Geographical and temporal Understanding In place of Death in England (1984-2010): analysis of trends and associated factors to improve End of Life Care (GUIDE_CARE)

Article history

The research reported in this issue of the journal was funded by the HS&DR programme or one of its proceeding programmes as project number 09/2000/58. The contractual start date was in April 2011. The final report began editorial review in May 2013 and was accepted for publication in July 2014. The authors have been wholly responsible for all data collection, analysis and interpretation, and for writing up their work. The HS&DR editors and production house have tried to ensure the accuracy of the authors’ report and would like to thank the reviewers for their constructive comments on the final report document. However, they do not accept liability for damages or losses arising from material published in this report.

Declared competing interests of authors

none

Copyright statement

© Queen’s Printer and Controller of HMSO 2014. This work was produced by Gao et al. under the terms of a commissioning contract issued by the Secretary of State for Health. This issue may be freely reproduced for the purposes of private research and study and extracts (or indeed, the full report) may be included in professional journals provided that suitable acknowledgement is made and the reproduction is not associated with any form of advertising. Applications for commercial reproduction should be addressed to: NIHR Journals Library, National Institute for Health Research, Evaluation, Trials and Studies Coordinating Centre, Alpha House, University of Southampton Science Park, Southampton SO16 7NS, UK.

Data source

The death registration data collected by the Office for National Statistics (ONS) constitute the main data source for this report. Death registrations are widely recognised and used as a reliable data source to study and monitor the PoD in society. Death registration is the legal recording of all deaths in England and Wales. By law in England, a death must be registered within 5 days of its occurrence, although this period may be extended in certain circumstances, i.e. when a death is referred to a coroner. The death must be registered by a qualified informant and this person must be one of the following: (1) a relative, usually the closest; (2) someone who was present at the death; (3) someone who is instructing the funeral director; or (4) another person may qualify as an informant (in rare circumstances). It is possible to register the death as soon as the informant has obtained either a ‘Medical Certificate of Death’ from the hospital/doctor or a Form Part B from the coroner’s office, the only exception being when an inquest is being held. In this case the registration of death will occur only once the coroner has given his/her permission. A death can be registered only in the registration district in which the death occurred.

The death registration data reviewed in this report were based on the details collected after deaths have been certified and registered. There are some situations that resulted in the registration of the death being delayed. Those deaths referred to coroners may take many months or even years to be registered, and the ONS is not notified that a death has occurred until it is registered. Therefore, the death registration data collected by the ONS can only capture the deaths that are registered in a particular period, rather than the number of deaths that actually occurred in that period.

However, the cases of delays in death registration account for only a small minority of all deaths. For example, in 2011 there were 484,367 deaths registered in England and Wales, of which 463,450 occurred in 2011, representing 95.7% of the deaths registered; the proportion of registered deaths compared with those that occurred varies in terms of the underlying CoD.

In 2011, except for the causes of death [pregnancy, childbirth and the puerperium (chapter 15); external causes of morbidity and mortality (chapter 20)],31 the proportion of deaths registered in the same year as they occurred was consistently above 90% in all other causes of death. A further breakdown shows 98.1% for neoplasms, 96.1% for the nervous system, 97.1% for the circulatory system, and 96.6% for the respiratory system. 23

The information contained in the ONS database includes details from the death certificate: PoD, the deceased’s usual residential address, gender, date of death, date of birth, country of birth, the deceased’s occupation, marital status, informant relationship, underlying CoD and any mentioned CoD. Information on derived area-level variables based on place of usual residence, for example area-level deprivation indicator, local authority (LA) and primary care organisation (PCO) were also available. Prior to 2001, the CoD was coded according to the Ninth Revision of the International Classification of Diseases (ICD-9);32 from 2001, deaths were coded using the Tenth Revision of the International Classification of Diseases (ICD-10). 31

The CoD recorded on the death certificate consists of two parts: Part I and Part II. Part I is used to show the immediate CoD and any underlying cause or causes. Part II is used for any significant condition or disease that contributed to the death but which is not part of the sequence leading directly to death. In the death registry database, an underlying CoD is accompanied by a varied number of causes of death mentioned on the death certificate. The rules of determining the underlying CoD have been changed over time, introducing challenges to the interpretation of the data particularly for the time trend analysis. The rule changes mainly concern the method of selecting the underlying CoD, which has a different impact on disease-specific mortality statistics. 33

Study sample, design and settings

This is a population-based cross-sectional study. All deaths registered between 1984 and 2010 in England were included in the data set (n = 13,264,769) but excluding the following two groups: (1) all external causes of death, and (2) all of those who died at the age of ≤ 24 years, which, in total, account for 0.83% of all deaths. The focus on non-external causes of death is driven by the fact that these are the deaths that could have potentially benefited from palliative and end-of-life care planning. Children and young people have very different end-of-life care needs, so have been excluded to maintain sample homogeneity. 34

Outcome variables

The outcome variable was the PoD. The classification of the PoD was different according to the time period, detailed as below:

1. From 1984 to 1992 There were four categories, consisting of hospital, own home, other communal establishment and elsewhere. Hospitals included UK NHS and non-NHS non-psychiatric and psychiatric hospitals, and NHS and non-NHS nursing homes. ‘Hospice’ was not classified as a separate category but was usually included as a nursing home in the hospital category. ‘Other communal establishment’ comprised non-hospital communal places (e.g. residential home, social services home). All other deaths in places not included in the listed categories were grouped into ‘elsewhere’.

