Parts of this chapter are reproduced or adapted with permission from a protocol paper published by the research team in BMJ Open.96 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/4.0/. The text below includes minor additions and formatting changes to the original text.
Overview
The overall aim of WP2 was to provide quantitative evidence to enable further testing of our programme theory about what works to deliver consistently safe, timely transfers of care. The underpinning logic of the HHDF was that safe, timely transfers of care will have an impact on a range of outcomes, including those for improved health and well-being of patients and the wider health and social care economy. In this chapter, we use data linkage findings to explore HHDSs’ impact on outcomes and different patterns of service use (objective 5). The primary outcomes are (1) time from discharge to next hospital inpatient admission, (2) time from discharge to next A&E attendance and (3) 28-day emergency readmission. We also consider a range of secondary outcomes.
In WP2, we broaden the focus of study, collecting data for 17 HHDSs. To test our programme theory about the importance of systems having all the ‘jigsaw pieces’ (including access to a clinically led multidisciplinary homeless team), we analyse and report the data linkage findings, comparatively exploring differences in the following:
Standard care compared with specialist care.
HHDSs that are clinically led (multidisciplinary) compared with those that are housing led (uniprofessional).
HHDSs with ‘step-down’ (intermediate) care compared with those without ‘step-down’ care. We focus on all types of intermediate care as the key mechanism of interest because of its relative importance, as evidenced in the previous chapter, and also because time and resources precluded any further deeper analysis of different typological configurations. This third comparator for step-down care was not outlined in the original protocol; however, reflecting good practice for realist quality standards, we have made allowances in the overall design of the study to collect additional data for further refinement of programme theories as the evaluation unfolds.
The main advantages of employing the data linkage as part of the wider realist mixed-method approach were twofold. First, it enabled us to gather data on outcomes that would be more geographically representative than previous analyses. For example, Hewett et al.’s78 RCT of the Pathway schemes was limited to two hospital sites, whereas this present analysis focuses on 17 hospital sites and a broader range of scheme typologies. Second, the use of administrative records may offer a complimentary and/or more realistic evaluation of the benefits of the intervention than measurements taken in an interventional study setting, especially as there is the risk of bias when HHDS staff are involved in data collection.
Methods
Comparator groups
Three groups of individuals admitted to hospital were included in the analysis. The first group comprised homeless individuals admitted to hospital at any one of 17 sites with a HHDS between 1 November 2013 and 30 November 2016. An initial audit of HHDSs established the date on which they were implemented.
The second group comprised individuals seen by a community homeless service in London, UK [Find & Treat (F&T)] and subsequently admitted to a hospital that did not have a HHDS. This group was, therefore, selected to provide a comparison group of people experiencing homelessness, admitted to hospital, but not seen by a HHDS. F&T is a specialist outreach team that works alongside over 200 NHS and third-sector front-line services to prevent or treat tuberculosis among homeless people. It is primarily based in London. We looked for hospital admissions during the study period for all individuals seen by F&T between 1 November 2013 and 30 November 2016.
The third group was a random sample of individuals equal in size to the F&T group and living in lower-layer super output areas in England in the most deprived quintile, as measured by the Index of Multiple Deprivation (IMD), who were recorded as having a fixed address. These individuals came from hospitals offering HHDSs with an admission during the study period but without being seen by the HHDS. This group was used to confirm or refute our hypothesis that the homeless people included in the study have worse outcomes than non-homeless people living in deprived areas. This group is subsequently referred to as IMD5 (Index of Multiple Deprivation quintile 5).
Data linkage
Data were linked in respect of each group from HES, which includes dates, causes and length of admission, and the Office for National Statistics (ONS) mortality database, which includes dates and causes of death.
Data collection, processing and linkage
Data flows used in this analysis are described in . We obtained demographic-identifying variables under our legal and ethics approvals. Where available, NHS numbers were collected from the 17 HHDS sites. NHS Digital (formally the Health and Social Care Information Centre) used the personal demographics service to identify and add NHS numbers to as many individual records as possible.
Study data flows. PDS, personal demographics service; UCL, University College London.
