Variation in Compulsory Psychiatric In-Patient Admission in England: A Cross-Sectional, Multilevel Analysis

Research aim 1

To establish if, and to what extent, compulsory admission rates vary across different places and between providers and commissioners of (mental) health care in England, and, if so, where this variance occurs.

Research objective 1

The geography of mental health care is complex and comprises overlapping, partially nested spatial clusters that include individuals, general practices, commissioners (previously PCTs), providers (NHS trusts or independent providers) and regional health authorities (SHAs). We will use multilevel statistical models to measure and compare variance in the rate of compulsory admission across England at each of these levels.

Research aim 2

To explore and quantify the extent to which variance in the rate of compulsory admission can be explained by:

• the individual characteristics of people using services (e.g. age, sex and ethnicity)

• the local area (i.e. contextual) socioeconomic and sociodemographic factors (e.g. small area socioeconomic deprivation, urbanicity and ethnic density)

• the features of local mental health services [e.g. bed availability, number of admissions, length of stay (LOS) in hospital and investment in mental health services].

Research objective 2

There are many factors that may influence (i.e. increase or decrease) the risk of a patient being compulsorily admitted to hospital. We will undertake substantial data linkage using multiple routine (secondary) data sources in order to create models that incorporate information about people, places and health-care organisations. We will use multilevel statistical models to examine the impact of specific risk factors at both the individual (patient) level and at each of the spatial levels in the Mental Health Minimum Data Set (MHMDS) for England in 2010/11, simultaneously, on variance in the rate of compulsory admission across England.

Finally, we estimated associations between risk factors operating at different spatial levels and compulsory admission, after adjusting for other model covariates.

These research aims were addressed through secondary analyses of routinely collected national mental health-care data, described in detail in the next chapter.

Ethics

Ethical approval for this study was obtained from Warwick Medical School Biomedical Research Ethics Committee on 12 January 2012 (REF 166-01-2012).

Brief overview of the Mental Health Minimum Data Set

The MHMDS was initially developed to support monitoring of the mental health element of the then government’s ‘Health of the Nation’ strategy (1992–7). 56 Since 2003 the MHMDS has become a mandatory quarterly data return for all providers of NHS-funded specialist mental health services for adults, and for some independent sector providers since 2011. 54,57 The MHMDS provides data on a wide range of mental health services, including those provided in hospital, in outpatient clinics and in the community (where the majority of people in contact with these services are treated). 54 The MHMDS is an approved NHS Information Standard that aims to deliver robust, comprehensive, nationally consistent and comparable person-based information on patients in contact with specialist secondary mental health services. It captures key information from the mental health-care pathway. 58 The MHMDS supports a variety of secondary use functions such as commissioning, clinical audit, research, service planning, inspection and regulation, monitoring key government policies and legislation including the National Service Framework for Mental Health and MHA, performance management and benchmarking at both local and national levels, and national reporting and analysis. 58

Systems for MHMDS data collection, collation and curatorship have improved dramatically since the data set was first created, and this is reflected in increasing reliance on MHMDS for the national governance of NHS mental health services. Nevertheless, significant data quality issues have been a constant feature, and pose challenges to its use. It became clear in the course of our research that, while the 2010/11 MHMDS was of sufficient completeness to support the cross-sectional examination of variance in compulsory admission rates, the data from previous years were much less reliable and of far inferior quality.

The 2010/11 MHMDS was designed to support the legal requirement of the NHS to administer and apply the 2007 MHA. 55 Providers make quarterly data returns which are then assembled to create an annual database, covering a reporting period from 1 April 2010 to 31 March 2011. During this 2010/11 reporting period, a total of 72 providers of mental health services submitted data to the MHMDS: three were independent sector providers, 10 were also NHS commissioners and 59 were (solely) NHS mental health providers.

The MHMDS is a large and complex data set. In 2010/11, a total of 1,287,730 individuals had contact with specialist mental health services in England. Each record in the MHMDS represents the complete, continuous spell of care for an individual, from initial referral to final discharge (referred to as a mental health-care spell – or care spell for short). This could cover any type of treatment from a short community or outpatient episode to an extended care spell over many years and covering all aspects of mental health services. 59 The care spell is assembled from separate records of care episodes, patient activity and contacts with mental health professionals. The vast majority of patients (89%) have only one care spell and therefore one record in the 2010/11 MHMDS; the remaining patients have more than one care spell and more than one record. Although multiple records are likely to arise for a variety of reasons (e.g. transfer of care from one mental health provider to another), the data set does not include this level of detail. This means that there are more records in the 2010/11 MHMDS (n = 1,443,858) than there are patients (n = 1,287,730). For the purposes of this study, each patient was counted once only using specific selection rules, which are outlined in detail later in this report. This is common practice in national analyses of the same data by the HSCIC. 54

Defining compulsory admission

The main study outcome is compulsory admission, and this terminology will be used throughout the report. Those who were admitted but not subject to the MHA (2007) are referred to as voluntary inpatients. We have not used the terms ‘informal’ and ‘formal’ (or ‘involuntary’) in relation to inpatient admissions, although these are sometimes used to denote voluntary and compulsory admissions, respectively.

Compulsory admission was defined as time spent in an inpatient mental illness bed subject to the MHA (2007) during the 2010/11 reporting period. We did not include patients who were detained only under sections of the MHA (2007) that apply only to conveyance to, and assessment in, a Place of Safety (S135 and S136), or for the purposes of assessment only [e.g. Sections 4 and 5(2)] where detention for longer than 72 hours does not occur. Likewise (because it does not relate to inpatient admission per se) we did not include sections of the Act relating only to guardianship or the supervision of community treatment. Patients who were detained under these short-term assessment sections and subsequently admitted under Sections 2 or 3 were included as cases of compulsory admission; those discharged from Sections 135 or 5(2), or admitted to hospital voluntarily, were not. The data did not allow us to distinguish between patients who were detained under the MHA (2007) at the time of admission and those admitted voluntarily and subsequently detained.

No single variable in the MHMDS data set described the study outcome. It was therefore necessary to derive this outcome using a number of other variables, starting with those concerned with inpatient admission.

The MHMDS contains several treatment and care variables which indicate if the patient was admitted to hospital during the reporting period, including number of days spent as an inpatient, number of admissions to hospital and number of discharges from hospital. There are also a number of variables relating to the nature of the admission to hospital (e.g. whether or not the patient was subject to the MHA), some of which apply to the reporting period (a year) and others to the entire course of the care spell (no definitive time period). These variables will be described in detail in the next sections.

Treatment in the community

Most patients in the MHMDS received all of their mental health care in the community, and were not admitted to hospital (either voluntarily or compulsorily). MHMDS collects information relating to what type of treatment the patient received in the community, for example the number of contacts a patient had with a community psychiatric nurse or the number of times a patient attended outpatient appointments. Combining these data with information from the number of days spent in hospital (bed-days) and number of admissions/discharges enabled us to identify patients who received only community care during the study year, with a high degree of accuracy.

Admission to hospital

A number of variables in the MHMDS provide information on admissions to hospital. The variable ‘bed-days’ is the number of days a patient spent as a mental health inpatient, including any days spent in a secure mental health unit. The number of admissions and discharges occurring during the reporting period are also recorded. However, the HSCIC discovered that there were inconsistencies and missing data, such that information on bed-days and on admissions/discharges does not concur for every patient. This is noted in the MHMDS Data Quality and Methodology Manual:

A very small number of trusts are failing to record consistent information about Inpatient spells in the Hospital Provider Spell (IPEP) and Ward Stay within Hospital Provider Spell (WARDSTAYS) tables in the input data and this affects the accuracy of the trust level results for these organisations . . . discrepancies in the source data result in discrepancies between the reported admissions, occupied bed days and patients categorised as ’admitted’. It also means that some trusts that certainly have beds have no patients in the ‘admitted’ category.

p. 959

However, IPEP and WARDSTAYS variables are generated by the HSCIC data assembler when the provider trust makes a submission but are not available to external users of the MHMDS. We were therefore unable to conduct any missing data analysis on these variables and had to conduct our own exploratory analyses to ascertain if a patient spent time as an inpatient subject to the MHA during the 2010/11 reporting period. We were not able to distinguish the type of mental health service providing inpatient treatment, and therefore these data will include people receiving specialist treatment in old age, learning disability (LD) and forensic mental health services.

Information from all relevant variables was used to optimise certainty about whether or not a patient spent time as an inpatient during the reporting period. Patients were therefore classified as follows:

• ‘admitted’ if there was any evidence that the individual spent time in hospital as an inpatient during 2010/11

• ‘only non-admitted care’ if there was evidence that the individual received treatment in the community but there is no evidence that the individual spent time as an inpatient during 2010/11

• ‘no care’ if there was no information about care provided during 2010/11.

A small proportion of patients (6.7%) had no care information in their 2010/11 record. This group may include patients recently referred to mental health services but not yet seen. It is more likely, however, that data-uploading errors were responsible. We assumed that these patients received some care in the community but no hospital admissions.

Table 1 provides a breakdown of the number of patients in each of the defined care status categories for the 2010/11 reporting period.