2. From 1993 to 2010 There were six categories, comprising hospital, own home, hospice, care home (including residential homes with and without nursing), other communal establishment and elsewhere. The proportion of hospital deaths in this period was therefore smaller than the earlier period, as hospice and nursing homes were counted separately. Accordingly, the proportion of deaths in ‘other communal establishment’ was also smaller, as these previously were communal places excluding care home with nursing.

In both periods, we grouped psychiatric hospitals with all other hospitals, as we assume that both types of hospitals had similar facilities for end-of-life and palliative care. We did not have sufficient information to identify the deaths that happened in palliative care wards within a hospital. Therefore, a death that occurred in hospital palliative care ward was counted as the death in hospital.

Explanatory variables

Two types of explanatory variables have been used for this analysis: individual level and area level. Individual-level variables included age, gender, underlying CoD, marital status, year of death, and the location of usual residence. Area-level variables were obtained through record linkage of the usual residence at the lower layer super output area (LSOA) level or other appropriate geographic boundaries. LSOAs are a geographic boundary designed to improve the reporting of small area statistics in England and Wales. Each LSOA has an approximately similar population size, ranging from 1000 to 3000 (average 1500) and contains 400–1200 households. There are a total of 32,482 LSOAs in England. LSOA boundaries have been subject to minor changes over time. The impacts and implications of these changes on this analysis are varied depending on purposes. 35

Age was predominantly analysed as a categorised variable in five groups: 25–54, 55–64, 65–74, 75–84 and 85+ years. The grouping took into account the commonly used age cut-off values to facilitate cross-study comparisons, and also the usefulness for policy development and targeting improvement. Gender was a binary variable (female vs. male).

The underlying CoD was classified into seven classes: all CoDs, excluding external CoDs; all non-cancer CoD; cancer; CVDs, CBDs, neurological conditions and COPDs. The International Classification of Diseases (ICD) codes used to identify these causes of death are listed in Table 1.

We planned to use the ‘country of birth’ variable but, when we looked into the details of this variable, we found coding inconsistencies that made its use problematic. The main problems were:

1. No country of birth information for pre-1988 data.

2. Pre- and post-1993 had substantially different recording schemes – the pre-1993 scheme had only 100 distinct codes recording country of birth, whereas the post-1993 scheme had a much more detailed set of country of birth codes, making the grouping and comparability with the earlier period significantly compromised.

3. Post-1993 codes were also problematic. In 2006, there was a coding scheme change by the ONS, with some codes in earlier periods overlapping with the later periods. For example, 012 was ‘Austria’ in the old scheme but was defined as ‘Algeria’ in the new scheme; 090 was ‘Hong Kong’ before and then as ‘Solomon Islands’ after. This made the 2006 data unusable.

Furthermore, there was a very imbalanced distribution of the country of birth variable across different disease groups. In some subgroup analysis it was hard to make sense or maintain statistical efficiency if we were to keep this variable as an explanatory variable. Therefore, we decided to use this section of information only in a specific subgroup analysis (see Appendix 2, Conferences abstracts and presentations, item 10).

Notwithstanding the difference in pre- and post-coding schemes, the data show that those who died in the study period were predominantly born in the UK (overall 91.3%; range 91.1% to 96.3%). On average, the pre-1993 coded data produced a consistently higher proportion of UK-born deaths than the post-1993 scheme (96.1% vs. 92.7%), suggesting a systematic discrepancy between the two coding systems. However, we acknowledge that the pre- and post-1993 difference may partly reflect a genuine consequence of changing levels of immigration over the middle decades of the twentieth century. We also tested the disease-specific distribution of the country of birth using the 2007–10 data. Post-1993 data also showed that 92.8% of those who died were born in the UK.

For the aforementioned reasons and findings from the empirical data, we do not feel there was merit to include this variable in the standard set of explanatory variables in the report. After consulting and discussing with the Project Advisory Group and the project team members, we decided to separately analyse the ‘country of birth’ variable in a subanalysis, through which we investigated the PoD with a special focus on ethnicity. To allow for full-scale exploration of the ethnicity issue, we chose London – a metropolitan city with high ethnic diversity – as the primary study sample (see Appendix 2, Conferences abstracts and presentations, item 10).

Marital status was included as an individual-level variable and grouped into five classes: married, widowed, divorced, single, and not stated or unknown. Marital status was not available in the data before 1988, and before 1993 the variable was not consistently coded, therefore analysis using this variable was carried out only for post-1993 data. Year of death was directly extracted from the date of death, as recorded on the death registration record. This was therefore the year that the death occurred.

The location of usual residence was used to derive in which health authority region the deceased had lived: North East, North West, Yorkshire and the Humber, East Midlands, West Midlands, East of England, London, South East Coast, South Central and South West.

An area-level variable – the Index of Multiple Deprivation (IMD) – was used in this report as a proxy measure of socioeconomic status (SES). The IMD is a measure of multiple deprivation for small areas in England. 36 Four versions of the IMD have been developed over the study period: IMD 2000, IMD 2004, IMD 2007 and IMD 2010. The IMD 2000 was measured at ward level, whereas the other IMD versions were measured at the LSOA level. The IMD is a composite of domains related to different aspects of deprivations. The IMD 2000 was based on six domains (income; employment; health deprivation and disability; education, skills and training; housing; geographical access to services). The later versions of the IMD were based on a consistent seven domains: income; employment; health and disability; education, skill and training; barriers to housing and services; living environment; and crime. A higher score in the indices indicates a higher level of deprivation. The areas were ranked by the IMD score and split into quintile groups (1 = most deprived, 5 = least deprived). Although the original area-specific composite scores of the IMD may have changed over time, the change in the quintile ranking score was not significant. There were significant changes in IMD ranking between IMD 2000 and the later versions of IMD; however, the IMD ranking movement was minimum between IMD 2004, IMD 2007 and IMD 2010 (Table 2). Therefore, we used IMD 2000 for the 1984–2000 data and IMD 2010 for the 2001–10 data.

