The association between primary care quality and health-care use, costs and outcomes for people with serious mental illness: a retrospective observational study

Measuring quality

There are many different conceptions and definitions of quality. One commonly used definition in health care is the degree to which services increase the likelihood of desired outcomes, consistent with current professional knowledge. 25 More specifically, health care can be considered to have six key aims: (1) safety, (2) effectiveness, (3) patient-centredness, (4) timeliness, (5) efficiency and (6) equity. 26 For practical purposes of measurement, health-care quality can be further divided into three components: (1) structures (e.g. levels of staffing), (2) processes (e.g. monitoring of cholesterol levels) and (3) outcomes (e.g. mortality rates). 27 There are multiple measures of quality, some of which are routinely recorded by health services and government agencies, and others that are measured for specific purposes by researchers or interest groups. These measures are often disease specific, reflecting the division of medicine into discrete specialties.

For primary care, a disease-specific focus is problematic because most patients have a combination of physical, psychological and social problems, and one of the core requirements of primary care providers is to provide integrated care that accounts for this complexity. 28 This approach often requires co-ordination across different disciplines and agencies, some of which may lie outside the health-care community. Measuring these aspects of care is challenging and they are therefore often neglected in favour of more mechanistic measures of quality.

Primary care quality in England

The first systematic attempt to measure quality of primary care in the UK was the Collings29 study in 1950. This was damning about the quality of care provided by many general practices, describing some as ‘bad enough to require condemnation in the public interest’. 29 The outcry raised by the report stimulated reform within the profession, leading to the establishment of the Royal College of General Practitioners in 1952. This was followed in subsequent years by mandatory vocational training in general practice as a specialty; incentives for physicians to work in groups; government financial support for practice facilities, staff costs and information systems; and limited pay for performance (P4P) (e.g. for conducting cervical screening). However, by the 1990s, quality of care was still largely unmeasured, despite accumulating evidence of wide variations in care, and the main formal check on substandard care was provided by the profession itself through the General Medical Council, which dealt with reports of negligent care. A series of high-profile scandals and subsequent public inquiries in the 1990s, including the Bristol Royal Infirmary Inquiry30 and the Shipman Inquiry,31 damaged trust in the medical profession, which finally accepted the need for reform and for more robust methods to monitor and improve quality of care.

From 1997 the government implemented a series of reforms aimed at improving quality of care, supported by additional funding for the NHS. These included the creation of the:

• National Institute for Clinical Excellence [later the National Institute for Health and Care Excellence (NICE)] to provide clinical guidelines and to conduct cost-effectiveness analysis of medical interventions

• Commission for Health Improvement (replaced by the Healthcare Commission and then the Care Quality Commission) to regulate and inspect health and social care services

• National Service Frameworks to provide guidelines and standards of care for some common chronic conditions.

Against this background of quality improvement efforts, in 2003 the national contract between the NHS and GPs was renegotiated in an effort to address an emerging recruitment and retention crisis. 32 It was recognised that existing methods of remuneration (including capitation payments, which constituted most of practices’ income) were at best neutral towards the quality of care provided and, in some cases, provided perverse incentives for under- or overtreatment. At that time, P4P schemes linking remuneration directly to performance on quality indicators were emerging in the USA as a potential method for improving quality of care and incentivising efficiency. It was therefore agreed that funding for practices in the UK would be increased, but that this would be achieved through a national P4P programme aimed at rewarding high-quality care.

The mental health domain of the Quality and Outcomes Framework

One of the clinical domains of the QOF is mental health, specifically SMI, which is the focus of our study. Indicators have been revised over time, but since 2006/7 practices have been required to record the number of patients with SMI on their practice list, document a care plan (CP), conduct an annual review of the patient’s physical health (AR), and monitor therapeutic levels and renal and thyroid health for patients treated with lithium. We excluded the lithium-related indicators from our analysis because they typically relate to patients with bipolar and mood affective disorder who are using this mood stabiliser, a subset of all people with SMI. Our study therefore focuses on the CP and AR indicators that are applicable to all patients on the practice SMI register. We describe these indicators in more detail in Chapter 3, Care quality indicators.

Deviations from study protocol

When considering the potential mechanisms by which each of the four quality indicators we examine (CP, AR, polypharmacy and continuity of care) might impact on outcomes, we recognised that some combinations of quality indicators and outcomes are highly unlikely to arise in practice, and so decided not to test them all explicitly. We outline our approach for which indicators are analysed against which outcomes in Chapter 3 and the rationale for the outcomes used for the respective indicators is provided in the empirical Chapters 4–7.

We deviated from the protocol in our analysis of the HoNOS, an outcome measure used by clinicians in secondary mental health services (see Chapter 6). On closer examination of the potential mechanisms through which primary care quality might impact on HoNOS outcomes within a secondary care setting, it became apparent that the causal pathway was unclear: any effect would need to be mediated by secondary care. We therefore changed this outcome to re-entry to mental health specialist community care (from primary care). Our original intention was to examine the level of severity of patients re-entering specialist mental health care using a clinician-rated outcome measure, the HoNOS. However, we faced practical challenges in implementing this analysis. HoNOS is recorded only for individuals treated in secondary care. Therefore, if primary care prevents relapse in people with SMI, the impact on HoNOS for these individuals is unknown. In addition, HoNOS is not recorded at regular intervals for people who access secondary care services and many individuals in our data set had only one HoNOS score. In theory, all patients have a HoNOS score completed on entry to mental health care, or soon after, but we found this not to be the case in the data. We looked at episodes when a patient re-entered specialist secondary mental health care after at least 12 months without secondary care activity. A HoNOS score was documented within a month either side of that date in only 23% of such episodes, and in only 36% of such episodes was a HoNOS score documented within 6 months either side. Such a sample of patients would not have been selected at random, as we could not assume that HoNOS completion was random. Rather, we suspected that those who had a HoNOS completed were different in some unobservable but systematic way from those who did not, and that any results we generated would not be generalisable to the wider SMI population.

We also deviated from the protocol in our analysis of costs (see Chapter 7). Initially, it was unclear that we would be able to obtain linked data to analyse mental health community care, but this became available during the project and we were therefore able to examine costs in primary, secondary and community mental health care.

We also added an additional outcome to the set of health-care utilisation measures, by including all unplanned admissions.

Aim

It was determined from the outset that patient and public involvement (PPI) would be a key aspect of this project. The benefits of involving experts by experience, particularly people with SMI, were threefold: (1) it allowed us to address questions that matter to service users; (2) it enhanced the interpretation of results, ensuring that the findings are plausible; and (3) it allowed continuing support for the dissemination of findings of this research in a manner suitable for lay people (e.g. production of a video showcasing the results of the study).

Methods

Two of our team members, LA and CD, have lived experience of SMI. These experts, by experience, have actively engaged throughout the project, with LA participating in early discussions with the team to help shape and clarify the research questions, including during the grant-writing phase. This helped focus on quality indicators that matter to service users. LA and CD were also members of the SSC, which provided them a further opportunity to steer the research programme.

Both experts by experience have played a pivotal role in the research team. They attended regular research team meetings, in which research findings were discussed, allowing them the opportunity to relay their interpretation of the results. This in turn influenced further analyses. Meeting minutes were distributed to these individuals, alongside the rest of the team, so that they could keep up to date with the progress of the project.

The experts, by experience, were able to liaise with and seek the expertise of various team members, as and when needed. This enabled them to feel supported and have adequate training for the role. The principal investigator was also available to help clarify things, when desired. Both individuals were reimbursed, based on guidance provided by INVOLVE. 38

Study results

As part of the team, LA and CD actively engaged with the literature review and contributed to the identification of quality indicators in primary care. The involvement of these individuals, and their unique perspective, allowed the team to keep sight of what is most important to service users. Both experts, by experience, have felt part of the team and have co-authored four articles during the project, with other research papers and dissemination activities to follow. 39–42

Discussion and conclusions

The insight provided by LA and CD has enhanced other team members’ understanding of SMI and the nature of the illness, and has had a positive impact on the project. They have participated in the project from its inception to conclusion and have felt genuinely engaged throughout.

Reflections and critical perspective

It was beneficial to have two experts by experience working on the project, as this facilitated the engagement of at least one individual at all times, particularly during periods in which one or the other may have been unable to attend because of illness.

Inclusion and exclusion criteria

We sought published examples of potential quality indicators that could readily be collected in primary care with reference to routine data. Search terms were identified by an information specialist, in conjunction with the research team. Included papers had the terms serious mental illness AND primary care AND quality indicator, including alternative spellings and synonyms. Studies on children, and on mental illnesses other than SMI, were excluded. All studies from January 1990 to February 2015 were considered for inclusion. No language restrictions were applied, although all search terms were in English, and all studies in English, German, Dutch and Afrikaans were considered due to the authors’ combined language knowledge.

Study selection

All titles were first reviewed by MG, TK, TD, RJ and Christoph Kronenberg (CK), a former member of the research team. All studies that two members of the research team indicated as potentially relevant were included in the abstract screening process. All abstracts were screened by LA, MG, TK, TD, RJ and CK, and full papers were obtained if two members of the team judged the abstract potentially relevant or in scope (i.e. covering SMI, primary care and quality indicators). Full papers were divided into four groups and independently reviewed by MG and CD, TD and LA, TK and SG, and RJ and CK. The focus of the selection was to identify papers that included relevant quality indicators that could be applied in primary care. It was evident that the definition of primary care varies between different countries, so we included indicators with elements of shared care between primary and specialist settings (e.g. prescribing and monitoring of antipsychotic medication), while acknowledging that, in some countries, those indicators may be more applicable to secondary care. Our search strategy complied with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) checklist.

Data extraction and analysis

A short description of each indicator from each paper was extracted and the descriptions for similar indicators were merged. The indicators were grouped into six relevant domains [(1) continuity of care, (2) substance abuse, (3) access to care, (4) medication management, (5) mental health assessment and care and (6) physical health]. The domains were selected by the research team as representing broad areas that would encompass all the chosen indicators. Given the main focus of our study, we decided if each indicator could, in principle, be measured from routine data or if primary data collection would be necessary. Furthermore, we checked if the identified indicators had ever been included in the QOF. We also assessed the quality of the evidence of the included studies using an adaptation of the Grading of Recommendations Assessment, Development and Evaluation (GRADE) guideline,51 and rated the quality of the evidence as high [systematic reviews or randomised controlled trials (RCTs)], moderate (non-randomised control studies or unsystematic reviews), low (expert opinion or uncontrolled studies) or not applicable (measure was extracted from the grey literature).

