people get older, they are more likely than young people to develop conditions such as heart disease and cancer; and (b) there is a negative correlation between the geographical pattern of social deprivation and age in England, with deprived areas tending to have younger populations and affluent areas tending to have older populations.

Yet disease patterns are changing. Rising rates of obesity among children and younger people are already altering the epidemiology of chronic diseases such as diabetes and coronary heart disease (CHD), and the prevalence of obesity is positively associated with deprivation. 88–92 There has also been a significant increase in levels of anxiety and depression in adolescents93 and in the number of people seeking treatment for mental health problems. 94 Between 2008–9 and 2010–11, the number of registered patients suffering from depression increased by 11.5% nationally and the number of prescriptions written for antidepressants rose by one-fifth. Psychological distress is associated with both low socioeconomic status and unhealthy behaviours such as tobacco and alcohol consumption,95 which, in turn, increase the risk of physical health problems such as CHD, hypertension and diabetes. There is strong evidence of a relationship between comorbidity and socioeconomic status, with people in deprived areas having a higher prevalence of both physical and mental health disorders96–98 and incurring higher health-care costs. 99

The significance of these trends is that they are particularly likely to be managed in general practice. In 2006, the Royal College of General Practitioners (RCGP) estimated that up to one-third of the 280 million consultations in primary care annually have a significant mental health component. 100 RCGP Scotland has been particularly vocal about the challenges of working at the ‘deep end’, where patients suffering from physical, emotional, psychological, financial and social problems, including problems related to substance misuse, present additional demands. 101 Given the level of demand, it is argued that there is a mismatch between need and resource, GPs having insufficient time to get to the bottom of their patients’ problems. The RCGP102 is one of a number of important bodies, including The King’s Fund,103 to have argued that there are fewer GPs, relative to need, in areas of deprivation.

Notwithstanding concerns about the medicalisation or ‘public healthification’104 of social problems and life events,105,106 there is no doubt that highly deprived populations can have complex and demanding service needs. However, the quantification and, by implication, understanding of the gap between needs and provision in deprived areas has been problematic. For example, the 2010 White Paper Healthy Lives, Healthy People107 reproduced two maps of Birmingham. The first showed the prevalence of CHD in 2008–9, according to GP QOF data. The second showed mortality (standardised) from CHD in 2007–9, according to data from the Office for National Statistics (ONS). On both maps, deprivation is indicated. The title of the figure suggests ‘stark differences’ between the locations of CHD patients who have been diagnosed by their GPs and the locations where CHD mortality is highest, implying inverse care. In fact, a comparison of QOF prevalence with crude CHD mortality shows a good fit, suggesting that GPs in deprived parts of Birmingham are providing good access to diagnosis for those at risk of heart disease.

Similarly, the widely cited ‘Darzi Report’ based its analysis of inverse care on the ratio of GPs per head of AREA-weighted population. 108 This found that areas such as Mid-Devon PCT had over twice as many GPs per head of weighted population as Oldham PCT. Subsequent analysis using condition-specific QOF prevalence rates as denominators drew rather different conclusions. 109 Thus, for some conditions, such as chronic obstructive pulmonary disease (COPD) and mental health, GPs in deprived areas did indeed face much higher workloads than those in affluent areas. For other conditions, including cardiovascular disease (CVD) and asthma, there was no significant difference, while with respect to cancer the pattern appeared to be reversed. Once age was factored in, the picture became more complex; thus, GPs serving demographically older populations – both affluent and disadvantaged – had higher caseloads of CVD, COPD, cancer, dementia and chronic kidney disease (CKD) than GPs working with young deprived populations. It was nevertheless the case that the worst problems arose where deprivation and demography reinforced one another. Such practices, which tended to cluster in the northern cities and some rural and coastal areas, had the highest workloads of all.

Review framework

As noted in Chapter 1 (see Shifting patterns of morbidity: implications for working at the ‘deep end’), evidence of the direction of inequality in access or use can vary according to the method used to establish denominators of health service need. This is not always acknowledged in reviews of evidence. Indeed, very little attention has been paid to the possibility of statistical artefact. We therefore propose that, in reviewing evidence of variations in cardiovascular care and mental health, studies are examined for methodological quality.

We decided to restrict the search period to 2004 onwards. This is, in part, in acknowledgement of key policy changes in the NHS, not least with respect to the distribution of funding. It also reflects our assessment of significant methodological progress in research over the past decade or so, particularly in developing more direct measures of disease prevalence which, when applied to the investigation of health-care equity, may address the fundamental difficulty of disentangling legitimate need from other influences (see Methodological issues). Furthermore, devolution of the NHS has resulted in significant differences in the health systems of the home countries. 114 Consequently, we have excluded evidence that focuses exclusively on Scotland, Wales or Northern Ireland. As Scottish researchers have shown an active concern with ‘inverse care’,96,115–121 it is important to acknowledge that different search decisions (such as admitting international evidence) are likely to have yielded different conclusions.

In addition to methodological concerns, some commentators have questioned whether the focus on socioeconomic dimensions has given rise to an incomplete account of inequalities in access to and use of NHS services. 109 The key concern is that, owing to the ‘ongoing concern to demonstrate that socioeconomically disadvantaged people are disadvantaged in their access to health care’,3 other dimensions of inequality may have been overlooked. In their impressively wide-ranging review of evidence, Dixon-Woods et al. 3 noted a relative dearth of research evidence on ethnic variations in health service use and of access to health care during childhood and older age, and the relative neglect of sex as a dimension in research on access to health services. This does not accord with the findings of the Centre for Reviews and Dissemination (CRD) scoping review,2 which suggested that a larger number of studies on cardiac care had focused on inequalities by age and sex than focused on socioeconomic status. Given this uncertainty about the possibility of research and, indeed, publication bias (inconclusive or negative results being less likely to appear in the published literature than evidence of inequality), we propose structuring the review around different dimensions of inequality: socioeconomic, sociodemographic (sex and age), ethnic and geographical.

With exceptions,122,123 interpretation of variations in access and use is often insufficiently nuanced. For example, lower-than-expected rates of specialist interventions such as coronary artery bypass grafts (CABGs) and percutaneous transluminal coronary angioplasties are usually interpreted as a ‘bad thing’. An alternative explanation is that the health of populations exhibiting lower-than-expected rates of secondary care is being adequately managed in primary and community settings. 124 It therefore becomes important to establish exactly where on the care pathway inequalities arise.

Several authors propose that the model ‘candidacy’ should be used to conceptualise access to health care at different points. 3,113 This captures the idea that people need to recognise their eligibility as candidates for health care and then have their candidacy assessed and acted on (adjudicated). The approach recognises that variations in treatment arise from interactions between supply and demand and accommodates the fact that barriers to access can occur across the pathway to care. For example, the decision to seek help in the first place may be influenced by individual patients’ knowledge, information, their evaluation of the seriousness of their problem, their judgement of the ability of the health service to respond, psychological factors such as lack of embarrassment or fear, and practical issues such as the need to rely on public transport or to arrange childcare/time off work. Once patients have gained entry to the system, the categorisation and disposal of their health needs depends to some extent on their ability to present in ways that health professionals find credible and legitimate. In turn, the way in which health professionals categorise health needs may be affected by their perceptions of patient preferences, technical eligibility and moral or social ‘deservingness’. Capacity factors (such as length of time available for consultations, availability of tests and perceptions and/or experience of poor local capacity) may also play an important role in shaping clinicians’ decisions to open up the pathway to treatment.

Concepts of candidacy and adjudication are not easy to operationalise in a review of largely quantitative research. Thus, in order to examine how variations in access may manifest at different points on the care pathway, we have structured the reviews of primary studies around presentation, primary management and specialist treatment.

Finally, reviews of existing research (see Review evidence of variations in access to and use of NHS care) suggest that there is uneven evidence with respect to clinical condition. There has been a strong concentration on cardiac care and, within this specialty, on determining whether access or use varies according to socioeconomic status. Evidence in other areas appears to be weak, though there has been a focus on ethnic variations in use of mental health services. Thus, although the decision to focus on cardiac care and mental health was made at the outset (with a subsequent decision to broaden the first category to cardiovascular care), we nevertheless propose that these present good contrasting areas with respect to existing evidence.

Review strategy

We have essentially drawn together four sets of reviews. These are (a) summaries of existing review studies, (b) evidence of geographical variations in access to and use of NHS care, (c) primary studies of variations in access to and use of cardiovascular care and (d) primary studies of variations in access to and use of mental health services. Our aim was to retrieve articles or reports that provided quantitative evidence of variations in access to and use of cardiovascular care or mental health services in England (though qualitative evidence that shows a clear direction of inequality has been admitted).

Early studies of variations in access and use

Until the 1990s, research on inequalities in access to and/or use of NHS care was hampered by the lack of readily available indicators of need. With the exception of bespoke surveys, such as the British Regional Heart Study (BRHS) (see Health survey data), studies were limited to examining variations in general utilisation. The General Household Survey (GHS) was used by several researchers to this end. 167–171 As this collected evidence on self-reported morbidity [acute sickness and limiting long-term illness (LLTI)], use of primary, outpatient and inpatient care and socioeconomic status, it allowed ratios to be calculated for the relationship between need and use by different socioeconomic groups. Different studies using the GHS nevertheless drew different conclusions about whether there was pro-rich or pro-poor bias in use of care. This mainly reflected the way in which researchers chose to adjust for differences in morbidity.

By the 1990s, there was a growing interest in undertaking what Dixon et al. 136 refer to as ‘micro-studies’, or studies of variations in access to particular services. In some cases, individual-level data were used to link ‘need’, usually defined in terms of initial presentation (diagnosis or admission), to subsequent use of investigations or treatment. 172–175 At the time, however, national sources of data lacked sociodemographic or residential details of patients or were insufficiently complete and accurate for use. 176 As a result, most studies of this kind were based on locally negotiated access to data.

Proxies for need: the use of mortality and deprivation

To make studies more generalisable, many researchers turned to area-level proxies of need, most particularly mortality and deprivation. Mortality rates have been widely used in the investigation of inequalities in access and use, particularly to cardiac care, reflecting the fact that mortality data are routinely available, can be decomposed by age and sex and reflect cumulative morbidity. Against this, mortality is a better proxy for diseases where case fatality is high than for conditions such as CHD where mortality statistics will fail to reflect the full extent of non-fatal morbidity and, by implication, health service need. The validity of using mortality as a need indicator may also be affected by social differences in survival rates, studies finding that the least affluent patients tend to experience the worst outcomes. 177–181 This may reflect poorer access and use of health services. However, if factors such as comorbidity, smoking status and obesity are more important in influencing survival outcomes, mortality statistics will illegitimately weight need in favour of deprived groups.

Such objections aside, one fundamental consideration in using mortality to investigate inequalities in health service use relative to need is the importance of ensuring comparability in the expression of the denominator (need) and nominator (use). Unfortunately, research using mortality rates as a proxy for CHD service need has been rather inconsistent in this respect, which means that findings should be interpreted with caution. Some studies145,147 have appropriately explored whether or not the relationship between standardised CHD/ischaemic heart disease mortality and standardised intervention rates varies by deprivation. Others have not directly adjusted intervention rates by need in order to examine variations in use relative to need by population characteristics, but have compared standardised intervention rates with (a) standardised mortality rates (SMRs) and (b) socioeconomic measures separately,146,172,182 while still others have concluded that a lack of association between SMRs and crude intervention rates is indicative of inequity. 183–186 However, for many diseases, there is a poor relationship between crude and standardised CHD mortality. This means that research examining the association between SMRs and crude rates of procedures is not comparing like with like and should not be cited as yielding evidence of the inverse care law.

It would appear that some evidence on inequalities in access to cardiac care – which has been a particular focus for researchers seeking to investigate variations in access to NHS care – has been subject to statistical artefact. This seems to stem from an assumption that standardised measures are synonymous with need for health care. According to Goddard and Smith,122 the weight of evidence relating to the treatment of CHD suggests that ‘admissions, rates of investigation and revascularisation do not match the higher levels of need experienced by the most disadvantaged groups compared with more affluent groups’ (our emphasis). Indeed, the fact that SMRs are subject to profound social gradients has led to the assumption that deprivation itself may act as a valid proxy for health service need. For example, a lack of general association between deprivation and CHD prescribing187 has been interpreted as suggestive of inequity due to an expectation that the most deprived areas have the highest levels of CHD need. Yet because deprived populations also tend to be young populations, they do not necessarily have the highest burden of disease in crude terms. This makes the simple comparison of deprivation scores against crude intervention rates, as in the influential Acheson Report,188 problematic.

Health survey data

As noted above, one advantage of surveys such as the GHS is that they offer a way to directly link self-reported morbidity with health service use at the individual level. Large population surveys are costly to administer. Thus, although some, such as the Health Survey for England (HSfE), are conducted annually, bespoke studies are rare.