NHS Digital undertook all data linkage to HES data using patient identifiers obtained (in combination with personal demographics service tracing) and we securely uploaded a de-identified copy of the data to the University College London (London, UK) Institute of Health Informatics’ Data Safe Haven, which is a robust infrastructure certified for processing and analysing identifiable data in accordance with international and national information security standards (ISO/IEC 27001:2013 and NHS Information Governance Toolkit). The deterministic linkage process was undertaken in three steps to ensure that the three groups were mutually exclusive:
HES linkage to data collected at the HHDS sites
HES linkage to data collected from F&T services, with admissions stratified into those sites with and without HHDSs
identification of people in the most deprived IMD quintile from HES at the HHDS sites who were admitted after the implementation date at a given HHDS site.
In the main analysis, individuals identified in more than one of the steps described above were classified using the hierarchy 1 > 2 > 3, such that data on homeless patients who were admitted multiple times to a combination of hospitals with, and without, HHDSs were analysed according to admission to the HHDS site only. Patients identified as both homeless (i.e. HHDS or F&T) as well as those in the non-homeless (IMD5) group were assigned to the homeless HHDS category. In addition to collecting data on admissions after being seen by an HHDS, we obtained data on all individuals from HES in respect of admissions to any hospital in England from 1 January 2008 to 31 October 2013 to create a profile of their existing health conditions.
Sample size
Our sample size calculation was based on historical data from the two types of HHDS, suggesting an average of 92 patients per month across the 17 sites. We estimated that each type would have data from approximately 2208 patients for the duration of the study.
From previous health service evaluations, we expected 0.7 hospital episodes per person-year for homeless individuals. A clinically important reduction in readmission rates would be 10%. To undertake a sample size calculation to determine the study size required to detect such a reduction in readmission rates, the following variables were defined:
µ_0 – mean readmission rate in the baseline (F&T) group = 0.7 episodes per person-year.
µ_1 – mean readmission rate in the HHDS group = 0.6 episodes per person-year.
v = 1.96 – percentage point of the normal distribution corresponding to a 5% two-sided significance level.
u = 0.84 – one-sided percentage point of the normal distribution corresponding to {100% – type II error [false negative/(true positive + false negative)]} at 80% power.
Using the following equation, we estimated minimum sample size (per group) required for comparison of two rates (readmissions per person-year):
Adjustment for confounders (i.e. age, sex, ethnicity, calendar quarter of admission and existing comorbidities), as far as is possible within the analysis, resulted in a required doubling of sample size (additional 10% per confounder) required. Therefore, 2038 person-years were required per group. Assuming that the HHDS sites see, on average, 90 patients each month during the study period (which, for the purposes of the initial calculation we assumed to be November 2013 to November 2015, as many schemes may have been shut down before this time period and in some data may have been collected at an earlier point), we estimated that this translated to a total of 2160 individuals at each of the two types of scheme (e.g. housing led and clinically led). Given that the average follow-up period was likely to exceed 1 year, we anticipated that the study would be powered to detect a 10% difference in readmission rates between HHDS sites and non-HHDS sites for both the clinically- and the housing-led HHDSs.
Outcomes
A series of primary and secondary outcomes were included to ensure that our analysis was both consistent with outcomes used in previous published analyses and included outcomes that previous studies were not powered to collect. We chose several secondary outcomes that collectively reflect the priorities of health policy-makers and individuals attending patient engagement workshops. Full definitions of primary and secondary outcomes are provided in and .
Definition and methodological approach for primary outcomes
Definition and methodological approach for secondary outcomes
Throughout the report we refer to three terms, avoidable,96 preventable and amenable, for which we use the following definitions and their associated ICD-10 (International Statistical Classification of Diseases and Related Health Problems, Tenth Revision) codes:
Avoidable mortality – avoidable deaths are all those defined as preventable, amenable or both, where each death is counted only once. Where a cause of death falls within both the preventable and amenable definition, all deaths from that cause are counted in both categories when they are presented separately.
Amenable mortality – a death is amenable if, in the light of medical knowledge and technology at the time of death, all or most deaths from that cause (subject to age limits, if appropriate) could be avoided through good quality health care.