These estimates match those reported in the annual MHMDS Bulletin,59 providing evidence of the validity of our methodology.

In Table 1, the ‘admitted’ category includes all patients who were admitted to hospital during 2010/11, regardless of whether they were subject to the MHA or not. This ‘admitted’ category required further classification according to legal status (voluntary versus compulsory admission) using additional information relating to the patient’s MHA legal status classification.

Specifying the reference group

The reference group for this study comprised all patients receiving any type of care other than compulsory admission to hospital under the MHA (including informal admission and community care only). Patients were only counted once; for patients with more than one care spell during the 2010/11 reporting period, the care spell with the most restrictive level of care was then selected.

Legal Status Classification

The MHMDS contains a number of variables which provide information on the patient’s legal status under the MHA (2007) during 2010/11. Legal Status Classification (LSC) specifies whether or not, on the last day of the reporting period (i.e. 31 March 2011), the patient was subject to any section of the MHA, including sections which allow for assessment or the removal of a person to a place of safety. If a patient was not formally detained, the LSC provides information on whether or not the patient was being treated in hospital on a voluntary basis (i.e. informally admitted). If the patient was not an inpatient on a voluntary or compulsory basis on the last day of the reporting period, LSC is not applicable.

Legal Status Restrictiveness

The original stated intention of the MHMDS was to collect information on the most restrictive legal status (referred to as LSR) during the reporting period. 55 In a major data incident, the HSCIC discovered that data-uploading instructions contained an error that resulted in all provider trusts recoding the most restrictive legal status at any point during the care spell rather than during the reporting period. 59 This meant that the LSR code could refer to any point during the care spell (which has no definitive time period), rather than in the annual period of interest (1 April 2010 to 31 March 2011). This posed a major and unexpected challenge to the study.

The HSCIC estimated that ‘around 11% (4,650) of patients categorised as being formally detained (in 2010/11) probably experienced detention in an early reporting period’ (p. 13). 59 This estimate was obtained using linked data (at patient level) from the 2009/10 and 2010/11 MHMDS. Although there is a detailed description (and syntax) provided in the MHMDS Data Quality and Methodology Manual,59 key variables that are necessary to replicate this estimation are not routinely provided to general external users of the MHMDS. Our request for access to these data was declined. Of the 55,852 patients who had a LSR of Section 2 or Section 3 in their current care spell, 19,631 (35%) had no recorded inpatient bed-days during the study year.

We therefore had to establish a means of ascertaining which inpatients in 2010/11 had been compulsory inpatients during this time.

Identifying patients who were admitted compulsorily in 2010/11

Further to establishing if patients had received treatment as an inpatient in 2010/11 reporting period, we had to ascertain if this was subject to the MHA (2007).

Providers are required to specify start and end dates for each care spell. The mental health-care spell began during the 2010/11 reporting period for around one-half of patients (54%) in the 2010/11 MHMDS. This means that the LSR referred to the reporting period under investigation (i.e. 1 April 2010 to 31 March 2011).

We accessed data from the 2009/10 MHMDS to assist in ascertaining LSR in 2010/11 for patients whose care spells began before this reporting period. We used unique MHMDS identification numbers to link care spell records from 2009/10 and 2010/11. We used all variables relating to admission and detention, including LSR and LSC (as described above) discharge, bed-days, revocation of CTOs, Section 17 leave and the start and end dates for care spells to determine whether or not patients were admitted compulsorily in the 2010/11 reporting period. Two main complications were evident: (1) there was often inconsistency between the 2009/10 and 2010/11 MHMDS databases in terms of information relating to care spell start and end dates; and (2) missing data in one or both of the reporting periods made it difficult to get a clear picture of the patient’s admission history.

Information relating to the patient’s ‘care status’ during 2010/11 (see Table 1) was used together with the corresponding LSR and LSC codes to ascertain which inpatients spent time in hospital during 2010/11 compulsorily, i.e. subject to the MHA (2007). The main classification rules were as follows:

• Anyone with a qualifying 2009/10 LCS code (i.e. for a compulsory admission other than for an assessment section or guardianship) on 31 March 2010 was assumed to have still been compulsorily admitted on 1 April 2010 (i.e. the first day of the 2010/11 reporting period).

• Anyone who was an inpatient during 2010/11 and had a qualifying LSR code in the 2010/11 MHMDS was coded as having had a compulsory admission during 2010/11.

• Anyone who was an inpatient during 2010/11 and had a LSR code of a compulsory admission to hospital in the 2010/11 MHMDS, but only for an assessment section or guardianship, was coded as a voluntary admission in 2010/11.

• Anyone who was an inpatient during 2010/11 but who was not subject to the MHA on the last day of the 2009/10 or 2010/11, and whose LSR for 2010/11 was the same as the LSR for 2009/10, was coded as a voluntary admission in 2010/11. This decision was based on the assumption that the 2010/11 LSR code relates to a time before the 2010/11 reporting period.

This exercise was critically important for the main aims of this study but it also served as a quality assurance task for the project. As will be discussed in greater detail in Chapter 3, Results and discussion, the number of patients deemed to be compulsorily admitted to hospital during the 2010/11 reporting period (n = 44,442) was consistent with the HSCIC estimates54 for compulsory admissions during the same reporting period (n = 43,941).

Patient

Each patient record (n = 1,287,730) in the MHMDS contains information about their contact with mental health services during the care spell. This clinical record also includes fields for diagnosis and Health of the Nation Outcome Scale (HoNOS), a global clinician-rated measure of clinical status. The MHMDS record includes fields for a number of demographic variables including patient age, sex, ethnicity, marital status, housing and employment status.

Lower layer super output area

Lower layer super output areas (LSOAs) are a set of stable geographical areas developed to facilitate estimation and dissemination of neighbourhood statistics and are also used for the reporting of Census data. On average, LSOAs typically include 672 households and 1614 residents. In 2001, there were 34,378 LSOAs across England (n = 32,482) and Wales (n = 1896) (www.ons.gov.uk/ons/guide-method/geography/beginner-s-guide/census/super-output-areas--soas-/index.html). The 2010/11 MHMDS used the 2001 LSOA boundaries. LSOA data were missing for 12,557 patients (1.0% of the sample). Almost all (98.1%) of the LSOAs in England were represented in the study sample. A small number of patients (n = 1014) (0.08% of the sample) received treatment from an English provider trust but had a LSOA in Wales; these patients were retained in the analysis. The mean number of patients per LSOA (i.e. in contact with an English mental health-care provider) was much higher (39.2) in England than in Wales (1.8).

General practice

A total of 9492 of 10,203 GP practices in England in 2011 (93.0%) were represented in the data set.

Primary care trust (commissioner)

In 2010/11, there were 152 PCTs in England; 146 (95.4%) of these were represented in the MHMDS. At the time the data were collected, PCTs were responsible for commissioning health care, including specialist mental health services but not forensic services. Several variables were present in the MHMDS regarding PCTs: Code of Commissioner, Code of PCT of GP practice and Code of PCT of residence were available and in agreement for ∼1.1 million records (86% of patients). Code of GP practice, an indirect PCT identifier, was also available for these patients. The Code of Commissioner was therefore used to specify PCT for these patients. Where one or more of these variables were missing or not in agreement with the others, we prioritised Code of Commissioner and Code of PCT of GP practice over Code of PCT of residence, since we were primarily interested in the effects of investment in mental health services and judged that these were more likely to indicate the organisations that commissioned care. Data on all PCT variables (including Code of GP practice) were missing for just 3128 patients (0.2% of the sample). No other information in the MHMDS could be used to identify PCTs for these patients and therefore these patient care spells were excluded from the analysis when PCT was included in multilevel models.

Provider trust

The Code of provider trust variable was present for all patients, since it is automatically derived at the time of annual data submission. This meant that there was 100% certainty about where the patient received their mental health care during the care spell. Patients receive mental health care from providers, mainly NHS trusts but also a small number of independent sector providers. In total, there are 72 mental health service providers (both NHS and independent sector) in the 2010/11 MHMDS. Independent sector providers are units registered with the Healthcare Commission to accept patients detained under the MHA 198360 (as amended by the MHA 2007), as specified under Section 2(3)(b) of the Care Standards Act 2000. 61 These independent hospitals that are registered to receive detained patients are considered to be hospitals for the purposes of the MHA (2007). In contrast to NHS provider trusts, these independent sector providers were coded by the service provider company in the MHMDS and not by individual units or according to geographical location.

Strategic Health Authority

The SHA was derived from Code of provider trust, and was likewise present for all patients. There were 10 SHAs in England in 2010/11. SHAs were responsible for improving health services in their local area and ensuring these were of a high quality and performing well.

Individual level variables

The MHMDS contained reasonably complete data on a limited number of (individual-level) patient characteristics, namely age, sex and ethnicity. Levels of missing data across these variables varied considerably, however: age was missing for 84 patients (< 0.01%); sex was missing for 454 patients (0.03%); and ethnicity was missing for 119,392 patients (9.6%).