An important exception is the BRHS, a prospective survey of middle-aged men drawn from general practices in 24 British towns. Originally recruited in 1978–80, surviving participants continue to be followed up by the University College London team managing the study. The vast majority of the team’s outputs relate to risk factors and outcomes. However, several important papers have been published on inequalities in access to and/or use of care. 189–195

These and other survey-based studies have mixed results. Thus, although Morris et al. 193 reported that men of lower socioeconomic status in the BRHS who experienced angina or myocardial infarction (MI) had a lower incidence of revascularisation, analysis of the Whitehall II prospective cohort study of civil servants found no association between individual socioeconomic position, undiagnosed angina196 and use of cardiac procedures or secondary prevention drugs in relation to need. 144 No statistically significant socioeconomic differences were found in the uptake of secondary prevention medication relative to need in the BRHS192 or the 1998 round of the HSfE. 197 Evidence of ethnic inequalities is also equivocal. In the Whitehall II study, South Asians tended to be more likely to have cardiac procedures and to be taking more secondary prevention drugs than white participants, even after adjustment for clinical need,144 while analysis of four waves of HSfE data found little evidence of ethnic inequalities in the use of primary and secondary health services. 198 Against this, another analysis of HSfE data found that low-income individuals and those from ethnic minorities are more likely to consult their GP but less likely to receive all forms of secondary care. 199

The validity of deriving needs estimates from self-reported survey data has been debated. Reviews of self-rated health research suggest statistically strong associations with future mortality200,201 and many studies comparing health service records with self-reports find high levels of agreement. 202–205 However, there are concerns that the ability and/or willingness of people to make accurate accounts of their health status may differ according to personal characteristics such as sex206 and education. 202,204,207–209 If this is the case, use-to-need ratios could be overestimated for less advantaged groups.

Such considerations make the HSfE particularly valuable. This does not comprise only self-reported health (in some cases using internationally validated instruments such as the Rose Angina Questionnaire) and health care; it supplements self-reported measures with various physiological measurements, such as cholesterol level, blood pressure, lung function tests and waist-to-hip ratio (taken during a nurse visit). Linkage to cancer registrations and Hospital Episode Statistics is expected to make this an even more powerful tool with which to examine relationships between need and use. Against this background, it is not surprising that the HSfE has been used as the data source for several prevalence models that have been developed for the Association of Public Health Observatories (APHO) – now part of PHE – including COPD, diabetes, CHD, hypertension and CVD. 210–215

Administrative data

In addition to proxy measures and survey-based indicators (individual and modelled), researchers examining variations in access and use have turned to existing administrative data, the quality of which has vastly improved in the past decade, to establish denominators of need. Here, the focus is on whether or not patients with a particular diagnosis receive the same level of care, regardless of personal characteristics. For example, a study in Nottinghamshire used admission rates for MI and angina as a proxy for practice-level prevalence. 216 This found that although deprived practices had a higher estimated prevalence of severe ischaemic heart disease, they had lower angiography rates. By contrast, research undertaken in East London which used nitrate prescriptions and admissions for MI as measures of need found no evidence that practices serving more deprived populations had lower-than-expected rates of angiography. 217 The NHS Atlases of Variation also use several denominators based on presentation (e.g. patients with a ST-elevation MI diagnosis, patients admitted with stroke). 151

The use of admission rates as a proxy for prevalence is not without its limitations. This approach is valid only if all patients with a given level of morbidity access hospital services to a uniform extent. Given what we know about socioeconomic and other differences in the relative use of elective and emergency care,218 it is unlikely that this condition is met. If, for example, deprived patients are more likely to be admitted to hospital with more severe disease, then there may be clinically appropriate differences in the subsequent uptake of hospital procedures. The impact of supply on expressed demand should also be acknowledged.

Insofar as QOF data offer an estimate of disease prevalence at the primary level, they may be less likely to be influenced by biases in supply and demand than are data on hospital-level admissions. However, because they can capture only patients who access primary care and who are accurately diagnosed, QOF data are not immune to these problems. If deprived patients are less likely to consult their GPs, or if the availability and/or quality of primary care is poorer in deprived areas, then GPs may be more effective at picking up and recording disease burden among more affluent patients and their families. 219,220 The extent to which these factors affect QOF-derived prevalence estimates is unclear. As noted above, data on primary care utilisation suggest that socially disadvantaged people make as good, if not better, use of general practice as other population cohorts. 193 At the same time, there is a mismatch between, for example, QOF-recorded CHD and APHO-modelled CHD estimates,151 suggesting that there may be random or systematic differences between community and recorded prevalence.

Overview of studies

In total, we found 123 studies that had investigated inequalities in access to and/or use of cardiovascular services. 109,138–140,144,184,185,191–195,198,215,217,219–326 These are described in Appendices 4–14 and summarised here in Table 2. The bracketed numbers refer to the number of studies that found evidence of inequality in access or use (for people of lower socioeconomic status, older people, women, non-white patients and by geography). The total number of counts in the table is higher than the number of studies because, for example, some studies examined more than one dimension of inequality. Although our search is not directly comparable with that of CRD,2 which examined research on cardiac services only, it is interesting to note that more studies were published in the decade 2004–14 (n = 188 counts) than in the previous decade 1995–2003 (n = 105 counts). This suggests that, notwithstanding the overall sway of interest in health inequalities, health-care equity remains a focus of concern among health services and related researchers.

There has, however, been a refocusing with respect to dimensions of inequality. As noted in Review evidence of variations in access to and use of NHS care, the CRD review found that a higher number of studies had explored inequalities with respect to age (30%) and sex (28%) than socioeconomic status (19%), geography (15%) and ethnicity (8%) of 165 counts, excluding studies on undefined ‘vulnerable groups’. 2 In our review, the respective breakdown was 15%, 20%, 35%, 19% and 11% of total counts (n = 187). Thus, we note a significant shift in focus away from sex and age inequalities and towards socioeconomic status and ethnicity.

With respect to type of service, our results are not comparable with those of the CRD study. The latter study categorised studies according to whether they examined primary/secondary prevention, emergency hospital care, tertiary cardiac services, cardiac rehabilitation or other services. In order to capture the concept of candidacy and adjudication at different stages of the care pathway, we have instead categorised studies with respect to a focus on presentation, primary management and specialist treatment of CVD. The studies we retrieved on patients accessing emergency services tended to focus on whether or not there was a relationship between rates of elective and unplanned admissions and quality of primary care. We have thus classed these as primary management studies. Further, we have included studies on cardiac rehabilitation in the group on specialist treatment.

If we group the CRD studies on the same basis (i.e. include studies on emergency hospital care within primary management and studies on cardiac rehabilitation within specialist treatment), there is a similar breakdown of focus. A total of 16% of total studies in our review related to presentation of CVD, 45% related to primary management and 39% related to specialist treatment. In the CRD review, approximately 51% and 49% of studies examined primary management and specialist treatment, respectively. 2

The direction of inequality

The tables of included studies in Appendices 4–14 are colour-coded to indicate (under lead author) whether or not the study found inequality in access/use, with lower rates among those of lower socioeconomic status, older people, women and non-white patients. Evidence of geographical variation is also flagged. Light green indicates evidence of inequality, blue indicates no evidence of inequality and dark green indicates evidence that rates are higher for those of lower socioeconomic status, older people, women and non-white patients (i.e. inequality in a different direction from that flagged as light green). Where methodological observations are made on studies, these are also highlighted in blue, though we have not changed the ‘colour’ of the study’s main findings, even if we consider the methodology to be flawed.

Given the powerful meta-narrative of inverse care in the field, the results are surprising. Of the 65 studies examining variations according to socioeconomic status, 46% (n = 30) reported evidence of poorer access/use among socially disadvantaged groups. 109,138,140,184,193,219,220,224,226,229,233,242–244,249,252,253,261,262^,263,268,269,284,287,288,291,293–296 Forty-four per cent (n = 16/36) of the studies examining ethnic variations found poorer access/use among non-white groups. 223,234,237,241,243,252,258,261,278,283,297,306,308,309,311,312 By contrast, 73% (26/37) of studies investigating sex variations found that women were less likely to enjoy the same level of access/use as men139,223,225,231,232,234–236,258,263,270–273,290,291,294,296–302,305,306 and 81% (22/27) of studies investigating age differentials found that older people had poorer access/use than younger people. 191–195,252,253,258,260,263,273,275,284,290,291,294–296,301,303,305,306

Presentation of cardiovascular problems

More studies examining variations in the presentation of CVD focused on socioeconomic status than other categories. Fewer than half found evidence of inverse care. This may reflect improved detection of disease within the community over time. For instance, using 2005–6 QOF data, the gap between estimated (APHO) and reported prevalence of CHD and hypertension increased with population deprivation and was higher among practices in more deprived areas. 224 By contrast, a comparison of NHS Health Check coverage (2011–12) against expected (APHO) cardiovascular health need found that coverage was significantly higher in PCTs in the most deprived areas than in those in the least deprived. 222 Some studies225,228 agree that people from less advantaged communities are more likely to access CHD risk-screening programmes, while others disagree. 219,226,227 Thus, the direction (if any) of socioeconomic inequality in presentation of CVD remains unclear.

Although the wider literature on sex differences in help-seeking for CVD is conflicting,236,327 the evidence of this review suggests that women delay help-seeking longer than men. This appears to be the result of biological differences rather than sex constructions, with, for example, women experiencing more atypical symptoms of CHD. 139,231,235,236 Against this, there is also evidence of the underpresentation of hypertension in men. 139 A review article219 also suggests that men (socially disadvantaged men, single men, non-white men) are less likely to attend health checks, though much of the evidence drawn on in this review comes from outside the UK. Youth also appears to be a factor lowering rates of presentation, at least among men. 222,233

With respect to ethnicity, several studies suggest high levels of presentation among South Asian patients,227,233 perhaps because of awareness about increased risk. Galdas et al. 238 attributed a greater willingness to seek help among South Asian men to different attitudes towards masculinity. During semistructured interviews, white men linked a reluctance to disclose symptoms and a delay in seeking treatment to fears of being seen to be weak. By contrast, Indian and Pakistani men emphasised wisdom, education and responsibility for the family and their own health as more valued masculine attributes, and this contributed to a greater willingness to seek medical help. However, other studies found no ethnic differences in presentation228,239 while still others found higher delays between the onset of symptoms of acute ST-elevation MI and arrival time at hospital among South Asians. 234 One study examining delay in presentation after an acute stroke also found that patients of black ethnicity had higher pre-hospital delays. 237

Overview of studies

In total, we found 101 studies113,328–427 that had investigated inequalities in access to and/or use of mental health services in England (or which presented review articles comprising a large proportion of English studies). These are described in Appendices 15–25 and summarised in Table 3. Again, the bracketed numbers refer to the number of studies that found evidence of inequality in access/use (for people of lower socioeconomic status, older people, women, non-white patients and by geography). With respect to mental health, however, the interpretation of ‘inequality’ is not straightforward. For example, there have been long-standing concerns that people from ethnic minorities have been preferentially ‘psychiatrised’, with excessive numbers of patients inappropriately being given a diagnosis of schizophrenia in particular, and, having received this diagnosis, a disproportionate number have been forcefully detained, medicated with antipsychotic drugs and involuntarily admitted into secure hospitals under sections of the Mental Health Act. 428,429 Against this background, evidence of higher rates of hospitalisation has been interpreted in the review as evidence of inequality.

As Table 3 shows, the breakdown of mental health studies with respect to type of service, type of inequality and direction of inequality is very different from that of studies of variations in access to and/or use for CVD. First, the largest category of interest is that of ethnic variations in access (35% of the total), followed by age variations (23%). Only 18%, 14% and 11% of studies focused on socioeconomic, geographical or sex variations, respectively. Second, a significantly higher proportion of studies explored variation in access to and/or use of specialist services (55%). This reflects the way in which mental health services are configured; even those provided in the community tend to be delivered by specialist teams as opposed to GPs. Finally, it should be noted that reviewed studies on variations in access to/use of mental health services included a far higher proportion of local studies (e.g. based on one mental health centre) than did research on CVD.

The direction of inequality

Mental health services also appear to be characterised by different dimensions of inequality. As with CVD, a high proportion of studies examining age variations found evidence of inequality (64%) in mental health services. Sixty-four per cent of studies examining ethnic variations also found evidence of inequality, supporting concerns about cultural misunderstandings, coupled with institutional racism. 430 Thirty-nine per cent of studies examining socioeconomic differences reported poorer access/use among socially disadvantaged groups, suggesting that the classic understanding of inverse care is not as strong a dimension of inequality. There is, moreover, little evidence that women experience poorer access to/use of mental health services (22% of studies). Care should be taken in interpreting these summary results, given the methodological difficulties of examining mental health service use relative to need. It is, nevertheless, interesting that a similar proportion of studies of mental health services report inequality (60%) as studies of access/use for CVD (62%).

Large-scale survey data sets

The key evidential constraint in small-area estimation is that the factors used to model whether or not an individual exhibits a particular response characteristic, for example reporting a particular health condition or meeting some biometric threshold such as blood pressure > 140/90 mmHg, must align with covariate data available for the populations for which estimates are to be produced. The practical implications of this in terms of modelling are discussed below, but it does demand that any survey data used in small-area estimation includes information on the sociodemographic characteristics of respondents, which aligns with data provided by 2011 census, the principal source of information on the sociodemographic characteristics of local populations. This does not pose any general problem, as most large-scale surveys aim to obtain such data, although there are particular issues, discussed in Modelling large surveys, concerning the use of different response options for the general health question and difficulties matching LLTI data from the census with cognate survey data.