We derived two temporal variables for this report: seasonal period (spring, summer, autumn, winter) and holiday period (Christmas, New Year and Easter). The seasonal period was defined by month of death according to the following cut-off: spring (March to May), summer (June to August), autumn (September to November) and winter (December to February). The Christmas, New Year and Easter holiday periods were defined as 3 days before and after the specified holiday. For example, a Christmas period was 3 days before and after 25 December, totalling seven calendar days (1 week). Weeks outside the holiday periods were defined as a normal period.

The pre-1993 data did not have the ‘informant relation’ variable, whereas 10.3% of the post-1993 death records had a ‘missing’ value for this variable. The ‘informant relation’ was noted in the database as ‘loosely coded’ short text, and therefore had little value as a proxy for indicating the social support of the deceased. There was no further information in the database available to compensate this. Therefore, we decided not to use ‘informant relation’ in this report.

Statistical analysis

Before the analysis, data were checked for completeness, problems and errors. Records with missing, illogical and invalid values on aforementioned variables were deleted. These included illogical range of data values, for example age > 300 years, year of death outside 1984–2010, month of death > 12 or < 1 and missing data on PoD.

Data were described for their characteristics, mainly using frequency, percentage and 95% confidence interval (CI). Continuous data were categorised as a categorical variable but also estimated for mean, median, standard deviation (SD) and range. The data were described for all listed causes of death in Table 1. The bivariate relationship between explanatory variables and PoD was plotted using appropriate graphs and the difference across groups was investigated using Pearson chi-squared statistics. The regional variation in PoD was displayed using geographic information system (GIS). PoD was described as a four- or six-category variable, depending on the time period.

Variations in PoDs at the aggregate LSOA level were analysed and modelled using a weighted linear regression model. The dependent variable was the area-specific percentage of hospital deaths, with the total number of deaths at the selected area level as one of the adjusted variables. The proportion that can be explained by the model was derived from the model and measured by a R square. The explanatory variables were derived from the data and introduced into the models as the percentage of defined subcategory (e.g. percentage female) or the average value for continuous data (e.g. age). The models were constructed using the data for the 1993–2010 period, as there were a limited number of demographic variables for the period 1984–92, and the geographic boundaries had changed significantly from the earlier period to the later periods. The explanatory variables included average age, percentage male, percentage married, year of death, average IMD score, percentage winter deaths or percentage deaths in the Christmas period. Area was defined at the PCO and LA levels, and modelled separately.

Factors associated with PoD were investigated using binomial regression models, or modified Poisson regression models in the case of convergence problem. 37,38 The dependent variable was a binary indicator of PoD. We considered the least preferred PoD [hospital (1)] compared with the number two or number three most common places of death (0), which was determined by the CoD of interest. All selected independent variables, as informed by the area-level modelling analysis and also checks for multicollinearity using variance inflation factor, were forced to stay in the model to adjust for potential confounding effects. The top three factors most strongly and significantly associated with PoD were tested for significance of the two-way interaction effects. The interaction effect was introduced into the model as a multiplicative term (e.g. age*cancer, age*marital status*cancer) along with the main effects. The prevalence proportion ratio, referred to in this report as proportion ratio (PR), derived from the multivariable models, was used to measure the strength of the association. The PR is a more appropriate relative risk measure than the widely misused odds ratio (OR); the latter tends to overestimate effect size, particularly when the outcome event of interest is common (≥ 10%). 39

We also constructed individual-level risk assessment models for listed CoDs, to assess an individual’s risk of dying in the hospital or non-hospital settings. A non-hospital was a place other than a NHS- or non-NHS hospital. Given the changing patterns we have found in a recently published study on cancer deaths in England,18 we used the recent 2006–9 data as the training set to build the model, and its predictive performance was tested using the 2010 data (validation set).

The model construction followed similar procedures, using binomial models or Poisson regression models in case of convergence problem of the former. Individual-level geographical and temporal variation modelling analysis involved additional steps to include all of the statistically significant or practically meaningful interaction effects. We used a stepwise selection method to select variables to be included in the final risk assessment model. 40 All variables or interaction terms that stayed in the final model needed to have a p-value of < 0.05 in the type 3 score statistics test. 41 Where possible, continuous variables were introduced, as this was found to be a more efficient way of using the available information. 42–46 We did not test for collinearity, as the primary concern for the model building was the prediction of accuracy for the risk of dying in hospital compared with a non-hospital setting. 47 The model performance validation was evaluated using (1) a Wald chi-squared test and (2) the area under the receiver operating characteristic curve (AUC) and its 95% CI. The AUC is a widely used statistical measure for assessing the discriminatory capacity of a predictive model. It represents the overall accuracy of the model, independent of the cut-off value. The AUC range goes from 0.5 to 1, with 0.5 showing no discrimination and being equivalent to coin tossing; an AUC value of 0.7 indicates good discriminating ability. 48

All analysis was carried out separately for all causes, non-cancer, cancer, CVD, CBD, neurological conditions and COPD. As this study is exploratory in nature, we did not adjust p-values for multiple testing. A p-value of < 0.05 is considered to be statistically significant. Data manipulation and statistical analysis were implemented by SAS 9.3 (SAS Institute Inc., Cary, NC, USA). Graphical presentations were completed by R 3.0.0 (The R Foundation for Statistical Computing, Vienna, Austria) and SAS 9.3.