Preventable mortality – a death is preventable if, in the light of understanding of the determinants of health at the time of death, all or most deaths from that cause (subject to age limits, if appropriate) could be avoided by public health interventions in the broadest sense.
Analysis plan
We undertook the analysis in two phases. In the first phase, we analysed baseline characteristics () of all participants to describe the characteristics of each of the study groups at or before the index admission (). With the exception of ethnicity, all baseline characteristics were anticipated to be fully observed (chronic disease is presumed to be absent unless recorded). Missing values of ethnicity were analysed grouped as ‘not recorded’.
Patient characteristics in the time prior to the index admission were collated as baseline measurements
Schematic outlining hypothetical patients. Data on the characteristics of patients before their index admission were collated from admissions occurring before the implementation of HHDSs at a given site (or for comparator groups a randomly selected date (more...)
We then summarised each of the primary and secondary outcomes by comparison group and explored the geographical spread of our 17 sites to explore their representativeness. This work enabled us to confirm the suitability of the proposed statistical methods and analysis protocol proposed. Where applicable, we compared our baseline mortality data with those in the recent analysis of homeless deaths by the ONS (see Blackburn et al.,96 table S1, supplementary file, for definitions and full ICD-10 code lists used).
In the second phase, we identified evidence of differences in the baseline characteristics, including age, sex, chronic disease and reason for hospital admission at the time of their index admission. We estimated the crude association between each of the primary and secondary outcomes and the study population groups. The baseline comparator group was IMD5. We then re-estimated the association between each of the outcomes and the study population group after adjusting for the following characteristics at the time of admission: age, sex, chronic disease and reason for hospital admission. We undertook supplementary subgroup analyses to evaluate evidence of a difference in the outcomes of people (1) admitted to clinically- compared with housing-led schemes and (2) schemes with step-down care compared with schemes without access to step-down care.
Crude models were fitted prior to adjustment for ‘baseline’ measurements at or before the index admission. We then wrote-up the analysis in accordance with the RECORD (REporting of studies Conducted using Observational Routinely collected Data) statement.
Protocol deviations
The data linkage study was significantly delayed at commencement because of the launch of the Health Research Authority and changes to the procedures for securing ethics approvals. It was further delayed by NHS Digital in relation to approving and returning the linked data set and then, finally, because of NHS Digital initially returning a data set without an IMD5 comparison group or linked death data. As a result, we did not have time to complete the analyses for the following secondary outcomes:
time from discharge to next elective admission
all-cause mortality expressed as a SMR
ICD-10 chapter-specific SMR.
Having completed the main literature synthesis (some 6 months after the study began), we developed our original realist hypothesis to capture new evidence about the importance of HHDSs having direct access to ‘step-down’ intermediate care. To enable us to interrogate this revised theory, we introduced a further comparative element across the evaluation to establish the effectiveness and cost-effectiveness of HHDSs with access to ‘step-down’ care (MIR4) compared with those with no direct access to ‘step-down’ care.
Results
Parts of this section are reproduced or adapted with permission from Aldridge et al.97 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: https://creativecommons.org/licenses/by/4.0/. The text below includes minor additions and formatting changes to the original text.
To create an analysis data set, we joined the HHDS data (12,931 records) with F&T comparator group data (47,569 records). The combined data set (60,500 records) was cleaned, assigned gender and underwent validation of NHS numbers. We then applied age and date exclusions, and expanded the data set on names. We then validated NHS numbers within the data set. This was then split into an identifiable linkage data set with 73,432 records. Identifiable records were destroyed prior to receiving HES data from NHS Digital. The second retained unidentifiable data set was kept within the Data Safe Haven. This data set was linked to HES data from NHS Digital and underwent cleaning after linkage. After the cleaning process, we created three separate cohorts of patients. First was the HHDS group for whom we had 3894 unique patients, 6985 admission records and a record level linkage rate of 54% (6985/12,931). Second was the F&T–HES group with 7226 unique patients, 20,206 hospital admission records and a linkage rate of 42.5% (20,206/47,569). Third was the IMD5 group with 34,450 unique patients. Linkage rates were, therefore, higher in the HHDS group [54.0%, 95% confidence interval (CI) 53.2 to 54.9] than in the F&T group (42.5%, 95% CI 42.0 to 42.9).