Indicator selection

We presented the final list of indicators to our SSC and discussed the list within the research team, as well as with our PPI team members. Based on these discussions, the evidence from the literature review and how readily the indicators could be translated into appropriate quality measures from routine data, we decided to choose two quality indicators that were not already included in the QOF for analysis in this study. Although some of the remaining potential indicators were, in theory, measurable from routine data, they would be unable to be measured in the primary care data used for this study owing to limitations of the clinical coding (e.g. quality of the data, missingness).

Summary of findings

To our knowledge, this was the first attempt to identify potential indicators of quality of primary care for people with SMI in a systematic way. Although we identified > 50 indicators that could potentially be captured and monitored using routine data, crucially, we note that the quality of the available evidence underpinning many of the indicators is relatively weak. We identified two indicators for further analysis on their association with outcomes using routine data: (1) continuity of GP care and (2) antipsychotic polypharmacy.

Strengths and limitations

The feasibility of collecting data for any set of quality indicators will vary across different health-care systems. Many countries have insurance or other systems, which routinely collect activity data in primary care. Some indicators are likely to require more effort to collect (e.g. patient questionnaires for perceived continuity of care). Our study focused specifically on finding indicators that could be monitored at relatively low cost to the UK health-care system.

The list of quality indicators identified here is much broader and more encompassing than the current list of indicators contained in the QOF SMI domain. However, even if good-quality metrics were identified for use, criticisms would remain around their adoption. These include measuring only what can be measured (in routine data), instead of focusing on measuring what matters, and the risk of unintended consequences arising from prioritising some activities at the expense of other non-incentivised activities. Moreover, there are gaps in the literature and in the indicators identified, as the service user perspective is not well represented in the literature. There is also an absence of quality indicators around aspects of the social environment, such as the stability of housing for people with SMI. Although such factors are important and may well influence health outcomes, the extent to which primary care could influence these factors may be very limited and hence it may not be appropriate to hold primary care practitioners responsible for improving quality in these domains. Finally, our search excluded non-published indicators and those written in languages other than those listed earlier (see Inclusion and exclusion criteria).

Implications for research and practice

Donabedian’s27 conceptual framework of quality of care suggests that indicators can usually be divided into three subcategories: (1) structure, (2) process and (3) outcome measures. As noted above, the vast majority of indicators included in this review relate to processes of care and although aspects of process are highly relevant, especially to patients, it is important to establish whether or not quality indicators also promote improved health outcomes. If so, there is a case for their inclusion in the QOF and other initiatives aiming to improve the care of people with SMI. For physical conditions, improvements in processes of care in primary care settings have been found to be associated with modest improvements in intermediate outcomes (e.g. cholesterol levels78) and quality of life,79 but associations with patient outcomes, such as emergency hospital admission, are weaker. 80 For SMI, the evidence is much more limited and suggests that higher provider performance on processes may not be associated with better patient outcomes. 43

This evidence gap is something that this research report explicitly addresses. We chose two areas in which we could construct a range of indicators from routine data (continuity of care and polypharmacy) and examine their respective associations with outcomes in Chapters 4–6.

Quality and Outcomes Framework indicators

Non-Quality and Outcomes Framework indicators

Clinical Practice Research Datalink

The Clinical Practice Research Datalink (CPRD) is a UK-based research service, jointly sponsored by the Medicines and Healthcare products Regulatory Agency and the National Institute for Health Research, as part of the Department of Health and Social Care. The CPRD collects fully coded and deidentified patient electronic health records from a network of general practices using the Vision^® (In Practice Systems Ltd, London, UK) software system. Data for this study were extracted from the CPRD GOLD database, which is drawn from practices using the Vision software system. We used the following CPRD GOLD data files:

• patient file – basic demographics and registration details

• practice file – collection and region information

• staff file – staff role and sex

• consultation – type of consultation

• clinical – medical history events, including symptoms, signs and diagnoses

• additional clinical details – information entered in the structured data areas in the Vision software

• referral – referrals to external care centres (e.g. hospitals), including specialty and referral type

• test – qualitative and quantitative test results

• therapy – prescriptions for drugs and devices issued by the GP.

Clinical information is captured as Read codes, which are recorded by practice staff (doctors, nurses, administrative staff, etc.) as part of routine data entry. Read codes94 are a hierarchical clinical data coding system used in primary care in the UK that classifies diseases, patient characteristics, procedures and tests. 95

A study conducted in collaboration between the CPRD and the London School of Hygiene & Tropical Medicine96 concluded that patients in the CPRD GOLD database are broadly representative of the English general population, in terms of age, sex, ethnicity and BMI. This was in line with previous research by the CPRD,98 which also demonstrated slightly lower mortality rates in younger age groups than in national rates, and a larger median practice size than the national average. 96 Owing to the geographical distribution of Vision software use, areas in the East Midlands, Yorkshire and the Humber and the north-east of England are under-represented relative to areas in Wales, Scotland and Northern Ireland.

We had access to CPRD data for the period of 1 April 2000 to 31 March 2014.

Hospital Episodes Statistics

Hospital Episodes Statistics (HES) comprises detailed records for all patients who received publicly funded care in public and private acute hospitals in England. The study data set includes HES data for those patients who had admissions and/or A&E department presentations during the study period. The HES data used in this study do not include hospital outpatient attendances. Admissions recorded in HES include those for physical health problems and for mental health problems, when the patient was admitted to an acute hospital rather than a specialist mental health inpatient facility.

Hospital admissions were classified using International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10), codes to identify SMI and ACSC admissions. (The codes used to classify admissions are listed in Appendix 2.)

Each linked HES record provides detailed information about the patient’s demographic characteristics, medical condition and care pathway. Inpatient data are reported at the level of finished consultant episodes and a new finished consultant episode is created every time a patient is discharged from the care of one consultant to another consultant. To capture the entire care pathway and derive correct admission numbers, we converted finished consultant episodes to continuous inpatient spells, which cover the entire period from admission to final discharge. Continuous inpatient spells also allow for transfers between providers of inpatient care.

We had access to HES inpatient data for the period of 1 April 2000 to 31 March 2014 and to A&E department data for the period of 1 April 2007 to 31 March 2014.

Mental Health Minimum Data Set

The Mental Health Minimum Data Set (MHMDS) (now called the Mental Health Services Data Set) is a nationally mandated routine data collection covering adults who receive care in specialist NHS mental health services, both inpatient- and community-based (outpatient) services. It comprises individual-level, administrative patient data, including data on visits, hospital admissions and other measures of service utilisation, such as contacts with health-care professionals. The study data set includes MHMDS data for all patients who received care in specialist NHS mental health services during the study period.

Mental Health Minimum Data Set data were used for the period of 1 April 2011 to 31 March 2014. MHMDS data prior to this date did not include all specific dates of activity (such as outpatient visits) required for the study.

Linked data sets

The anonymised CPRD GOLD primary care patient data can be individually linked to secondary care and other health and area-based data sets to provide a fuller picture of the patient care record. Linkages are available for English practices that have consented to participate in the linkage scheme, covering patients with a valid NHS number recorded in their GP record, who have not opted out. The CPRD is expanding its health-care data and research services to increase the number of data sets that are linked and made available on a routine basis to the research community. This study is one of the first to use the linked HES A&E data and the MHMDS. Approximately 75% of English practices participating in the CPRD permit linkage of their patients’ data in this way. 96 We test below whether there were systematic differences between practices with and without linkage (see Testing for linkage bias).

In addition to HES and MHMDS linkages, for this study, the CPRD data were also linked to the Index of Multiple Deprivation (IMD) data from 2010, for patients’ area of residence at the lower-layer super output area (LSOA) level. This provided the ranking of the patients’ area of residence according to twentiles of IMD data at the national level. Information was also linked at the practice level. Based on 2011 Census data,97 the practice postcodes held by the CPRD allowed practices to be classified in rurality. They were also classified based on postcodes in terms of estimated distances to the nearest hospital and nearest mental health inpatient facility. Exact IMD scores or distance information were not provided to the study team to avoid disclosing location information. Finally, mortality data from the Office for National Statistics (ONS) were also linked to the CPRD GOLD database.

To preserve anonymity, the data linkages were carried out by NHS Digital as the trusted third party of the CPRD. To avoid the risk of reidentification, the CPRD permitted only four data sets to be linked simultaneously. We therefore obtained the following two sets of linkages: CPRD–A&E–MHMDS–HES and CPRD–HES–ONS–MHMDS.

Information was provided by CPRD for all patients who were eligible for linkage and had an incident or prevalence diagnosis of SMI in their GP record on or before 31 March 2014.

We had two data use agreements with the CPRD approved by the International Scientific Advisory Committee, one for the CPRD GOLD database linked with HES A&E data, the MHMDS, patient IMD, distance to provider and rurality indicators (protocol number 15_213), and one for the CPRD GOLD database linked with HES inpatient data, ONS data, distance to provider and rurality indicators (protocol number 14_168).

Testing for linkage bias

We compared the set of patients eligible for linkage to HES and MHMDS with a random sample of 3000 patients with SMI attending practices that participated in CPRD but had not permitted the linkages, to ensure that our practice populations were not biased in observable ways. Figure 4 shows that the linked and non-linked samples were similar in age and sex distribution.

FIGURE 4.

Clinical Practice Research Datalink samples of patients based on eligibility for linkage to other data sets. (a) Not linked; and (b) linked.

Sample

Patients were included in the sample if the conditions listed in Chapter 3, Study samples were met. Continuity is measured over defined periods of time, so the observation period for each patient was divided into periods of 3 months, dating from their first date of observation. Patients were followed until outcome or censoring [with censoring due to the patient changing general practice, death or the end of the study period (31 March 2014)].

Outcome measures

We separately examined time to three types of unplanned hospital use: (1) A&E department presentations, (2) unplanned admissions for SMI and (3) unplanned admissions for ACSC (as outlined in Chapter 3, Outcomes). For each type of outcome, we considered only the first observed instance (presentation or admission), as this could have influenced subsequent care quality in unobserved ways. For example, CPs may be updated in secondary care at the end of an admission period and this may not be reflected appropriately in the GP electronic records.

The occurrence of the outcome was measured in the 3-month period, t, and continuity of care was measured over the 12 months prior to the outcome period, t-1 to t-4. The outcome variable was a binary variable for each 3-month period, indicating whether or not the event occurred in that period.

Quality and Outcomes Framework indicators of quality

The analysis included time-varying care quality indicator variables for CPs and ARs. These binary variables indicated if a CP or AR had been recorded within the 12 months prior to the current 3-month period. The choice of the 12-month window to determine expiration status was based on the QOF guidance that a CP or AR should be undertaken and reviewed annually.