Several other individual-level variables were not included in our analyses because there were high levels of missing data. For example, marital status was missing for almost 15% of patients and psychiatric diagnosis was missing for around 81% of patients. 59 Accommodation status was missing for 920,288 (64%) and employment status was missing for 1,099,311 (75%) of all care spells. Recent clinical status (as measured by the HoNOS) was recorded for only 25% of all patient care spells. Other information of potential salience to our research questions included medication history (e.g. use of antipsychotics), drug and alcohol use, family and household composition (including if a patient lives alone, or has children) and social support were not recorded in the 2010/11 MHMDS. Although we were able to ascertain reliably whether or not a patient had been admitted during the reporting period, we did not have information about the number of admissions per patient. The same consideration applied to ascertaining the number of compulsory admissions during the reporting period, or if a person was admitted to hospital voluntarily and subsequently detained under the MHA.

The following section describes the external data sources that were linked to the MHMDS at different spatial levels to facilitate the study of factors that influence variation in compulsory admission across England.

Linkage at lower layer super output area level

Linkage at primary care trust level

Linkage at provider trust level

Multilevel models

As outlined earlier, there are six discrete and identifiable spatial levels in the MHMDS at which variance in rates of involuntary admissions might occur. Multilevel modelling is a widely used method for analysing nested (or clustered) data because it mirrors the structure of the data to be analysed and allows explanation to be correctly assigned to the various ‘levels’ within the data set (e.g. individual, LSOA, provider trust). The structure of the MHMDS is complex, and the various spatial levels do not nest or cluster neatly within each other. For example, approximately two-thirds of LSOAs were not nested completely within (i.e. served exclusively by) a single provider trust. This is known as a cross-classified data structure and is illustrated schematically in Figure 1. This illustrates a three-level model with patients (level 1) nested within LSOA (level 2) and a non-nesting hierarchy between LSOA and provider trust (level 3). ‘Non-nesting’ is because LSOAs straddled provider trust boundaries and therefore more than one provider trust delivered care to people living in some LSOAs. These level 2 and 3 units can be conceptualised as ‘classifications’: combinations of particular LSOAs and provider trusts. This classification structure is recorded for each individual within the data set. Ignoring cross-classification complexity would result in incorrect estimates of area-level variance and biased estimations of statistical validity. We refer to individuals, and the geographical or health-care areas with which they are associated, as ‘levels’ but it is important to acknowledge that variation across these levels takes into account this non-nested, cross-classified data structure. A fuller description of cross-classified models can be found elsewhere. 70

FIGURE 1.

An example of a cross-classified data structure.

The specification of multilevel models (MLMs) depends on the research question or hypothesis being tested; details of these models are provided in Chapter 3, Results.

Despite the pre-eminence of MLwiN (MLwiN 2.27, Centre for Multilevel Modelling, Bristol, UK) as a multilevel modelling program, it has finite limitations in computational power (when the sample size is very large) and model stability (when estimating more than three levels). Given the size of the database (∼1.2 million patients), we were aware that estimating cross-classified models with more than three spatial levels would be difficult because of these program limitations.

Null models

The most basic MLM is the null model, which estimates the total variance in compulsory admission in the study sample, without the inclusion of any explanatory or potentially confounding variables. This variance is equal to the sum of the variances at the different spatial levels. Further models can then be developed which attempt to reduce and explain this initial geographical variability in the rates of involuntary admission by including explanatory variables relating to the characteristics of individuals or the higher-level geographies (e.g. LSOA, provider trust).

We ran all permutations of four-level models (using combinations of individuals, LSOAs, GP practices, provider trusts and PCTs) to identify consistent cross-model patterns and compare goodness of fit. However, where model stability allowed, we included a fifth level (i.e. SHA) to ensure that we had derived the most parsimonious model for the data set. This stage provided an estimate of the variance in compulsory admission to hospital at these various different spatial levels, before characteristics of the given levels were taken into account. It also provided a clear rationale for building more complicated four-level models with explanatory variables, ensuring that only spatial levels with significant proportions of variance were considered for inclusion.

Markov chain Monte Carlo (MCMC) estimation methods were used to estimate the variance of our outcome variable across these geographic classifications (i.e. LSOA, provider trust, PCT, etc.). Analyses of the cross-classified multilevel data structure were undertaken using a logit link function to model the binary response of compulsory admission (denoted by a ‘1’) or not (denoted by a ‘0’). 71 Furthermore, we assumed that our binary observations were derived from an underlying continuous latent variable so that (in this case) compulsory admission to hospital under the MHA during 2010/11 (denoted by ‘1’) is assumed to occur after a certain threshold has been exceeded in the severity of illness. Level 1 variance on this latent variable was the standardised logistic variance of π²/3 = 3.29. 71

The number of iterations used for each model depended on the number and arrangement of levels in the null model, but this ranged from 500 to 100,000 iterations and was checked using the Raftery–Lewis diagnostic. 72 The mean variance estimate was recorded for each higher level along with the 95% credible interval. The total variation in each model was calculated by adding up the estimated variance at each level, and the relative percentages occurring at the various spatial levels were calculated. A commonly used statistic to compare the fit of alternative MLMs is the Bayesian deviance information criterion (DIC) goodness-of-fit statistic. 73 The DIC compares alternative models based on a trade-off between the fit of the data to the model and the corresponding complexity of the model. 74 It is considered analogous to the Bayesian information criterion (BIC)75 or the Akaike information criterion (AIC)76 fit statistics for models estimated using MCMC. Models with lower DIC values are preferred to those with larger values and, as a general rule of thumb, a difference of 10 or more is considered to be substantial. The main limitation is that the DIC value is comparable only across models with exactly the same observed data (nesting of models, however, is not required). 77 DIC statistics were therefore estimated using only cases that had complete data for all of the model variables, to ensure identical samples.

Modelling and mapping residuals

After selecting the optimal four-level null model (research aim 1), we controlled for basic demographic characteristics of the patient population served by the provider trust (i.e. individual age and sex) and examined the provider trust-level residuals produced by the model. A residual at provider trust level is the difference between the actual (observed) risk of compulsory admission and the risk predicted by the (age- and sex-adjusted) model. Positive residuals indicate that the model is underpredicting the risk of admission and negative residuals indicate that the model is overpredicting risk (compared with the actual risk). Residual results were reported graphically in two ways: using ‘caterpillar’ plots [showing the raw residual for each trust in rank order, as a point estimate with 95% confidence interval (CI) error bars] and by mapping the normal score of the residual (i.e. the normal deviates of each residual). Upper and lower CIs were obtained by multiplying the standard error of each residual by 1.96 and adding and subtracting this value from each estimate, respectively. Normal scores were used to allow comparison of residual size across different models. To produce a digital map of the residual normal scores, we derived the postcode of the provider trust headquarters and, using standard lookup tables available from The UK Data Service,62 we derived the associated National Grid co-ordinate for the postcode centroid. The normalised residual value and its associated co-ordinate were then ‘attached’ to a digital boundary of England and Wales (available from Edina Digimap Ordnance Survey Service) and mapped using ArcMap version 10.1 (Redlands, CA, USA). 63 This exercise was repeated for the fully adjusted, final model (research aim 2) as described next.

This was essentially an exploratory diagnostic tool designed to help us identify ‘missing’ explanatory variables, i.e. hypothetical variables whose existence might be posited to explain observed variance in the study outcome. If negative residuals are clustered together and positive residuals clustered together (i.e. model residuals are positively spatially auto-correlated), this would suggest a missing spatially clustered explanatory variable. If residuals are not spatially auto-correlated, then we would expect to find a random arrangement of positive residuals and negative residuals and there will be no clustering. Such a finding would direct us away from a search for a ‘missing’ explanatory spatial variable.

Modelling explanatory variables

We extended null MLM to include explanatory variables at different spatial levels. Our aim was to estimate the effect of doing so on the variance in compulsory admission rates identified in our null models (including the residuals at provider trust level).

Extending the MLM in this way increased the complexity of the model and placed additional demands on the computing capacity of the statistical software MLwiN. Given the sample size and the potential number of linked variables (from external data sources), multivariate models with more than three (cross-classified) levels proved unstable. For this reason, we had to choose the most important spatial levels for inclusion in these more complex models, and we did so this on the basis of our null model results.

Adding explanatory variables across one or more spatial levels allowed us to quantify the proportion of variance explained by these risk factors for compulsory admission. This involved modelling the predicted value of the (assumed) underlying continuous latent variable by using an equation with both fixed [intercept plus regression coefficient(s) for explanatory variable(s)] and random parts. 78 The proportion of variance explained at, for example, level 1 was estimated as the proportion of total variance attributable to the fixed part of a model that included individual-level explanatory variables, divided by the total variance in the underlying latent variable. The total variance is equal to the sum of the variances of the fixed part of the model and the (unexplained) variances at each of the higher levels. Variance terms were derived from the MLwiN output.

Spatial clustering of primary care trusts and provider trusts

Reasons for the differences in variance estimates when PCT and provider trust levels were both included were unclear. To investigate this further, we explored the distribution of PCTs and provider trusts. Although most provider trusts were associated with three or four local PCTs (commissioners) at most, some appeared to be treating (a few) patients from many different PCTs around the country. This was probably because of specialised tertiary care services, as well as movement of patients around the country for personal reasons.