To examine possible biases introduced into the linkage, we undertook several descriptive analyses. We also compared the age and sex characteristics of each cohort before and after linkage (). This analysis revealed that after linkage the data sets contained slightly more women and older patients, on an admission record basis, than the data sets prior to linkage. The HHDS sites from which data were collected covered seven out of nine statistical regions in England (), with London having the largest number of participants. We examined comorbidities across comparator groups and, with few exceptions, we found these to be markedly high in those seen by HHDS sites (). HHDS patients had the highest levels of comorbidities across all ICD-10 chapters, except for endocrine and maternal disorders.
Age and sex distribution of comparison groups at the hospital admission record level (e.g. not unique) before and after linkage to HES
Map of number of individuals seen by HHDSs by ONS statistical region. There were no HHDS data linkage sites in the greyed areas.
Levels of comorbidities by ICD-10 chapter comparison group
These preliminary descriptive analyses showed that the F&T group had lower linkage rates and levels of comorbidity than the HHDS groups. These differences in linkage rates and comorbidity between the F&T and the HHDS groups, and lack of data on timing of homelessness exposure in the F&T data set, provided cause for concern about the validity of our assumption that this group was still homeless during their hospital admission during the study time period. For this reason, we took the decision not to include the F&T population as a comparison group in analyses of the primary and secondary outcomes.
Causes of death
In addition to understanding the geographical spread of the data, the demographic differences in each group after linkage and the levels of comorbidities, we also examined causes of death before undertaking a full analysis of the primary and secondary outcomes. shows the demographic characteristics of decedents in each group. Males made up 77.7% (466/600) of deaths in the HHDS group and 56.4% (1416/2512) of deaths in the IMD5 group. The median age of death was 51.6 [interquartile range (IQR) 42.7–60.2] years for the HHDS group and 71.5 (IQR 60.67–79.0) years for the IMD5 group. Ethnicity was broadly similar across comparison groups, but 12.9% (403/3,112) of ethnicities were either not known or not stated. The top three underlying causes of death by ICD-10 chapter in the HHDS group were external causes of death (21.7%, 130/600), cancer (19.0%, 114/600) and digestive disease (19.0%, 114/600) in the unweighted analysis.
Baseline demographic characteristics and percentage of deaths in each comparison group
In an age- and sex-weighted analysis ( and ), the top three underlying causes of death in the HHDS group were cardiovascular disease (30.1%, 180/600), cancer (21.8%, 131/600) and respiratory disease (16.7%, 100/600). In this weighted analysis, the underlying causes of death for each cohort by ICD-10 chapter showed a greater proportion of deaths with an underlying cause in ‘External Causes of Morbidity and Mortality’ chapter in the homeless group (7.4%, 44/600) than in the IMD5 group (3.9%, 99/2,512). This chapter covers a range of causes, including accidents, intentional self-harm, assault and events of undetermined intent (e.g. poisoning). Weighted alcohol- and drug-related causes were higher in the HHDS group (9.8%, 59/600 and 3.6%, 22/600, respectively) than in the IMD5 group (2.7%, 67/2512 and 0.9%, 23/2512, respectively). The percentage of deaths with an amenable cause of mortality was higher in the HHDS group (30.2%, 181/600) than in the IMD5 group (23.0%, 578/2512). This means that after weighting for age and sex, nearly one in three deaths of people in our homeless hospital discharge cohort were due to conditions, many of which are common in the general population (e.g. heart disease), that are amenable to timely health care.
Underlying causes of death by ICD-10 chapter and amenable mortality
Underlying cause of death by ICD-10 group and subgroup, comparing homeless weighted and IMD5 groups. COPD, chronic obstructive pulmonary disease; CVD, cardiovascular disease; IHD, ischaemic heart disease. Adapted with permission from Aldridge et al. This (more...)
