Sage Journals: Discover world-class research

Abstract

Background

There is increasing evidence that access to critical care services is not equitable. We aimed to investigate whether location of residence in Scotland impacts on the risk of admission to an Intensive Care Unit and on outcomes.

Methods

This was a population-based Bayesian spatial analysis of adult patients admitted to Intensive Care Units in Scotland between January 2011 and December 2015. We used a Besag–York–Mollié model that allows us to make direct probabilistic comparisons between areas regarding risk of admission to Intensive Care Units and on outcomes.

Results

A total of 17,596 patients were included. The five-year age- and sex-standardised admission rate was 352 per 100,000 residents. There was a cluster of Council Areas in the North-East of the country which had lower adjusted admission rates than the Scottish average. Midlothian, in South East Scotland had higher spatially adjusted admission rates than the Scottish average. There was no evidence of geographical variation in mortality.

Conclusion

Access to critical care services in Scotland varies with location of residence. Possible reasons include differential co-morbidity burden, service provision and access to critical care services. In contrast, the probability of surviving an Intensive Care Unit admission, if admitted, does not show geographical variation.

Keywords

Health equity healthcare disparities geography systems of care Bayesian disease modelling

Introduction

Equity of access is a central tenet of a modern healthcare system.^1,2 Equity requires conditions whereby those with equal needs have equal opportunities to access healthcare. Intensive Care Units (ICUs) provide the setting to care for critically ill patients, requiring extensive resources. These cannot be provided in all locations and therefore patients who develop critical illness in remote areas often have to be transferred to regional centres.

There is a recognised variation in the provision of critical care services internationally³ and a growing interest in the subject of geographical equity within national healthcare systems.⁴ A recent study from Denmark concluded that, whilst there was only minimal geographical variation in Intensive Care admissions and interventions at the regional level, there was pronounced variation at the municipality level.⁵ A further study, from Scotland, specifically of elderly patients, has shown regional variations in admission rates.⁶ However, neither of these studies employed methods that allow direct probabilistic comparison of the areal units used in their study.

Scotland has large remote and rural areas, including many islands. Accessing critical care services from such locations requires the referring doctor to identify the need for ICU admission, the receiving intensive care physician to accept the patient and a retrieval service to transfer the patient. The aim of this project was to investigate whether location of residence in Scotland impacts on the probability of admission to an ICU and on outcome.

Comparing rates from geographical areas is methodologically complex,^7,8 but recent developments in statistical methods and computational power have paved the way for more comprehensive, spatially adjusted models which provide better insights into geographical variations in service utilisation patterns.

We hypothesise that location of residence has no impact on the risk of admission to an ICU in Scotland. Additionally, we also hypothesise that geographical location of residence will not influence outcomes, if admitted to an ICU.

Methods and materials

Study design

This was a population-based Bayesian spatial analysis of patients who required admission to an ICU in Scotland, between January 2011 and December 2015.

Setting

Scotland has a land area of 78,770 km², including 95 inhabited islands, and in 2011 had a population of 5,299,900.⁹ There are 24 ICUs, which are all part of the National Health Service (NHS) in Scotland,¹⁰ and all contribute data to a continuous national audit, administered by the Scottish Intensive Care Society Audit Group (SICSAG). The SICSAG database has been used in multiple previous investigations,^6,11,12 and is routinely audited at the local level via local audit coordinators to ensure accuracy and completeness.^13,14

Data sources

All patient-related data were obtained from SICSAG. The geographical analysis was based on Council Areas, which are geographic and administrative subdivisions of Scotland, overseen by local elected governments, similar in nature to American and English Counties. Geographic and demographic data on Council Areas are freely available from the Scottish Government.¹⁵

Permissions

The study was approved by SICSAG and the Public Benefit and Privacy Panel for Health & Social Care (ref. 1617-0131).¹⁶ All data were accessed and analysed using the North of Scotland “Safe Haven” (ref DaSH-179), provided by the University of Aberdeen, which is a secure remote server that hosts data, minimising the risk of breaches of patient confidentiality.