Ethics

All analyses were based on fully anonymised patient-level records; therefore, no ethical approval was required according to the Information Commissioner’s Office guidelines, ONS procedures and those of the King’s College London’s (KCL) Research Ethics Committee. We have obtained an ethical clearance letter from the KCL ethics office. Following ONS procedures, a data access agreement was signed by both parties (researchers and ONS), and copies of all required forms were provided in a formal agreement of data management and protection for the project. Individual members of the project team had data access approvals from ONS.

Characteristics of the study sample

In total, 1716 (0.037%), 1471 (0.037%) and 1535 (0.034%) records with missing, illogical and invalid values on the standard sets of variables were deleted in periods of 1984–92, 1993–2000 and 2001–10, respectively. From 1984 to 2010, there was a total of 13,154,705 deaths (Table 3). The average annual death remained relatively stable at around 500,000 before 2001 but with a declining trend in the most recent 10-year period (n = 456,770). People died at an increasingly older age, with average age at death of 75.3 years in 1984–92, 76.8 years in 1993–2000, and 77.9 years in 2001–10. The proportion of those who died at > 75 years increased from 58.2% in 1984–92 to 67.9% in 2001–10. Slightly more women than men died in the study period (52.4% vs. 47.6%). CVDs were the most common CoD before 2001 (34.5%) but were overtaken by cancer (27.8%) and became the second most common CoD (26.9%) in the most recent 10-year period. CBD deaths also showed an overall reducing trend (12.7% in 1984–92, 10.4% in 2001–10). Those who were ‘widowed’ made up the largest proportion of the deceased (43.8%). ‘Married’ people accounted for around two in five deaths (40.2% and 38.5% in 1993–2000 and 2001–10, respectively).

Deaths were clustered around the more deprived end of the population before 2001 (53.5% below the second quintile); however, the deceased were more evenly distributed across IMD quintiles between 2001 and 2010 (ranging from 17.1% to 21.3%). The regional distribution of deaths was broadly representative of the underlying population structure. 49 Deaths in winter accounted for the highest number of deaths, whereas summer had the lowest number of deaths. Christmas had higher-than-average proportions of deaths (2.3% vs. 1.9% in normal periods, given 52.1775 weeks in 1 year); the pattern remained similar throughout the study period. Hospital, home and care home were the three most common places of death; the deaths in hospital accounted for nearly three times the number of deaths at home or in care homes. From 1993 to 2010, although there was a slight increase in hospital deaths (55.0–57.3%) and a reduction in both home (20.4–19.0%) and care home deaths (18.2–17.2%), there was a nearly 25% increase in hospice deaths (4.1–5.1%).

People who died from cancer were the youngest, at 71.8 years, throughout the study period; CBDs had the oldest average age at death of 80.3 years among the selected causes of death (Tables 4–9). Although most CoDs showed a pattern of increasing age at death, neurological conditions showed a decrease from 1993 to 2010 (78.2 years in 1993–2000 to 76.3 years in 2001–10). The proportion of those who died at > 75 years varied across CoDs, with cancer having the lowest (51%) and the CBDs the highest (82%) in 2001–10. More men died from cancer, CVDs and COPD than women, whereas the deaths from non-cancer, CBDs and neurological conditions had an overall opposite gender profile. The deaths from neurological conditions showed a distinct pattern from other CoDs, with more men than women dying from this condition in 2001–10 (53.9% vs. 46.1%) and more women dying in earlier years.

Although over half of those who died from cancer were married, nearly half of patients dying from a non-cancer cause were widowed (47.9% and 48.5%); this remained the case for the other CoDs as well. Patients with COPD were more likely to live in the most deprived areas, although this had significantly reduced from 38.3% before 2001 to 28.7% after 2001. Other CoDs were more evenly distributed across the IMD quintiles. The regional distributions of deaths were following the pattern seen in the data for all of CoDs combined.

The EWDs were more pronounced in non-cancer causes, particularly in COPD: nearly one in three deaths occurred in winter, although the unequal distribution across seasons improved over time with the lowest at 19.2% and 20.1% in summer and 36.4% and 32.7% in winter for 1993–2000 and 2001–10, respectively. Deaths during Christmas were highest across the CoDs, again, COPD deaths were more likely to happen during the Christmas period (2.8–3.4%). The combined total deaths in holiday period were highest for COPD (6.7%) and lowest for cancer (5.0%). New Year had lower than average deaths irrespective of the CoD.

Place of death varied by CoD. Hospital was the most common PoD regardless of the underlying CoD. Neurological conditions had the lowest rate of hospital deaths (46.1% in 2001–10), whereas COPD deaths most commonly occurred in hospital (68.3% in 2001–10). The second most common PoD in cancer deaths was home (24.1%) but it was care home in non-cancer deaths (20.0%). In total, 27.4% of deaths from CVDs were at home but the figure for CBD deaths was only 6.7%. Neurological conditions had the highest proportion of care home deaths (35.2%); this was around 10% for the other conditions. A significant number of cancer deaths took place in hospices (17.1%). Less than 1% of deaths from the other CoDs occurred in a hospice, except for neurological conditions, which were the second largest group to most commonly die in a hospice (3.5%).