From 2006/7 to 2010/11 the QOF annual review indicator entailed the patient being given appropriate health promotion and prevention advice. As outlined in Chapter 3, The annual review indicator and Table 3, in 2011/12 this broad indicator was split into more specific indicators (a record of alcohol consumption, checks of blood pressure, BMI, blood glucose or glycated haemoglobin, ratio of total to high-density lipoprotein cholesterol and, when appropriate, cervical screening). To explore ARs for the full period of our analysis, we formulated an aggregate indicator. This signified that the patient had at least three of the four ‘health risk’ checks (blood pressure, BMI, cholesterol and glucose) documented within a 3-month period. 101 The date of the aggregate indicator was the date of the final check. The AR variable in the analysis was equal to 1 for the current 3-month period if there was an AR (using the original Read coding) or an aggregate indicator date within the 12 prior months. Appendix 3 provides an overview of how the rates of these indicators changed over the period of analysis.

Non-Quality and Outcomes Framework measures of quality: practitioner continuity

We used three indices measuring different dimensions of GP continuity: the COC, UPC and SECON. 102

The COC captures dispersion of visits across GPs as follows:

in which N is total number of visits and n_j is number of visits with GP j.

The UPC is the proportion of a patient’s visits that are with the GP most frequently seen by the patient in that year:

The SECON is the proportion of consecutive pairs of visits which are to the same physician:

in which i is the visit number, N is the number of visits, N – 1 is the number of sequential pairs of visits. s_i = 1 if visit i and visit i + 1 are to the same provider, and s_i = 0 if otherwise.

Each index ranges from 0 (lowest continuity) to 1 (perfect continuity). Illustrative examples are shown in Table 5.

We measured continuity over 12 months (four 3-month periods), considering only face-to-face visits with GPs. There is no standard level for ‘high’ and ‘low’ continuity, so we classified practitioner continuity as ‘high’ if the level of continuity was above the median for the index and ‘low’ if at or below the median level.

A minimum of two visits is required to calculate COC and SECON, but to improve index stability we set the minimum to three visits and set continuity as ‘undefined’ below this level. We constructed a set of categorical variables based on visit frequency and on whether continuity was low or high. This allowed for different effects of continuity for frequent and less frequent users of family practice, as suggested by previous research. 103 Visit frequency was classified into low (zero to two visits), moderate (three to five visits) and high (six or more visits). These categories correspond to tertiles of the full visit distribution. Continuity indices were defined as low or high based on the median value of each index: COC low (0–0.35) and COC high (> 0.35); UPC low (0–0.67) and UPC high (> 0.67); SECON low (0–0.17) and SECON high (> 0.17).

Periods were then classified into five categories according to continuity level and visit frequency in the prior 12 months: low visit frequency (with continuity undefined – the base category); moderate visit frequency with low continuity; moderate visit frequency with high continuity; high visit frequency with low continuity; and high visit frequency with high continuity.

Statistical analysis

The necessity of creating periods for continuity measurement led us to employ discrete-time survival analysis to model all three care quality indicators (continuity, CP and AR) simultaneously. Although the outcomes of interest were (effectively) continuous measures (as we had day-level data on when the events occur), these were converted into discrete outcomes for each period, to match the measurement of continuity. The model evaluated the association between risk of the outcome in a particular 3-month period and each of the three care quality indicators in the prior 12 months.

The hazard rate at period t is the probability of observing the outcome for individual i in period t, conditional on the individual ‘surviving’ in the sample to period t (i.e. no censoring and the outcome was not observed in prior periods for that individual). A complementary log–log (cloglog) model, which is the discrete-time analogue of the continuous-time proportional hazards model, was estimated for each outcome. This model is useful when the occurrence of the outcome is rare.

A flexible piecewise constant baseline hazard function was applied by specifying dummy variables for each 3-month period. This assumes that the hazard function is constant within each period, but can vary across periods and avoids imposing a more restrictive functional form on the model. The resulting exponentiated coefficients can be interpreted as HRs, the discrete-time counterpart of the hazard from a continuous-time proportional hazards model. 81 The HR is equal to the proportional change in the hazard of the outcome for a unit change in the variable.

The hazard rate (h) is specified as:

in which T_i is the discrete survival time variable for individual i; t represents the time period of interest, X_it is the vector of time-varying factors, including a constant term; Z_i is a vector of time-invariant factors; D is a series of time period dummy variables representative of the baseline hazard; and vi∼N(0,σv2) is normally distributed individual unobserved heterogeneity.

Our main modelling approach accounted for individual unobserved heterogeneity, v_i. Owing to the incidental parameter problem104 that arises through specifying individual fixed effects to represent such heterogeneity in non-linear models, we instead assume unobserved heterogeneity is normally distributed and specify this as a function of the means of time-varying variables. This is often termed a correlated random-effects model. Following Mundlak,105 we assume:

in which X¯i is the mean of the time-varying variables, D¯ the mean of the time dummies, a is a constant and c_i represents the remaining unobserved heterogeneity assumed to be normally distributed. The mean variables (X¯i,D¯) capture confounding by unobserved time-invariant patient factors (e.g. long-standing illness, health-seeking behaviour) that drive both primary care quality and use of hospital services. The coefficients of the period-specific levels (X_it, D_t) can be interpreted as the effect of the quality indicator specific to that 3-month period.

To allow comparison of our results to previous studies examining the effect of continuity of care, we also estimated models that did not specify individual heterogeneity (v_i) as a function of the means of the time-varying variables (i.e. by estimating Equation 4 directly). These models allow for normally distributed individual heterogeneity, but it is assumed to be uncorrelated with the explanatory variables contained in the model.

All models included observed patient characteristics as explanatory variables and adjusted standard errors (SEs) for clustering at the practice level. We estimated separate models for each of the three continuity indices.

Robustness checks

We tested the sensitivity of the results to the level of visit frequency at which continuity was classified as ‘undefined’. The minimum level for measuring continuity (and corresponding categories for low vs. moderate visit frequency) were set to two or four visits rather than three visits as in our main analysis.

Sample

The sample consisted of 19,324 patients attending 215 practices, observed for 15.8 3-month periods on average (range 1–28 periods). Table 6 presents the characteristics of patients in the sample. Half of the sample (50.3%) had an A&E department presentation at some point during the observation period, 13.1% had an admission for SMI and 12.8% had an ACSC admission. Using a three-visit minimum to define continuity, median (mean) values for each continuity index were 0.35 (mean 0.46) for COC, 0.67 (mean 0.65) for UPC and 0.17 (mean 0.26) for SECON. A CP had been documented in the previous 12 months for 40% of the patient-periods observed and an AR (or the alternate aggregate indicator) in 69% of periods. The Spearman’s rank-order correlation between COC and UPC was 0.94 (p < 0.001), between COC and SECON it was 0.55 (p < 0.001) and between UPC and SECON it was 0.47 (p < 0.001). Mean COC in periods with a CP in the previous 12 months was 0.47, compared with 0.45 in periods without a CP in the previous 12 months (two-sample t-test of difference in means: p < 0.001). The equivalent for UPC was 0.67, compared with 0.66 (p < 0.001), and for SECON it was 0.24 compared with 0.23 (p < 0.001).

Association between quality of care and outcomes

Table 7 presents the association between the quality measures and each outcome from the discrete-time survival analyses, with practitioner continuity measured by the COC. The results in the second column are HRs for the key variables of interest from our preferred specification, the correlated random-effects model, which accounts for unobserved confounding. Results from a model that does not account for unobserved confounding, the standard random-effects model, are presented in the third column for comparison. Full results from the correlated random-effects models are presented in Table 8.

Having a CP or AR documented in the previous 12 months was associated with lower hazard of all three outcomes. The relative magnitude of the hazard reduction associated with CPs was 26% for A&E department presentation [HR 0.74, 95% confidence interval (CI) 0.69 to 0.80], 33% for SMI admission (HR 0.67, 95% CI 0.59 to 0.75) and 27% for ACSC admission (HR 0.73, 95% CI 0.64 to 0.83). The reduction associated with ARs was 20% for A&E department presentation (HR 0.80, 95% CI 0.76 to 0.85), 25% for SMI admission (HR 0.75, 95% CI 0.67 to 0.84) and 24% for ACSC admission (HR 0.76, 95% CI 0.67 to 0.87).

For those with moderate visit frequency (three to five visits in the previous 12 months), higher GP continuity was associated with 11% lower hazard of A&E department presentation (HR 0.89, 95% CI 0.83 to 0.97) and 23% lower hazard of ACSC admission (HR 0.77, 95% CI 0.65 to 0.92), but not with hazard of SMI admission. For frequent attenders (six or more visits in the previous 12 months), higher GP continuity was associated with 26% lower hazard of ACSC admission (HR 0.74, 95% CI 0.62 to 0.87), but not with risk of A&E department presentation or SMI admission.

The standard approach (random effects) to modelling continuity, which does not account for unobserved confounding and so is not our preferred specification, produced different results (see final column of Table 7). Compared with the correlated random-effects model, the standard approach found larger risk reductions associated with higher GP continuity, none with annual reviews and CPs was associated with higher (rather than lower) risk of SMI admissions.

Some patient characteristics were also associated with risk of unplanned hospital utilisation. Living in a more deprived area was associated with a higher hazard of all three types of hospital care, as was white ethnicity (vs. black or minority ethnicities). Age had different patterns of association with the three types of hospital care: those in the middle range (aged 36–65 years) had lower risk of A&E department presentation than younger or older patients, whereas younger patients (aged < 45 years) had the highest risk of SMI admission, and the risk of ACSC admission steadily increased from the age of 46 years. Physical comorbidities were associated with higher risk of A&E department presentation and ACSC admissions, but lower risk of SMI admission. Those with a diagnosis grouped with bipolar disorder (and not a diagnosis grouped with schizophrenia) were at lower risk of SMI admission.

Table 9 shows that using a different measure of GP continuity, either UPC (measuring density of visits to a single provider) or SECON (measuring sequential continuity), produced a similar pattern of results to those obtained using the COC (measuring dispersion), although with some differences. There was an association between higher continuity and lower risk of SMI admission for frequent attenders when using the UPC, and for those with moderate visit frequency using the SECON index. In addition, there was an association between higher continuity for frequent attenders and lower risk of A&E department presentations when using the SECON. The use of different continuity indices made no difference to the results for CPs or ARs.

Robustness results

Varying the minimum number of visits deemed necessary to measure continuity, from three visits in the main analysis to two or four visits, did not substantially change the overall findings. These results are presented in Table 10.

Summary of findings

The two QOF indicators, CPs and ARs, were both associated with substantial reduction in the risk of unplanned hospitalisation for patients with SMI. Among those who had a CP documented in primary care in the previous 12 months, the hazard of an A&E department presentation, ACSC admission or SMI admission were 26%, 27% and 33% lower, respectively, than for patients without a recent CP. The corresponding reductions associated with ARs were 20%, 24% and 25% lower, respectively. All of these effects were statistically significantly different from 1 when compared against a critical value of alpha = 0.05.