Participants

We included patients admitted to ICU with a diagnosis other than traumatic injury (the provision of critical care for trauma patients differs), aged 18 years or older, who required tracheal intubation and ventilatory support on the day of admission and whose main residence was in Scotland. We limited the study to patients who required ventilatory support on admission in order to select patients with “level 3” critical illness and exclude patients with “level 2” (high dependency unit, rather than ICU) needs. “Level 3” patients represent a relatively homogeneous group of critically unwell patients that could be compared across multiple sites. This method also eliminated the issue of up-transfer of patients from level 2 to level 3 care within a hospital, which could have impacted on the results. We extracted demographic details, home post codes, admission diagnoses, APACHE II score, Acute Physiology Score (APS) and probability of survival, and hospital outcome (death or survival).

Statistical methods

Data were analysed by Council Area. There are 32 such areas in Scotland, which vary in size, population and demographics, seen in Table 1. We used spatial data in the form of “shapefiles” provided by the Scottish Government and the “sp” package of the statistical programme R¹⁷ to produce aggregate data on the number of patients from each Council Area. Standardised mortality ratios were calculated for each Council Area by dividing the number of observed deaths by the number of expected deaths.

Table 1.

Demographic and spatial details of the 32 Council Areas in Scotland.

Council Area	Population (2011)	Area (km²)
Aberdeen City	222,460	205.6
Aberdeenshire	253,650	6338.8
Angus	116,200	2204.2
Argyll and Bute	88,930	7163.8
City of Edinburgh	477,940	273.0
Clackmannanshire	51,500	163.9
Dumfries and Galloway	151,410	6676.0
Dundee City	147,200	62.2
East Ayrshire	122,690	1270.3
East Dunbartonshire	105,000	174.5
East Lothian	99,920	700.9
East Renfrewshire	90,810	174.2
Falkirk	156,250	314.9
Fife	365,300	1373.9
Glasgow City	593,060	176.4
Highland	232,730	26484.4
Inverclyde	81,220	173.6
Midlothian	83,450	355.3
Moray	93,470	2257.1
Na h-Eileanan an Siar	27,690	3268.6
North Ayrshire	138,090	903.9
North Lanarkshire	337,720	472.2
Orkney Islands	21,420	1086.2
Perth and Kinross	146,850	5418.9
Renfrewshire	174,700	269.2
Scottish Borders	113,880	4742.6
Shetland Islands	23,240	1656.6
South Ayrshire	112,980	1234.7
South Lanarkshire	313,900	1774.0
Stirling	90,330	2254.8
West Dunbartonshire	90,610	182.8
West Lothian	175,300	431.6

Statistical comparison of one or more geographical areas is complicated by spatial autocorrelation (dependence between areas) and the potential for unstable summary data,⁷ where areas with low predicted or observed cases result in extreme and unreliable estimates of parameters. Bayesian hierarchical models overcome these issues through “spatial smoothing”, taking prior information from neighbouring areal units to mitigate the effects of instability and spatial dependence.¹⁸

All data cleaning, analysis and cartography was performed using R.¹⁷ A description of our model and prior specifications¹⁹ can be found in the online supporting information (Supplementary Digital Content – Document 1).

The objective of the first analysis was to estimate the risk of being admitted to an ICU in Scotland, for residents of a given Council Area, relative to that expected if all Council Areas had the same admission rate as the national Scottish average. Given the demographic differences between Council Areas, we used age- and sex-adjusted five-year ICU admission rates. We performed direct age–sex standardisation using the European Standard Population (2013) as our reference population. This produced an age–sex-adjusted expected Scotland wide ICU admission rate. This Scotland wide ICU admission rate was subsequently applied to each individual Council Area to produce estimated counts of expected ICU admissions.²⁰

We used a Besag–York–Mollié (BYM) model,²¹ which is recommended for hierarchical disease mapping studies,⁸ and employs a conditional autocorrelation (CAR) prior distribution for the spatially structured random effect.²² This is based on a pre-defined neighbourhood structure, which allows the model to smooth local rates relative to neighbouring area values. We chose a neighbourhood structure that assumed areas to be neighbouring if any part of their border touches another area.²³ The response variable in this model was the number of ICU admissions per Council Area (count), thus necessitating the use of a Poisson distribution, with an offset provided by the expected number of admissions to an ICU for each Council Area.