The principal determinant of which large-scale surveys were included in this study was thus whether they contained recent data relevant to cardiovascular and/or mental health. Having interrogated the range of data sets available through the UK Data Service, we list in Table 4 those that have been used to provide data relevant to a range of response items which, in principle, might be expected to relate to CVD and CMHDs. 435,454–468

The total number of individuals and households listed in Table 4 is, however, somewhat moot, as the size of the samples extracted for modelling each response variable varies hugely. These samples include, as discussed below, a maximum of one adult respondent (aged ≥ 16 years) per English household for whom valid data are available. For instance, almost all adult respondents to the HSfE provide information on LSI, whereas only a subset are visited by a nurse and tested for blood pressure. As a result, whereas the HSfE LSI data set comprises 28,340 individuals, the data set for ‘nurse-measured’ blood pressure comprises only 12,103 respondents. Full details concerning the 24 CVD and CVD-related response items listed in Table 5, and the 41 mental health-related items listed in Table 6, are provided in Appendices 26 and 27. The appendix entries include a detailed description of the precise question or biometric measure used to derive each response variable, along with a full description of predictor variables included in each model and the size and demographic composition of the underlying sample.

2011 census data and the 2011–12 Attribution Data Set

Model specification and fitting

In all cases except modelling the risk that individuals aged 25–74 years will develop CHD over the next 10 years, the response variables in which we are interested are binary, for example whether or not an individual reports a LSI or provides a high blood pressure reading. The predictor variables are generally categorical (e.g. the seven age bands or the five general health bands), the only exception being the actual IMD scores, as opposed to IMD score quintiles, that are available for models utilising UKHLS survey data. As a result, the vast majority of the models constructed here are logistic regression models (i.e. generalised linear regression models using a binomial link function). The ‘risk of CHD over the next 10 years’ model was also a generalised linear model, this time generalised by use of a reciprocal gamma link function.

All models were fitted to the survey data using the R Statistics glm function (The R Foundation for Statistical Computing, Vienna, Austria). Importantly, all potential individual-level main effects (i.e. excluding those which were not, as discussed above, sufficiently cognate with equivalent census variables) have been included in all models. These variables are statistically significant across most models, suggesting that they are closely associated with the general health status of individuals, but for some conditions traditional approaches to model specification would exclude one or other of the main effects. They key here is to recognise that the models being constructed are for predictive rather than explanatory purposes. Generalised linear regression is commonly used as a tool for identifying statistically significant relationships between potential predictor variables and a particular response variable. In such circumstances the rigorous use of model selection criteria is a crucial and widely used method for ensuring model parsimony, with the researcher then focusing on the effect (‘other things being equal’) that individual predictor variables have on the response variable. The goal is to ensure that overfitting the model does not result in the identification of potentially spurious causal relationships.

In this study the goal – specifying a suitable predictive model rather than identifying plausible casual relationships between individual predictor variables and the response variable – suggests a somewhat different approach. The local areas for which predictions are required vary hugely, and often quite uniquely, in terms of their sociodemographic composition. To exclude information about relationships between predictor and response variables, even if not formally significant, risks dampening the predictive power of the models with respect to specific areas. For instance, ethnicity and tenure are not uncommonly statistically insignificant predictors of variations in morbidity and would be removed using standard model selection criteria, but this would mean that potentially useful predictive information regarding, say, predominantly Asian populations living in rented accommodation would be lost. There may be very few such areas, but only if evidence of the individually weak relationships is retained is it possible to capture the potential impact of unusual social circumstances on disease prevalence. Moreover – and this is key – nothing is risked by including such variables in predictive models. If there is uncertainty surrounding the effect of a particular parameter, then that uncertainty will be taken into the final prevalence estimates by means of bootstrapping, as described in Capturing uncertainty: a bootstrapping approach.

All of the generalised linear models thus include all available main effects, but thereafter the potential relevance of all potential one-way interaction effects was tested using the Akaike information criterion. The set of one-way interaction effects which minimised the Akaike information criterion were thus used, alongside the available main effects, in the final predictive model for each of the response items being considered. A full description of each model is provided in Appendices 26 and 27, with a table giving (to four decimal places) the estimates, standard errors and upper and lower 95% CIs associated with each factor (relative to its reference). More immediately useful in terms of judging the relative importance of the various factors, we also in each case provide a plot of the parameter estimates. Interpreting these parameter plots is relatively straightforward, with the width of the 95% CI illustrating the degree of uncertainty associated with each parameter estimate, and whether or not it ‘crosses zero’, indicating whether or not the factor is statistically significant. Thus, in Figure 1, which replicates the parameter plot given in the appendix for ‘diabetes as a LSI’, it is clear that, relative to people aged 16–24 years (the reference group), increasing age is very strongly associated with an increased likelihood of reporting diabetes as a LSI (at least until the ≥ 85 years cohort). Males are also more likely to report diabetes as a LSI, as are people of black and Asian ethnicity, those with poor general health, those who report a LLTI, and those who live in rented accommodation (relative to owner–occupiers). As with most of the models, the effect of deprivation is muted once individual-level sociodemographic variables are included, although it is clearly present.

FIGURE 1.

Parameter plot: self-reported long-standing diabetes problem.

It is important to recognise that no direct causal relationship (or direction) is being claimed. The proposed model ‘merely’ offers the best means of utilising data (available for local areas) for predicting whether or not an individual will exhibit or develop a particular disease, condition or health-risk behaviour. To that extent, it does not matter, for instance, that it is having diabetes that, in causal terms, means that an individual is more likely to have a LLTI. From a predictive perspective, when one is able to use only data available for local areas, if an individual has a LLTI then this increases the likelihood that they will have diabetes – as it does, of course, across many of the other CVD/CVD-related and mental health problems considered in this study. At an individual level this is of limited utility, but when likelihood estimates (derived using all available sociodemographic data) are applied to populations, this allows us to predict how many adults are likely to have diabetes, or any of the other conditions in which we are interested.

The only complication when interpreting this, and most other parameter plots, is allowing for the possible impact of interaction effects. In this instance, although there is much uncertainty as to the exact scale of the effect, various combinations of age and LLTI, and of general health and LLTI, are included as having a statistically significant relationship with the likelihood that an individual will report diabetes as a LLTI. The complication is that these interact with the respective main effects, although in this instance the overall picture remains quite clear.

One final aspect of model fitting needs to be noted. All the surveys which provided data to this study utilised a stratified random probability sample to select households. This approach incorporates a geographic ‘primary sampling unit’ (PSU) in the sampling framework which, in theory, could result in some covariance. To test whether or not PSU membership explains any of the variation in each response variable, we fitted a ‘mixed’ model equivalent to the final generalised linear model. This generalised linear mixed model, implemented in R Statistics using the glmer function of the lme4 package,471 incorporated the PSU as a random effect and the size of that effect was considered. This was invariably statistically insignificant and could be discounted.

From parameter estimates to disease likelihoods

An exploration of the various tables and plots in Appendices 26 and 27 will illustrate the nature of the relationships which underpin the estimation of local disease prevalence rates, but it should be reiterated that these are essentially ‘predictive’ rather than ‘explanatory’ models. Indeed, in a very real sense it is quite unimportant whether or not the particular phenomena represented in the final models are, in themselves, causally related to morbidity. Rather, they represent the best fitted solution for predicting variations in morbidity using only variables for which cognate census-derived covariate data are also available. The issue is now how best to utilise the information captured by the survey-based models for predicting disease prevalence in small populations.

As outlined in the introduction to this chapter, the approach adopted is to use the parameter estimates to calculate the likelihood that different types of person, as defined by each model, will have a particular condition or disease. Thus, the model for ‘diabetes as a LSI illustrated by Figure 1 comprises six individual-level variables and a total of 24 factors: age band with seven factors (16–24 years, 25–34 years, 35–49 years, 50–64 years, 65–74 years, 75–84 years and ≥ 85 years); sex with two factors (male and female); ethnicity with five factors (white, mixed, black, Asian and other); general health status with five factors (very good, good, fair, bad and very bad health); LLTI with two factors (yes and no); and tenure with three factors (owner–occupier, social renting and private renting/other). This, in fact, defines 7 × 2 × 5 × 5 × 2 × 3 = 2100 ‘person-types’ for whom a specific disease likelihood can be calculated; this rises to 10,500 person types once one includes reference to the IMD quintile of the LSOA in which an individual lives. Different models will generate likelihood estimates for a different number of person-types depending on the number of factors in the model, but in all cases a rich description of how disease likelihood varies across the sociodemographic spectrum is provided. The particular advantage of this approach is that, in association with the use of interaction effects, the full compositional diversity of different populations can be incorporated into the final estimates of disease prevalence.

Microsimulation for small-area estimation describes how suitable local covariate data are obtained, and Small-area estimation: prediction describes how the person-type disease likelihood estimates are attached to those covariate data to finally derive the required local prevalence estimates, but a key difficulty with this process is how to incorporate parameter uncertainty into those prevalence estimates. Figure 1 makes clear that, even for clearly statistically significant parameters such as age, there is still considerable uncertainty around the actual parameter estimate.

Capturing uncertainty: a bootstrapping approach

Although this study commenced with the intention of using an explicitly Bayesian approach, in practice this proved problematic, principally because of the computational demands of Markov chain Monte Carlo which, in the context of fitting multiple models with many parameters and with respect to relatively large data sets, proved too time-consuming. The perceived advantage of a Bayesian approach had been that, in explicitly generating posterior distributions for small-area estimates, it offered a suitable framework for capturing model uncertainty in the final estimates. Retaining this focus on understanding estimate uncertainty was deemed essential, and thus we turned to an alternative method of inferring the precision of the estimates based on a well-established technique known as bootstrapping.

The classic illustration of bootstrapping concerns estimating the mean height of people worldwide. We cannot measure the heights of all people in the global population, so instead we sample only a tiny part of the population and measure the height of the people within that. From this sample of n individuals a single mean height is obtained, but we have no sense of how reliable that mean estimate actually is. The bootstrap approach involves multiple resampling (with replacement) from the original sample to obtain a large number of bootstrap samples, from each of which we derive a mean height. From this set of ‘bootstrapped’ means we can calculate both an overall ‘point estimate’ and 2.5% and 97.5% quantiles to indicate the range over which we are 95% confident that the actual global mean lies (assuming, of course, that we had taken a suitably representative sample in the first place).

The key feature of the bootstrap approach as applied here is that, following standard practice, multiple bootstrap (‘with replacement’) samples were taken of the original survey data set (retaining the PSU structure to reflect the underlying sampling methodology) and each was then used to generate a new set of model parameter estimates. Given the time available, 220 bootstrap samples, and, thus, 220 independent sets of parameter estimates, were obtained for each model. This is a very small number compared with most bootstrap procedures (which typically extract thousands of samples) and the precision of the upper and lower 95% CIs will be affected, but it is worth emphasising that each bootstrap sample is used to generate 220 sets of likelihood estimates for each and every one of the 10,500 person-types, and that these are then aggregated to generate the 220 sets of estimates for all LSOAs, MSOAs and higher geographies. Although this massively increases the computational burden, it enables us to extract summary statistics to capture how model-based uncertainty affects the final estimates at all scales. Thus, as detailed in Small-area estimation: prediction, the range within which 95% of the 220 bootstrap-based estimates lie provides a good, albeit approximate, empirical measure of the 95% credible interval. (The term ‘credible interval’, rather than ‘confidence interval’, is preferred, as it represents a region, sometimes known as the ‘credible region’, of the predictive distribution defined by the 0.025 and 0.975 quantiles.)

The lack of a gold standard

The fact of the matter is that there is no ‘gold standard’ measure of disease prevalence against which to assess the prevalence estimates. Indeed, if there were, those estimates would immediately become redundant. Legitimacy can only flow from methodological rigour (the subject of the previous chapter) and, problematically, a somewhat circular comparison with available utilisation data. This is circular (and problematic) in the sense that, although the prevalence rates are to be used as a measure of underlying illness (need) in an assessment of how health service use varies relative to need, the only evidence we have on variations in illness lies with data collected by the NHS about people who have actually accessed some form of health care. This is not optimal, but it is inescapable.

The fundamental problem is, thus, that it is impossible to prove whether or not discrepancies, whether systematic or stochastic, are due to inadequacies in the need estimates or because of variations in uptake relative to need. However, although definitive proof may be elusive, we would argue that confidence in the legitimacy and usefulness of the prevalence estimates generated can be obtained by comparing them in detail against a variety of measures of health service utilisation. We acknowledge that, for any particular scenario, contradictory interpretations can always be placed on the same observations – discrepancies between need estimates and utilisation data could be due to inadequacies in the needs estimates or because of local variations in access – but it is the overall weight of evidence in the context of a coherent explanatory narrative that matters.

Quality and Outcomes Framework data

In this section, therefore, we compare our CVD need estimates with QOF disease registration data. The QOF database72,440 publishes counts of people on each practice disease register and, using denominators based on overall or age-specific patient populations, practice-level prevalence rates. As practices are paid for keeping the registers, and as the number of people on their disease registers is used to calculate payments within each of the clinical indicator groups, the policy intention is that registers are properly kept and that efforts are made to ensure that as many patients as possible with the specified conditions are identified and registered. As a result QOF-based prevalence rates have been used as a source of proxy data on the level of health-care need in different populations. 140,243,247,259,262,289,477

There are, however, significant problems with the use of QOF disease registers, both generally and as a comparator against which to assess the reliability of our model-based need estimates. The first is their provenance as a tool designed to influence and reward GP behaviour. This can have very specific implications, such as the fact that an individual can be included on the diabetes register only if the practice has established not only that a patient has diabetes, but whether the diabetes is type 1 or type 2. A more fundamental problem is that QOF prevalence rates will reflect variations in case finding. One consequence is that they tend to increase over time (inclusion of registers within the QOF payments system was, in part, intended to improve case finding). More significantly, case finding will also vary in response to factors such as the time, commitment and resource GPs direct towards such activity as well as, no doubt, the responsiveness of the populations being served.