In an unweighted analysis that compared the HHDS group with the ONS homeless causes of death analysis, accidents and suicides contributed to a smaller percentage of all underlying causes of death in the HHDS homeless group (19.5%, 117/600) than in the ONS homeless group (54.9%, 319/581) (). Conversely, we observed a larger number of deaths due to cancer (18.2%, 109/600 vs. 4.6%, 27/581), diseases of the liver (13.8%, 83/600 vs. 9.0%, 52/581) and chronic lower respiratory disease (6.0%, 36/600 vs. 2.2%, 13/581) in the HHDS group than in the ONS analysis.
Underlying cause of death across the three comparator groups. Adapted with permission from Aldridge et al. This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others (more...)
Primary outcomes
Time from discharge to next hospital inpatient readmission (any cause)
In the unadjusted analysis comparing clinically led schemes with housing-led schemes for time from discharge to next hospital inpatient readmission (any cause), male sex, older age group, larger number of comorbidities and larger number of admissions in the 1 year before index admission were all associated with an increase in readmissions (). In the multivariable analyses, there remained an association between older age, larger numbers of comorbidities, larger numbers of numbers of admissions in the 1 year before the index admission and an increase in readmissions. We found no evidence in either the unadjusted or adjusted analyses for a difference between the housing- or clinically led services for readmissions [hazard ratio (HR) 1.07, 95% CI 0.88 to 1.30; p = 0.4827]. When grouping the HHDS sites into those with step-down care and those without step-down care, in a multivariable analysis, we found no evidence for a difference in readmission rates between the two groups (HR 1.10, 95% CI 0.92 to 1.33; p = 0.2863) (see ). In an adjusted analysis, hazard of readmission was higher in the combined HHDS group than in the IMD5 group (HR 3.14, 95% CI 2.81 to 3.51; p < 0.0001) (see ).
Baseline characteristics and univariate and multivariate analysis of time from discharge to next hospital inpatient readmission (any cause) for housing- vs. clinically led sites
Summary table of adjusted analyses
Time from discharge to next accident and emergency attendance
In the unadjusted analysis comparing clinically led services with housing-led services for time from discharge to next A&E attendance, sex, age group, number of comorbidities and number of admissions in the year before index admission were all associated with an increase in emergency attendance (). In the multivariable analyses, older age, larger numbers of comorbidities and larger numbers of admissions in the 1 year before the index admission were all associated with an increase in emergency attendance. We found no evidence in either the unadjusted or adjusted analyses for a difference between the housing- or clinically led services for emergency attendance (HR 1.05, 95% CI 0.83 to 1.32; p = 0.6835) (). We grouped HHDS sites into those with step-down care and those without step-down care and found no evidence for a difference in emergency attendance between the two groups in a multivariable analysis (HR 1.01, 95% CI 0.82 to 1.26; p = 0.9063) (). In a multivariate analysis, the hazard of emergency attendance was higher in the combined HHDS group than in the IMD5 group (HR 7.34, 95% CI 6.31 to 8.54; p < 0.0001) (see ).
Baseline characteristics and univariate and multivariate analysis of time from discharge to next A&E attendance for housing- vs. clinically led sites
Baseline characteristics and univariate and multivariate analysis of 28-day emergency readmission for housing- vs. clinically led sites
28-day emergency readmission
The overall percentage of patients with a 28-day emergency readmission in the housing-led group (19.5%, 196/1006) was slightly lower than in the clinically led group (21.2%, 613/2888) (see ). Male patients (21.1%, 600/2840) had slightly higher readmission rates than female patients (19.8%, 209/1054) and readmission rates were highest in the 55- to 64-years age group (22.4%, 134/599). In both the unadjusted and the adjusted analyses, only number of comorbidities and number of admissions in the year before index admissions were associated with 28-day emergency readmissions. In an adjusted analysis, no evidence was found for a difference in readmission rates between housing- and clinically led services [odds ratio (OR) 1.08, 95% CI 0.89 to 1.29; p = 0.4420]. In an adjusted analysis, we found that the HHDS group had higher odds of 28-day emergency readmissions than the IMD5 group (OR 3.60, 95% CI 3.24 to 3.99; p < 0.0001) (see ). When comparing services with step-down care with those without step-down care, the percentage of patients with 28-day emergency readmissions was similar across both groups (20.2% for step-down care vs. 21.1% for no step-down care). There was no evidence in either the unadjusted or adjusted multivariate analysis for a difference between these groups in terms of 28-day emergency readmission (OR 1.05, 95% CI 0.88 to 1.25; p = 0.5821) (see ).