The same method was employed to estimate the risk of mortality, if admitted to an ICU. The APACHE II score²⁴ was used to calculate the number of expected deaths of patients admitted to an ICU by Council Area of residence. The APACHE II score is a commonly used prognostic scoring tool, combining acute physiological parameters, age and chronic health issues. Patients with a predicted mortality of greater than 50% were classified as “expected deaths”.^24,25 The APACHE II model has been extensively validated.¹³

The classical BYM model we used is defined below:

γ_{i} \sim poisson (E_{i} θ i)

log (γ_{i}) = a + μ_{i} + v_{i}

where γ_i is the observed number of ICU admissions, with E_i being the estimated number of ICU admissions per area and θ_i is the relative risk of ICU admission. α is the intercept term representing the mean risk of admission across Scotland. ν_i is the Council Area specific random effect term that represents the heterogeneity of the relative risk and µ_i is the spatially structured spatial random effect term. The same model was used for the risk of ICU mortality.

We used Integrated Nested Laplace Approximation (INLA), via the R package R-INLA,²⁶ which provides a computationally efficient and sufficiently accurate way to sample posterior means to produce Bayesian output. All risk estimates are presented with 95% credible intervals (CrI).

While the model used is not able to produce a summary statistic to quantify the variation in effect size, the “exceedance risk”, as described by Kang et al.¹⁸ allows us to identify Council Areas in which there is a greater than 95% chance that the effect observed differs from the Scottish mean.

A recognised limitation of Bayesian hierarchical models with spatially smoothed conditional autoregressive priors is that they are not able to identify the spatial extent of clusters.¹⁸ Our disease mapping method used aggregates of admissions and mortality per unit area. The areal unit used therefore influences the model output and data interpretation.²⁷ Our model involves a trade-off between granularity, statistical validity and clinically interpretable outputs. We believe that our unit area – Local Council Areas – is appropriate due to their mid-size, freely available demographic data and reproducibility throughout Scotland and the United Kingdom.

The APS is the sum of the values applied to the 12 physiological parameters used to produce the APACHE II score.²⁴ We have used this score to allow direct and objective comparisons of acute disease severity between patients.

Results

There were 34,756 patients admitted to ICUs in Scotland between January 2011 and December 2015. Of these, 18,213 patients were intubated and ventilated on the first day of admission, indicating “Level 3” critical illness; 617 patients had no home address recorded, or lived outside Scotland and were therefore excluded, leaving a study population of 17,596. The baseline characteristics of these patients are shown in Table 2. The median age at admission was 61 years (interquartile range 44–69 years); 43.2% of patients were female. The median APS was 18 (interquartile range 13–24), and the median probability of mortality was 26% (interquartile range 10%–50%). And 76% of patients survived to hospital discharge.

Table 2.

Baseline characteristics of patients included in the analysis.

Baseline characteristics
Total study population	17,596
Hospital mortality, n (%)	4363 (24.0)
Male, n (%)	9990 (56.8)
ICU length of stay (days), median (IQR)	2.09 (1.21–6.02)
Age, years, median (IQR)	61 (44–69)
Acute Physiology Score, median (IQR)	18 (13–24)
Percentage probability of mortality median (IQR)	26.0 (9.9–49.7)

ICU: Intensive Care Unit; IQR: interquartile range.