The association of continuity of care and unplanned hospitalisations varied with the underlying frequency of patients’ visits to their GP. For those with moderate visit frequency (three to five visits in the previous 12 months), higher GP continuity was associated with 11% lower hazard of A&E department presentation and 23% lower hazard of ACSC admission, but not with risk of SMI admission. For frequent attenders (six or more visits in the previous 12 months), higher GP continuity was associated with 26% lower hazard of ACSC admission, but not with risk of A&E department presentation or SMI admission.

Strengths and limitations

The primary strength of this analysis was its use of detailed information on the timing of care events, which allowed us to study the association between time-varying measures of care quality and hospital utilisation, and to ascertain the action–outcome ordering necessary for meaningful causal inference. Furthermore, by using statistical modelling techniques that separate within- and between-patient variation, we controlled for time-invariant patient characteristics, when possible, and reduced the scope for unobserved confounding. To our knowledge, this is the first time this modelling approach has been used to analyse the effect of care quality indicators on outcomes for patients with SMI.

There are a number of limitations to our analysis. First, although we attempted to account for time-invariant unobserved patient characteristics, we cannot rule out time-varying confounding bias that remains uncontrolled, which may have contributed to our findings. For instance, during periods of deterioration leading to admission, GPs may have less opportunity, or put less emphasis on, spending time on preventative measures, such as CPs. This would induce a negative correlation between having a current CP and risk of unplanned hospitalisation. Our results should therefore be interpreted as showing association with many sources of confounding avoided, but not strictly causation.

Second, we restricted the sample to patients registered with the same practice for the year preceding diagnosis, to allow us to include measures of historical health-care utilisation and medical history. However, the restriction may have excluded individuals with a more severe form of SMI if this led them to move practice. In addition, the sample may have included some patients who were not newly diagnosed if their diagnostic information had not been transferred to a new practice or to a new electronic recording system.

Third, patients in England can register with only one general practice and may face greater administrative barriers to changing practices than in other countries. This may result in less apparent benefit of continuity of care in our results than would be observed in other health-care systems. When patients are free to choose their provider, high practitioner continuity may reflect a strong, valued therapeutic relationship, which may in turn improve outcomes. In England, patients may have a constrained choice of general practice, so that higher continuity could, in principle, be less beneficial for some individuals.

Fourth, the collection of HES A&E data commenced on an experimental basis in April 2007 and captured 62% of national A&E department attendances in the first financial year. The experimental label was lifted by April 2013 after capture had increased to > 80%. However, there is no reason to consider that the capture of attendances was related to if a patient had a CP or AR, meaning that this should not impact adversely on our comparative analysis.

Finally, the clinical outcomes we examined are important, as they represent some of the excess health risks for people with SMI and carry substantial health-care costs. However, they are not the only outcomes that matter. Both people with SMI and GPs value continuity of care in itself, as part of how they experience giving and receiving care. 106,107 Regular physical reviews may uncover health issues that require treatment, but which would not necessarily have led to an ACSC admission. Broader outcomes important to people with SMI may also be affected by the different measures of care quality, including social functioning and health-related quality of life. 108

Implications for research and practice

The association of the QOF indicators with reduced hazard of unplanned hospital use might suggest that these activities help patients avoid hospital for conditions that can be effectively managed in primary care. As hypothesised, proactive reviews of physical health (ARs) may lead to earlier detection of physical health problems, whereas documentation of patients’ current health status, early warning signs, triggers, social support needs, co-ordination arrangements with secondary care and preferred course of action in the event of a clinical relapse (CPs) could improve the management of the patient’s health overall. Both of these indicators could thereby prevent deterioration, reducing the need for urgent care, with CPs also having the potential to direct patients into appropriate community-based services during periods of deterioration, and thereby avert hospital use.

Similarly, the finding that practitioner continuity is associated with lower hazard of A&E department presentations and ACSC admissions supports the hypothesis that seeing the same GP over time could improve the doctor–patient relationship, facilitate communication and trust, reduce fragmentation of care, and allow for earlier detection of problems and/or more effective and appropriate management of problems as they arise.

The potential for better-quality care by GPs to reduce A&E department attendances and hospital admissions suggests that improving quality may help reduce demand on secondary care services and hence possibly reduce NHS expenditure, a question we explore in Chapter 7.

From a methodological point of view, the results presented in this analysis differed substantially from a previous study addressing similar research questions, but using practice-level aggregate data of hospitalisation rates and QOF achievement rates. 43 Although the practice-level analysis identified a positive association between AR and the risk of unplanned SMI hospitalisation (i.e. high-quality care increases unplanned hospital use), the current analysis found an opposite effect. The current results may be more plausible given the hypothesised mechanism by which high-quality care affects patients’ health and health-care needs. Although the studies are not directly comparable – as they are based on different samples – it is reasonable to assume that the more aggregate-level study suffers from aggregation bias and reverse causality bias.

Our results were also informative about trends in practitioner continuity in English primary care over time and across patient groups. We found slightly lower levels of continuity than those in an earlier study of family physician continuity for patients with long-term mental illness in the UK,109 but much lower than those found in studies looking only at specialist mental health care. 110,111 Higher, and rising,112 rates of consultation in family practice may contribute to these differences. Practitioner continuity in English family practices may be affected by reductions in full-time working and increasing practice size in future. Average UPC scores for all patients in 2011–13 were 0.61. 103 Comparison with our results suggests that this dimension of family physician continuity is not lower for patients with SMI than for patients overall. Based on our results, we predict that further decreases in practitioner continuity in primary care may have negative spillover effects onto the secondary care sector.

Study design

The study employs an observational cohort design. We constructed the antipsychotic prescribing profile of patients from primary care records, which we linked to hospital and mortality data.

The argument underpinning a cohort study design is that effectiveness is assessed under the usual circumstances of health-care practice rather than the possibly atypical circumstances of a RCT, and in many situations it is not possible to randomise to evaluate an intervention. As with all observational studies, validity relies on rigorous design and adjustment of confounding factors to minimise selection bias. Significant progress towards this direction has been made by studies from Denmark121 and Finland,122 which focused on the effect of polypharmacy on mortality. However, studies that explore associations between polypharmacy and inpatient hospitalisations123,124 and A&E department attendances125 suffer from important weaknesses that stem from failure (or inability due to lack of data) to model the timing of polypharmacy episodes and outcomes. The analysis presented in this chapter42 improves on the fundamental issue of confoundedness, by employing a Cox survival analysis model that analysed time to each outcome, adjusting for both time-invariant confounders and time-dependent polypharmacy and monotherapy. 126

Sample

Patients were included in the sample if they had at least one antipsychotic drug record during the observation period and all conditions listed in Chapter 3, Study samples, were met. For reasons explained in Chapter 3, Study samples, we use two samples: one for the analysis of A&E department attendances and one for the analysis of all other outcomes.

Patient outcomes

We investigated the association between polypharmacy and the occurrence of three outcomes: (1) unplanned hospital admissions (all cause), (2) A&E department attendances and (3) mortality.

Definition of antipsychotic polypharmacy

There is no consistent definition of polypharmacy in the literature. We defined polypharmacy as the concurrent use of two or more antipsychotic substances for at least 30 days. The overlap period allowed for cross-tapering between substances. A longer overlap period has a higher risk of misclassifying polypharmacy as monotherapy, whereas a shorter overlap may misclassify switching between substances as polypharmacy.

We considered 33 antipsychotic substances covering first-generation antipsychotic drugs or typical antipsychotics, second-generation antipsychotics or atypical antipsychotics, and depot antipsychotics127,128 (see Appendix 4 for substance product codes used).

The CPRD data provide the date a prescription was issued, but the duration of prescriptions is not always recorded. We therefore inferred treatment duration from the total quantity (number of units) prescribed and the numeric daily dose (ndd) (number of units per day) when duration was absent. Furthermore, the daily dose was missing for 23% of prescriptions. For these prescriptions, we imputed the dose using an imputation strategy explained in Appendix 4. Less than 0.02% of prescription records were dropped from the analysis, as they had implausibly large estimated duration. From the prescription dates and durations, we constructed the patient’s medication profile (times at which the patient was on any antipsychotic medication and on polypharmacy).

We calculated two measures of polypharmacy prevalence. First, we calculated the annual prevalence of polypharmacy as the number of patients with at least one polypharmacy episode in a year divided by the total number of patients observed during that year. Second, we calculated the rate of polypharmacy, defined as the sum of all patients’ polypharmacy days in a year over the sum of all patients’ days at risk of polypharmacy in that year. The latter measure is an improvement over the commonly reported point estimates of polypharmacy prevalence, which measure the proportion of eligible individuals on polypharmacy on a given day (see Appendix 4 for a proof).

Statistical analysis

Semi-parametric Cox hazard models129 were applied to estimate the effect of polypharmacy on the time to the first occurrence of each of the three outcomes. The model adjusts for censoring, which may occur because (1) a patient dies, (2) registration with the practice ends or (3) the study period ends, prior to the outcome of interest occurring. The follow-up period – time from entry to the sample until the outcome occurs or censoring – was different from the observation period for outcomes other than death.

An individual could have multiple polypharmacy episodes. On each day during the study period the patient was in one of three states: (1) received no antipsychotic medication, (2) monotherapy (on one antipsychotic or on more than one, but for < 30 days) and (3) polypharmacy. To model this, we introduced two time-varying binary variables: (1) ‘no antipsychotic substance’, which took a value of 1 during periods the patient was not on an antipsychotic drug and zero otherwise, and (2) ‘polypharmacy’, which took a value of 1 during periods the patient was on two or more antipsychotic substances for > 30 days and zero otherwise. The results were interpreted with regard to monotherapy, which was the reference category.

Robustness checks

We performed two checks. First, to assess the impact on the results of a shorter or longer overlap period, we estimated our models assuming overlap periods of 14, 60 and 90 days. Second, we explored whether or not SMI admissions, which are a subset of all unplanned admissions, were affected differently by polypharmacy.

Robustness check results

Table 14 shows the results of sensitivity analyses that explored the impact of changing the length of overlap in the definition of polypharmacy. The estimated relationships were generally insensitive to the length of overlap. The only exception is unplanned admissions: when the lower boundary of the overlap duration was reduced to 14 days, polypharmacy was associated with an increased hazard of unplanned admission of about 21%.

The results for SMI admissions are reported in Appendix 4. The results show no association between SMI admissions and polypharmacy.

Summary of findings

The principal findings of the study are threefold. Annual polypharmacy prevalence fluctuates over time by between 5% and 6%. Relative to those on monotherapy, patients who are not on antipsychotic medication have an 8.2% (18.6%) higher hazard of an unplanned hospital admission (A&E department presentation). Relative to monotherapy, polypharmacy is not significantly associated with mortality, unplanned hospital admissions or A&E department presentations.