Secondary outcomes
Time from discharge to admission with ambulatory care sensitive conditions
Being of female sex appeared to be somewhat slightly protective against readmission with ambulatory care sensitive conditions () in the unadjusted figures (HR 0.82, 95% CI 0.68 to 1.0; p = 0.0480); however, this was not true of the adjusted HR 0.95 (95% CI 0.78 to 1.16; p = 0.6227). There was no overall difference in likelihood of readmission when comparing the housing-led group with the clinically led group, with the unadjusted HR being 1.11 (95% CI 0.86 to 1.42; p = 0.4174) and adjusted HR being 1.11 (95% CI 0.86 to 1.44; p = 0.4159).
Baseline characteristics and univariate and multivariate analysis of time from discharge to next admission owing to an ambulatory care sensitive condition for housing- vs. clinically led sites
Overall and unscheduled readmission rates
We looked at the 12-month risk of readmission by ICD-10 chapter of index admission (). The 12-month risk of unscheduled readmission (often referred to as emergency readmission) was 59% (95% CI 57% to 61%) for homeless patients and 20% (95% CI 19% to 21%) for housed patients, for planned readmission it was 18% (95% CI 17% to 20%) for homeless patients and 28% (95% CI 26% to 29%) for housed patients and for A&E visits it was 92% (95% CI 91% to 93%) for homeless patients and 57% (95% CI 55% to 59%) for housed patients.
Twelve-month risk of readmission, stratified by the ICD-10 chapter of index admission: (a) emergency readmission; (b) planned readmission; and (c) A&E.
Unplanned and A&E readmissions were extremely high for the HHDS group, and these unplanned readmissions are high regardless of the type of index admission. This is in contrast to the IMD5 group, where the readmission risk was lower than in the HHDS group and varied according to the type of index admission. In addition, the readmission risk was particularly high for those admitted for ‘mental and behavioural’ causes. Planned readmissions were generally lower in the HHDS group than in the IMD5 group.
Unadjusted incident rate ratios from a negative binomial model that examined the risk of readmission showed that homeless patients had 5.14 (95% CI 4.71 to 5.62) times the rate of unscheduled readmission, 1.08 (95% CI 0.95 to 1.23) times the rate of planned readmission and 5.38 (95% CI 5.04 to 5.72) times the rate of A&E visits (). Adjusting for comorbidities and the ICD-10 chapter of the index admission partially attenuated these values, with adjusted rate ratios of 3.77 (95% CI 3.46 to 4.10), 0.71 (95% CI 0.63 to 0.81) and 3.76 (95% CI 3.53 to 4.01), respectively. As an exploratory sensitivity analysis (not prespecified), we fit the model for planned readmissions, excluding ‘series’ of admission (defined above, mainly for renal failure and dialysis), which gave an unadjusted rate ratio of 0.69 (95% CI 0.62 to 0.78) and an adjusted ratio of 0.62 (95% CI 0.55 to 0.69).
Incident rate ratios of readmissions and A&E visits, comparing homeless with housed patients (results of negative binomial regression)
We then constructed negative binomial regressions models to examine the risks of readmission, comparing homeless patients at sites with an associated step-down service with homeless patients at sites without a step-down service (baseline group). Among homeless patients, the rate of unscheduled admissions was similar when comparing those discharged at a site with or without a step-down service, but the rate of planned admissions was 56% lower (incident rate ratio 0.44, 95% CI 0.35 to 0.55; p < 0.001) and the rate of A&E visits was 21% lower (incident rate ratio 0.79, 95% CI 0.71 to 0.87; p < 0.001) (). These results were partially attenuated when adjusting for comorbidities and the ICD-10 chapter of the index admission. After adjustment, there was evidence that A&E visits were 18% lower among homeless patients discharged at a site with a step-down service than among those without a step-down service (incident rate ratio 0.82, 95% CI 0.75 to 0.91; p < 0.001).