A further issue is the fact that no information is available on the age–sex profile of patients included on the various QOF disease registers. 478 The published QOF prevalence rates are ‘crude’ rates affected by differences in, or changes to, the demographic composition of practice populations. Practices, PCTs and CCGs have very different populations and variations in ‘crude’ QOF prevalence rates are as likely to reflect varying demographic structures as any genuine difference in the health status of populations. Indeed, the stronger a disease’s demographic gradient, the more problematic will be the QOF prevalence rate.

Such problems notwithstanding, the QOF registers explicitly count the number of people ‘known’ by GPs to have a series of specific conditions and, as such, represent a key measure of primary presentation. We can thus simultaneously ‘test’ our estimates by assessing whether or not they broadly align with the number presenting and, by focusing on the differences between presentation (QOF registration) and underlying need (modelled prevalence estimates), can begin to investigate the nature and scale of variations in the utilisation of primary care.

Determining appropriate need indicators

Although we are principally concerned (as detailed below) with comparing how counts and rates vary at more local levels (practices, PCTs and CCGs), we must first establish which of the modelled prevalence rates are likely to provide the most reliable guide to variations in the underlying prevalence of the conditions for which suitable QOF data are available, namely diabetes, CHD, hypertension, stroke and, in Chapter 5, depression. As argued in Chapter 3 (see Methodology overview), it was not possible to determine a priori which of the prevalence rates that could be derived from large-scale survey data would provide the most ‘useful’ indicators of health-care needs. Thus, a range of alternative prevalence rates have been modelled – distinguished by definitional differences and/or because they have been drawn from different surveys using different predictor variables. This is of particular relevance to depression (see Chapter 5, Variations in the use of mental health services relative to need), where a large and very diverse range of prevalence rates is available and where there is scant a priori evidence on which to choose between them.

The only sensible approach to deciding between the alternative prevalence rates as potential indicators of ‘health service needs’ is to incorporate an assessment of how well they compare with the QOF registration data. This is not the only consideration, for there are often definitional differences that affect how the estimates should be interpreted (such as the distinction between biometric, diagnosis-based and self-reported LSI measures), but it is nonetheless important that a convincing explanatory narrative can link any putative measure of ‘health service need’ with the only proxy source we have on the number of people known to have such a need, namely the number of people on QOF registers.

Hypertension (high blood pressure)

High blood pressure is also a well-defined and widely recognised condition which respondents to the HSfE might be expected to report reliably, although it should be noted that the QOF threshold for high blood pressure is 150/90 mmHg, compared with the 140/90 mmHg threshold used in the HSfE. Also blurring precise comparability is the fact that clinical hypertension requires elevated readings to be obtained on more than one occasion, whereas the survey-based designation of high blood pressure is based on a one-off reading.

Once again, as illustrated in Table 14, considerably fewer people report high blood pressure as a LSI than appear in the QOF hypertension disease register, presumably for the same reasons as given above. What is most notable, however, is the very much higher number of people estimated to have high blood pressure when that estimate is derived from variables relating to nurse measurement or diagnosis. Here, the alternative estimates are based on:

1. ‘nurse-measured high blood pressure’ – the number of survey respondents with a systolic reading of > 140 mmHg and/or a diastolic reading of > 90 mmHg

2. ‘nurse-measured high blood pressure or blood pressure drugs’ – which extends the definition to include respondents taking diuretics, beta-blockers, ACE inhibitors, calcium channel blockers or ‘other blood medication’ specifically intended to control high blood pressure

3. ‘clinical evidence of high blood pressure’ – which further incorporates respondents diagnosed with current high blood pressure

4. ‘any evidence of high blood pressure’ – which finally includes respondents who have ever been diagnosed with high blood pressure or who report it as a LSI.

TABLE 14 - Quality and Outcomes Framework register data and comparator modelled estimates: hypertension

QOF register	Practice population (all ages)	Patients on the register	QOF prevalence rate, %
Hypertension 2010–11	55,169,643	7,460,497	13.5
Hypertension 2011–12	55,525,732	7,567,965	13.6
Potential hypertension needs indicators (2001 ADS population ≥ 16 years = 44,813,241)
Modelled prevalence	Predicted cases (≥ 16 years)		Prevalence rate, % (95% CI)
As a LSI	3,010,018		6.7 (6.5 to 7.0)
Nurse-measured	9,251,361		21.5 (20.2 to 21.5)
Nurse-measured or on blood pressure drugs	13,408,487		30.1 (29.4 to 30.7)
Clinical evidence	14,473,610		32.5 (31.9 to 33.0)
Any evidence	16,101,059		36.1 (35.4 to 35.7)

These definitions may appear progressively inclusive, but it is important to recognise that they also differ in terms of potential utilisation and social bias and their interpretation is not entirely straightforward. For instance, although nurse-measured blood pressure data may seem free of bias – they are effectively based on blood pressure readings taken from a random sample of household residents – in fact a proportion of people will have normotensive blood pressure only because they are on medication. As an individual’s recognition that they need to seek treatment may not be socially neutral, and may furthermore reflect the availability and accessibility of appropriate health care, this measure would, somewhat perversely, tend to underestimate the needs of populations which are actually receiving high levels of support through prescribing and the associated primary management of hypertension. It would seem reasonable, therefore, to include people who report that they are currently taking prescription drugs to control blood pressure. But, if so, what about individuals who have been diagnosed with hypertension but who (possibly because of non-pharmacological treatments and/or lifestyle changes) do not provide a high blood pressure reading to the nurse administering the HSfE? They too have demonstrably made demands on health services (at the very least having been diagnosed by a health professional), even though to include them risks incorporating additional supply-side bias. Finally, what about those who return a normal blood pressure reading but who report high blood pressure as a LSI? There may be socioeconomic and/or demographic bias in how individuals respond to the LSI question, but, arguably, if a respondent states that a condition has ‘troubled’ them over a period of time, they have, by definition, a health-care need.

Given the possible sources of bias it is perhaps surprising that, as shown in Table 15, all of the definitions generate estimates that prove to be extremely powerful at predicting PCT-level QOF registration counts. From a public health perspective there may be a case for adopting the most inclusive ‘any evidence’ definition, but marginally the best predictor of variations in PCT-level registration is based on ‘clinical evidence’ of diagnosed or undiagnosed high blood pressure. This predicts a remarkable 99.2% of the variation in PCT-level QOF registrations. The advantage of using this prevalence rate as a ‘needs indicator’ for hypertension is illustrated effectively by Figure 2. This contrasts scatterplots of QOF registrations against (a) at-risk populations (people aged ≥ 16 years) and (b) the estimated number of people with hypertension. The former is relatively effective (predicting 94.3% of the variation in QOF registrations), but use of the ‘clinical evidence’-based measure of needs drags the outliers into an almost straight line. Using this indicator of needs has an equally dramatic impact with respect to practices (n = 8181), for which 77.8% of the variation in QOF registration counts can be explained by variations in the size of the underlying ‘at-risk’ practice populations, as opposed to 94.1% if one uses the number of people expected to have hypertension in each practice.

TABLE 15 - Primary care trust-level 2011–12 ratios of QOF: modelled prevalence – hypertension

Denominator population	Prevalence rate, %	Registrations per expected case			Count correlation
		Mean	Minimum–maximum	Variation	R	% R2
Total QOF population (all ages)	13.6	–	–	–	0.971	94.3
Total population ≥ 16 years	16.9	–	–	–	0.974	94.9
Adults reporting hypertension as a LSI	6.7	2.50	1.99–2.88	1.45	0.994	98.8
Adults with nurse-measured hypertension	21.5	0.81	0.59–0.95	1.61	0.995	99.0
Adults with nurse-measured or taking drugs for hypertension	30.1	0.56	0.42–0.66	1.55	0.996	99.2
Adults with clinical evidence of hypertension	32.5	0.52	0.40–0.61	1.53	0.996	99.2
Adults with any evidence of hypertension	36.1	0.47	0.35–0.54	1.57	0.995	99.1

FIGURE 2.

Primary care trust-level QOF registrations against population and modelled estimate denominators: hypertension. (a) Population (≥ 16 years) against QOF registrations, 2011–12; and (b) estimate (≥ 16 years) against QOF registrations, 2011–12.

Interestingly, although we have evidence of high levels of unmet need (the QOF ‘all-ages’ rate of 13.6% is broadly equivalent to about 17% for adults aged 16 years and older, which is far short of the estimated 32.5% of people aged ≥ 16 years with hypertension), there is, nevertheless, a very strong relationship between, on the one hand, the modelled estimate of the number of people with hypertension and, on the other hand, the number of patients with hypertension who are ‘known’ to primary care services. This, we would argue, is strongly indicative that the modelled estimates are fundamentally robust and that case finding is relatively consistent at both PCT and practice level. It does not, in other words, seem plausible that any misestimation of the number of people with hypertension should be so effectively balanced by an opposite and equal variation in rates of case finding.

Stroke

The final QOF register which provides relevant comparator data on CVD-related morbidity is the Stroke and Transient Ischaemic Attack (TIA) Disease Register. In this case only one prevalence rate is available as a potential needs indicator, based on the number of people reporting a stroke (including cerebral haemorrhage and cerebral thrombosis) as a LSI. As shown in Table 16, the resulting prevalence rate is significantly lower than the number of patients on the QOF stroke and TIA disease register. This reflects the pattern found with diabetes, CHD and hypertension, and, once again, no doubt reflects how respondents have interpreted whether an illness ‘has troubled you over a period of time, or that is likely to affect you over a period of time’.

TABLE 16 - National QOF register data and comparator modelled estimates: stroke

QOF register	Practice population (all ages)	Patients on the register	QOF prevalence rate, %
Stroke and TIA 2010–11	55,169,643	944,099	1.7
Stroke and TIA 2011–12	55,525,732	964,273	1.7
Potential stroke needs indicator (2001 ADS population ≥ 16 years = 44,813,241)
Modelled prevalence	Predicted cases (≥ 16 years)		Prevalence rate, % (95% CI)
As a LSI	424,768		0.95 (0.78 to 1.14)

This disparity does not, of course, mean that local estimates of the number and proportion of people living with stroke as a LSI cannot be used to proxy variations in the need for appropriate health-care services. Indeed, as with respect to the other conditions for which we have ‘LSI’ estimates, the very fact that the condition ‘troubles’ respondents implies a need for health care. More generally, however, the issue is whether or not the relationship between reporting stroke ‘as a LSI’ and the need for relevant health-care services is socially and demographically invariant.

In fact, as shown in Table 17, the number of people estimated to be living with stroke as a LSI condition is very strongly predictive of PCT-level variations in the number of people on the QOF register. Thus, whereas the number of people in ‘at-risk’ populations aged ≥ 16 years predicts only 88.3% of the variation in QOF registrations, the estimated number of people living with stroke as a ‘LSI’ predicts 98.3% of the variation. Figure 3 graphically illustrates the far better relationship achieved by the needs indicator compared with using a simple ‘at-risk’ population. This is also apparent at practice level, where practice-level counts of people aged ≥ 16 years explains 69.4% of the variation in the number of people QOF registered, compared with 91.4% of variation explained by modelled estimates of the number of people in each practice living with stroke as a LSI.

TABLE 17 - Primary care trust-level 2011–12 ratios of QOF: modelled prevalence: stroke

Denominator population	Prevalence rate, %	Registrations per expected case			Count correlation
		Mean	Minimum–maximum	Variation	R	% R2
Total QOF population (all ages)	1.74	–	–	–	0.933	87.0
Total population ≥ 16 years	2.15	–	–	–	0.939	88.3
Adults reporting stroke as a LSI	0.95	2.23	1.69–2.76	1.63	0.991	98.3

FIGURE 3.

Primary care trust-level QOF registrations against population and modelled estimate denominators: stroke. (a) Population (≥ 16 years) against QOF registrations, 2011–12; and (b) estimate (≥ 16 years) against QOF registrations, 2011–12.

Comparison with existing prevalence estimates

It is necessary to conclude this assessment of the ‘competence’ of our modelled needs indicators by comparing them with the only available alternative measures of modelled prevalence, namely the various prevalence estimates published by PHE. Prevalence models for hypertension and CVD were produced at around the same time that we commenced this project. 213,214 As in our approach, these models were developed from data from the HSfE. However, different modelling strategies have been used (see Chapter 3, Introduction). Moreover, the existing PHE estimates have been made available at PCT level and are in the process of being converted to CCG level. However, they are not available for LSOAs and MSOAs and, therefore, cannot be used to model variation in utilisation at a finer geographical level.