Strengths and limitations

This study is a step forward towards understanding the links between polypharmacy and health-care utilisation and mortality. As with all observational studies, validity relies on rigorous design and adjustment of confounding factors to minimise selection bias. We addressed this fundamental issue employing a three-step strategy.

First, we constructed the antipsychotic prescribing profile of patients from primary care records. In the UK, family practices provide the majority of care for patients with SMI,24 including the management of long-term prescribing. Therefore, unlike previous studies that used solely hospital data to investigate polypharmacy,130 we defined polypharmacy and monotherapy from primary care data. Second, we linked primary care data with hospital and mortality data at patient level, to determine the sequence of polypharmacy episodes, hospital utilisation and mortality. Third, we employed a Cox survival analysis model that analysed time to each outcome, adjusting for both time-invariant confounders and time-dependent polypharmacy and monotherapy. 126 By specifying polypharmacy as a time-dependent variable, we addressed the statistical challenge arising in cases in which the exposure is not present throughout the entire time of observation. The use of a large, linked data set, coupled with a suitable survival analysis model, provided more robust estimates of the effects of polypharmacy on outcomes than would be possible with aggregate data or a cross-sectional design.

There are three main limitations to the study. First, the measures of health status and health-care utilisation prior to diagnosis of SMI may not fully depict the complexities of health status, including severity of the condition. Second, imputing the treatment duration for a number of prescriptions may introduce measurement error in the calculation of polypharmacy. Last, we explored the effect of polypharmacy on broadly defined outcomes. Future research could investigate whether effects vary by reason for admission or for particular combinations of antipsychotic medication.

UK guidelines36 recommend against combining antipsychotic drugs, except as a last resort. These recommendations are based on limited supportive evidence for superior efficacy of polypharmacy over monotherapy, as well as concerns that combined antipsychotics are associated with an increased risk of side effects. Our study cannot draw conclusions on the polypharmacy effect in terms of efficacy and tolerability and, bearing in mind the limitations of an observational study (despite its advanced design), cannot substitute for RCTs. Its contribution lies in providing real-world evidence on the effectiveness of polypharmacy.

Implications for research and practice

Our study examined the overall effectiveness of polypharmacy relative to monotherapy by investigating associations between polypharmacy and three patient outcomes. We found no evidence of a positive or negative effect of polypharmacy on mortality, inpatient hospitalisations or A&E department presentations. At a policy level, these findings do not rule out polypharmacy options, but highlight the need for further research on the appropriateness of polypharmacy.

Current UK guidance36 recommends antipsychotic monotherapy as a treatment option, and our results provide further supportive evidence by establishing a negative association between antipsychotic monotherapy and hospitalisations. This may be because drug therapy helps to stabilise the patients’ condition and allows better management of their physical health. Being prescribed an antipsychotic may be associated with closer or more regular clinical monitoring in the primary care setting, as set out in the guidelines, which recommend that prescription of an antipsychotic should be considered as ‘an explicit individual therapeutic trial’,117 accompanied by detailed requirements for monitoring. The latter may facilitate timely diagnosis and treatment of health problems, avoiding the need for hospital care.

The lack of significant association between polypharmacy and any of the three outcomes provides some reassurance that the safety profile of polypharmacy may be comparable to that of monotherapy. For a shorter overlap period (14 days or longer), which captured more cross-tapering in the definition of polypharmacy, we observed an increase in the risk of unplanned admission. One explanation is that patients who changed drugs might have had more unstable disease profiles and/or that changing drugs further destabilised their condition. This suggests a need for close monitoring in the first few weeks of cross-tapering, when the risk of unplanned hospitalisation is higher.

Study design

We restricted our sample to individuals whose mental health had been managed fully within primary care for at least 12 months (described as ‘stable’) (see Chapter 3, Study samples). This observational cohort analysis used discrete-time survival analysis to explore the effect of primary care quality on the risk that a ‘stable’ patient with SMI would re-enter specialist mental health care. The outcome was measured in 3-month periods and the primary care quality indicators were measured over the prior 12 months.

Sample

We limited the sample to people who had not been treated in specialist mental health care for at least 12 months prior to the current quarter, as captured in the MHMDS, to try to isolate the effect of primary care quality. The MHMDS includes specialist inpatient and community care. We hypothesised that if the patient was seeing both specialist mental health services and their GP within a 12-month period, the specialist care would be more influential on their subsequent mental health and use of specialist services.

Care quality indicators

The two primary care quality indicators examined in this analysis were CPs and GP continuity, as outlined in Chapter 3, The care plan indicator, for CPs and Chapter 3, Continuity of general practitioner consultations, and Chapter 4, Non-Quality and Outcomes Framework measures of quality: practitioner continuity, for GP continuity. These were hypothesised to influence patient mental health and the probability that a patient would remain stable in the care of the GP rather than requiring specialist care. This analysis uses the UPC to measure GP continuity, as re-entry to specialist mental health-care services reflects deterioration of mental health and being seen by the same GP who knows the patient could help stabilise the patient and prevent re-entry. As in Chapter 4, periods were classified according to both number of visits in primary care and level of GP continuity in the 12 months prior to the start of the period. The classifications were low visit frequency (with continuity undefined – the base category); moderate visit frequency with low continuity; moderate visit frequency with high continuity; high visit frequency with low continuity; and high visit frequency with high continuity (as outlined in Chapter 4, Non-Quality and Outcomes Framework measures of quality: practitioner continuity).

Outcomes

Re-entry to specialist care, although a measure of health-care utilisation rather than health itself, was used as a potential proxy for deterioration in the patient’s mental health. We did not examine first entry to specialist care, as most patients are treated in a specialist context when first diagnosed with SMI. Based on clinical advice, we based our approach on the assumption that only patients whose SMI was stable and well managed would be discharged from specialist services back to the care of their GP for ongoing management. If a patient then re-entered specialist mental health care, this was considered likely to indicate a deterioration in their mental health and a need for escalated care beyond that provided by the GP.

The outcome variable was a binary variable, indicating that within a 3-month period, the patient had some recorded activity in either inpatient- or community-based specialist services, having had no secondary care activity recorded in the previous 12 months. The MHMDS classifies periods of patient engagement with services into cluster episodes (CEs), intended to depict the dates during which the patient receives treatment while allocated to a particular category of mental health problem or needs (termed a ‘cluster’). However, we found that it was common for activities recorded in MHMDS to be outside the bounds of CEs, suggesting recording of errors either of the CE dates or of activities that did not actually occur. We assumed that the former was more likely and based our classification of periods of specialist mental health care on the dates of activity, rather than CE dates.

Statistical analysis

The analysis applied discrete-time survival analysis, with 3-month periods as the unit of time, following essentially the same approach as that outlined in Chapter 4, but with a different outcome variable (see Chapter 4, Statistical analysis for additional detail on this method). Note that, as the outcome indicator is re-entry to secondary care, a HR < 1.0 on a quality measure implies that the measure is associated with a better outcome (a reduced probability of re-entry to secondary care). This approach accounted for the different timing of entry and exit to the observation period for individuals in the sample, and different timing of the outcome. It captured the sequence of events (primary care quality preceding re-entry to specialist care), as well as the duration of time spent in each state. We used a panel-complementary log–log (cloglog) model with a correlated random-effects specification (Mundlak correction),105 to account for unobserved patient-specific effects.

The hazard of re-entering specialist care in a 3-month period – either in the community or as an inpatient – was defined as a function of primary care quality in the previous 12 months, and of patient and practice characteristics as common to all the analyses presented in this report. SEs were adjusted to account for clustering of multiple patients within each practice.

Summary of findings

We explored whether or not better quality of primary care was associated with reduced risk of re-entering specialist mental health-care services, either in the community or as an inpatient. Our results indicate no statistically significant associations between either of the two primary care quality measures and the risk of re-entry to specialist care. This suggests that neither the continuity of GP visits to the same doctor nor the presence of a CP in the previous 12 months reduced the chance of deterioration of mental health status, implied by the need for re-entry to specialist care.

However, a lower risk of re-entry was significantly associated with other explanatory variables: being aged between 45 and 65 years (compared with a younger age group), having had a SMI diagnosis for > 2 years and having two or more physical comorbidities. It is possible that individuals who have had a diagnosis for a longer period may potentially have a more stabilised condition than newly diagnosed patients, which may reduce their risk of re-entry to services. Those with a larger number of physical comorbidities may have been at lower risk if management of physical health took precedence over referral to secondary mental health services for the patient’s mental health condition.

A higher risk of re-entry was associated with greater socioeconomic deprivation, being white and having a diagnosis that included schizophrenia-type classifications, in addition to bipolar-type classifications. The latter may suggest that patients with more complex mental health conditions are more at risk of re-entry and the link with deprivation may reflect the commonly observed greater utilisation of many types of health-care services by socially disadvantaged populations experiencing higher levels of need. 131

Overall, only 7% of the sample categorised as being ‘stable’ re-entered specialist secondary care services over a 12-month period, with very few of these patients re-entering inpatient care directly (seven patients). Despite the lack of association with the indicators of primary care quality, re-entry is therefore an uncommon occurrence and suggests that the management of stable patients in primary care is, by and large, successful.

Strengths and limitations

The main strengths of the analysis relate to the longitudinal nature of the data and the methods employed, which improve on standard cross-sectional analyses based on aggregate data. First, by linking primary care and secondary care data together at the patient level we can observe the timing and sequence of events. This allowed us to identify a sample of patients who were categorised as having been ‘stable’ for a period of time and then observe their utilisation of specialised secondary care from that time onwards. As with the other analyses presented in this report, the use of survival analysis techniques allowed us to exploit fully this data linkage to investigate the relationship between primary care quality and utilisation of secondary care. Owing to the rarity of such data linkages, there has been very little previous research on this topic that uses the richness of individual-level data available to this study.

One of the main limitations of the analysis relates to our inability to use the HoNOS scores as a measure of severity of mental health condition, as originally planned (as described in Chapter 1, Deviations from study protocol). This also means that our results may be confounded, given that severity is likely to be an important factor in determining the degree to which the condition can be controlled solely in primary care. Some patients who were deteriorating (and, therefore, went on to appropriately re-enter secondary care) may have been offered more continuity of care and care planning as a priority by their GPs (indicating good-quality care), rather than re-entry indicating a failure of primary care quality.

Re-entry to specialised services may not be a reliable indicator of deterioration of SMI condition, as we do not know from these routine data the reason for and circumstances leading to re-entry. Most patients re-entering the secondary care sector in our sample entered community care services rather than going directly into inpatient care. Depending on the nature of the community services received, they may not be targeted at addressing deterioration, but more at enhancing the quality of general support received. Again, it is feasible that re-entry into secondary care is not always a negative outcome that signals poor management. Last, we had limited measures of primary care quality and it is feasible that there were other aspects of that care that influenced the risk of deterioration of condition, as well as re-entry into secondary care.