Effect of geography on the probability of admission to an ICU

The crude Scottish five-year admission rate (number of ICU admissions per population) was 332 per 100,000 residents, and the five-year age- and sex-standardised admission rate was 352 per 100,000 residents. Table 3 shows the number of admissions, by Council Area, as well as the mean APS of each Council Area, which ranged from 16.9 (Shetland) to 20.1 (Dundee City). Interquartile range of the APS by area was 18.4–19.2. Table 3 also shows area-specific standardised admission rates, observed counts, expected counts and the relative risk of ICU admission, with 95% CrI. The ordered spread of admission risk, with 95% CrI are seen in Figure 1(a). The geographical relationships are demonstrated in Figure 2(a), which shows a model-derived risk map of admission to an ICU after spatial dependence has been accounted for. The Five Council Areas with the lowest admission risk (lightest green) are all located in the North East of Scotland. The five-year age- and sex-adjusted admission rates for each of the 32 Council Areas is shown graphically in Figure 3, with the size of the dot proportional to the population of each Council Area.

Figure 1.

Council Area spatially adjusted risk of ICU admission (a) and in-hospital mortality (b) ordered by effect size. If 95% credible interval error bars do not cross one, the effect is statistically important.

Figure 2.

Model derived map of spatially adjusted risk of ICU admission (a) and in-hospital mortality (b).

Figure 3.

Crude 5yr age–sex-adjusted ICU admission rates by Council Area. Size of dot on dot plot is proportional to population size of Council Area.

Table 3.

Area specific observed and expected ICU admissions with disease severity, admissions per population and risk of admission with 95% credible intervals.

Council Area	Observed admissions	Expected admissions	Disease severity (APS)	Admissions per population	Risk of admission	Lower CrI	Upper CrI
Aberdeen City	500	782	18.6	0.00225	0.702	0.634	0.774
Aberdeenshire	459	892	18.9	0.00181	0.578	0.521	0.637
Angus	248	409	18.8	0.00213	0.677	0.595	0.766
Argyll and Bute	317	313	19.8	0.00356	1.100	0.978	1.232
City of Edinburgh	1811	1681	17.3	0.00379	1.183	1.102	1.267
Clackmannanshire	188	181	18.7	0.00365	1.115	0.964	1.279
Dumfries and Galloway	615	532	18.0	0.00406	1.261	1.148	1.380
Dundee City	389	518	20.1	0.00264	0.823	0.737	0.915
East Ayrshire	425	431	18.6	0.00346	1.083	0.974	1.199
East Dunbartonshire	258	369	18.6	0.00246	0.793	0.698	0.894
East Lothian	350	351	18.7	0.00350	1.102	0.983	1.228
East Renfrewshire	270	319	19.2	0.00297	0.946	0.835	1.064
Falkirk	546	549	19.2	0.00349	1.089	0.988	1.196
Fife	1174	1285	18.2	0.00321	1.002	0.927	1.081
Glasgow City	2337	2086	18.5	0.00394	1.227	1.147	1.310
Highland	858	819	17.8	0.00369	1.137	1.045	1.236
Inverclyde	264	286	19.8	0.00325	1.022	0.901	1.152
Midlothian	364	294	18.5	0.00436	1.349	1.205	1.503
Moray	205	328	18.8	0.00219	0.697	0.605	0.796
Na h-Eileanan an Siar	106	97	19.5	0.00383	1.160	0.962	1.381
North Ayrshire	495	486	17.8	0.00358	1.117	1.010	1.230
North Lanarkshire	1065	1188	18.8	0.00315	0.986	0.911	1.065
Orkney Islands	83	75	18.5	0.00387	1.164	0.947	1.409
Perth and Kinross	336	516	19.9	0.00229	0.731	0.652	0.815
Renfrewshire	630	615	19.6	0.00361	1.123	1.024	1.229
Scottish Borders	446	401	19.9	0.00392	1.219	1.099	1.347
Shetland Islands	62	82	16.9	0.00267	0.879	0.702	1.077
South Ayrshire	390	397	18.1	0.00345	1.081	0.968	1.200
South Lanarkshire	1077	1104	18.6	0.00343	1.072	0.991	1.158
Stirling	269	317	18.0	0.00298	0.939	0.829	1.056
West Dunbartonshire	374	319	18.4	0.00413	1.264	1.131	1.407
West Lothian	685	617	18.4	0.00391	1.215	1.110	1.326

The likelihood of Council Areas having a higher risk of admission to an ICU than the Scottish mean is shown graphically in Figure 4(a). The darkest green areas have the highest likelihood of exceeding the Scottish average admission rate of between 99% and 100%, with the lightest green areas have a probability of 0% to 80%.