Table 18 compares our modelled estimates with those of PHE for each of the four conditions for which we have appropriate QOF registration data. The ‘need estimate only’ correlation (r) and associated ‘proportion of variance explained’ (% R²) statistics describe how well (a) the PHE and (b) our Plymouth estimates of health service need predict PCT-level variations in QOF registration rates. In each case the Plymouth estimate significantly outperforms the equivalent PHE estimate, often markedly so. For instance, whereas published PHE modelled estimates of the prevalence rate for stroke explain 52.4% of the variation in PCT-level QOF registration rates, the Plymouth estimate explains 87.4% of variation.

As they are, in and of themselves, better predictors of variations in QOF registrations rates, it is highly likely that the Plymouth estimates are better indicators of underlying morbidity across all four conditions. This is reinforced by the evidence given on the right-hand side of Table 18, which describes the outcome of stepwise-selection-based linear regression modelling of QOF registration rates relative to underlying per-capita morbidity (‘need’) and a series of measures which attempt to capture the sociodemographic and geographical characteristics of PCTs. Specifically, the percentage of PCT populations of Asian or British Asian ethnicity (‘Asian’); the percentage of black or black British ethnicity (‘black’); the percentage aged < 50 years (‘pop < 50’); a measure of average deprivation for the PCT; and, finally, the percentage of PCT populations living in LSOAs defined as ‘rural’.

The first column describes the total variation in PCT-level QOF registration rates that is explained by the final linear regression model, while the remaining columns give the standardised coefficient values for all model parameters. For instance, with respect to diabetes (the only one of the conditions for which PHE estimates are available for CCGs), when using currently available PHE prevalence estimates, an initial ‘need estimate only’ % R² value of 58.9% rises to 67.3% once ‘Asian’, ‘pop < 50’ and ‘IMD’ factors are included in the model. In this case the parameter coefficients suggest that QOF registration was higher than expected in PCTs serving more deprived populations (with higher IMD scores) and with a higher percentage of people of Asian or British Asian ethnicity, and lower than might have been expected in PCTs with younger populations.

This can be compared with the model based on Plymouth estimates, which explains a somewhat larger 75.1% of the variation in PCT-level QOF registration rates and, crucially, incorporates a somewhat different set of parameter estimates in the final predictive model. The effect of deprivation (IMD) disappears, although rates remain higher than expected in PCTs serving populations with a higher percentage of people of Asian ethnicity, and lower than expected in PCTs serving younger populations. There is now also some evidence of relative under-registration in PCTs with a higher percentage of people of black or black British ethnicity.

There is much to glean from this table, but two sets of interrelated observations stand out. First, the Plymouth need estimates are always better predictors of variations in QOF registration rates than the PHE estimates and, moreover, whereas the former are generally incorporated as the most important factor in the linear regression model, the PHE need estimates behave erratically. The PHE need estimate is thus excluded from the hypertension model, and incorporated as a negative parameter in both the CHD and the stroke models. Second, although the final linear regression models tend to explain a similar amount of the variation of PCT-level QOF registration rates, the factors included in the models do differ. Most significantly, deprivation is always incorporated as a positive factor in the PHE models, but appears in only one of the Plymouth models (for stroke). The point here is that if, as seems extremely likely, the PHE estimates fail to fully capture variations in the underlying prevalence of these four conditions, then little can be concluded from any analysis of variations in use relative to need. In this case, for instance, the apparent relatively higher-than-expected rates of QOF registration in areas of high deprivation are almost certainly spurious.

In contrast, using the Plymouth estimates offers a very much more satisfactory explanation of variations in QOF registration rates across all four conditions. In particular, the fact that one gets consistently large and positive standardised ‘need’ parameter values and, conversely, almost invariably much smaller parameter values associated with all other parameters except ‘pop < 50’ is strongly suggestive of measures which genuinely capture variations in underlying morbidity.

There is every reason, therefore, to believe that these needs indicators can be used to study variations in access to, and uptake of, health services relative to need. This is the subject of Variations in the use of cardiovascular care relative to need, though it should be noted that the foregoing discussion also serves to demonstrate the intrinsic value of the estimates to the public health community and commissioners of local NHS services. With precisely the same methodology being used to produce prevalence rates for a wide range of geographical and organisational units (LSOAs, MSOAs, upper- and lower-tier LAs, regions, general practices, PCTs, CCGs and SHAs), many of which are to be made available via PHE’s Local Health website (www.localhealth.org.uk/), these prevalence rates constitute a robust evidence base for future investigations into health service equity at a variety of scales and using a variety of measures of service use, performance and outcomes.

Conclusion

This part of the chapter has been primarily concerned with (a) identifying which of the range of prevalence rates available for diabetes, CHD, hypertension and stroke should be used as needs indicators, and (b) establishing whether or not there is prima facie evidence that the selected indicators are sufficiently robust and unbiased to provide a suitable basis for assessing variations in health service utilisation relative to need. As noted in The lack of a gold standard, in one sense the latter is an impossible goal, as there can be no definitive empirical proof that the modelled estimates accurately describe variations in the underlying morbidity of populations. Nevertheless, the extent to which variations in the number of patients on the various QOF registers are so strongly related to estimates of the underlying number of people with relevant conditions lends some initial support to the veracity of the estimates.

The count-based correlations and associated R² statistics considered above are, however, of only limited utility to this study. The overwhelming significance of the relative size of different PCTs and practices on the number of people who are QOF registered (or who utilise other health services) means they impart an impression of strength of association which is not matched when rates, rather than counts, are compared. Standard measures of inequality such as the Gini coefficient (or, for that matter, other measures, such as the Theill and Atkinson Indices which similarly aim to describe how an outcome, such as use of health-care services, is distributed between cohorts481–484) are also of only limited analytical utility.

In terms of Gini coefficients, access to primary health care at PCT level (as measured by QOF registration relative to our condition-specific needs estimates) appears very equitable. With a scale that runs from 0 (complete equality) to 1 (complete inequality), the coefficients for diabetes, CHD, hypertension and stroke are 0.0497, 0.0884, 0.0358 and 0.0549, respectively. Indeed, in each case the ‘Lorenz curve’, which plots the cumulative proportion of PCT-level QOF registrations against the cumulative number of people in PCTs estimated to have each condition, is very close to the 45° ‘line of equality’. As might be expected, greater inequalities of access to primary health care is apparent at practice level, with Gini coefficients for diabetes, CHD, hypertension and stroke of 0.1105, 0.1290, 0.0948 and 0.1393, respectively.

Gini coefficients and associated plots of health-care inequality (using Lorenz curves) can thus be used to illustrate how, in general terms, equity of use relative to estimated need varies between different conditions (Figure 4) or, for each condition, between different stages of the care pathway (Figure 5), but this does not help us to understand the sociodemographic dimensions of inequalities. To that end ‘concentration indices’ have often been used which, in the present context, would describe the extent to which inequalities in use-to-need ratios are associated with particular factors,483,485 generating, for instance, measures of ‘deprivation-related’ health-care inequality or ‘age-related’ health-care inequality. Our preferred approach, however, has been to utilise regression models to isolate and quantify the independent effect of the different factors, that is, once the role of other factors in predicting variations in use-to-need ratios has been allowed for.

FIGURE 4.

Condition-specific inequality (Lorenz) curves for QOF registrations relative to need.

FIGURE 5.

Condition-specific inequality (Lorenz) curves for QOF registrations and prescribing and secondary programme budget expenditure. (a) Diabetes; (b) CHD; (c) hypertension; and (d) stroke.

Thus, in the next section, which turns to consider evidence of equity with respect to the uptake of services associated with the care of patients with diabetes, CHD, hypertension and stroke, we examine how per-capita QOF registration rates, along with other per-capita utilisation rates, compare with per-capita estimates of underlying morbidity. Linear regression modelling is used to establish whether measures summarising the age composition, ethnicity, deprivation or rurality (a systematic as opposed to random measure of geography) of practices and PCTs can account for any of the variation in the relationship between estimated service need and recorded service use. An important aspect of this modelling is to construct a coherent and convincing narrative which demonstrates that the modelled estimates capture as much as possible of any systematic sociodemographic and spatial variation in underlying morbidity, ensuring that any remaining unexplained variation in the relationship between use and needs is attributable to patterns of unmet need rather than to inadequacies in the underlying need estimates. This is crucial to the entire project, as any systematic failure of the needs indicator will result in spurious relationships between use and need.

Introduction

As noted in Chapter 1 (see Implications for research), in order to capture the possibility that inequalities in access and use occur in different ways at different stages of the care pathway, we have explored evidence with respect to presentation, primary management and secondary management (Table 19). For each of the disease conditions under consideration, variations in presentation are explored by examining variations in QOF-recorded prevalence relative to modelled prevalence. Primary management is measured by examining rates of prescribing and Programme Budget Category (PBC) expenditure on prescribing. 486 British National Formulary (BNF)487 codes for drugs examined in each category are listed in Tables 20–23. Finally, variations in secondary management are examined with respect to elective and emergency admissions, the utilisation of specific procedures, and PBC secondary expenditure.

Variations in use relative to need: diabetes

Diabetes has been included in the study for reasons outlined in Chapter 1 (see Aims and objectives), not least because it is an important risk factor for CHD, hypertension and stroke. Table 20 shows the results of full linear regression modelling to predict per-capita service use for diabetes at PCT level, and Table 21 predictions at general practice level (n = 6984). The ‘need’ parameter is the estimated per-capita rate of diabetes, using the LSI definition. The first thing to note is that, taken on its own, the underlying prevalence of diabetes explains 62.1% of the variation of PCT-level QOF registration rates, suggesting that a good deal of the variation in presentation can be explained by underlying need. As might be expected, the percentage of variation that is predicted at practice level is smaller (44.5%).

Although underlying prevalence accounts for the greater proportion of presentation relative to need at both PCT and practice levels, additional factors increase the variance explained (to 75.6% between PCTs and 57.6% between practices). As indicated by the standardised coefficients given in Table 20, both the proportion of people of Asian or British Asian ethnicity and the percentage of people under the age of 50 years contribute to the explanation of variations in PCT-level QOF registration rates. At practice level (see Table 21), diabetes registration is also significantly higher than one might expect among Asian populations. However, registrations are also higher in deprived and rural communities, and lower in practices with larger percentages of black populations.

As one moves up the care pathway, the percentage of variance that can be explained by either estimated prevalence or additional factors falls. With respect to diabetic prescribing items, for example, < 50% of the variation at PCT level can be explained by underlying need alone. However, once additional variables are factored in, prediction improves [e.g. to 60.1% for PCT-level variations in prescribing of diabetic drugs (BNF 6.1487)]. Prediction is far poorer at practice level (25.9% for diabetic drugs). Again, significantly higher rates of prescribing among Asian populations appear to play an important role in improving explanation of overall variance. Here, however, additional factors emerge as explanatory variables. Prescribing rates are significantly higher in deprived and rural communities, and lower in practices with larger percentages of black populations. However, younger communities continue to appear to be underserved, albeit with mixed results, with practices with a higher percentage of adults under the age of 50 years being less likely to prescribe diabetic drugs (BNF 6.1487) and other antidiabetic drugs (BNF 6.1.2487) but more likely to prescribe insulin (BNF 6.1.1487). It is important to note that, although these coefficients are significant (i.e. p < 0.05), their contribution, compared with underlying need, is small.

With regard to secondary care, the direction of some of these coefficients changes. Importantly, PCTs and practices with a higher percentage of Asian populations now have lower-than-expected rates of admission (elective and emergency) for a main diagnosis of diabetes and lower PBC expenditure on the disease. By contrast, PCTs with a higher percentage of black populations have higher PBC expenditure on hospital care (data available at PCT level only). This suggests that relative mix of care may be a key factor in understanding variations in the appropriateness of NHS care provided. High rates of presentation and prescribing among Asian populations and lower-than-expected rates of secondary care suggest that, among this population, diabetes may be better managed within primary and community settings than among other population groups. Against this, more deprived PCTs and practices have higher-than-expected rates of both prescribing and secondary care (particularly with respect to emergency admission for diabetes as a main diagnosis); and PCTs and practices with higher black populations have lower utilisation rates than expected in both primary and secondary care.

Finally, it is worth pointing out that, although full linear regression modelling can throw light on variations in access and use according to population characteristics, a significant percentage of variation remains unexplained. This may be indicative of random variation (i.e. a ‘postcode lottery’). However, as we explore with respect to CHD, systematic regional differences in use relative to need may complicate the picture. This is discussed in the following section.

Variations in use relative to need: coronary heart disease

Is there a ‘geography’ to variations in cardiac care?

As noted above, although the modelled estimates predict a good proportion of variance, particularly at earlier stages of the care pathway, a significant percentage of variation remains unexplained. Although this may be indicative of random variations in medical practice, we propose that there may also be regional variations in utilisation relative to need, which cannot be explained by their compositional (i.e. socioeconomic, sociodemographic and ethnic) characteristics.

Maps of QOF registrations (Figure 6), PBC prescribing (Figure 7), PBC secondary expenditure (Figure 8) and total CHD expenditure (Figure 9) by expected CHD case suggest that the North East has relatively high levels of use of cardiac care, as do parts of the South East and South Central regions. By contrast, the East of England, South West and East and West Midlands have relatively low levels of use. Interestingly, although rurality did not emerge as a significant factor explaining variation in use relative to need, several ‘shires’ (Shropshire, Herefordshire, Worcestershire, Warwickshire, Leicestershire, Northamptonshire, Staffordshire, Derbyshire, Norfolk and Devon) have lower levels of CHD expenditure than expected by their caseloads. This does not appear to be explained by higher levels of prescribing, which might be expected to reduce levels of secondary use in these areas (see Figure 4).