In one-third of provider trusts, over 80% patients originated from a single PCT. Indeed, examination of the data matrix mapping patient counts for all provider trusts against all PCTs revealed that about one-quarter of the cells had counts of 0 or 1. This could account for the decrease of variance at the individual level in models 3 and 4, as there may be possible conflation between individual and higher-level variances when the cell counts contain just one observation – a problem that is also exacerbated by the cross-classification occurring in the data structure. 81,82 We therefore considered the estimates of higher-level variance in models 1 and 2 to be more reliable than those of models 3 and 4.

Comparing model fit using deviance information criterion statistic

As noted in Chapter 2, Methods, DIC statistics can be compared only when the sample used in each model is exactly the same. The four null models were therefore re-estimated using complete cases only (i.e. by including only patient records that had complete data across LSOA, GP practice, PCT and provider trust; n = 1,149,541). This reduced the sample size but ensured that the DIC statistic could be used to compare model fit (see Table 3). Results in Table 4 revealed that models 2 and 3 were superior to models 1 and 4, with model 3 having a slightly better fit than model 2.

Given the greater reliability of estimates of higher level variances in models 1 and 2 (i.e. those models that contained either PCT or provider trust but not both together), we selected model 2 as the preferred four-level model. This choice was strengthened by the a priori importance of provider trusts to the study of compulsory admission rates. Provider trusts deliver mental health care and are ultimately the locus at which decisions about whether or not an individual patient should be admitted to hospital compulsorily are made. In the same vein, future interventions to reduce compulsory admission rates will almost certainly have to involve provider trusts.

Estimating the effect of including Strategic Health Authority as a fifth level in the null model

We re-estimated the four-level model of choice (model 2) by including SHA as a fifth spatial level (model 2a) to test whether or not this improved model fit using data from the same number of patients. This meant that it was appropriate to use the DIC statistic to compare the fit of the models and to statistically test whether or not including SHA as a spatial level enhances the model fit. The DIC statistic for model 2 (336,537.7) was smaller than that of model 2a (336,538.2). The proportion of unexplained variance at SHA level was not statistically significant (2.4%, 95% CI 0.0% to 5.9%; χ² = 2.04; p = 0.15). SHA was therefore not considered for inclusion in further models.

Mental Health Foundation

The Mental Health Foundation (MHF) (www.mentalhealth.org.uk) is the UK’s leading mental health research, policy and service improvement charity. Service user involvement was coordinated by DC-K, Head of Empowerment and Social Inclusion at the MHF (and co-applicant on this project). The main aim of this partnership was to ensure meaningful patient and public involvement (PPI) such that the service user perspective shaped the research questions, influenced analysis and interpretation, and assisted in disseminating the findings.

Background

This aspect of the project comprised three meetings with people with direct personal experience of compulsory admission, or of caring for someone who had been admitted compulsorily. These meetings were held at key points during the research process and focused on relevant issues at each stage of the project.

Aims

The aim of the first meeting was to seek service user input into the analysis by encouraging service users to suggest reasons for different rates of compulsory admission. The intention was to use the content of these discussions to guide the researchers, particularly in selecting explanatory variables through data linkage.

The second meeting was used to reflect on interim findings and help guide their interpretation by exploring experiences of compulsory admission, including those that occur before, during and after the inpatient spell, upon discharge. At the third meeting, people examined how the study findings might be used to reduce rates and variability of compulsory admission, and improve service user experiences in future.

Methods

Results

Underlying trends

The participants concurred with the report of historic trends (1988–2008), and felt certain that rates of detention were likely to have increased in the meantime; one participant noted that these rates had increased by 6% in the last year alone. It was also noted, and agreed generally, that the majority of inpatients across the country are now detained patients.

Professionals’ views about reasons for increased rates of detention

This was one of the major areas of discussion, and produced a comprehensive view of possible factors to be investigated.

Useful and desirable outcomes of the research

Following this lengthy discussion, the participants were aware of the main aims and objectives of the ENSCA project. They were then asked to give their opinions about what specific evidence they would like to see being derived from the project; in other words, how they felt the evidence from this research study might be useful for professionals. The following points were raised:

• What is the association between LOS in hospital and the compulsory admission rate?

• What implication does the increasing compulsory admission rate have on the availability of resources in other areas of care (e.g. services to help prevent relapse)?

• What evidence is there to support the ‘reinvention’ of inpatient services and ‘upskilling’ of ward staff in an attempt to reduce the compulsory admission rate?

• What impact does the commissioning of CRHT services have on the rate of compulsory admission to hospital?

As outlined in the MHF PPI section, it was not be feasible within the current project (given the limitations of the MHMDS and external data sources) to address all of these proposed avenues for analyses raised in this consultation event. The content of the discussion, however, did inform the selection of explanatory variables as described in Chapter 2, Methods. These recommendations were also considered as opportunities for future research studies, as covered in Chapter 5, Discussion and conclusions.

Patterns of variance in the null model

The choice of an optimal four-level, cross-classified null model was complicated, and involved careful comparison of competing models (including model fit statistics) and further examination of the data structure in respect of PCTs and provider trusts. We began with the a priori assumptions that individual (patient) and provider trust should remain in all models, since the former is the level at which outcome is classified (and where most variance occurred) and the latter represents the locus at which decisions about admission and detention are made, and therefore represents an important target and/or location for future interventions.

The relatively small number of PCTs (commissioners) per provider trust meant that higher-level variances were likely to be unreliable in models that included both PCT and provider trust. We therefore selected the four-level cross-classified null model comprising individuals, LSOA (based on Census area geography), GP practice and provider trust. Although around 84% of the (unexplained) variance in compulsory admission in this model reflected differences between individuals, substantial and statistically significant variance (∼7% in both instances) was found between areas (LSOA) and between provider trusts. The variance at GP level was smaller (2.7%) but statistically significant. The model fit was not improved by including SHA, and there was no evidence of statistically significant variance at this (higher) level.

Examining and mapping null model residuals

The limitations of even the best-fitting null model were evidenced by examining the provider trust-level residuals. These differed from the modelled average to a statistically significant degree for 45 out of 64 (70%) provider trusts included in the study sample even after adjusting for patients’ age and sex.

We mapped these residuals to explore their geographical distribution, and particularly the spatial distribution of positive and negative values. This was an exploratory diagnostic tool designed to help identify potentially ‘missing’ explanatory variables. Missing variables would be considered more likely if negative residuals were clustered together and positive residuals clustered (i.e. if there were evidence of spatial auto-correlation among residuals). Although there are no formal statistical tests with which to assess this, inspection of the map based on null model residuals did not suggest clustering of this type, with the possible exception of London, where (notwithstanding the limitations of the map itself) there appeared to be a cluster of large positive model residuals.

The next stage (research aim 2) involved attempting to explain this variance by introducing variables that characterised the different spatial levels within the data set. To maintain model stability while accommodating explanatory variables, we were forced to reduce the model from four to three spatial levels, and retained individual, LSOA and provider trust. The distribution of variance in this three-level null model was broadly in keeping with that of the null four-level model described above. In the fully adjusted model we estimated that individual-level risk factors explained 8.0% of the risk of being compulsorily admitted, while LSOA- and provider trust-level variables each explained 1.1% of the total variance in the study outcome.

On examining the fully adjusted model residuals we found that the number of provider trusts whose observed compulsory admission rate differed from the model average to a statistically significant extent had fallen from 45, in the null model, to 20. Mapping of these residuals did not reveal any striking evidence of spatial auto-correlation.

In our adjusted models, we found significantly higher rates of compulsory admission for men, and for patients of black, Asian and mixed ethnicity (compared with white patients). The finding that patients of black ethnicity were almost three times more likely to be admitted compulsorily than white patients is consistent with evidence from elsewhere. 28–30 The rate of compulsory admission was highest in patients aged 18–35 years, and fell with age thereafter. The lowest rate was observed in those aged under 18 years.

We found statistically significant associations with area-level deprivation (which appeared to show a dose–response effect) and ethnic density, but not with population density. The risk of compulsory admission was inversely associated with the proportion of white British LSOA residents. This finding highlights, and raises further questions about, ethnic inequalities in the experience of mental health care.

We found very little evidence of significant associations between compulsory admission and provider trust-level factors. Only one such covariate was a statistically significant predictor of compulsory admission: patients receiving care in provider trusts with community mental health services rated as average (‘same as other trusts’) had a significantly higher risk of being admitted compulsorily than in those rated as poor (‘worse than other trusts’) on the survey of patient experience.

One important finding concerns bed occupancy and rates of compulsory admission. Before adjusting for other covariates (and potential confounders), we found a statistically significant association between (higher) average annual trust-level mental illness bed occupancy and compulsory admission. However, this association no longer reached statistical significance after adjusting for the other covariates in the model. Although there was a trend towards lower rates of compulsory admission in trusts outside than in London, this association did not reach statistical significance in the fully adjusted model.

For the reasons described we did not include PCT (commissioner) in our null or adjusted models. Instead we carried out separate analyses in which we examined associations between PCT investment in mental health services and rates of compulsory admission. Our results indicated that inclusion of these variables explained almost no additional variance in the rate of compulsory admission after adjusting for individual- and LSOA-level variables. No statistically significant association was found between total investment in mental health services and compulsory admission in adjusted models. A small but statistically significant association was, however, found for investment in community mental health teams, but not with investment in assertive outreach or crisis resolution teams.