Bivariate association of place of death and selected explanatory variables

The plotted bivariate associations for all causes of death are shown in Figures 1–5. The geographical variations in PoD for all CoDs are shown in Figures 6–10. We also produced similar graphs for non-cancer deaths, cancer deaths, CVD and CBD deaths, neurological condition deaths and COPD deaths (see Appendix 1, Figures 18–41). All pairwise associations were statistically significant at the level of p < 0.001.

FIGURE 1.

Place of death by age group in all causes of death, England 1984–2010.

FIGURE 2.

Place of death by gender in all causes of death, England 1984–2010.

FIGURE 3.

Place of death by cause of death, England 1984–2010.

FIGURE 4.

Place of death by marital status in all cause of death, England 1988–2010.

FIGURE 5.

Place of death by quintile of IMD (1 = most deprived, 5 = least deprived) in all causes of death, England 1984–2010.

FIGURE 6.

Percentage of hospital (a) vs. home deaths (b) for all causes combined by geographical areas, England 1984–92.

FIGURE 7.

Percentage of hospital (a) vs. home deaths (b) for all causes combined by geographical areas, England 1993–2000.

FIGURE 8.

Percentage of hospital (a) vs. home deaths (b) for all causes combined by geographical areas, England 2001–10.

FIGURE 9.

Percentage of care home deaths for all causes, combined by geographical areas, England (a) 1993–2000; and (b) 2001–10.

FIGURE 10.

Percentage of hospice deaths for all causes, combined by geographical areas, England (a) 1993–2000; and (b) 2001–10.

Area-level regression analysis

Regional-level models on the basis of selected demographic and clinical variables together with a temporal variable (see Table 10) explained a statistically significant (p < 0.001) proportion of the variation in hospital deaths, ranging from 24.3% to 26.7% in the data with all causes combined. The models constructed at the PCO level tended to perform slightly better than those at the LA level; models incorporating holiday periods marginally outperformed those including the seasonal temporal variable. For every year increase in age, there was a 1.686% (95% CI 1.814% to 1.559%) to 1.613% (95% CI 1.829% to 1.397%) reduction in hospital deaths; every 1% increase in male proportion was associated with a 0.013% (95% CI 0.007% to 0.018%) to 0.005% (95% CI 0.000% to 0.010%) increase in deaths in hospitals; married marital status was negatively related to hospital deaths (−0.009%, 95% CI −0.012% to −0.006% for the PCO-level models; −0.012%, 95% CI −0.015% to −0.009% for the LA-level models). Year of death was positively related to proportion of hospital deaths, the more recent years’ deaths were more likely to occur in hospitals: every year was associated with 0.277–0.279% increase in hospital deaths for the PCO models with holiday and seasonal variables, respectively. The corresponding figures for the LA-level models were 0.332% (95% CI 0.300% to 0.364%) and 0.339% (95% CI 0.307% to 0.371%).

A higher level of area deprivation was related to a higher proportion of hospital death at the PCO level. Every increasingly deprived IMD quintile was associated with a 0.415% (95% CI −0.043% to 0.874%) and 0.419% (95% CI −0.039% to 0.878%) increase in hospital deaths for the models with percentage winter deaths and percentage Christmas deaths. For the LA-level models, every increasingly deprived IMD quintile was associated with a 0.041% (95% CI −0.285% to 0.203%) reduction in proportion of hospital deaths in the seasonal model but an increase of 0.037% (95% CI −0.206% to 0.280%) in the holiday period model, although none of those was statistically significant. People who died in the winter had slightly lower but statistically non-significant chance of dying in hospital (−0.001%, 95% CI −0.006% to 0.004% for PCO-level model; −0.002%, 95% CI −0.008% to 0.003% for LA-level model); the Christmas period tended to be associated with an increased chance of people dying in hospital, with the −0.007% (95% CI −0.020% to 0.007%) of associated hospital deaths in the PCO-level model and −0.012% (95% CI −0.015% to −0.009%) in the LA model; the latter was statistically significant (p < 0.001).

The disease-specific models varied widely between diseases (Tables 10 and 11). The selected variables explained the least proportion of variation in the area-level models for cancer (5.4–6.9%), and the highest proportion for CBDs (38.2–39.0%) for PCO-level models and in non-cancer (36.5–37.2%) for LA-level models. The explained variation in hospital deaths by selected explanatory variables in CVDs, neurological conditions and COPD CoD accounted for between 6.6% and 26.1% of the observed variation. Overall, the holiday models performed equally well or slightly better than seasonal period models, the exception being in cancer deaths, for which seasonal models were slightly better, although the difference was minimal (6.9% vs. 6.8% for the PCO-level models and 5.5% vs. 5.4% for the LA models). The direction of travel for the individual explanatory variables in disease-specific models was consistent with that in all causes combined models for most disease groups, although the magnitude differed.

Age was consistently associated with lower hospital deaths in disease-specific models; this remained the case for PCO- and LA-level models: a 1-year increase in age was linked with a 0.204–2.650% reduction in hospital deaths; the link was strongest in CBDs. However, COPD deaths had inconsistent age patterns for PCOs (0.012%, 95% CI −0.209% to 0.233% in the seasonal model; −0.073%, 95% CI −0.299% to 0.154% in the holiday model), although none of them were statistically significant and the associations were weak. At the LA level, age was consistently associated with lower hospital deaths in COPD (−0.402%, 95% CI −0.546% to −0.257% in the seasonal model; −0.305%, 95% CI −0.467% to −0.143% in the holiday temporal model).