Implications for research and practice

The absence of a statistically significant positive association between aspects of quality of primary care and re-entry to specialist secondary care services suggests that prevention of a deterioration of SMI is not solely within the control of GPs. This is not surprising, given the complexity of SMI conditions and the range of services and settings appropriate to meet these needs. The main implication for research arises from the challenges arising from the HoNOS data, which are detailed in Chapter 1, Deviations from study protocol. Good measures of health outcome for people with SMI are central to the ability to undertake research that goes beyond consideration of process measures. Improved recording of HoNOS scores in secondary care would facilitate this and further allow more generalisable conclusions to be drawn from research. Further research to investigate the underlying reasons for re-entry to specialist secondary care services would provide greater insight into the role of primary care in supporting patients to avoid deterioration of SMI conditions.

Study design

Quarterly costs of health-care utilisation for primary care, inpatient admission, A&E department attendances and community mental health services were estimated using a bottom-up costing approach for 3 financial years (12 quarters), from 1 April to 31 June 2011 (quarter 1 of 2011/12) through to 1 January to 31 March 2014 (quarter 4 of 2013/14), applying unit costs for the 2013/14 financial year. 132 We examined the association of costs in a financial quarter with primary care quality in the previous 12 months, accounting for a range of individual and practice characteristics, and unobserved time-invariant heterogeneity.

Sample

The sample consisted of 150,748 quarter observations across 16,485 adults with SMI for the financial years 2011/12 to 2013/14, as outlined in Chapter 3, Study samples.

Table 19 shows the distribution of characteristics across the sample. Just over half of the patients in the sample (56%) were observed for all 12 quarters. The mean age in the sample was 52 (median 50) years. The majority (71%) of patients were white. Around one-third of patients had at least one physical comorbidity and two-thirds had comorbid depression. Three-quarters of patients were current or ex-smokers and 20% had been recently diagnosed with SMI. People in the sample were more likely to live in urban areas than rural areas and in more deprived areas.

Care quality indicators

As noted in the Introduction, we focus in this chapter on the two QOF indicators (described in Chapter 3, Quality and Outcomes Framework indicators) as measures of primary care quality, as practices were financially incentivised on these measures. We also include GP visit continuity (described in Chapter 3, Continuity of general practitioner consultations).

Costing methods for primary care data

Table 20 provides an overview of all the sources of utilisation data and unit costs (not just primary care). Primary care utilisation data were extracted from CPRD records based on Read codes recorded by practice staff, as part of routine data entry. Three types of primary care activity were included in the analysis: (1) consultations, (2) drugs prescribed and (3) diagnostic tests.

Unit costs for consultations were obtained from national estimates of the cost per minute of staff time for GPs and practice nurses for 2014. 133 We used individual visit duration data to calculate visit costs rather than a standard unit cost because of evidence suggesting that consultations for patients with mental health problems are likely to be longer than average consultations. 137–140 We considered multiple visits in a day for a patient to a single staff member as duplicates, but included visits to different staff members on the same day.

The CPRD visit duration variable was based on the time that a patient’s electronic record was open for use by a member of practice staff, and our use of this variable assumed that GPs and nurses would open the electronic file when they were about to see the patient and close it at the end of the consultation. However, files may also be opened for review or other purposes at times that do not correlate to patient care, or may not be closed at the end of a consultation. We therefore applied restrictions when visit duration was recorded as zero minutes or > 120 minutes. A recording of zero minutes indicates that the file was opened for < 1 minute. The CPRD data set included two date variables: (1) the date of data entry and (2) the date associated with the consultation, as entered by the practice staff member. If the duration was zero minutes and if the date of data entry was later than the visit date, we assumed that this reflected retrospective recording of a visit that should be costed and applied the sample median duration for that type of visit (see Table 20). Otherwise, we assumed that the entry did not reflect an episode of patient care and dropped these observations from the data. We capped the maximum duration at 120 minutes, assuming that this reflected a file being left open but without ongoing patient care.

A cost of £3.60 per minute of patient contact for GPs included travel and qualification costs, but excluded nursing care staff costs, as we were accounting for these by counting nurse visits separately. A cost of £0.88 per minute of patient contact for practice nurses included qualification costs.

Table 21 outlines the average visit duration for different types of primary care visits observed in the sample.

The costs of drug prescriptions were derived from the prescription costs analysis for 2014,134 which applied the British National Formulary (BNF) edition 66141 classification system. This provided the net ingredient cost per item according to the drug tariff. It did not include any discounts, dispensing fees or adjustment for income, and was based on the usual quantity dispensed for that drug. Numbers of drug prescriptions were taken from the therapy section of the CPRD data set. We used data at the BNF subparagraph level when available or otherwise at the most specific BNF level available (from paragraph, section or chapter level). This approach did not account for other items listed in CPRD as ‘therapy’ that were not included in the BNF, such as dressings or glucose test strips.

National average unit costs for diagnostic tests were taken from the NHS reference costs for the financial year 2013/14,135 covering diagnostic imaging, diagnostic services, pathology services and outpatient procedures. The reference costs listed the average cost across the NHS of providing a particular health-care service. Utilisation volume of each test was taken from Read terms (descriptive terms attached to Read codes) recorded in the test section of the CPRD data set. These were classified, following clinicians’ advice, into groups that corresponded to the NHS reference cost categories. For example, Read terms for blood tests recorded in CPRD data included ‘liver function test’, and ‘serum creatinine’, which were both classified as ‘clinical biochemistry’ and allocated to that unit cost from the reference costs. A supplementary file is available online and lists the Read terms included in the CPRD data and how they were classified into reference cost categories (see Report Supplementary Material 1).

Costing methods for community mental health care

We used a subset of the activity recorded in MHMDS as measures of utilisation of community mental health services. Specifically, we counted health-care professional contacts (consultations), CPA reviews, electroconvulsive therapy (ECT) sessions and day attendances, and assigned units costs reported by the Personal Social Services Research Unit (PSSRU)133 and the Electronic Staff Record (ESR). 142 The cost of consultations was dependent on the staff type and the duration of the contact.

We used staff groups defined in the PSSRU unit costs 2014,133 namely hospital-based scientific and professional staff, nurses and doctors. Each of these groups was divided into subcategories (called bands) with different costs per hour. Other types of staff (ambulance staff, administration and estates staff, health-care science, general payments) had cost equal to zero. When the occupation code was not recorded in MHMDS, we used the job role to assign a consultation to one of the three staff groups. The PSSRU unit costs reported costs per hour for these groups and bands, which included not only the staff members’ salary but also cost of training and overheads, and were therefore a measure of the total cost of 1 hour of work by the different staff groups.

However, the occupation codes available in MHMDS identified the three staff groups defined in the PSSRU unit costs 2014,133 but not the bands within them. So, we needed a way to calculate an average cost for each of the three staff groups, which we could then match to MHMDS.

We used the ESR to allocate each staff type to a staff group and a band, and then calculated the (weighted) average cost of each staff group (which took into account the composition, in terms of bands, within the staff group).

The ESR had two parts, one containing information regarding the number of workers of different types (measured as full-time equivalents, rather than as head counts) and one containing information regarding the salary for those types of staff. The data for staff numbers were available for each month, so we calculated the average over the financial year (1 April–31 March). The salary data were available by financial year. We used data for 2013/14. ESR contained information for different types of organisations; we focused only on providers that have submitted data to MHMDS (mental health trusts). We used the occupation code and the salary to allocate each staff type to a staff group and a band (the occupation code identified the staff group and the salary was used to identify the band). We then attached the PSSRU unit cost for each staff group and band, and calculated the average cost per hour for the three staff groups (hospital-based scientific and professional staff, nurses and doctors) as a weighted average of the different types of staff in the mental health trusts (i.e. the average cost per hour of, for example, a nurse reflected the number of nurses in the different bands across all mental health trusts). We also calculated an overall average (including staff groups with zero cost), which we used as a cost for consultations in MHMDS without staff type.

As the duration of the contact was an important part of the calculation of the cost of a consultation, we needed to make sure that there were no observations with zero or missing duration. If the duration was zero, we used the information regarding attendance to adjust this: if the patient had attended we initially recoded the duration to missing (which we later replaced) and if the patient had not attended (or the information regarding attendance was missing) we recoded the duration to 1, a minimum amount of staff time that would be required to record that the patient did not attend. When the duration was missing, we replaced it with the average for those from the same staff group for that consultation medium.

Table 21 shows the mean duration of outpatient mental health visits observed in the data. We matched the occupation code of each consultation to the staff groups in PSSRU and assigned to them the cost calculated using ESR and PSSRU. For observations without staff type, we used the overall average cost per hour.

We restricted the number of consultations per day to a maximum of two for each type of staff member. Regarding CPA reviews, MHMDS provided the date only and not the staff involved or duration. We assumed, following clinical advice, that reviews (which were supposed to be multidisciplinary) involved a psychiatrist and a nurse and lasted 30 minutes, a lower-bound estimate. Cost per ECT session was based on past costings136 updated to 2013/14 values using the hospital and community health services pay and price inflation series. 143 Day attendance cost was based on the cost of day care provided by local authority social services for people with mental health problems. 133

Costing methods for hospital data

Admissions to specialist mental health inpatient facilities were recorded in MHMDS and admissions to general hospitals (including admissions to non-specialist mental health providers) were recorded in HES. We costed these separately.

Statistical analysis

We calculated costs for each patient for each quarter for each sector – primary care, general hospital costs, specialist mental health care (including inpatient- and community-based care) – and for total costs (as the sum of the three sector costs). To provide some indication of the proportion of health-care costs that are related to mental health as distinct from physical health for this population, we identified a subset of costs that could be clearly labelled as mental health ‘related’, comprising all specialist mental health-care costs (inpatient and outpatient), and mental health drug costs in primary care.

The aim of the empirical analysis was to relate the quality of care for a patient to the costs the patient incurs in primary and secondary care. Individual patient health-care cost data are usually highly skewed, with a long right tail. They are also usually characterised by a large mass at zero (patients who do not incur health-care costs in a particular period). These features need to be accounted for in the empirical modelling. This was undertaken by the use of two-part models. 145 The first part used a probit model to estimate the probability of incurring a positive cost. The second part used a generalised linear model (GLM)146 to estimate the level of cost for those patients who incurred a positive cost. The GLM allowed for flexible forms for the distribution of costs to take into account its skewed nature. Based on the results of the modified Park’s test147 and the Pregibon link test,148 we selected a gamma distribution and log-link function for the second part of the model.