Research questions

1. How variable are rates of involuntary admission across different places, namely Strategic Health Authorities (SHAs), NHS Trusts and independent sector hospitals, and Primary Care Trusts (PCTs) in England?

2. To what extent is this variance explained by local factors, such as small area socio-economic deprivation, urbanicity, ethnic density, PCT investment in mental health services, and by characteristics of local mental health services including the availability and occupancy of NHS mental illness beds?

3. To what extent is variance in rates of involuntary admission explained by local factors, such as small area socio-economic deprivation, urbanicity, ethnic density, PCT investment in mental health services, and by features of local mental health services and the experiences of those who use them?

4. To what extent is variance in rates of involuntary admission explained by the characteristics of people using services, including socio-economic status and ethnicity?

5. Is there evidence of geographical variation in the use of Community Treatment Orders (CTOs), and if so, how is this associated with bed numbers and occupancy, length of stay, area-level factors (like deprivation and PCT investment), experience and quality of local services, and rates of involuntary admission?

6. Using the model of best fit based on local services and settings and including historical data on trends in bed numbers, can we predict the future need for involuntary treatment with sufficient accuracy to support commissioners and NHS managers in determining optimal local bed numbers?

How the proposed research will add to the body of knowledge with reference to current NHS policy and practice

See Final Study Protocol.

Importance of the proposed research

See Final Study Protocol.

Design

Secondary analysis of routinely collected data. Data will be organised to permit analysis as cross-sectional survey, cohort and (nested) case control studies.

Data sources

Most of the data for the proposed study will be obtained from the NHS Information Centre (http://www.ic.nhs.uk), which is responsible for collecting and collating national health and social care data in the UK. The Information Centre also provides support and advice to end users through its Secondary Uses Service (SUS) and Medical Research Information Service (http://www.connectingforhealth.nhs.uk/systemsandservices/sus and http://www.ic.nhs.uk/services/medical-research-information-service), in partnership with NHS Connecting for Health. The NHS Information Centre ensures appropriate governance checks for the use of these data, to protect the confidentiality of those on whom data is collected, through the National information Governance Board for Health and Social Care (NIGB). The NHS Information Centre also collates geographical data, including deprivation and other neighbourhood statistics, and provides support in ensuring appropriate linkage to other data sets (http://www.ic.nhs.uk/services/population-geography-information/geographical-information).

The data archive at the NHS Information Centre includes historic information on geographic boundaries and organisation changes in the NHS, which will enable us to track this over time and to make adjustments for changes over time in the configuration of boundaries relating to the regulation, commissioning and delivery of NHS services. One example of this is the PCT mapping tool (http://www.ic.nhs.uk/statistics-and-data-collections/population-and-geography/pct-mapping-tool), which enables older service use data to be correctly mapped onto current PCT boundaries (http://www.statistics.gov.uk/geography/england_health.asp).

The significant investment on the part of the NHS in software resources and staffing to support secondary users through the NHS Information Centre represents a major strength of the proposed research. Although the proposed data linkages will be methodologically challenging, we are confident that the study aims are achievable.

The data sets which we propose to access and link are set out below in Table 9, followed by more detailed descriptions about each of these. Outcomes will be ascertained using the first two of these data sets (Mental Health Minimum Dataset and Hospital Episode Statistics); the remainder will be used to provide exposure (risk factor) information via data linkage. The data on In-Patients Formally Detained in Hospital will be used to validate data on involuntary admissions in MHMDS and HES, and to generate study weights if necessary.

Participants

See Final Study Protocol.

Setting

See Final Study Protocol.

Primary outcome

See Final Study Protocol.

Strategy of analysis

1. Cross-sectional multilevel analysis using patient-level data at the lowest level. MHMDS will be used for this (rather than Hospital Episode Statistics) since the data comprise information about all those using mental health services rather than just those admitted to hospital. The reference group will therefore be those receiving mental health care but not admitted to hospital involuntarily. This will have the advantage of adjusting for the overall prevalence of mental illness in particular areas, and for the accessibility of mental health services generally (36). This is not the case for HES data, where the reference population would be those using hospital services (in-patients admitted voluntarily and those seen in hospital out-patient clinics) but not including those accessing community mental health services. We will explore the possibility of pooling two years’ worth of MHMDS data to optimise the sample size.

Data from other sources will be linked at the appropriate level, e.g. by PCT and/or Trust. This will enable us (for example) to explore the effects of age, sex, ethnicity (at the individual level), bed numbers (and types of available beds), overall bed occupancy rates, average length of in-patient stay (at Trust level), investment in services (at PCT level) and socio-economic deprivation, urbanicity and ethnic composition (at area level) on the likelihood that a user of mental health services will be admitted to hospital involuntarily. As well as assessing the main effects of these risk factors, we will explore variance at different spatial scales and test for cross-level interactions, ie the extent to which associations vary between places.

2. Cross-sectional multi-level analysis using aggregated involuntary admissions by PCT. The data set of choice for this analysis will be Hospital Episode Statistics (HES) for 2009/10, validated against data for the same year from In-Patients Formally Detained in Hospital (IPFDH) (and weighted if necessary). IPFDH data are obtained from providers in statutory returns. HES is preferred to IPFDH for these analyses because data can be aggregated at PCT (and even Census Output Area) level rather than provider (Trust); and HES is preferred to MHMDS because of its more detailed recording of involuntary in-patient spells, and because it contains information about numbers of in-patient episodes and length of stay. Unlike MHMDS, HES distinguishes between admissions under the Mental Health Act (from the outset) and those where admission was initially voluntary and where the person was later detained. Both HES and IPFDH include data on admissions to independent sector facilities that are funded by the NHS; these data are not yet available in MHMDS. We will explore variance at different spatial scales, assess the extent to which this is accounted for by a range of explanatory variables and test for cross-level interactions, i.e. the extent to which associations vary between places.

3. Longitudinal analyses. The primary aim of this research is to understand the determinants of changing rates of involuntary admission. Therefore the preferred data set for our longitudinal analyses will Hospital Episode Statistics (HES), which are available from 1998 to the present. As for our cross-sectional analyses, we will undertake a validation exercise comparing HES with data from IFDH, and where necessary we will derive appropriate weights.

Longitudinal analysis of HES data will explore and assess the extent to which increases in rates of involuntary admission over time vary between PCTs and between different areas of England (including comparing urban versus other types of area), or whether these increases are similar across the country. We will also test hypotheses about changes in rates of involuntary admission and prior (time lagged) changes in bed numbers, bed occupancy, and/or in the experiences of, and investment in, mental health services. Data linkage and the number of years for which data are available will enable us to explore and model the effects of variables at different spatial levels, and over different intervals. These analyses will also allow us to explore the effects of changes to the Mental Health Act in 2007, and the introduction of Community Treatment Orders in 2008. We will also model the growth in involuntary admissions in both secure (forensic) units and the independent sector, to estimate their contribution to overall increases in the use of detention and to see if changes in these rates differ from those pertaining to adult mental illness beds in NHS general psychiatric settings.

4. Predictive modelling. Finally, we will use longitudinal models to predict future rates of involuntary admission based on multivariate, multilevel models. In particular, we will model the effects of reducing bed numbers, increasing bed occupancy and rising numbers of Community Treatment Orders to predict future changes in involuntary admissions. We will do so using multilevel synthetic estimation, a methodology developed by one of the applicants (37,38) and validated independently by the Department of Health (39).

Multilevel modelling

Datasets will be described using standard statistical methods including frequency tables and chi-squared tests. Further analysis, and testing of the study hypotheses, will be undertaken using multilevel modelling (MLM) (40). MLM enables variances at different spatial levels to be modelled simultaneously and accurately, and in a way that reflects the structure of the data (i.e. ensuring that standard errors are adjusted to reflect spatial clustering within the hierarchical data). We will estimate how much of the observed variation is explained by individual and/or ‘compositional’ factors (such as the age, gender and ethnic composition of the local population), and how much is due to area characteristics, otherwise known as ‘contextual’ factors such as area deprivation or service investment. Multilevel approaches will also allow us to test for independent cross-level interactions, ie the extent to which individual level associations with involuntary admission vary between types of places (i.e. in terms of service investment, quality, socio-economic deprivation etc.).

Multilevel models will be developed using MLwiN software (41). In each instance a null, random effects model will be derived that estimates variance at each level: for example, individuals nested in hospitals (and/or Trusts), with Trusts nested in PCTs and PCTs nested in SHAs (or regions). These null models will serve to indicate the degree to which variance in rates of involuntary admission vary within and between these different levels, before the introduction of any (potential) explanatory variables. Subsequent modelling will be used to explore the extent to which this variance is explained by factors that include individual characteristics (like age, sex, diagnosis and ethnicity, where these are available) and area-level exposures such as urbanicity, socio-economic deprivation and the size of minority ethnic populations. Each of these variables will be allocated (and introduced to the model) at the appropriate level. Likewise, modelling will explore associations between rates of involuntary admission and the availability (and occupancy) of different types of inpatient beds (at provider level). A major advantage of multi-level modelling is the opportunity to measure cross-level interactions (e.g. between bed numbers and PCT mental health spend, for instance).