Figure 4.

Model derived map showing the probability of the effect size of the Council Area exceeding that of the Scottish mean for ICU admission (a) and in-hospital mortality (b).

Effect of geography on mortality

Of the 17,596 patients included in our study, 4363 died while in hospital, giving a hospital mortality rate of 24%. The risk of hospital mortality in patients admitted to an ICU did not vary by Council Area. The relative risk of mortality, with CrI, is shown in Table 4. The values ranged from 0.94 (Dundee City) to 1.05 (in Dumfries and Galloway), with the ordered spread of values found in Figure 1(b). Figure 2(b) shows the geographic relationships of the Council Areas in a model derived map. The probability of Council Areas effect size exceeding that of the Scottish mean is seen in Figure 4(b). Two Council Areas (Dumfries and Galloway and City of Edinburgh) have a probability of their difference exceeding the Scottish mean of between 80% and 99%.

Table 4.

Areas specific observed and expected ICU deaths with risk of mortality and associated 95% credible intervals.

Council Area	Observed deaths	Expected deaths	Probability of mortality	Lower CrI	Upper CrI
Aberdeen City	142	142	0.971	0.873	1.064
Aberdeenshire	142	129	0.995	0.922	1.075
Angus	64	60	0.980	0.890	1.068
Argyll and Bute	94	97	0.982	0.903	1.052
City of Edinburgh	429	363	1.041	0.976	1.121
Clackmannanshire	48	52	0.979	0.891	1.056
Dumfries and Galloway	159	121	1.049	0.974	1.157
Dundee City	110	129	0.940	0.834	1.022
East Ayrshire	111	109	1.004	0.927	1.082
East Dunbartonshire	73	58	1.016	0.941	1.111
East Lothian	104	90	1.023	0.941	1.120
East Renfrewshire	77	73	1.004	0.927	1.083
Falkirk	148	145	0.987	0.910	1.058
Fife	310	285	0.994	0.922	1.069
Glasgow City	603	559	1.001	0.946	1.058
Highland	194	161	1.020	0.949	1.112
Inverclyde	80	82	0.988	0.905	1.064
Midlothian	103	95	1.014	0.930	1.106
Moray	67	53	1.018	0.928	1.137
Na h-Eileanan an Siar	29	27	1.001	0.913	1.094
North Ayrshire	109	94	1.016	0.944	1.102
North Lanarkshire	317	270	1.024	0.964	1.102
Orkney Islands	27	19	1.014	0.929	1.120
Perth and Kinross	109	109	0.978	0.904	1.046
Renfrewshire	182	176	0.993	0.921	1.062
Scottish Borders	125	134	0.990	0.906	1.061
Shetland Islands	11	12	0.997	0.905	1.090
South Ayrshire	111	89	1.037	0.957	1.145
South Lanarkshire	290	262	1.013	0.951	1.081
Stirling	59	68	0.978	0.895	1.046
West Dunbartonshire	97	85	1.004	0.934	1.084
West Lothian	175	172	0.994	0.920	1.063

Discussion

The recently published “National Clinical Strategy for Scotland”⁴ has emphasised the importance of equity of access to healthcare services. We hypothesised that location of residence would not affect the probability of admission to an ICU or outcomes following such an admission.

Our study has identified a potential disparity in access to critical care services in Scotland, with the risk of admission to an ICU in Scotland varying according to location of residence. However, for those patients who are admitted to an ICU, the severity-adjusted probability of survival to discharge from hospital is similar, regardless of location of residence.