The probit model to estimate the probability of incurring costs can be written as:

The GLM to estimate the level of costs for individuals who incur costs can be represented as:

For both Equations 6 and 7, y_it was the cost for individual i in quarter t (and follows a gamma distribution in Equation 7), Qual_itq are the three quality indicators (CP, AR and GP visit continuity), X_it is a vector of observed time-varying explanatory variables and Z_i is a vector of observed time-invariant explanatory variables. c_i and m_i represent unobserved individual-specific random effects. In Equation 6, Φ() is the standard normal cumulative density function required to estimate a Probit model.

Equations 6 and 7 are non-linear models, which do not allow for the unobserved individual-specific effects to be specified as fixed effects via the use of a set of dummy variables to represent individuals (often referred to as the incidental parameter problem). Instead, we accounted for the individual-specific effects, c_i and m_i, using the Mundlak approach105 (as outlined in Chapter 4, Statistical analysis), by modelling these as functions of the within-individual means of the observed time-varying variables, as follows:

in which X¯ is the mean of X_it for each individual and a_i is a random error term assumed to be normally distributed and uncorrelated with X¯i. We adopted an equivalent specification for m_i in Equation 7.

Marginal effects of each explanatory variable on overall estimated cost were calculated by obtaining predictions from the first and second parts of the model for each individual at the base level of the variable and at an alternative level, and then taking the mean of these effects over all individuals. For example, for a binary variable, such as a CP, the predicted probability [P_i] of non-zero costs for individual i was obtained from the first part of the model with the CP variable set at zero and at 1 (P^i0 and P^i1), with all other variables at their observed levels. The equivalent predictions of the level of cost with CP set at zero and 1 were obtained from the second part of the model (y^i0 and y^i1). The point estimate of the average marginal effect of a CP on overall cost was then estimated as:

Standard errors were calculated by accounting for clustering of patients within practices. For the average marginal effects SEs (seAME¯) were obtained by bootstrapping using 1000 replications,149 for which:

and R is the number of replications, AME_r is the estimated marginal effect in replication r and is the mean of the bootstrapped estimates.

Robustness check

As an alternative to using the consultation duration data in CPRD and MHMDS, we estimated costs using the average duration for different types of consultations derived from the 2006/07 General Practice Workforce Survey150 that were applied by Curtis. 133

Resource use and costs

The mean quarterly health-care cost per patient was £1309.96, with £653.38 (50%) identifiable as related to SMI. Figure 6 shows that total costs decreased with age until middle age and then increased for patients aged > 65 years, whereas the proportion of costs related to SMI steadily decreased with age. For patients aged 19–35 years, 64% of total costs were related to SMI, whereas for patients aged > 65 years the proportion was 34%, reflecting a greater emphasis on management of physical comorbidities.

FIGURE 6.

Costs per quarter for people with SMI by age group: total and SMI related.

Table 22 shows that primary care costs (£235) made up approximately 18% of total mean quarterly costs, general hospital costs (£481) constituted 37% and specialist mental health-care costs (£594) accounted for around 45%.

Table 22 also provides mean cost per quarter for subgroups of patients. The mean total costs increased with age, level of deprivation and number of comorbidities and were higher for white patients (vs. black or minority ethnicities) and for patients diagnosed with SMI more recently. Patients with both schizophrenia- and bipolar-type disorders had higher costs than those with schizophrenia-type disorders alone, with bipolar-type disorders associated with the lowest costs.

Of the quarterly observations, 8.4% had zero total cost, 12.0% had zero primary care cost, 62.8% had zero community mental health cost and 85.5% had zero hospital cost. Approximately one-third (33.3%) of the quarterly observations had no CP or AR recorded in the prior 12 months, 5.4% had a CP but no AR recorded, 40.3% had an AR but no CP recorded and 20.9% had both recorded. Overall, just over one-quarter of observations (26.3%) had a CP recorded in the prior 12 months and nearly two-thirds (61.3%) had an AR recorded.

Table 23 outlines the mean and median quarterly utilisation of the different types of service making up these costs across all individuals in the sample.

Regression results

The average marginal effects of quality on cost estimated from the two-part models are presented in Table 24. CPs in the 12 months prior to the start of the current 3-month period were associated with lower health-care costs in the current 3-month period, in terms of overall costs and in primary care and hospital costs, but not mental health-care costs. ARs were associated with lower primary care, mental health care and total costs, but not with hospital costs. High GP visit continuity in the prior 12 months was associated with lower primary care costs in the current 3-month period, but not with total costs or any of the other sectors of costs.

Other factors were also associated with costs. Greater socioeconomic disadvantage was associated with higher costs across sectors, as were white ethnicity (vs. black or minority ethnicities) and physical comorbidities. Older age was associated with higher primary care and hospital costs, but with lower mental health-care costs. Patients whose general practice was further from the nearest acute facility had lower hospital costs. Men had lower costs overall and in primary and hospital care. Smoking was associated with higher costs overall and in primary care. SMI diagnosis of a bipolar-type disorder was associated with lower costs than diagnosis of a schizophrenia-type disorder or both categories, and longer duration of SMI was associated with lower costs also.

Table 25 shows how the two parts of the model contributed to the estimated marginal effects for total costs. It reports the coefficients from the estimated models for having positive cost (see Equation 6) and for costs given that they were positive (see Equation 7). In the first part of the model, a negative coefficient implied that the probability of incurring any cost is reduced. In the second part of the model, the coefficients were the proportional change in cost (β = (∂y/∂x)/y). CPs were associated both with a higher probability of having positive costs and with a lower level of cost if a cost was incurred, whereas ARs showed no statistically significant association with the probability of having non-zero costs, but were associated with a lower level of cost if a cost was incurred. Overall, both CPs and ARs were associated with lower costs (as shown in Table 24). High GP visit continuity was associated only with a lower probability that a non-zero cost would be incurred, but overall is not associated with a statistically significant difference in cost.

Older age was associated with a higher probability of incurring non-zero costs and age > 65 years was associated with a higher cost, if a cost was incurred. White ethnicity was associated with both a higher probability of incurring cost and with a higher level of cost if any cost was incurred, and the same pattern was seen for physical comorbidities. The reverse pattern (lower probability and lower cost) was observed for men.

Robustness check

The robustness check supports the findings from the core analyses, assuming the average duration for each type of consultation, as applied in Curtis133 (instead of using the CPRD duration data), did not make a substantial difference to the estimated marginal effects of the quality indicators, as shown in Table 26 (compare with the results in Table 24).

Summary of findings

We calculated the costs of health-care utilisation in each quarter for people with SMI using a bottom-up costing approach for 16,485 patients and a total of 150,748 patient-quarter observations. Using 2013/14 unit costs, the mean quarterly health-care cost per patient was £1309.96, with £653.38 (50%) identifiable as related to SMI (rather than to physical health). Converted to 2018 terms (using the health domain of the Consumer Price Index151), mean quarterly health-care cost per patient was £1432.81, with £714.65 related to SMI. Total costs decreased with age until middle age, and then increased for patients aged > 65 years, whereas the proportion of costs related to SMI steadily decreased with age. For patients aged 19–35 years, 64% of total costs were related to SMI, whereas for patients aged > 65 years the proportion was 33%, reflecting a greater emphasis on management of physical comorbidities. Primary care costs (£235) made up approximately 18% of total mean quarterly costs, general hospital costs (£481) constituted 37% and specialist mental health-care costs (£594) accounted for around 45%.

We estimated two-part models of costs, examining the effect of patient characteristics and quality of care on (1) the probability that they had any (i.e. non-zero) NHS cost in a quarter and (2) the level of cost for those with positive cost. We estimated that having a CP in the previous 12 months was associated with lower health-care costs in the current quarter. On average, the overall cost per quarter per patient was reduced by £53.40, primary care cost by £9.47, hospital costs (excluding those due to SMI) by £25.53 and specialist mental health costs (including inpatient care) by £12.50. Having an AR in the previous 12 months was associated with lower primary care costs (£9.21), mental health-care costs (£14.97) and total costs (£33.32), but not with hospital costs in the current quarter. High practitioner continuity in the prior 12 months was associated with lower primary care costs (£2.81) in the current 3-month period, but not with total costs or with hospital or mental health service costs.

Cost implications

We found in previous chapters that CPs and ARs for people with SMI were associated with reduced risk of hospital SMI-related admissions, ACSC admissions and A&E department attendances, and that greater continuity reduced the risk of hospital ACSC admissions and A&E department attendances. It was therefore to be expected that CPs, ARs and greater GP continuity were associated with reduced hospital costs.

General practices were financially incentivised via the QOF to provide CPs and ARs to people with SMI and these payments had an opportunity cost in terms of forgone care elsewhere in the NHS. It was beyond the scope of our study to attempt a comparison of the value of the gains to people with SMI from CPs and ARs with the opportunity costs of providing them. However, we could compare the QOF payments to practices for providing CPs and ARs with the reductions in costs for people with SMI.

Having a CP in the previous year reduced total costs in a quarter for a person with SMI by £53.40, or £213.60 per year (Table 27). There were 353,695 CPs in all practices in England in 2012/13, so that if the cost saving estimates were valid for all 8020 practices in England taking part in the QOF, and not just those in our sample, the total annual cost saving from CPs for patients with SMI would be £75.5M. Practices were awarded 47,965 points under QOF indicator MH10 for achievement of CPs and, at a national average price per point of £133.76, the total QOF payment for CPs was £6.4M. The total reduction in costs for people with SMI was thus around 11 times as large as the total QOF payments to practices for CPs and the net saving of NHS expenditure was £69.1M. As the other columns in Table 27 suggest, the reductions in NHS costs for people with SMI in each of the three sectors were considerably smaller than the total QOF payments for CPs.

We can make similar calculations for ARs, although because of the change in the definition of ARs from 2011/12, we used data on QOF payments for the AR indicator MH9 in 2011/12, rather than in 2012/13. As ARs were associated with smaller reductions in total costs for people with SMI, the net cost implications were less dramatic and were, indeed, only positive in relation to total costs.

Although the calculations in Table 27 suggested that the cost savings for people with SMI across all three sectors for CPs and ARs were much greater than the corresponding QOF payments to practices, this does not imply that withdrawing these payments would have increased total costs for people with SMI, as we did not attempt to estimate the effects of the payments on the number of CPs. However, the magnitude of the cost saving associated with CPs suggests that even a small reduction in CPs, due to removal or reduction in the QOF incentive payment, would have had an opportunity cost in terms of care forgone in other parts of the NHS. Alternatively, to deliver the same level of care overall, the budget would have needed to increase.