Male gender was associated with lower hospital deaths for non-cancer, cancer, CBD and neurological condition CoDs (0.003–0.870, all significant at the level of p < 0.001), but not in deaths from COPD and CVDs, for which these deaths had a negative parameter estimate but only COPD’s parameters were statistically significant (−0.054%, 95% CI −0.085% to −0.022% for the seasonal model; −0.057%, 95% CI −0.088% to −0.025% for the holiday temporal model at the PCO level; and −0.057%, 95% CI −0.100% to −0.014% for the seasonal model; −0.061%, 95% CI −0.104% to −0.018% for the holiday temporal model at the LA level).

The proportion of married people who died from a non-cancer, cancer or CVD cause was associated with lower hospital deaths (−0.002% to −0.038%), whereas among deaths from CBDs, neurological conditions and COPD, marital status was linked with an increased proportion of hospital deaths (0.023–0.301%). Again, the link was strongest for deaths from neurological conditions (0.223–0.301%). Along with the year, the proportion of hospital deaths increased in all selected CoD (annual increase 0.361–0.690%), with the only exception being cancer deaths (annual decrease −0.282% to −0.202%). The hospital deaths in non-cancer and CVD causes were linked with increased area deprivation (0.601–1.280%); however, deaths in hospitals from cancer, CBD, neurological conditions and COPD were related to reduced area deprivation (−4.570% to −1.457%). The regional models showed most of the temporal trends were not significant, although the direction of effect varied by CoD. The only significant trend was that deaths over the Christmas period had a lower chance of taking place in hospital (−0.013%, 95% CI −0.017% to −0.009%) for the PCO-level model.

The visual inspection of overall fit of area-level models, influence of outliers, normality, homoscedasticity and linearity assumption showed a consistent misfit pattern with the R-squared statistics: models with a higher value on R squared (e.g. non-cancer, CBD vs. cancer and COPD) had a better visual fit. We carried out transformation of data in an attempt to improve less well-fitted models but no tangible improvement was observed. For the interests of simplicity and ease of interpretation, we presented raw data-based models only.

Factors associated with place of death: multivariable modelling

In this section, we report the multivariate adjusted association between factors and PoD. The measure of effect is expressed as the PR, a more conservative but more appropriate measure of association than OR when the outcome of interest is common (> 10%). It is useful to bear this in mind when interpreting these results, as the effect presented by PR tends to be smaller than that of OR. Bivariate and area-level modelling analyses indicate that the factors associated with PoD are very different by cause of death, therefore we ran the multivariable regression modelling analysis separately for all causes, non-cancer causes, cancer, CVDs, CBDs, neurological conditions and COPDs.

Risk assessment

The results for the Wald tests, and the AUCs and their respective 95% CIs for PoD were as follows: all causes of deaths 0.553 [95% CI 0.551 to 0.554; p < 0.001, χ² = 3581.4, degrees of freedom (df) = 1], non-cancer deaths 0.592 (95% CI 0.590 to 0.594; p < 0.001, χ² = 7579.1, df = 1), cancer deaths 0.554 (95% CI 0.551 to 0.557; p < 0.001, χ² = 1098.1, df = 1), CVD deaths 0.576 (95% CI 0.573 to 0.580; p < 0.001, χ² = 1815.3, df = 1), CBD deaths 0.637 (95% CI 0.632 to 0.643; p < 0.001, χ² = 2344.0, df = 1), deaths from neurological conditions 0.599 (95% CI 0.586 to 0.612; p < 0.001, χ² = 227.1, df = 1) and deaths from COPD 0.552 (95% CI 0.551 to 0.554; p < 0.001, χ² = 3344.5, df = 1) (Figures 11–17).

FIGURE 11.

Receiver operating characteristic curve of risk assessment model for PoD in all-cause deaths, AUC = 0.5527.

FIGURE 12.

Receiver operating characteristic curve of risk assessment model for PoD in non-cancer, AUC = 0.5919.

FIGURE 13.

Receiver operating characteristic curve of risk assessment model for PoD in cancer, AUC = 0.5540.

FIGURE 14.

Receiver operating characteristic curve of risk assessment model for PoD in CVDs, AUC = 0.5763.

FIGURE 15.

Receiver operating characteristic curve of risk assessment model for PoD in CBDs, AUC = 0.6371.

FIGURE 16.

Receiver operating characteristic curve of risk assessment model for PoD in neurological conditions, AUC = 0.5988.

FIGURE 17.

Receiver operating characteristic curve of risk assessment model for PoD in COPD, AUC = 0.5522.

Recommendations for future research

Our work has identified the following priority research questions:

1. What is the pattern of health service utilisation in the last year of life and how is it related to where patients die?

2. How are clinical characteristics in the last year of life related to where patients die?

3. Is the distribution of end-of-life care facilities related to where patients die?

This report researched in detail the temporal and geographical patterns in PoD in several important causes of death, and found varied levels of inequality in PoD. However, with the cross-sectional death registry data we cannot know what happened before the final PoD, the transition of place of care and how these transitions related to PoD. We also believe it would be helpful to explore the distribution of end-of-life health-care facilities and its relationship with PoD.

We found a consistent inequality pattern in age, marital status and deprivation. The patients who died at older age (e.g. > 85 years); were single, divorced and widowed; and lived in deprived areas were more likely to die in hospital. The inequality in PoD is also evident in CoD, for example deaths from haematological and lung cancers were more likely to have occurred in hospitals.