Our primary outcome (involuntary admission) will be analysed in two ways: (i) as an individual-level outcome, coded as the number of involuntary admissions per year; and (ii) as the rate of involuntary admission by PCT (for example). The latter can be treated as a continuous outcome across PCTs and analysed using a normal iterative generalised least squares hierarchical regression. The former outcome is measured at an individual level, whereby the individual risk or likelihood of being detained over a given year is treated as a dichotomous variable (i.e. detained or not). This type of outcome will lend itself to multilevel logistic regression where analyses will be undertaken using a logit link function and assumed non-constant, between-individual variance based on a Bernoulli distribution (42). In such models, the constant term will be the logit (log_e of the odds) of a person in the base (reference) category experiencing an involuntary admission in a given year. The proportion of each involuntary admission will therefore be estimated from the constant term in the null model, which is equal to ln(p/1+p). Although unlikely, it is possible that some people may be detained more than once in a year and if this is found to be sufficiently common to model, then it is possible to develop multinomial multilevel models (i.e. modelling the risk of being detained once, twice, three times etc.).

Sample size

It is not possible to derive precise sample size estimates for our analyses, and nor are these needed in the design of our research since no new data are to be collected. The data sets at our disposal are extremely large, and are arguably the most comprehensive population-based sources of data on the use of mental health services anywhere in the world. This reflects both the nature of nationalised health care provision in the UK, and high levels of investment in the collection and collation of routine activity data.

Our analyses will cover all of England, and will include information from NHS and independent sector providers of in-patient mental health care. In 2009/10 this comprised 10 Strategic Health Authorities (signifying region), 151 Primary Care Trusts (as commissioners), 58 NHS Mental Health Trusts (including Care Trusts and Mental Health Foundation Trusts) and 211 registered independent mental hospitals. In 2009/10 more than 1.25m people accessed mental health services in England (2.7% of the population). (36) In the same year, 107,765 people were in-patients during the year and 42,479 were detained in hospital under the MHA (39.4% of all admissions). Of these, 2717 (4.8% of all detentions) were in independent mental hospitals and 760 (1.8% of all detentions) were under court or prison disposal orders.

Whilst sample size is unlikely to be a problem for the multilevel rates analyses, the individual logistic analyses may have relatively few observations nested within small areas. To overcome this problem, parameters will be estimated using second-order Taylor expansion with predictive quasi likelihood (PQL). This estimation procedure is considered superior to first or second order marginal quasi-likelihood (MQL) when observations within clusters, such as individuals within small areas are small (42). Markov chain Monte Carlo (MCMC) methods will further improve the accuracy of such estimates (43).

Justification of support requested

See Final Study Protocol.

Research questions

1. To what extent are national rates of involuntary admission applicable to lower level health administrative areas? In other words, how variable are rates of involuntary admission across different places, namely Strategic Health Authorities (SHAs), NHS Trusts and independent sector hospitals, and Primary Care Trusts (PCTs) in England?

2. What proportion of the total variance in rates of involuntary admissions occurs between SHAs, trusts/hospitals and PCTs, respectively?

3. To what extent is variance in rates of involuntary admission explained by local factors, such as small area socio-economic deprivation, urbanicity, ethnic density, PCT investment in mental health services, and by features of local mental health services and the experiences of those who use them?

4. To what extent is variance in rates of involuntary admission explained by the characteristics of people using services, including socio-economic status and ethnicity?

5. After taking into account characteristics of areas, services, commissioner (PCT) investment and individual service users, to what extent can the variance in involuntary admissions be attributed to the availability and occupancy of NHS mental illness beds?

Our original proposal included two additional research questions:

1. Is there evidence of geographical variation in the use of Community Treatment Orders (CTOs), and if so, how is this associated with bed numbers and occupancy, length of stay, area-level factors (like deprivation and PCT investment), experience and quality of local services, and rates of involuntary admission?

2. Using the model of best fit based on local services and settings and including historical data on trends in bed numbers, can we predict the future need for involuntary treatment with sufficient accuracy to support commissioners and NHS managers in determining optimal local bed numbers?

Given our knowledge and experience with the MHMDS at this stage in the project, it is not feasible to explore these two research questions due to data quality issues. The complexities of the MHMDS structure and issues with data quality will be outlined throughout this revised protocol and have already been provided in detail in our interim progress reports submitted in September 2012 and February 2013.

Increasing rates of involuntary psychiatric in-patient admission and falling bed numbers

Rates of involuntary admission to psychiatric in-patient beds have been rising in many Western European countries,1 including in England since the introduction of the Mental Health Act (MHA) in 1983. 3,4 The reasons for this remain unclear, particularly since this trend has continued despite the development of a range of community-based services such as Community Mental Health Teams, Assertive Outreach, Crisis Resolution Home Treatment and Early Intervention service. Despite being patient- and carer-centred, responsive to need, recovery-focused, timely and popular,22,47 none of the developments in community mental health services has been shown to reliably reduce involuntary admissions. 48–51

The upward trend in rates of detention appears to have accelerated since the introduction of the Mental Health Act (2007). Formal admissions alone (i.e. patients admitted to hospital under the Act as opposed to those detained after removal to a place of safety, or subsequent to informal admission) increased by 7.3% from 2008/9 to 2009/10, the largest increase in three years. 92 This increase was due solely to increased rates of formal admission to NHS facilities and concealed a decline in numbers of formal admissions to private sector units. 86 Formal admissions to NHS facilities under Section 2 of the Act (the most commonly used section, which allows for assessment rather than treatment) increased by almost 15% in 2009/10. By comparison, this number fell by 10% in the private sector over the same period.

The increasing rate of compulsory admission in England has coincided with an equally dramatic reduction in the number of mental illness beds. 20 Analysis for England shows that the number of NHS involuntary admissions increased by over 80% between 1988 and 2008, while the number of mental illness beds fell by around 60% during the same period (Keown et al. , personal communication based on submitted manuscript). 5 While closure of mental illness beds has been part of the deinstitutionalisation of care for the mentally ill for decades in high income countries,25 the United Kingdom has taken this further than its neighbours and most other developed economies. There are substantially fewer psychiatric in-patient beds in the UK (5.8 per 10,000 population in 2005) than the median for Europe (8.0 per 10,000 population) or high income countries as a whole (7.5 per 10,000 population). 26 These figures should be seen in the context of high overall levels of funding for mental health services in the UK, which currently account for 10% of the total health budget. 27 The United Kingdom has greater numbers of psychiatrists (11 per 100,000 population) and psychiatric nurses (104 per 100,000 population) than the medians for Europe (9.8 and 24.8, respectively) and high income countries as a whole (10.5 and 33.0, respectively).

People from black and minority ethnic groups experience poorer mental health services compared to the rest of population in the UK. 28–32,34,36,87 Disparities in rates of admission, including compulsory detention31,34 and more coercive pathways to acute psychiatric care, involving the police and criminal justice system disadvantage black service users in particular. 30,36,37 These inequalities have persisted despite major policy initiatives,38–40 additional investment, workforce developments and the general improvements in the coverage and quality of mental health services over the last decade. 30,39,41

Supervised Community Treatment (SCT) in the form of Community Treatment Orders (CTOs) was introduced as part of the new Mental Act in 2007. Since November 2008 it has been possible for patients detained under Section 3 of the Act to be discharged from hospital by means of a CTO, subject to the possibility of recall to hospital if necessary. This order requires the patient to adhere to specific conditions, most often including taking their medication. In the first full year of operation, 4103 such orders were made – significantly greater numbers than anticipated – while only 1789 were revoked, suggesting that some people are being kept on these orders for long periods of time. 86 SCT remains controversial, and its effect on and associations with bed numbers, occupancy, detention rates and therapeutic outcomes remain unknown. It has been suggested that SCT is both a means to ensure early discharge where beds are scarce (particularly where therapeutic to medication are restricted or unavailable), and a defensive and alienating response to managing risk. 93

The increase in the use of compulsory detention sits uneasily with most professionals, who have spent two decades attempting to improve engagement with care and to reduce the fear and stigma associated with mental health services. 22–24,31,45 It is also a source of concern to commissioners and service providers, given the high costs of in-patient care. The increasing rate of involuntary treatment represents a major financial obstacle to investment in community services, particularly at times of austerity. The acute care pathway in adult mental health services (the most expensive of all service areas) is particularly vulnerable to budgetary pressures. The growing strain on these services may explain recent acceleration detention rates. Understanding the determinants of this trend – and how to identify and therefore prevent potentially catastrophic disinvestment – is therefore imperative.