Whilst these differences appear to be clinically important, the interpretation of such data must be approached with caution. Disease mapping studies are designed to identify areas of potential disparity at the population level – not to identify causes of such disparities.²⁸ The spatially adjusted risk of admission to an ICU varies along a continuum by Council Area, although there is a cluster of Council Areas in the North East of Scotland that have lower admission rates than the Scottish mean, as derived from our model.

Patients from Aberdeenshire have around half the chance of admission to an ICU, as compared with the Scottish population as a whole. At the other end of the spectrum, patients from Midlothian have a one-third greater risk of admission to an ICU, compared to the rest of Scotland. Reassuringly, we have shown that the APS of patients admitted to ICUs in Scotland is similar and has no clinically important spatial variation – supporting evidence that Scottish ICUs have similar access thresholds for patients.¹⁰

We recognise that we have undertaken a hypothesis generating study, and therefore explaining such effects seen in our data is not possible without additional, focused investigation. Future analyses of value may include a direct comparison of the spatial areas at either end of our admission rate spectrum with a focus on reason for ICU admission and demographics. Additionally, identifying the geographical pattern of true critical illness in Scotland, regardless of whether a patient is admitted to an ICU or not would provide valuable information. With this data it may be possible to investigate whether there is a cohort of patients that might benefit from ICU who are not accessing this service.

We believe our results are important for two key reasons. At the national level, we have added to the evidence base that geographic difference in ICU utilisation exist as shown by Cooke in the USA²⁹ and Vestergaard in Denmark.³⁰ Additionally at the national level, we have identified areas in which the risk of ICU admission is lower than the national average. Our data provide a stimulus to undertake additional investigation into quantifying the reasons behind these results, potentially at a local level.

The effects we have observed could be artefactual, representing demographic characteristics that were not accounted for by the age- and sex-adjustment, or differing admission practices (such as tertiary referrals). However, previous studies have also shown variation in admissions, for example Docherty et al. had previously shown age-related and possible regional variation in admission rates and intensity of treatment among the elderly population in Scotland.^6,11

A limitation of our study is the use of Council Areas as the areal unit. These areas vary in size, but can be very large, and thus could hide detail – an issue sometimes referred to as the modifiable areal unit problem.³¹ We would have preferred to use a smaller areal unit, which would have provided greater geographical granularity. However, the next available geography with georeferenced demographic data is the “intermediate datazone”, of which there are more than a thousand, resulting in unstable event rates and zero inflation issues.

The strength of this study is the use of Bayesian disease mapping methodology, which allows direct probabilistic statements to be made regarding the effect size of one areal unit compared to another. Bayesian disease mapping is an important means of identifying areas of disparity regarding disease and healthcare utilisation.

All health systems face challenges regarding equity of access, across the healthcare spectrum, and particularly so in remote and rural areas.³² Establishing and identifying the cause of such disparities in access to healthcare services is not easy – particularly within time sensitive environments such as critical care. The reasons for potential disparities are likely to be wide ranging and multi-factorial. We argue that Intensive Care services, despite being expensive and resource intense should receive the same degree of prioritisation and attention regarding equity of access. Our study and the methodology employed can provide direction to further research in this area with the goal of ensuring each critically ill patient in Scotland gets the same high quality, timely management of their critical illness.

Conclusion

Access to critical care services in Scotland varies with location of residence. In contrast, the probability of surviving an ICU admission, if admitted, does not show geographical variation. This present study cannot explain the data patterns described, and further investigations are warranted.

Supplemental Material

INC914462 Supplementary material - Supplemental material for Equity of access to critical care services in Scotland: A Bayesian spatial analysis

Supplemental material, INC914462 Supplementary material for Equity of access to critical care services in Scotland: A Bayesian spatial analysis by Philip Emerson, David R Green, Steve Stott, Graeme Maclennan, Marion K Campbell and Jan O Jansen in Journal of the Intensive Care Society

Footnotes

Acknowledgements

We would like to thank the Scottish Intensive Care Society Audit Group for giving us access to the data.