We believe that enhanced community care and primary care facilities would be helpful in reducing the inequalities in PoD. Currently, our end-of-life health-care facilities are not optimised to enable people to die at their preferred place, which usually is at home or in a hospice. This research is important, as it can deepen our understanding of how we can make better use of existing resources to deliver quality care at the end of life; it can also help us to prioritise and plan for resource allocation.

There was a scarcity of UK data evaluating the relationship between end-of-life care patterns in the last year of life and PoD, although limited information on the topic is available from the USA, Canada and some Western European countries. We found two UK-based studies that assessed the service use in the last year of life and its relationship with PoD; however, both studies were conducted in specific settings (e.g. hospice) or subpopulations (e.g. aged > 75 years).

To the best of our knowledge, there are no data from wider end-of-life care settings or populations in the UK that address the proposed research questions. We believe that linked health-care data facilities can supply information to answer these important research questions and make improvements for end-of-life care policy.

Contribution of authors

Dr Wei Gao (Senior Lecturer in Statistics and Epidemiology, KCL) Co-leader of the project and project manager; undertook the analysis, and wrote the first draft of the report.

Mr Yuen K Ho (Research Assistant, KCL) Managed and maintained the database, prepared and linked the data, produced the bivariate graphs and GIS graphs. Assisted and supervised by WG and IJH.

Dr Julia Verne [Director for Knowledge & Intelligence (South West) and Clinical Lead – National End of Life Care Intelligence Network] Contributed to all phases of the project, and provided critical comments on the draft of the report and interpretation of the data.

Dr Emma Gordon (Head of Life Events Analysis, ONS) Contributed to all phases of the project, and provided critical comments on the draft of the report and interpretation of the data.

Professor Irene J Higginson (Professor of Palliative Care, KCL) Co-leader of the project; provided senior support for project management and significantly contributed to writing the first draft of the report;

All authors approved the final version.

The investigators of the GUIDE_Care project: Irene J Higginson (Principal Investigator), Wei Gao, Julia Verne, Myer Glickman and Barbara Gomes.

The Members of the Project Advisory Group (PAG): Tony Bonser, Shaheen Khan, Jonathan Koffman, Katie Lindsey, Roberta Lovick, Tariq Malik, Julie Messer, Carolyn Morris, Andy Pring, Stafford Scholes and Katherine Sleeman.

We thank the ONS for supplying data; ONS staff Claudia Wells and Vanessa Fearn for their advice and support in the preparation of the data for analysis, and for critically commenting on data-related issues.

We thank Miss Joanna Davies for proofreading the report and administrative support.

Patient and public involvement

Service users were involved through an existing group – the COMPASS Collaborative Consumer Group – with whom we have developed a close working relationship through other projects. Patients and service users have been involved in the project in the following ways: (1) provided lay feedback and comments; (2) monitored progress of the research as a member of the PAG; (3) helped to develop and implement dissemination plans; and (4) contributed/commented on reports/publications/findings establish links with wider (i.e. non-cancer) user groups.

Disclaimers

This report presents independent research funded by the National Institute for Health Research (NIHR). The views and opinions expressed by authors in this publication are those of the authors and do not necessarily reflect those of the NHS, the NIHR, NETSCC, the HS&DR programme or the Department of Health. If there are verbatim quotations included in this publication the views and opinions expressed by the interviewees are those of the interviewees and do not necessarily reflect those of the authors, those of the NHS, the NIHR, NETSCC, the HS&DR programme or the Department of Health.

Publications

Gao W, Ho YK, Verne J, Glickman M, Higginson IJ; project GUIDE_Care. Changing patterns in place of cancer death in England: a population-based study. PLOS Med 2013;10:e1001410. doi:10.1371/journal.pmed.1001410

Sleeman KE, Ho YK, Verne J, Glickman M, Silber E, Gao W, et al. Place of death, and its relation with underlying cause of death, in Parkinson’s disease, motor neurone disease, and multiple sclerosis: a population-based study. Palliat Med 2013;27:840–6. doi:10.1177/0269216313490436

Sleeman KE, Ho YK, Verne J, Gao W, Higginson IJ; GUIDE_Care project. Reversal of English trend towards hospital death in dementia. A population-based study of place of death and associated individual and regional factors, 2001–2010. BMC Neurol 2014;14:59.

Koffman J, Ho YK, Davies JM, Gao W, Higginson IJ. Does ethnicity affect where people with cancer die? A population based 10 year study [published online ahead of print. PLOS One 2014 April. doi:10.1371/journal.pone.0095052.

Evans CJ, Ho YK, Daveson BA, Hall S, Higginson IJ, and Gao W on behalf of the GUIDE_Care project. Place and cause of death of centenarians dying in care homes: a population-based observational study in England. PLoS Med 2014;11:6. e1001653.

Peer-reviewed journal articles

Gao W, Ho YK, Verne J, Glickman M, Higginson IJ. Changing patterns in place of cancer death in England: a population-based study. PLOS Med 2013;10:e1001410. doi: 10.1371/journal

Sleeman KE, Ho YK, Verne J, Glickman M, Silber E, Gao W, et al. , on behalf of the GUIDE_Care Project. Place of death, and its relation with underlying cause of death, in Parkinson’s disease, motor neurone disease, and multiple sclerosis: a population-based study. Palliat Med 2013;27:840–6.

Sleeman KE, Ho YK, Verne J, Gao W, Higginson IJ, on behalf of the GUIDE_Care project. Reversal of English trend towards hospital death in dementia. A population-based study of place of death and associated individual and regional factors, 2001–2010. BMC Neurol 2014;14:59.