Different explanations for the increase in involuntary admissions have been proposed, including secular increases in the use of illicit drugs and alcohol. 3 But empirical evidence is extremely limited. An alternative hypothesis is that the rise in involuntary admissions is due to bed shortages arising from reductions in mental health bed numbers (and especially long stay or ‘rehabilitation’ beds) dating back to the 1950s in England. 18 It has been suggested that fewer beds results in greater pressure to discharge patients from hospital, leading to relapse and readmission. 3

An ecological study across eight hospital trusts in England provided some support for this view, finding that the rate of involuntary admission was associated with delays in accessing in-patient beds (as well as with socio-economic deprivation and the size of minority ethnic populations). 21 More recently, our own research based on time-series analysis of national data on mental illness beds and detentions for England between 1988 and 2008 found the strongest association between these variables was observed when a time lag of one year was introduced, with bed reductions preceding increases in involuntary admissions. This association increased in magnitude when analyses were restricted to civil (non-forensic) involuntary admissions and non-secure beds. Our findings support the view that that the annual reduction in mental illness bed numbers is directly associated with the number of involuntary admissions over the short to medium term. We have estimated that for every two mental illness beds that are closed there will be one additional involuntary admission in the following year (Keown et al. , personal communication based on submitted manuscript). 5 This research has also highlighted differential rates of bed closure by type of bed: while long stay bed numbers have fallen dramatically, there has been a small but steady increase in secure beds (i.e. beds in specialist forensic mental health units, sometimes referred to as ‘medium secure units’).

Despite these important findings, there are limitations to this type of ecological analysis based on national-level data. Not only can this overlook informative local variations occurring over smaller areas, but correlation at a national level may be confounded by controlling for other associations at the level of region, trust or even hospital. The most obvious potential confounder is (area-level) socio-economic deprivation. This has never been fully explored.

Studying the geography of involuntary admission

There is an urgent need to understand the determinants of the persistent and seemingly irrevocable trend towards ever higher rates of involuntary psychiatric admission in the UK. This is even more imperative in the light of evidence that this trend may have accelerated following changes to the Mental Health Act in 2007. 86,92 Studying the geography of involuntary admission offers the opportunity to explore variations in this trend at local level and therefore to identify factors associated with higher and lower rates of involuntary admission. It therefore represents the logical next step towards achieving the aim of understanding the causes of the upward national trend.

Preliminary investigation reveals marked variation in rates of involuntary admission between regions of England. In 2009/10, there was a two-fold difference between regions with the highest (London SHA, 146 detentions per 100,000 population) and lowest (South Central SHA, 70 per 100,000) rates of formal detention, after excluding Place of Safety orders. 86

There is substantial variation in investment by PCTs on mental health services per head of population. It was estimated that the crude ratio between the PCTs with the highest and lowest mental health spend was 4.6 in 2007, falling slightly to 2.9 after weighting for age, population need and overall PCT spending allocations. 42 It is not yet known whether and how this variation is associated with differences in mental health bed numbers, occupancy, length of stay, detention rates, and access to services or patient experiences of care.

We suggest that urbanicity (i.e. population density and settlement size), ethnic density, socio-economic deprivation and the availability (or otherwise) of in-patient beds are likely to be the most important local factors that contribute to needs for involuntary treatment. While there is now limited evidence of a direct association with (fewer) mental health beds, we propose to test the hypothesis that detention under the Mental Health Act is also related to the availability and quality of community mental health services, as well as the lack of barriers to using these (as indicated by area-level deprivation). One key hypothesis is that the gains associated with improved community services have not been enjoyed equally by all, and that rates of involuntary admission may be growing fastest in those places where social networks are weakest and community services are the hardest to deliver, namely the deprived inner cities with large minority ethnic populations.

We will also explore the extent to which urbanicity and ethnic composition interact (as opposed to acting independently of one another) in their associations with rates of involuntary admission. The rate of involuntary admission may be higher in areas with diverse minority populations than in places with greater ethnic density (i.e. concentrations of people of the same ethnicity), where social networks might be stronger. By establishing linkages to Census data, we will assess the independent and interactive effects of ethnic diversity and density, and urbanicity on rates of involuntary admission.

Design

Secondary analysis of routinely collected data. Data will be organised to permit analysis as cross-sectional survey, cohort and (nested) case control studies.

Data sources

As the time this protocol was submitted, the main data sources were to be obtained from the NHS Information Centre. In 2013, the NHSIC was re-organised to be the NHS Health and Social Care Information Centre, or the HSCIC (http://www.hscic.gov.uk/). As a result, many of the services (and corresponding webpage links) we outlined in our original protocol are no longer available and we have updated this section of the protocol accordingly.

The HSCIC is responsible for collecting and collating national health and social care data in the UK. The HSCIC ensures appropriate governance checks for the use of these data, to protect the confidentiality of those on whom data is collected, through the National information Governance Board for Health and Social Care (NIGB). It also collates geographical data, including deprivation and other neighbourhood statistics, and provides support in ensuring appropriate linkage to other data sets. The significant investment on the part of the NHS in software resources and staffing to support secondary users through the HSCIC represents a major strength of the proposed research. Although the proposed data linkages will be methodologically challenging, we are confident that the study aims are achievable.

The data sets which we propose to access and link are set out below in Table 10, followed by more detailed descriptions about each of these. Since the original protocol was submitted, our research team has become very familiar with the structure, content, and capabilities of the data sets we proposed to analyse to achieve the aims of this project. In light of this, a number of amendments and additions have been made to the data sources we aim to work with in this project. Specifically, we will focus on the Mental Health Minimum Dataset (MHMDS) as our main data source to derive the outcome variable for the study, and link in information from the Hospital Episode Statistics (HES) data as well as other data sources to provide additional contextual information relevant to the study of psychiatric admission rates. Table 10 below, which was included in the original protocol, has been modified to reflect these changes. As outlined previously (and in our interim reports submitted in September 2012 and February 2013), the quality and reliability of the MHMDS is of a relatively high quality from 2010/11 onwards but poor from 2007–2009. Given that the project will focus on comprehensive cross-sectional analyses of the most comprehensive data available, additional data sources relevant to this recent annual period (2010/11) and to the study of psychiatric admissions have been sourced for linkage to the MHMDS (Table 10).

Participants

Although there will be no primary data collection (field work), our subjects will be people who have used NHS and independent sector mental health services in England, about whom data is collected routinely during episodes of care. Other analyses will be undertaken at the level of service provider (NHS Trusts and independent sector providers) and commissioning Primary Care Trusts (PCTs).

Setting

The proposed research will be conducted in England, and will include analysis of data covering all Strategic Health Authorities (SHAs), Primary Care Trusts (PCTs) and all NHS and independent sector providers of in-patient mental health care.

‘Independent sector providers’ are those units that are registered with the Healthcare Commission to accept patients detained under the Mental Health Act 1983 (as amended by the Mental Health Act 2007), as specified under S. 2(3) (b) of the Care Standards Act 2000. These independent Hospitals that are registered to receive detained patients are considered to be ‘hospitals’ for the purposes of the Mental Health Act.

Primary outcome

As previously described, a great deal of time and effort has been spent on deriving the main outcome for the study, that is, categorising patients in terms of their admission to hospital under the Mental Health Act 2007 (MHA) during the 2010/11 reporting period. Details of this process have been documented in our interim progress reports. Briefly, the outcome variable will be dichotomous: patients who were admitted to hospital under the MHA during 2010/11 versus all other patients who received any other type of care (including voluntary admission to hospital, community treatment only, or no care). This categorisation will also permit us to determine the rate of involuntary compulsory admissions to hospital at, for example, Provider Trust and PCT level (as outlined in our original protocol).

Data linkage

The next main task will be to link the MHMDS to a wide variety of relevant external data sources. The purpose of data linkage is to create a data set that will specify multiple distinct levels at which variance in detention rates might be quantified and explained, including geography (e.g. physical environment and demographic composition), service provider (most commonly NHS Trust), commissioner (PCT) and General Practice. Creating such linkages is complex and involves finding suitable data sources and appropriate methods for linkage to the MHMDS. This linkage will take place at the smallest possible spatial scale: as indicated above in Table 10, the most common linkages will be by provider Trust and (commissioning) PCT.

Strategy of analysis

This section was included in the original protocol but has been redrafted to provide greater clarity and to describe the refocused emphasis on conducting details and robust cross-sectional analysis of the best data sources available to our team: the 2010/11 MHMDS linked to contextually-rich external data sources. Linking the MHMDS to external data sources will enable us (for example) to explore the effects of age, sex, ethnicity (at the individual level), bed numbers (and types of available beds), overall bed occupancy rates, average length of in-patient stay (at Trust level), investment in services (at PCT level) and socio-economic deprivation, urbanicity and ethnic composition (at area level) on the likelihood that a user of mental health services will be admitted to hospital involuntarily. As well as assessing the main effects of these risk factors, we will explore variance at different spatial scales and test for cross-level interactions, i.e. the extent to which associations vary between places.

Multilevel modelling (MLM)

Testing of the study hypotheses will be undertaken using multilevel modelling (MLM). 94 Multilevel models will be developed using MLwiN software. MLM enables variances at different spatial levels to be modelled simultaneously and accurately, and in a way that reflects the structure of the data (i.e. ensuring that standard errors are adjusted to reflect spatial clustering within the hierarchical data). We will estimate how much of the observed variation is explained by individual and/or ‘compositional’ factors (such as the age, gender and ethnic composition of the local population), and how much is due to area characteristics, otherwise known as ‘contextual’ factors such as area deprivation or service investment. Multilevel approaches will also allow us to test for independent cross-level interactions, i.e. the extent to which individual level associations with involuntary admission vary between types of places (i.e. in terms of service investment, quality, socio-economic deprivation etc.).