FIGURE 6.

Map of QOF CHD registrations72 per expected CHD case. Contains Ordnance Survey data © Crown copyright and database right 2012.

FIGURE 7.

Map of PBC CHD prescribing expenditure486 per expected CHD case. Contains Ordnance Survey data © Crown copyright and database right 2012.

FIGURE 8.

Map of PBC CHD secondary expenditure486 per expected CHD case. Contains Ordnance Survey data © Crown copyright and database right 2012.

FIGURE 9.

Map of PBC CHD total expenditure486 per expected CHD case. Contains Ordnance Survey data © Crown copyright and database right 2012.

One factor that may be worth further investigation is the role of per-capita funding. Referring back to Table 1, which compared allocations for PCTs with the youngest and the oldest demographies, it is worth noting that, apart from North Staffordshire and Devon (which, respectively, received per capita allocations of £1614 and £1506 in 2010–11), all of the shires listed above received < £1500 per head for their Hospital and Community Health Services. By comparison, all of the PCTs (n = 12) in the North East SHA, 12 of 14 in Yorkshire and the Humber, 23 of 24 in the North West, 6 of 8 in the South East, and 30 of 31 in London received higher funding allocations.

Variations in use relative to need: hypertension

Variations in use relative to need: stroke

At the PCT level, estimated prevalence of hypertension (based on the clinical evidence definition) explains a very high percentage (87.4%) of variations in QOF registrations (Table 27). Once additional factors are accounted for, prediction increases to 90.4% of explained variance. At practice level (n = 7839), variance in presentation explained by underlying need (57.1%) increases to 60.1% in the full linear regression model (Table 28).

Although stroke registration and prescribing at practice-level is higher than expected in more deprived areas, younger age and ethnicity emerge as consistently significant coefficients of lower use. In this case, both black and Asian populations appear to have lower QOF registrations and lower prescribing of oral anticoagulants and warfarin than expected. Antiplatelet prescribing is also lower than expected among black populations.

Hospital admissions for stroke are also lower than expected for Asian and younger patients, although the standardised coefficients for percentage black population do not reach significance. Again, admissions for stroke are higher than expected among deprived practices. This finding conflicts with that of Chen et al.,287 who found that deprived patients had statistically significant reductions in odds of being admitted to hospital.

Determining appropriate need indicators

Identifying a suitable indicator to describe how CMHDs vary between populations is complicated by the wide range of possible survey-based measures which, unlike the CVD prevalence rates considered in Chapter 4, are likely to be describing very different phenomena. Thus, although, for instance, CHD as a LSI represents a particular type of need which differs in important ways from evidence that an individual has had symptoms of CHD, it was nevertheless reasonable for us to treat both as potential proxies for the same underlying condition. This is less obviously the case with respect to mental health.

The nature of the problem is illustrated by the diversity of rates obtained using definitions ranging from rather narrowly conceived concepts such as ‘doctor-diagnosed clinical depression’ through to very broadly defined measures capturing how individuals respond to questions about their current level of anxiety and/or depression. Thus, although only about 2.3% of adults are estimated to be ‘extremely anxious or depressed’, and 4.0% are ‘currently diagnosed with clinical depression’, some 21.7% are estimated to be ‘moderately or extremely anxious and/or depressed’, a figure which rises to 24.2% if one includes people who, though not reporting being anxious or depressed, are nevertheless taking prescribed medication for anxiety or depression (see Tables 29 and 30).

The various indicators can also bear very little relation to one another in terms of their descriptions of how mental health needs vary between populations. For instance, the definition based on individuals scoring ≥ 12 on the revised Clinical Interview Schedule (CIS-R) returns an overall national prevalence rate of 6.4%, compared with the 21.5% prevalence rate with respect to people who report ‘ever having had a mental health issue’. Both are derived from an analysis of the 2007 APMS, but the resulting CCG-level prevalence estimates are completely unrelated (with a correlation coefficient of –0.009). Whatever phenomena they are describing, they are clearly very different phenomena. More to the point, it is quite uncertain which, if any, is likely to provide the best measure, or proxy, of variations in health service needs.

As with respect to our investigation of CVD and related conditions, the absence of any clear a priori criteria on which to base the choice between the alternative prevalence rates means that it is necessary to consider how well they predict variations in health service activity, once again, of course, acknowledging the essentially circular logic that the subsequent analysis of inequalities in utilisation will depend on use-to-need ratios that are based on a need indicator which, in part, has been chosen because it best mirrors variations in health service utilisation.

To that end we generated 41 different sets of prevalence rates across the whole range of geographies and organisational structures in which we are interested (LSOAs, MSOAs, upper- and lower-tier LAs, regions, general practices, PCTs, CCGs and SHAs). Apart from four very specific definitions, which are unlikely to relate to broader data on health service activity (namely generalised anxiety disorder, mixed anxiety/depressive disorder, symptoms of a depressive episode, and a combined measure which captures whether or not an individual has exhibited symptoms of any of these conditions), these are summarised in Tables 29 and 30.

In addition to a series of self-reported measures (e.g. ‘ever having had a mental health issue’) and number addressing doctor-diagnosed clinical depression (the full details of which are provided in Appendix 27), these include estimates based on a standard instrument used to assess the presence of a CMHD, the CIS-R, and one designed to identify the presence of non-psychotic minor psychiatric disorders, the 12-item General Health Questionnaire (GHQ12). But even here there are issues concerning (a) which thresholds should be used – for instance, a CIS-R score of ≥ 12 indicates the presence of a CMHD while a score of ≥ 18 has been taken to indicate a CMHD in ‘need of treatment’; (b) whether or not it is appropriate to ‘adjust’ responses by taking into account the fact that individuals may score below any given threshold because they are on medication and are thus, by definition, already receiving (presumably necessary) health care; and (c) evidence of survey-specific effects on scores (with, for instance, precisely the same GHQ12 instrument generating somewhat different results in the HSfE and the two waves of the UKHLS).

Comparisons with Quality and Outcomes Framework registration data

The overall national prevalence rates associated with each of the defined measures listed in Tables 29 and 30 can be compared with QOF disease register rates of 11.2% and 11.7% in 2010–11 and 2011–12, respectively, and then 5.8% and 6.5% in 2012–13 and 2013–14 (Table 31). The halving of the rate between 2011–12 and 2012–13 was owing to a change in the definition of who was to be included on the register, which now excludes any patients identified prior to April 2006. This change emphasises the fact that a comparison with national prevalence rates is somewhat moot; what matters is the extent to which variations in QOF registration can be explained. Here, in marked contrast to what was found with respect to cardiovascular and related conditions, it is apparent (with respect to 2011–12 PCT-level data, Table 32, or with respect to 2012–13 data, Table 33), that none of the modelled prevalence rates performs particularly well.

These tables describe how much interPCT and interCCG variation exists in terms of the ‘registration per expected case’ ratios constructed using the different prevalence rates, and this is invariably far higher than was found with respect to the cardiovascular and related conditions considered in the previous chapter. The degree of variation is generally lower in 2012–13, by which time the new criteria for QOF registration had been introduced, even though this was with respect to a greater number of CCGs (n = 211) than PCTs (n = 151). In 2011–12 it was the ‘extreme anxiety/depression or using mental health drugs’ definition which resulted in the least variation; in 2012–13 it was one of the ‘doctor-diagnosed clinical depression’ definitions. Also notable is how poorly correlated the PCT- and CCG-level QOF count data are with the variously defined modelled counts of people with mental health problems. As previously noted (see Chapter 4, Conclusion), correlations based on count data tend to be exceptionally high at PCT or CCG levels because of the overwhelming effect of population size on the number of people with any given condition. Yet the highest correlation statistic for 2011–12 was 0.959 (R² = 92.0%) and in 2012–13 it was 0.958 (R² = 91.7%). This contrasts markedly with, for instance, the fact that it was possible to explain 99.2% of the variation in the number of people on PCT-level QOF hypertension registers.

The fact of the matter is that most (though not all) of the modelled prevalence estimates prove to be very poor at predicting how many people in each PCT or CCG will be QOF registered with depression. Indeed, in 2011–12 only 10 of the 32 measures (Table 32) are more effective than a simple count of the ‘at-risk’ population (people aged ≥ 16 years) in each PCT, although in 2012–13 (and with respect to CCGs) this rises to 16 of the 32 putative measures of health-care needs (Table 33). The indicators perform equally poorly at practice level, with the best (‘ever had a mental health issue’) explaining only very slightly more of the variation in the number of people on the QOF depression register than does a simple count of practice ‘at-risk’ populations (63.9% as opposed to 61.6% in 2012–13).

Two quite different interpretations could be placed on this. First, although the method clearly proves perfectly adequate at capturing relative levels of underlying cardiovascular and related morbidity, some aspect of the way in which people respond to questions about their mental health undermines the use of the resulting data for predictive purposes. This is, of course, plausible, but it seems far more likely that the problem lies with variations in QOF case finding, particularly as all of the modelled prevalence rates are significantly more effective at predicting CCG-level variations in the number, and proportion, of people accessing community mental health and IAPT services. The vast majority are also far more effective at predicting CCG-level variations in the number of people using secondary mental health services.

Comparisons with other indicators of mental health service use

Moving away from a focus on count data, Table 34 summarises the extent to which the various per-capita modelled estimates predict CCG-level variations in per-capita (a) QOF registration; (b) prescribing [net ingredient cost for BNF 4.4.1–4.1.3 (hypnotics and anxiolytics) and BNF 4.3.1–4.3.4 (antidepressant drugs)];487 (c) uptake of specialist community mental health-care services (based on the combined number of people accessing Community Mental Health and IAPT services); and (d) use of specialist secondary care services (based on the number of people in contact with secondary mental health services). As intimated by the foregoing discussion, most of the modelled estimates simply do not correlate with CCG-level variations in QOF registration rates. This is also true with respect to CCG-level variations in per-capita prescribing costs. Nevertheless, a small number of modelled prevalence rates estimates do stand out. The estimate based on self-reported ‘ever had a mental health problem’ is the best predictor of CCG-level per-capita variations in QOF registrations (R = 0.551; R² = 0.304). This is the second best predictor of variations in per-capita prescribing costs (R = 0.678; R² = 0.460), with the best being self-reported ‘extreme anxiety and/or depression (r = 0.689; r² = 0.474). It seems reasonable to presume that these indicators, at least to a degree, are capturing variations in the CMHDs that are being picked up in general practice. What is intriguing is that these potential ‘need indicators’ are relatively poor at describing variations in the use of specialist community and secondary services. For these services, other prevalence rates are far more predictive.

Thus, in stark contrast to its lack of explanatory power with respect to QOF registration and prescribing, the most effective predictor of CCG-level variations in the number of people in contact with secondary services is that based on the number of people with a CIS-R score of ≥ 12 (R = 0.570; R² = 0.324). With respect to specialist community services, three self-reported measures stand out: from the APMS, having (a) ‘had a mental health issue in the last 12 months’ (R = 0.517; R² = 0.267) and (b) having a ‘mental health issue first experienced at least a year ago’ (R = 0.514; R² = 0.264); and, from the HSfE, (c) reporting ‘mental health’ as a LSI (R = 0.514; R² = 0.264).

It seems, therefore, that quite different prevalence rates ‘best’ predict variations in the use of services at different points on the health-care pathway and, moreover, that perhaps one should abandon the idea of testing equity of utilisation against a single measure (or proxy) of mental health needs. Arguably, ‘unexplained variation’ in health service use should instead be calculated with reference to service-specific needs indicators. This implies that to predict the underlying number of people with depression (and, thus, how many people expected to be on the QOF depression register) it is necessary to use the prevalence rate based on people reporting that they had ‘ever had a mental health issue’. In order to examine variations in net ingredient cost (NIC) prescribing, meanwhile, the most appropriate prevalence rate would be based on people reporting that they currently suffer ‘extreme anxiety and/or depression or are taking prescribed mental health medication’. With respect to number of people expected to require community mental health/IAPT services, the preferred ‘need indicator’ would be the prevalence rate based on self-reported ‘having had a mental health issue in the last 12 months’, while for an analysis of utilisation of specialist inpatient services it is to the prevalence rate based on people with a CIS-R score of ≥ 12 that we should turn as the most appropriate measure of underlying need.

This is the approach taken in Chapter 4 (see Cardiovascular disease prevalence estimates: testing for reliability) where, as in the analysis of cardiovascular and related conditions, the goal is to account for variations in use-to-need ratios in terms of the socioeconomic, sociodemographic and geographic characteristics of CCGs and practices. We recognise, however, that in taking this approach we risk confusing de facto relationships with what might be clinically desirable. Thus, with respect to each area of service activity Cardiovascular disease prevalence estimates: testing for reliability in Chapter 4 examines how use-to-need ratios vary using a needs indicator which, over the country as a whole, best predicts how that particular type of health service activity varies. Concepts of ‘higher-than-expected’ or ‘lower-than-expected’ use-to-need ratios must, thus, be interpreted as being relative to the current norm, rather with respect to, for instance, some agreed definition of the underlying mental health needs that such services should be addressing. Although this is entirely reasonable, as it allows us to identify whether there are socioeconomic, sociodemographic or other characteristics which influence the utilisation of care relative to current norms, the limitations of such an approach must be recognised.