Strengths and limitations

Tracking health-care resource use of people with SMI is particularly problematic because they receive NHS care in primary, secondary and community care. To assess accurately the implications of better quality of primary care on health-care expenditure, detailed information at an individual level is required over a substantial period of time. A strength of this study is in exploiting data linked across a number of administrative sources at an individual patient level. Successful linkage and management of such data is non-trivial and requires both substantial expertise and computational resources. The ability to observe health-care use in continuous time over a 3-year period provided a richness of data of sufficient detail to allow us to link resource use to practice quality accurately and robustly. Costing health-care services, particularly primary care consultations, was not straightforward and required a number of assumptions to be made, for example about the duration of a consultation when these were recorded in the data set as zero or > 2 hours. Although we believe that our assumptions were reasonable, better-quality recording in primary care would make cost estimates more reliable. However, complex data derived from different administrative systems will always require some assumptions to be applied when data are imputed and cleaned to prepare for analysis. Health-care cost data are typically highly skewed with long right-hand-side tails in their distribution, and often contain a large mass point at zero. These features were observed in our data and required non-standard regression models to estimate the link between costs and covariates. With a large mass point at zero, the two-part model applied here was the appropriate way to derive marginal effects of the impact of quality on costs.

Implications for research and practice

We found that younger patients with a SMI had a high level of need and generated additional costs for the NHS, mainly in specialist mental health care. There was also considerable variation between patients: most did not generate any hospital costs in a given year, but those who did were more likely to be admitted to general rather than mental health wards and may have had long and costly admissions. Further research to identify characteristics of patients who are more likely to develop comorbidities and to require admission may help to identify interventions that could improve outcomes for people with SMI and reduce secondary care costs. A better understanding of the characteristics of patients who attend general rather than mental health hospitals and variations in admission patterns may also help identify ways to manage costly resources more efficiently.

We found that the greatest volume of care for people with SMI was provided by specialist services, with patients having three times as many contacts as with primary care. As we used primary care records in this research, we had limited information on the quality of specialist services, and this warrants further investigation. In terms of quality of primary care, only one-fifth of patients received both a CP and an AR, and one-third received neither, despite the existence of financial incentives for practices to provide them. These low rates may have had serious implications for patient health and NHS resources: patients who were in receipt of a CP had fewer emergency contacts and generated lower primary and secondary care costs than those without, and patients who were reviewed annually generated lower primary and mental health-care costs. These interventions require further exploration, as the pathways to improved outcomes remain unclear (e.g. ARs were not, as might be expected, associated with reduced secondary care costs, but with reduced mental health-care costs).

According to the national QOF results,152 7832 patients nationally (12.7% of registered people with SMI) were ‘exception reported’ for the CP indicator in 2017/18, with wide geographical variation (from 6.6% in London North East and Central to 18.6% in Midlands Central). Another 41,784 patients (9.1%) did not receive a CP, despite not having a reason for exclusion recorded. There may therefore be potential to improve outcomes and substantially reduce costs for people with SMI by moving towards full implementation of CPs and ARs. Although most patients receive the recommended care in the context of the financial and reputational incentives offered by QOF, variation in achievement rates between regions and practices suggests that higher rates are achievable. Further research into the reasons why patients do not receive CPs and ARs could suggest potential solutions to this problem, and address the geographical variations in quality and costs of care.

Better evidence is needed on the impact of primary care quality on a broader set of health outcomes for people with serious mental illness

Most indicators included in our systematic review (see Chapter 2) relate to processes of care, and although highly relevant, especially to patients, it is important to establish whether or not they lead to improved health outcomes. With the exception of mortality (which we examined only in relation to polypharmacy), all our study outcome measures were based on health-care utilisation. We had planned to examine effects on the severity of SMI using the clinician-rated HoNOS outcome measure (see Chapter 6); however, HoNOS is not recorded in primary care and the quality of reporting in secondary care is patchy. We therefore used another measure of health-care utilisation instead: re-entry to secondary care. Although these outcomes are likely to impact on health status and are also important in terms of their implications for the efficient use of NHS resources, they are indirect measures of health outcomes.

To be included in the QOF, quality indicators must be supported by NICE evidence-based clinical guideline recommendations or evidence from systematic reviews. This is why the large majority of indicators we identified are not currently part of the QOF. Better measures of health outcome are essential to determine the effectiveness of current and potential QOF indicators.

Good measures of health outcome for people with SMI are central to the ability to undertake research that goes beyond consideration of process measures, for example measures of social functioning, patient experience and health-related quality of life. Routine recording of quality-of-life measures would enable the cost-effectiveness of incentive policies and other measures of primary care quality to be determined. Looking beyond direct health outcomes to the impact of high-quality care on wider aspects of mental health and well-being would mean including data on other non-health outcomes, such as employment and housing status.

The mechanisms by which primary care affects utilisation of hospital care need to be understood

Although not all hospital care can, or should, be avoided for people with SMI, the prevention of acute episodes of the condition is a key objective of good primary care. Our study assesses ‘real-world’ primary care quality, rather than care provided in accordance with strict protocols in a clinical trial context. Our study data are therefore likely to encompass wide variations in the content and delivery of primary care (e.g. an AR). A better understanding of the nature and quality of CPs and ARs, which might be best gained through qualitative research rather than using observational data, would help understand who receives CPs and ARs, why some patients do not receive incentivised care and how that may affect their utilisation of secondary care. It is unclear which components, or combination of components, of care are most effective in preventing avoidable hospital care. The timing and setting for providing care also needs to be understood. For example, further research to investigate the underlying reasons for re-entry to specialist secondary care services would provide greater insight into the role of primary care in supporting patients to avoid the deterioration of SMI conditions.

A clear conceptual framework would be an essential first step, identifying the potential mechanisms by which aspects of primary care quality impact secondary care use among this specific patient group. Interface issues with care provided in other settings, such as informal care provided by families or friends, or social care, could also be identified within the framework. Hypotheses could then be drawn and tested empirically. If suitable data do not exist – which is highly probable – primary studies may be needed.

Understanding variations in capacity to benefit could help to deliver person-centred care

Our study found that CPs and ARs were associated with significantly lower hazards of A&E department attendances and emergency admissions and, in some cases, with lower costs. However, these are average effects and mask a range of effects at individual level. A better understanding of the variations in outcomes could enable clinicians to target efforts accordingly. For example, capacity to benefit may vary according to each individual’s clinical and sociodemographic characteristics. Equally, certain subgroups may be underserved or face greater barriers to accessing care; evidence on supply side facilitators could be used to improve the delivery of person-centred care.

A better evidence base is needed on what outcome measures and quality indicators are most highly valued by people with serious mental illness, so that efforts can be prioritised according to what matters most to patients

There is a growing literature on what matters to people with SMI and how to better incorporate patient perspectives into outcome measurement in mental health, including the development and application of patient-reported outcome measures and experience measures. However, we need to know which outcome measures and quality indicators matter most to people with SMI, so that efforts to measure these in routine data can be prioritised. Focus groups with service users could provide views on what outcomes matter and various elicitation methods, including qualitative research and discrete choice experiments, can be used to prioritise outcome and quality measures using patient preferences. However, any new indicators should be properly consulted on before implementation. An evidence base on how patients prioritise outcome and quality measures could help to focus resources and efforts towards areas that matter most to patients, for example if people with SMI value having a CP renewed when nothing has changed over time, or whether they care about continuity with the same GP or would more highly value ease of seeing any GP urgently. These preference sets may also vary by different groups (e.g. age, sex or ethnicity) and it would be important to understand better this variation to facilitate a targeted approach to service delivery.

Evidence is needed on how patient outcomes are influenced by the interfaces between health care, social care and informal care

Serious mental illness is a complex set of diseases that typically affects individuals across the life course and in all areas of their lives. Focusing on only one or two parts of the health-care system when examining patient outcomes risks ignoring wider social determinants of health and well-being and increases the chances of drawing spurious inferences about cause and effect.

To the best of our knowledge, our study was the first to cover the full patient care pathway for people with SMI, and this was facilitated by the use of MHMDS, the collection of which is a welcome effort to improve the coverage and quality of data on people with mental health problems. However, details on the type of care provided by different health and social care professionals are sketchy and are reported inconsistently. There is almost no information on informal social support networks (friends and family). Many of these types of care may be substitutes or complements. The inter-relationships between these types of care need to be clarified through further research, so that care can be integrated around the individual. This is likely to require the collection of primary data, as such factors are not recorded in routine care data.

A fuller understanding of resource implications is needed to inform resource allocation

This study has indicated the potential for better-quality primary care to reduce costs in secondary care. However, a more detailed research agenda investigating the mechanisms for identifying, releasing or transferring and reinvesting resources ‘saved’, particularly across sectors, would be welcome. This would provide valuable evidence on which to base decisions about the allocation of resources, as well as informing the integration of care, which is often hampered by financial barriers.

Data coverage and linkage, specifically for mental health data, should be improved to address further research questions

Throughout this report, we have commented on some of the limitations of our analyses that arise from constraints in the nature of the data to which we had access. Improving the scope of routine outcome measurement in mental health services and facilitating linkages from diverse data sources would enhance the ability to address further important research questions. Facilitating researcher access to relevant data sets in a timely way is also a key aspect of enhancing both the quantity and quality of mental health research undertaken.

Overview

The practice-level analysis investigated whether or not better quality of primary care is associated with unplanned hospital admissions. It was an extension of our previous project1 and a precursor of the patient-level analysis in the current project.

Quality of primary care was captured by the practice performance on the CP and AR indicators discussed in Chapter 3, Quality and Outcomes Framework indicators. Although the patient-level analysis utilised patient-level information on the two indicators, in the practice-level analysis CPs and ARs were analysed at the practice level using two measures of achievement on these indicators: (1) the population achievement rate and (2) the reported achievement rate, which we define in the next subsection. Our two research questions were formulated as follows:

1. What is the association between the CP population and reported achievement rates with emergency hospital admissions for SMI?

   The analysis was carried out for the period 2006/7–2013/14.

2. What is the association between the AR reported achievement and emergency hospital admissions for ACSCs?

   The analysis was performed for the period 2011/12–2013/14. The impact of the AR indicator on all physical hospital admissions during the period 2006/7–2010/11 was assessed in our previous SMI project. For reasons explained in the next subsection, the population achievement rate was not analysed.

In a previous study,1 we conducted a practice-level analysis to investigate the association between achievement rates on the CP and AR indicators and hospital admissions for SMI and physical conditions for the period of 2006/7 to 2010/11. The current practice-level analysis differs in two ways: first, we analysed ACSCs instead of all physical conditions; second, we extended the study period to include 3 more years (2011/12–2013/14).

We combined patient- and general practice-level data for the study period of April 2006 to March 2014. Specifically, we merged QOF data from around 8500 general practices in England with admissions data from HES (aggregated to practice level) for the study period of April 2006 to March 2014, using a unique general practice identifier. We linked these data to publicly available information on general practice characteristics, characteristics of their patient population, such as disease prevalence, and population characteristics, such as deprivation and other potential confounders recorded at small area level. We also controlled for measures of access to care.