In a multilevel model, it is important to structure the data correctly. In a classic MLM, the various spatial levels nest neatly within one another and, in this sense, we can create a nested hierarchy – i.e. individuals (level 1) nested within LSOAs (level 2) within Provider Trusts (level 3). The structure of the MHMDS is quite complex and the data are not nested neatly. If we were to run a MLM that assumed the data were nested neatly then the resultant estimates of variance across the different geographies would not be correct. Therefore, rather than regard the geographical areas as ‘levels’ within a hierarchy, they are instead regarded as ‘classifications’ and for each individual within the data set, the classification structure is recorded, allowing for a non-nested, cross-classified data structure. Markov chain Monte Carlo (MCMC) estimation methods are available in MLwiN to estimate the variance of our outcome variable across these geographic classifications (i.e. LSOA, Commissioner, Provider Trust etc.), within these types of multilevel models.

Our first aim was is to estimate how much of the observed variation in involuntary admission rates across England is explained by individual factors such as gender and age and how much is due to factors operating at these different spatial levels. The most basic MLM is the null model, which estimates the total unexplained variance in involuntary detentions in the study sample (this is equal to the sum of the unexplained variances at the different spatial levels).

Given the size of the database (∼1.2 million patients), estimating models with more than three spatial levels would be difficult due to problems with model stability and computing power. A series of three-level models will be estimated, varying the structure of the spatial levels in an attempt to identify where the largest variance in involuntary admissions might be occurring, and in order to guide us in populating more definitive models. Subsequent modelling will be used to explore the extent to which this variance is explained by factors such that include individual characteristics (like age, sex, and ethnicity, where these are available) and area-level exposures such as urbanicity, socio-economic deprivation and the size of minority ethnic populations. Each of these variables will be allocated (and introduced to the model) at the appropriate level. Likewise, modelling will explore associations between rates of involuntary admission and the availability (and occupancy) of different types of inpatient beds (at provider level). A major advantage of multi-level modelling is the opportunity to measure cross-level interactions (e.g. between bed numbers and PCT mental health spend, for instance).

Sample size

It is not possible to derive precise sample size estimates for our analyses, and nor are these needed in the design of our research since no new data are to be collected. The data sets at our disposal are extremely large, and are arguably the most comprehensive population-based sources of data on the use of mental health services anywhere in the world. This reflects both the nature of nationalised health care provision in the UK, and high levels of investment in the collection and collation of routine activity data.

Our analyses will cover all of England, and will include information from NHS and independent sector providers of in-patient mental health care (as previously, these independent sector providers are coded by the service provider (company) and not by individual units or according to geographical location). In 2009/10 this comprised 10 Strategic Health Authorities (signifying region), 151 Primary Care Trusts (as commissioners), 58 NHS Mental Health Trusts (including Care Trusts and Mental Health Foundation Trusts) and 211 registered independent mental hospitals. In 2009/10 more than 1.25 M people accessed mental health services in England (2.7% of the population). 95 In the same year, 107,765 people were in-patients during the year and 42,479 were detained in hospital under the MHA (39.4% of all admissions). Of these, 2717 (4.8% of all detentions) were in independent mental hospitals and 760 (1.8% of all detentions) were under court or prison disposal orders.

Whilst sample size is unlikely to be a problem for the multilevel rates analyses, the individual logistic analyses may have relatively few observations nested within small areas. To overcome this problem, parameters will be estimated using second-order Taylor expansion with predictive quasi likelihood (PQL). This estimation procedure is considered superior to first or second order marginal quasi-likelihood (MQL) when observations within clusters, such as individuals within small areas, are small. 96 Markov chain Monte Carlo (MCMC) methods will further improve the accuracy of such estimates. 97

Timetable

This project was granted a no-cost time extension to extend the project end date from 31/05/2013 to 31/08/2013. The reasons for this extension were outlined in our interim reports (September 2012 and February 2013) and were largely due to administrative delays in hiring the Research Fellow to the project on a full-time basis. This section has been amended accordingly to reflect this extension.

Months –3 to 0: Applications for (a) research ethics and NIGB approval and (b) permission for access to anonymised patient-level data from the NHS Information Centre’s Security and Confidentiality Advisory Group; further data and software acquisition. Research Fellow will begin on-line training in multilevel modelling through the Centre for Multilevel Modelling (University of Bristol). First meeting of full study team.

Months 1 to 3: Induction of Research Fellow and completion of on-line multilevel modelling training; secure NIGB and NRES permissions and finalise acquisition of all study data; liaison with NHS Information Centre Secondary Uses Service to support data linkage. Plan & execute data management. Further meeting of the full study team.

Months 4 to 15: Data analysis. Regular study team meetings; regular supervision of Research Fellow (Warwick and Portsmouth). Initial meetings with service users (via Mental Health Foundation, MHF), and with NHS mental health service clinicians and managers (via Mental Health Network, MHN).

Months 16 to 18: Interpretation of findings, final meetings of full study team. Further consultation event with service users (facilitated by Mental Health Foundation) NHS mental health service Chief Executives (facilitated by Mental Health Network); report writing and dissemination (including formal briefing reports to be published by MHF and the NHS Confederation (governing body of the MHN).

Project management

Management of the project is of the utmost importance, given that the research team will be spread out across five sites within the UK. As we describe below, this project has arisen from strong research partnerships and collaborations that span the past decade. We have an established track record of working together very productively, and this is further evidenced by the significant research outputs that have accrued between the time of the outline submission and the present. We have published numerous papers together in high impact scientific journals, including several notable publications in the last year alone. We are a tried and tested collaborative of like-minded and committed academics who know one another very well, and we have established an effective modus operandi over many years.

SW, who is based at the University of Warwick, will be the Chief Investigator and Project Manager, and will have day to day responsibility for all aspects of the project. His time is budgeted at 10%. SW has considerable management experience, and is also Clinical Director of West Midlands (South) CLRN, where he is responsible for an annual budget of £6.7 m. He is also chair of the Psychological and Community Therapies Panel at the NIHR Health Technology Assessment programme.

PK (in Newcastle) will be Deputy Project Manager, and his time has also been costed at 10%. He has a wealth of experience of using a number of the study data sets, and of liaising with the NHS Information Centre.

There will frequent communication between members of the research team, and we have scheduled regular face-to-face meetings every two months at rotating venues. We will hold a teleconference (assisted where appropriate by Skype) between these meetings, to ensure that the full study team discuss progress at least once a month. These meetings will be minuted and will be used to monitor progress and to identify any necessary action points.

The proposed Research Fellow, Dr McBride, has a proven record of academic achievement and a well-deserved reputation for being able to work independently. Her CV (which we attach) testifies to her drive and ambition, as well as her expertise. She has undertaken very productive visiting fellowships overseas, where she has increased her knowledge and learning but has also produced significant research outputs. She has worked and published previously with SW, and the physical separation is unlikely to be a hindrance. We have budgeted for her to travel regularly to England where she will be supervised both by SW but also by Dr Twigg in Portsmouth, by Liz Twigg. Dr Twigg has worked extensively with SW and has been recruited to the project for her experience as a senior health geographer and her considerable expertise in multilevel statistical modelling.

We have also set out our proposals for working with our two partners, the Mental Health Foundation and the Mental Health Network below. The former will be enhanced by the David Crepaz-Keay’s membership of the project steering group.

Service users/public involvement

We propose to engage two sets of ‘users’: users of mental health services, and senior NHS mental health service managers and clinicians.

We will involve mental health service users through our partnership with the Mental Health Foundation (MHF) (http://www.mentalhealth.org.uk), the UK’s leading mental health research, policy and service improvement charity. Service user involvement will be coordinated by David Crepaz-Keay, Head of Empowerment and Social Inclusion at the MHF and a co-application on this proposal. The main aim of this partnership is to ensure that the service user perspective shapes the research questions, influences analysis and interpretation and assists in disseminating the findings.

Working through and with the MHF, we propose to circulate a summary of the study aims and methods to a network of mental health service users for comment. We will also seek feedback from service users on our preliminary findings.

We have also established a partnership with representatives of another key ‘user’ group, namely senior NHS mental health service managers. We will work with the Mental Health Network (part of the NHS Confederation) to recruit participants to two consultation events. The NHS Confederation represents more than 90% of the organisations that make up the NHS. Its members include the majority of NHS acute trusts, ambulance trusts, foundation trusts, mental health trusts, primary care trusts, special health authorities and strategic health authorities in England; trusts and local health boards in Wales; and health and social service trusts and boards in Northern Ireland.

Both mental health service users and NHS mental health managers (and chief executives) will be asked to comment on our research aims, hypotheses at the start of the project, and their views on how best to interpret our findings will be solicited at the end, but before formal dissemination. This will ensure that the research questions are framed to address key concerns of those responsible for delivering mental health care in the UK as well as those on the receiving end. Feedback from both sets of consultations will enable us to present our findings in the form of actionable conclusions.

Expertise and justification of support required