Comparison with existing prevalence estimates

Before reporting the analysis of use relative to need for mental health services, it is worth assessing the relative ‘competence’ of our modelled needs indicators by comparing them with the only currently available measure of community prevalence of CMHDs. 491 This, produced by Glover for the North East Public Health Observatory (NEPHO), is available via PHE’s Disease Prevalence Models website, although it has not been included on its Community Mental Health Profiles website. The latter presents QOF prevalence data as well as prevalence estimates based on responses to the GP Patient Survey, both of which, of course, capture only those who are ‘known’ to GPs. Glover’s estimates model community prevalence (i.e. underlying need) and, as such, are the only directly applicable comparator against which to assess the Plymouth mental health estimates.

Table 35 describes the extent to which the NEPHO and Plymouth needs indicators predict, either on their own or as part of linear regression models incorporating a variety of contextual factors, PCT-level per-capita variations in (a) the number of people on the QOF Disease Register and (b) NICs for BNF 4.4.1–4.1.3 (hypnotics and anxiolytics) and BNF 4.3.1–4.3.4 (antidepressant drugs). 487 The ‘need estimate only’ correlation (r) and associated adjusted proportion of variance explained (% R²) statistics describe how well the Plymouth and NEPHO estimates predict PCT-level variations. In each case the Plymouth-based estimate wholly outperforms the NEPHO estimate. Indeed, the NEPHO estimate is not statistically correlated with PCT-level per-capita variation in QOF registration, and stepwise linear regression model selection fails to include the NEPHO estimate with respect to both QOF registration and prescribing costs. Using this measure, factors such as ethnicity and deprivation become the sole explanatory variables.

By contrast, the Plymouth estimates of underlying need explain 37.2% of variation with respect to both QOF registration and PBC prescribing. Here, stepwise model selection retains the needs indicators, alongside the percentage black or black British population and, in the case of prescribing, the IMD score. Importantly, although using the Plymouth needs indicator makes very little difference to the amount of variation that can be explained, it significantly alters how one interprets use relative to need. With respect to QOF registration, for instance, the inference to be drawn using the NEPHO prevalence is that registration tends to be higher than expected in more deprived areas (i.e. with higher IMD scores). With recourse to the Plymouth needs indicator, IMD disappears as a significant effect in the linear regression model. Here, the higher needs of more deprived populations are being directly captured in the needs indicator and are not being left to be picked up in apparently higher rates of use relative to need. This is clearly a spurious relationship, and we conclude that the Plymouth estimates offer a far more robust indicator of the underlying prevalence of mental health needs than the only currently available alternative.

Introduction

As with CVD, we have explored evidence of variations in use with respect to presentation, primary management and secondary management in order to examine whether or not variations occur at different levels of the care pathway. Presentation is captured by QOF registrations for depression. NIC prescribing data have been extracted from GP-level monthly presentation data489 and refer to BNF 4.4.1–4.1.3 (hypnotics and anxiolytics) and BNF 4.3.1–4.3.4 (antidepressant drugs). 487 Data on patients attending community mental health services have been extracted from CCG Indicator 2.9; data on patients accessing IAPT services have been extracted from CCG Indicator 2.10; and data on people in contact with secondary mental health services have been extracted from CCG Indicator 2.17. All data sets refer to 2013–14 and all are available via the HSCIC Indicator Portal. 71

Variations in the presentation of common mental health disorders

At the CCG level, estimated prevalence of CMHD (based on ‘ever had a mental health issue’) explains 30.4% of variations in QOF registrations (Table 36); at PCT and practice levels, 37.2% and 14.9% are, respectively, explained. These are poorer predictions than achieved with respect to CVD. Once additional factors are accounted for, prediction increases to 35.3%, 42.9% and 17.0% at CCG, PCT and practice levels. Thus, a large proportion of the variation in QOF registrations is unexplained.

At all three scales of analysis, areas with higher percentages of black populations have significantly lower QOF-recorded depression than expected. This corresponds with the results of many studies that find that being a member of an ethnic minority is associated with lower rates of help-seeking behaviour for mental distress. 330,332,336,341,433 At practice level, however, we find that areas with higher percentages of Asian populations have significantly higher-than-expected rates of diagnosis. One qualitative study we reviewed found that GP consultation rates were higher in depressed people of Pakistani origin because they consulted more often for bodily symptoms. 338

At CCG level, more deprived areas have higher registrations than expected. This does not contradict the findings of the literature, in part because very few studies appear to have looked at socioeconomic differences in help-seeking behaviour for mental health problems. The two studies that did directly examine this found that the socially disadvantaged were more likely to seek formal help. 434,437

In the literature we reviewed, age emerged as a significant factor in help-seeking behaviour and studies suggested that older patients are sensitive to the stigma of formally diagnosed mental illness,431 are reluctant to recognise and name depression432,433 and, in consequence, are less likely to seek help. 434–436 Younger men are also reported to be disproportionately deterred from seeking help for mental health problems by stigma. 433,437 In our analysis, however, there is no evidence to suggest that younger or older populations have differential rates of diagnosis relative to need. This may be attributable to the use of large-scale survey data as the basis for modelling needs, insofar as it is plausible that whatever sensitivities affect individuals’ propensity to seek clinical help will also affect how they respond to questions about their mental health status.

Variations in the primary management of common mental health disorders

Estimated prevalence (based on those reporting extreme anxiety and/or depression or taking prescribed mental health medication) is a significant predictor of variations in net ingredient prescribing costs for hypnotics, anxiolytics and antidepressants (BNF 4.4.1–4.1.3 and 4.3.1–4.3.4487) at CCG level, explaining 47.2% of variation. At practice level, however, only 7.3% of the variation in prescribing is explained. This reflects partly the less precise modelled prevalence estimates available at the practice level, but also presumably significant local variations in prescribing practice. The PBC data for 2011–12 are for a broader ‘mental health’ category,486 but here, too, a significant proportion of variation can be explained by the prevalence-based need indicator (37.2%).

At all three scales, the percentage of population that is black is an important predictor in the linear regression of variations in per-capita prescribing (with areas with higher proportions of black/black British populations having lower-than-expected rates of prescribing). This corresponds with evidence from the National Psychiatric Morbidity Surveys, which found that black people were less likely to be taking antidepressants than their white counterparts after controlling for symptom severity. 332,355 Reviewed evidence also suggests that prescribing for anxiety and depression is significantly lower in areas with higher percentages of Asians. 345,362 This corresponds to our findings at CCG and practice levels.

Levels of prescribing are significantly higher than expected in more deprived CCGs, PCTs and practices. Again, this tallies with the conclusion of published studies that find no evidence of inverse care in the treatment of moderate and severe depression. 345,349,351 Unlike in the published literature, we find no evidence of systematic undertreatment of older populations.

Variations in the specialist community management of common mental health disorders

Expanding services in the community to address problems of anxiety and depression has been an important policy priority. Since its inception in 2008, the IAPT programme has expanded significantly. By 2012,493 more than 1 million people had been in receipt of IAPT services, over 680,000 completing a course of treatment, with recovery rates in excess of 45%.

Reviewed studies suggest that the IAPT initiative is addressing inverse care and that there is now a pro-poor socioeconomic gradient in public psychotherapy treatments which is widening over time. 365,367 We find no evidence of systematic bias – by socioeconomic status, age, ethnicity or rurality – in rates of use relative to need. Thus, IAPT would appear to be an equitable, if geographically varied (see The geography of variations in mental health service use relative to need), service.

Variations in the secondary management of common mental health disorders

As with respect to community/IAPT services, there is little evidence that the number of people in contact with secondary mental health services varies according to population characteristics at the CCG level. Thus, stepwise linear model selection rejected the inclusion of any parameters other than underlying need, implying that none of the unexplained variation in use-to-need ratios is systematically related to any of the sociodemographic factors offered to the model. In contrast, however, the stepwise linear regression model of variations in PCT-level PBC expenditure on secondary services in 2011–12 does reveal the IMD as a significant coefficient. This suggests that the PCTs serving more deprived communities are spending significantly less than expected. As the CCG data refer to numbers of patients, whereas the PCT data refer to expenditure, these results can be reconciled if levels of presentation are broadly as might be expected in areas with relatively deprived populations, but less is then spent on those who do access secondary services.

Estimating local disease prevalence

The decision to use a ‘bottom-up’ approach to estimating population-level disease prevalence through the aggregation of modelled individual-level disease risk estimates raised significant methodological and computational challenges. We were nevertheless concerned to produce estimates that were independent of issues of candidacy and adjudication. We have also sought to avoid a priori decisions about which particular survey questions (or biometric evidence) would best reflect variations in need and have instead made post-hoc assessments of the face-validity of the final estimates vis-à-vis a suboptimal but unavoidable comparison with available utilisation data. We have been clear about the fact that, in the absence of ‘gold-standard’ comparator data, confidence in the estimates must ultimately rest with the suitability and utility of the estimation methodology that underpins them. It is nevertheless encouraging that we were able to construct need estimates for CVD, diabetes, CHD, hypertension and stroke that are significantly better predictors of variations in QOF registration rates than the currently available PHE estimates. Furthermore, whereas the Plymouth estimates were generally incorporated as the most important factor in linear regression modelling of variations in health service utilisation, the PHE need estimates behaved erratically. We have little doubt that these new estimates offer a step-change improvement in our knowledge concerning the distribution of CVD and CVD-related needs.

The mental health estimates are more difficult to assess. Forty-one different estimates were produced, which give rise to a diversity of prevalence rates, not all of which even relate to each other. As predictors of QOF prevalence, they perform poorly. Indeed, in stark contrast to hypertension, CHD and diabetes, there is surprisingly little relationship between the proportion of patients on QOF Depression Registers and estimates of either (a) the number who report having ‘ever been diagnosed’ with depression or (b) any of a large number of self-reported or schedule-based assessments of the presence of a CMHD. For instance, at practice level, the strongest correlation with QOF registration rates is with the proportion of people who report that they ‘had ever had a mental health issue’ (R = 0.386; R² = 0.149).

Prevalence rates based on established schedule-based assessments of the presence of CMHDs are even less well correlated with QOF registration rates. For instance, correlation with GP-level estimates of the proportion of people with a CIS-R score of ≥ 12 (commonly taken to indicate the presence of a non-psychotic psychiatric disorder) is R = 0.088 (R² = 0.008), while correlation with the estimated proportion of people scoring ≥ 18 (indicative of a non-psychotic psychiatric disorder in need of treatment) is only marginally better at R = 0.121 (R² = 0.015). The CIS-R interview captures a wide range of non-psychotic mental health disorders, but the subset concerned with ‘symptoms of a depressive episode’ generates a set of estimates which are only marginally better correlated with QOF Depression Register rates (r = 0.124; r² = 0.015). Estimates of the proportion of people with a GHQ12 score of ≥ 4 (indicating possible psychological disturbance or mental ill health) are similarly poorly correlated with QOF rates (R = 0.134; R² = 0.018).

The overwhelming impression is that our modelled estimates of the prevalence of depression (and non-psychotic mental illness generally) bear very little relation to QOF registration rates; but is this because of a complete failure of our approach to capture variations in CMHD morbidity, or because QOF registration of depression is profoundly variable?

In the first place, it would be surprising if the underlying approach, demonstrably effective with respect to hypertension, CHD and diabetes, should fail with respect to mental health. Thus, although it is possible that the way in which people respond to questions about their mental health may undermine the use of the resulting data for predictive purposes, the fact of the matter is that the same data sources are used, the same method of deriving individual-level risk estimates is employed, and those estimates are then applied to the same microsimulated populations. Second, some of the modelled estimates for CMHDs perform much more creditably with respect to other measures of health service use. For instance, in stark contrast to their lack of association with CCG-level QOF registration (R = −0.005), estimates of the proportion of adults with a CIS-R score of ≥ 12 are much better correlated with estimates of the proportion of adults in contact with secondary services (R = 0.570; R² = 0.322). Similarly, although the proportion of adults reporting that they ‘had a mental health issue in the last 12 months’ is relatively poorly correlated with QOF registration rates at CCG-level (R = 0.379; R² = 0.143), it is a better predictor of the proportion of adults accessing community mental health or IAPT services (R = 0.516; R² = 0.263). CCG-level prescribing, meanwhile, correlates relatively well with CCG-level estimates of the proportion of adults per capita reporting ‘extreme anxiety and/or depression or taking prescribed mental health medication’ (R = 0.689; R² = 0.472).

These associations are not as close as those found with respect to diabetes, hypertension and CHD, but they do indicate that the proportion of adults QOF registered with depression is likely to be a poor indicator of underlying rates of depression, let alone of non-psychotic mental illness more generally. In other words, much of the discontinuity between our estimates of need and QOF registration rates is likely to lie with practice- and PCT-/CCG-level variations in case finding.

One cannot, of course, attribute ‘truth’ to the modelled estimates of need and we recognise that the mental health estimates are, by their nature, more problematic than those for CVD. Indeed, given ongoing debate about the construct validity and reliability of psychiatric diagnostic categories and criteria (diagnosis itself purportedly lacking a gold standard), the hope of producing clear and uncontested indicators of mental health service need was perhaps naive. Yet there were some encouraging findings, not least the fact that the modelled prevalence rates were significantly more effective than both QOF prevalence and the existing Public Health Observatory estimate at predicting CCG-level variations in the number and proportion of people accessing community mental health and IAPT services, as well as at predicting variations in secondary mental health service use.