Incident rate ratios of readmissions and A&E visits, comparing homeless patients at sites with an associated step-down service with homeless patients at sites without a step-down service (baseline group) (results of negative binomial regression) (more...)
Time from admission to death
In the unadjusted analysis for this outcome, male sex, older age group, larger number of comorbidities and larger number of admissions in the year before index admission were all associated with an increase in likelihood of death (). In the multivariable analyses, there remained an association only between older age and higher comorbidities. We found no evidence in either the unadjusted or adjusted analyses for a difference between the housing- or clinically led services for death (adjusted HR 1.01, 95% CI 0.84 to 1.21; p = 0.9285). In another adjusted analysis, hazard of death was also higher in the combined HHDS group than in the IMD5 group (HR 2.20, 95% CI 1.98 to 2.44; p < 0.0001) (see ). When comparing schemes with and without step-down care, we found no evidence for a difference in likelihood of death between the two groups (adjusted HR 1.05, 95% CI 0.88 to 1.24; p = 0.6103) (see ).
Baseline characteristics and univariate and multivariate analysis of time from death due to any cause for housing- vs. clinically led sites
Time from admission to mortality from causes amenable to health care
Male sex, older age group and larger number of comorbidities were all associated with an increase in likelihood of death from causes amenable to health care () in the analysis before and after adjusting. We found no evidence in either the unadjusted or adjusted analyses for a difference between the housing- and clinically led services for readmissions (HR 1.02, 95% CI 0.69 to 1.49; p = 0.9383). The HHDS group had a higher hazard of death from amenable causes than the IMD5 group in the adjusted analysis (HR 1.78, 95% CI 1.43 to 2.22; p < 0.0001) (see ). We found no evidence for a difference in likelihood of death between the step-down care group and the no step-down care group (adjusted HR 0.99, 95% CI 0.69 to 1.42; p = 0.9583) (see ).
Baseline characteristics and univariate and multivariate analysis of time from death causes amenable to health care for housing- vs. clinically led sites
Parts of this section are reproduced with permission from Aldridge et al.97 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: https://creativecommons.org/licenses/by/4.0/. The text below includes minor additions and formatting changes to the original text.
Summary
Overall, the data linkage results show that HHDSs were engaging with a population with high levels of unmet health-care needs that were often linked to common long-term conditions, such as heart disease, as well as conditions more usually associated with homelessness, such as drug and alcohol misuse. There is some evidence in support of our programme theory about the MIRs that enable HHDSs to deliver their intended outcomes; however, this is not supportive of all elements.
We found no evidence of a difference for our primary outcomes between HHDSs that were clinically led and those that were housing led.
We found no evidence of a difference for our primary outcomes between HHDSs that did and did not encompass ‘step-down’ intermediate care.
In a secondary outcome analysis, we found evidence that A&E visits were 18% lower among homeless patients discharged at a site with step-down service than among those discharged at a site without a step-down service.
In the HHDS cohort, regardless of scheme typology, readmission rates increased across the board with increasing comorbidities.
After weighting our analyses for age and sex, nearly one-third of deaths in the HHDS cohort were from conditions amenable to health care, suggesting a service shortfall in the HHDS and/or the wider complex adaptive system.
When making comparisons between the HHDS group (as a whole) and the IMD5 groups, the HHDS group consistently fared worse. We expected this finding, as we did not expect the HHDSs to improve the health outcomes to that of a deprived housed population. Instead, these results are helpful in terms of providing evidence on the extent of the remaining health gap between those experiencing homelessness and housed populations. Of note is the finding that the HHDS group were more likely than others to be readmitted in an emergency, with five times the rate of unplanned hospital readmission and five times the rate of A&E visits than housed people from deprived neighbourhoods. This was not explained by morbidity, although there is likely to be some residual confounding in which differences in health at the index admission are not captured by long-term conditions identified from previous hospital records (our measure of morbidity). The predominance of emergency rather than planned health care suggests that access to planned care and community support is poor, despite high levels of need.
In the next chapter, we begin to tease out the implications of these quantitative findings for our programme theory by situating them as part of the wider picture ascertained through both the qualitative and economic WPs. In the remainder of this section, we reflect on some of the limitations of the data linkage and how this may have an impact on these results.