Koffman J, Ho YK, Davies JM, Gao W, Higginson IJ. Does ethnicity affect where people with cancer die? A population based 10 year study. PLOS One 2014;9:e95052. doi:10.1371/journal.pone.0095052

Evans CJ, Ho YK, Daveson BA, Hall S, Higginson IJ, and Gao W on behalf of the GUIDE_Care project. Hidden needs of centenarians dying in care homes: a population-based observational study in England. PLOS Med 2014;3:11. doi:10.1371/journal.pmed

Conference abstracts and presentations

Sleeman KE, Davies JM, Ho YK, Verne J, Gao W, Higginson IJ. The changing demographics of inpatient hospice deaths. Population based study in England 1993–2010. European Association for Palliative Care, Lleida, Spain, 4–7 June 2014.

Sleeman KE, Ho YK, Verne J, Gao W, Higginson IJ. Reversal of English trend towards hospital death in dementia: a population-based study of place of death and associated individual and regional factors, 2001–2010. European Association for Palliative Care, Lleida, Spain, 4–7 June 2014.

Koffma, J, Ho YK, Davies JM, Gao W. Higginson IJ. Does ethnicity affect where people with cancer die? A population-based 10 year study. HERON Conference, London, 14 May 2014.

Gao W, Verne J, Glickman M, Higginson IJ. Place of death is associated with holiday periods: implication for end of life care. 13th World Congress of the European Association for Palliative Care, Prague, CZ, 30 May 2013. URL: www.congressinfo.org/filerun/weblinks/?id=e94550c93cd70fe748e6982b3439ad3b&filename=EAPC-Abstract-Book_FINAL%20Version_small.pdf (accessed 27 October 2014).

Evans CJ, Ho YK, Daveson B, Hall S, Higginson IJH, Gao W. ‘Prolonged dwindling’; an over simplification of dying for centenarians? 13th World Congress of the European Association for Palliative Care, Prague, CZ, 30 May 2013. URL: www.congressinfo.org/filerun/weblinks/?id=e94550c93cd70fe748e6982b3439ad3b&filename=EAPC-Abstract-Book_FINAL%20Version_small.pdf (accessed 27 October 2014).

Gao W. Changing patterns in place of cancer deaths in England, 2001–2010: Time trends and associated factors. Society for Social Medicine 56th Annual Scientific Meeting, London, UK, 12 September 2012.

Gao W, Ho YK, Verne J, Glickman M, Higginson IJ. Place of cancer deaths in England, 2001–2010: Time trends and determinants. Society for Social Medicine 56th Annual Scientific Meeting; 2012 Sep 12; London, UK. J Epidemiol Community Health 2012;66(Suppl. I):A27.

Ho YK, Gao W, Verne J, Glickman M, Higginson IJ. Where do cancer patients die? An 18-year trend in England. 2012 NCRI Cancer Conference; Liverpool, UK, 5 November 2012. URL: http://conference.ncri.org.uk/abstracts/2012/abstracts/A170.html (accessed 27 October 2014).

Gao W, Ho YK, Verne J, Glickman M, Higginson IJ. Place of death in tuberculosis: an analysis from Death Registry in England, 2001–2010. 6th Conference of the Union Europe Region: International Union Against Tuberculosis and Lung, London, UK, 4 July 2012.

Sleeman K, Ho YK, Verne J, Gao W, Higginson IJ. Sociodemographic determinants of place of death in dementia: whole population cross-sectional analysis in England, 2001–10. Spring Meeting for Clinician Scientists in Training, 2013 Feb 27; London, UK. Lancet 2012;381:S101.

Sleeman K, Ho YK, Verne J, Gao W, Higginson IJ. Where do people with dementia die? Population-based study of ten year trends, and associated individual and regional factors. Spring Meeting for Clinician Scientists in Training, London, UK, 27 February 2013.

Sleeman K, Ho YK, Verne J, Glickman M, Silber E, Gao W, et al. Trends in place of death, and the effect of death certificate classification and coding changes, in Parkinson’s disease, motor neurone disease, and multiple sclerosis in England: 1993–2010. J Neurol Neurosurg Psychiatry 2012;83(Suppl. 2):A17.

Sleeman K, Ho YK, Verne J, Glickman M, Silber E, Gao W, et al. Trends in place of death and death certificate coding for Parkinson’s disease, motor neurone disease and multiple sclerosis in England 1993–2010. Poster session presented at King’s College London, London, UK, 10 May 12.

Sleeman K, Ho YK, Verne J, Glickman M, Silber E, Gao W, et al. Place of death and death certificate coding for Parkinson’s disease, motor neurone disease and multiple sclerosis in England 1993–2010. Poster session presented at conference of the Association of British Neurologists, Brighton, UK, 29 May 2012.

Press releases

National Institute for Health Research & EurekAlert. Hospital Remains Most Common Place of Death for Cancer Patients in England. 2013. URL: www.netscc.ac.uk/hsdr/news27032013.html, www.eurekalert.org/pub_releases/2013-03/plos-hrm032613.php (accessed 21 October 2014).

Cicely Saunders Institute. Home and Hospice Deaths Increasing Since the Launch of National Programme. 2013. URL: www.csi.kcl.ac.uk/home-and-hospice-deaths-increasing-since-the-launch-of-national-programme.html (accessed 21 October 2014).

BBC News, Health. Centenarians ‘Outliving Diseases of Old Age’. 2014. URL: www.bbc.co.uk/news/health-27682376 (accessed 18 September 2014).