We thus argue that the Plymouth estimates are based on methodological rigour and offer useful measures of underlying need with respect to both CVD and CMHDs. They may not be definitive, but need estimates are required for planning, commissioning and evaluative purposes and, for the present, there is no plausible alternative means of quantifying the underlying ‘need’ for health-care services.

Understanding the distribution of health service need

To this end, an important finding is the significant variation that exists in the geographical distribution of CVD and CMHD. Reflecting the relative contributions of age, deprivation, ethnicity, etc. in disease risk, the prevalence of most CVD conditions is largely driven by demography and exacerbated by poverty. Thus, the highest prevalence rates of diabetes (a risk factor for CVD) are to be found in Lincolnshire, Newcastle and Tyneside, North Devon and Cornwall, areas in which older age profiles and deprivation combine to increase disease risk. Stroke and hypertension prevalence show similar patterns, although more coastal areas (with ageing demographies) show up on the maps of high prevalence. It is, however, important to note that not all CVD prevalence estimates are distributed in this way. For example, self-reporting a heart/circulatory disease as a self-reported LSI shows a distinct north (high)/south (lower) pattern.

Mental health, by contrast, has a strongly urban distribution. This is particularly the case when based on CIS-R ≥ 12 and ≥ 18 criteria, although most of the prevalence estimates show a similar pattern (although Lincolnshire also appears to be a high-prevalence area). Symptoms of a depressive episode slightly bucks this trend, with high rates being found in demographically older coastal areas such as Cornwall, Devon, Norfolk, Kent and, again, Lincolnshire.

The main conclusion to be drawn here is that the distribution of crude prevalence varies by clinical condition. Moreover, the pattern of CVD prevalence is not cognate with patterns of standardised morbidity and mortality which have been increasingly deployed as proxies for health service ‘need’. It is important that service planners and commissioners recognise this difference. As standardised rates effectively design out the effects of age, they are poor proxies for the crude prevalence of (and, by association, health service need for) diseases where risk is highly age-related. There is a need for greater clarity about the purpose and impact of age standardisation and we believe that our prevalence estimates and maps will make a contribution towards this end.

Existing literature on variations in access and use

As noted above, taken-for-granted assumptions can shape researchers’ choices about the way in which they pose research questions, design studies and interpret results. In our review of existing literature, we found that studies that had explored inequalities in access to and use of CVD care were more likely to have focused on socioeconomic status, while studies on access to/use of mental health services were more likely to have focused on ethnicity.

Study design can influence the eventual evidence base, though we note a distinct improvement in the methodological quality of studies since the 1990s. Accepting published literature at face value, reviewed studies suggest that older age and female sex are the key dimensions of inequality with respect to CVD care; and older age and non-white ethnicity are the key dimensions of inequality with respect to mental health services. For both CVD and mental health, very significant unexplained geographical variation exists in utilisation. Evidence of socioeconomic bias in access and/or utilisation is more ambiguous and contradictory.

The methodological quality of studies is better for CVD than for CMHD. Most accounts of mental health use do not adjust for underlying need, in part, no doubt, because of the hitherto lack of robust estimates of need.

Analysis of use relative to need

Enhancing prevalence estimation and exploiting the prevalence estimates

The final prevalence estimates rest on a proper understanding of the detailed sociodemographic composition of local populations, and this was established through microsimulation. Four issues arise with respect to microsimulation.

First, the ONS has recently published microdata for the 2011 census and these could significantly improve the iterative process which reconciles the aggregate characteristics of LSOA populations detailed in individual census tables to derive a description of the underlying, but unknown, ‘full joint distribution’ (see Chapter 3, Microsimulation for small-area estimation). As the microdata detail which of the myriad of theoretically possible person-types actually occur (albeit within a 10% sample), this would constrain the iterative allocation of individuals to cells so that cells known to be empty are not populated.

Second, microsimulation could be designed to reconcile tables across geographic levels (rather than just at LSOA level, as undertaken here). Census tables for higher geographies tend to be more precise (e.g. including all five general health categories in cross-tabulations rather than the three aggregated categories commonly used at LSOA level). The simultaneous iterative allocation of individuals to cells at multiple levels would add considerable computational burden, but this would ensure that the final microsimulation fits with a considerably wider evidence base. In a similar vein, the ONS will produce commissioned non-disclosive census tables (for a fee). A judicious choice of bespoke cross-tabulations at LSOA level or above could be used to further widen the range of ‘marginal distribution’ totals with which the final ‘full joint distribution’ would have to be reconciled.

Third, as noted in Chapter 3 (see Microsimulation for small-area estimation), in theory multiple solutions (i.e. full joint distributions) could meet the constraints defined by any given set of marginal distributions. It is intuitively likely that these will occupy a relatively small area within the multidimensional space defined by the marginal distributions (i.e. by the census cross-tabulation tables), but this is an issue we have not been able to explore in detail. Of particular concern is whether or not a variety of quite disparate solutions exist, and whether or not the solution to which the iterative ‘random walk’ method converges depends on its starting point. There is a need, in other words, to explore whether or not uncertainty surrounding the microsimulation of local populations is significant and, if so, whether or not that should be incorporated in the uncertainty surrounding the final prevalence estimates.

Finally, and more generally, a detailed description of the composition of local areas would be of value across many areas of research and planning. Implementing and disseminating (via a suitable web tool) a microsimulation covering all of the major sociodemographic variables, and reconciling a wide range of census cross-tabulations, could add significant value to the census. Care would be needed to ensure non-disclosure, but, given the methodology, this should not be an issue. The potential insights and utility of microsimulation is of growing interest, as recently summarised by Tanton and Edwards,499 and the experience of this project suggests that developing a web-based microsimulation tool would both be possible (if computationally demanding) and, if accompanied by an appropriate quantification of uncertainty, provide an invaluable resource to any analyses which need to use data on the detailed composition of local areas.

Turning to the modelling of survey data in order to estimate the likelihood that different individuals will exhibit or develop particular diseases, we are conscious that a number of essentially pragmatic decisions were made at various points. For instance, we decided not to request ‘special access’ survey data with additional information about precisely where individual respondents lived. This was partly because of the administrative difficulties and additional time involved, but primarily because, a priori, we decided not to utilise mixed-methods models. The problem, as we saw it, was that incorporating place (e.g. LSOA of residence) as a random effect in a mixed-effects model of needs risks introducing potential supply-side factors. Thus, if where you live influences how readily you can access a GP, and access to a GP influences the likelihood of being diagnosed with a condition (or, having been diagnosed, of reporting such as a long-standing illness), then using place of residence as a random effect in a mixed-effects model would, inevitably, incorporate this supply-side factor in the needs estimate. As a result it would most probably not appear as a relevant factor when variations in health service use are compared with the needs estimate. The point, however, is that the practical consequence of this, and the various other a priori decisions noted in Chapter 3, should be explored further in order to establish the sensitivity of the final estimates to those decisions. If the Plymouth estimates describing variations in the prevalence of CVD, CVD-related and mental health are to be as widely used by researchers, commissioners and public health practitioners as we believe they deserve, a thorough sensitivity analysis is possibly necessary.

Although, as detailed above, further research could be usefully directed towards the further improvement and evaluation of the methodology, more notable areas for further research concern the opportunities that arise now we have demonstrated that small-area estimation and microsimulation can be combined to generate robust prevalence estimates. First is a simple extension of the approach to cover a wider range of conditions. The various surveys used here, and particularly the HSfE, provide a wealth of evidence across a wide range of diseases. Indeed, the extent to which self-reported long-standing health problems predict variations in the use of health-care services for diabetes, hypertension, CHD and stroke is particularly encouraging, given that self-reported LSI data are widely collected using a standard reporting framework (Box 1). The methodology is perhaps not suited (or the results are maybe not useful) across all disease areas, but there is clearly considerable scope for extending the range of conditions for which prevalence estimates are produced and disseminated. In addition to that, of course, are a variety of doctor-diagnosed, biometric and other survey-based sources of information which could be used to provide a far broader understanding of how disease prevalence varies at various scales.

A second obvious use of the methodology would be to focus on risk factors. Responding to interest expressed by colleagues in PHE, we used the Framingham Risk Calculator in conjunction with data from the HSfE to produce estimates of the number and proportion of adults aged 25–74 years (a) with a > 10% risk of developing CHD in the next 10 years, and (b) who would be expected to develop CHD in the next 10 years. We also produced prevalence estimates of the number and proportion of adults with three or more of NICE’s five modifiable risk factors for CVD. Although detailed in Appendix 26, these have not been discussed in the main report as they were peripheral to our main focus. Yet they are indicative of the wider range of risk factors and health-risking behaviours, such as smoking, obesity and physical inactivity, for which prevalence estimates could be produced. Although in many cases similar estimates already exist,441 these are not always available at sub-LA level, are seldom based on recent data and rarely use methods comparable with that developed as part of this study. Extending the current methodology to cover such risk factors would, we believe, significantly strengthen the public health evidence base.

A third important avenue for future research would be to address, head-on, precisely what the prevalence of particular diseases might mean in terms of health-care needs. There are two issues here. First, what health-care resources are actually used by individuals with specific self-reported, doctor-diagnosed or biometrically assessed conditions? Second, to what extent is that resource use sociodemographically invariant? Understanding these issues would be of particular interest to commissioners/planners of local health-care services as well as to those responsible for the allocation of resources to CCGs and general practices. For instance, it was noted above that although the further targeting of funding towards GPs serving more deprived communities might help them to address the high level of mental distress that is characteristic of such populations, in a zero-sum resource environment this would mean taking funding away from practices serving less-disadvantaged, but older, populations. These practices are typically managing higher caseloads of degenerative chronic disease. Establishing the prevalence of different diseases in different practice populations, as we have done for mental health and CVD, takes us only part of the way towards finding an equitable balance between these conflicting demands. What is also needed is an understanding of the demands such patients typically make on various health-care services, and to that end the recent linkage of HSfE respondents to HES data offers exciting new opportunities. Although individuals’ responses to HSfE questions were unavailable when we undertook this project, it is now possible to compare these with their use of inpatient, outpatient and accident and emergency services both before and after they were surveyed. As this can be quantified using, for example, reference costs, this means there is an opportunity to (a) investigate whether or not different types of person with any given prevalence marker for various diseases or conditions (such as a self-reported LSI) make different demands on NHS secondary services, and (b) utilise that information to translate cohort-level prevalence counts into explicit resource need estimates which, crucially and uniquely, are independent of supply-side influences.

This concern with how ‘raw’ prevalence counts relate to actual resource needs, and the extent to which such might vary by age and/or socioeconomic factors, could also be addressed by focusing on comorbidity and, where possible, symptom severity. Here the aim would be to replace models predicting the likelihood that different person-types would have, say, CHD, with more nuanced models predicting the likelihood that their CHD was accompanied by more or less complex health-care needs, which would primarily reflect the extent to which their CHD was accompanied by other health problems. Although it would be difficult to express such information in terms of actual costs, the aim would again be to provide local and national commissioners with unbiased evidence on the relative needs of different populations. Determining an equitable allocation of resources between very different populations – as illustrated above in terms of the conflicting needs of young urban deprived populations with relatively high rates of mental illness as opposed to less-deprived but older populations suffering high rates of chronic degenerative disease – requires a far better evidence base than is currently available. This project’s approach to estimating the prevalence of health-care needs in local areas suggests how this evidence base might be constructed and, by incorporating HSfE-HES linkage and/or evidence on comorbidity, could be developed to provide the information needed for the equitable allocation of NHS resources to CCGs and below.

Further research into health-care equity

We have not been able to investigate all potential demand- and supply-side factors and the impact they may have on the utilisation of health-care services. As discussed at length in Strengths and limitations, this means that many of our findings, particularly with respect to mental health, must be treated as provisional. The problem is that very little of the substantial GP- and PCT-/CCG-level variation in mental health service use (relative to need) can be explained with reference to demography, ethnicity, deprivation or rurality. Other factors may, or may not, be influencing the utilisation of primary and secondary mental health services, but, until their role has been fully investigated, our conclusions regarding the significance, size and even direction of the sociodemographic factors that have been considered must be considered provisional. This, then, is our single most important recommendation for further research and, as such, it is where our own future research plans focus. The fact of the matter, though, is that as the Plymouth estimates will be publicly available, this is an issue which other researchers can also readily address. Indeed, we hope that the availability of these estimates will encourage others to investigate issues surrounding health-care equity in England, particularly with respect to mental health, for which evidence on local variations in needs has hitherto been available.

Finally, this project has made genuine progress in identifying factors which are associated with variations in health service use relative to need. Summary and discussion of findings summarises many of the principal findings. However, as this is a large-scale ecological study investigating relationships between modelled estimates of need and health service utilisation data, it has not been possible to investigate the underlying processes which give rise to observed variations in health service use. As any policy response to inequalities in the use of CVD and mental health services requires a detailed understanding of why they arise, it is now clearly necessary to take the findings reported here as a starting point for a detailed investigation of individual-level processes of candidacy and adjudication.