Limitations
As described above, when designing the study, we chose the F&T comparator group (for standard care), as we hoped that this would provide a control group of people initially identified as homeless in the community and subsequently admitted to hospitals without HHDSs (i.e. a standard-care comparator). However, with the benefit of hindsight, and our initial descriptive analysis of the data once we received it, we do not feel this to be a fair comparison. For this reason, we do not present results for our primary and secondary outcomes for the F&T comparator group. We were, therefore, not able to undertake a comparative effectiveness analysis that examined the outcomes for people experiencing homelessness admitted to hospitals with a HHDS and those admitted to a hospital without a HHDS.
As noted above, we found no evidence for a difference in our primary outcomes when comparing housing- and clinically led services, or those with step-down service or no step-down service. These findings may be correct, but there remain several important alternative explanations that mean there could still be a difference between these types of service that we were unable to detect. First, we gained data from the sites we planned to and were only slightly under the total number of patients required by our sample size calculation. However, there were more than twice the number of patients in the clinically led sites than in the housing-led sites. Our study is, therefore, underpowered to detect the clinically significant 10% difference in the primary outcomes that we set out to achieve. Taking one of our primary outcomes as an example – readmission owing to any cause – the CI in our final adjusted comparison between housing- and clinically led services (HR 1.07, 95% CI 0.88 to 1.30) is statistically consistent with housing-led services being up to 30% better or up to 12% worse than clinically led services. Our secondary outcome analysis of multiple readmission rates between different types of HHDSs had greater power because of the larger number of events included and it is, therefore, not surprising that this was the one analysis where we found statistical evidence of a difference between different typologies of HHDS for a reduction in A&E admissions. Second, the data we collected were for routine records and, therefore, there may be problems in the recording of outcomes and exposures, which may introduce random measurement error into our analyses, reducing power of the study to detect a difference between the groups further. Third, we were limited in the number of variables we could collect in our analysis on each group and there may be residual confounding in our comparisons. Fourth, there was substantial geographical differences in the way sites were run and the number of patients they saw, with capital city sites seeing larger numbers of patients and being mainly clinically led. This may introduce biases to our analyses in both directions. Finally, we were not always able to identify the date of index admission, therefore limiting our ability identify the exact intervention for each participant, which would further reduce our study power.
Despite these limitations, in being able to compare the effectiveness of different HHDSs, our linkage study had many strengths. The linkage between hospital and death records enabled us to determine long-term outcomes in our groups, which is evidence that has been sorely lacking from this population in England. We present some of the first analysis of the long-term morbidity and mortality outcomes in this socially excluded population and we also describe, in great detail, their health trajectory on admission to hospital and when subsequently admitted or at death.
A further strength of our study is the certainty with which the homeless groups were identified, as these individuals had been seen by services that provide support specifically for people with experience of homelessness. Although ‘no fixed abode’ can be entered as a patient’s address in NHS systems, the use of this coding is inconsistent and, as a result, has limitations as marker of homelessness within HES records.88
The deaths in this study for the HHDS and IMD5 groups relate to people who have had a hospital admission at centres providing a HHDS service. We were not able to identify patients accessing primary care services, which would be likely to provide an insight into deaths of patients with fewer acute health conditions, but may also include unexpected deaths that occur without prior hospitalisation. The IMD5 comparison group was older and included more women, and we stratified and adjusted our analyses to account for these differences.
The higher levels of deaths owing to alcohol-related causes and drug poisoning in the recent ONS data compared with our study may be as a result of the large number of death records in this study identified from coroners’ reports. Our study is not subject to this potential bias, but we recognise that deaths arising from those admitted under care of HHDSs may also not be representative of all deaths. We have data on deaths for people with experience of homelessness who were admitted to hospital only and, therefore, our results will exclude those people who died having never accessed health care, many of whom may be individuals who die because of external causes. Our results will likely over-represent people with chronic diseases admitted to hospital treatment, compared with the ONS analysis. All methods employed to measure the homeless population have limitations and it is difficult to envisage a method that would provide a fully generalisable population-level estimate for this excluded population. Data linkage for further services that work with homeless people is likely to offer the best opportunities for future research.