Author's Note

Philip Emerson is currently affiliated with Department of General Medicine, Royal Adelaide Hospital, Adelaide, Australia.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Ethics approval

Ethics approval for this project was obtained from the appropriate committee.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This project was supported by a grant from NHS Grampian Endowments.

ORCID iD

Philip Emerson https://orcid.org/0000-0002-1253-8700 Jan O Jansen

Supplemental Material

Supplemental material is available for this article online.

References

Chin

. How to achieve health equity. N Engl J Med 2014; 371: 2331–2332.

Oliver

Mossialos

. Equity of access to health care: outlining the foundations for action. J Epidemiol Community Health 2004; 58: 655–658.

Murthy

Wunsch

. Clinical review: international comparisons in critical care – lessons learned. Crit Care 2012; 16: 218.

The Scottish Government. A national clinical strategy for Scotland, www.gov.scot/Resource/0049/00494144.pdf (2016, accessed 25 September 2017).

Vestergaard

AHS

Christiansen

Nielsen

, et al. Geographical variation in use of intensive care: a nationwide study. Intensive Care Med 2015; 41: 1895–902.

Docherty

Anderson

Walsh

, et al. Equity of access to critical care among elderly patients in Scotland: a national cohort study. Crit Care Med 2016; 44: 3–13.

Lawson

. Bayesian disease mapping: hierarchical modeling in spatial epidemiology, 2nd ed. UK: Chapman and Hall/CRC Press, 2013.

Lawson

Biggeri

Boehning

, et al. Disease mapping models: an empirical evaluation. Disease Mapping Collaborative Group. Stat Med 2000; 19: 2217–2241.

National Records of Scotland Web Team. National Records of Scotland Statistics & Data. National records of Scotland, www.nrscotland.gov.uk/statistics-and-data (accessed 24 October 2017).

10.

The Scottish Intensive Care Society Audit Group annual reports, http://www.sicsag.scot.nhs.uk/publications/main.htm (accessed 5 October 2017).

11.

Docherty

Lone

Anderson

, et al. Epidemiology and outcomes of older patients admitted to Scottish intensive care units: a national database linkage study. Lancet 2015; 385: S33.

12.

Lone

Gillies

Haddow

, et al. Five-year mortality and hospital costs associated with surviving intensive care. Am J Respir Crit Care Med 2016; 194: 198–208.

13.

Harrison

Lone

Haddow

, et al. External validation of the intensive care national audit & research centre (ICNARC) risk prediction model in critical care units in Scotland. BMC Anesthesiol 2014; 14: 116.

14.

SICSAG. The Scottish Intensive Care Society Audit Group methodology, www.sicsag.scot.nhs.uk/Data/methodology.html (accessed 2 September 2018).

15.

statistics.gov.scot. Population estimates (current geographic boundaries), http://statistics.gov.scot/data/population-estimates-current-geographic-boundaries (accessed 4 October 2017).

16.

Public Benefit and Privacy Panel for Health and Social Care, www.informationgovernance.scot.nhs.uk/pbpphsc/ (accessed 4 October 2017).

17.

R Core Team. R: A language and environment for statistical computing, Vienna, Austria: R Foundation for Statistical Computing, 2016.

18.

Kang SY, Cramb SM, White NM, et al. Making the most of spatial information in health: A tutorial in Bayesian disease mapping for areal data. Geospat Health 2016; 11: 190–198.

19.

Gelman

. Prior distributions for variance parameters in hierarchical models (comment on article by Browne and Draper). Bayesian Anal 2006; 1: 515–534.

20.

ISD Services|Geography, population and deprivation analytical support team|Population|Standard populations|ISD Scotland, www.isdscotland.org/Products-and-Services/GPD-Support/Population/Standard-Populations/ (accessed 5 October 2017).

21.

Besag

York

Mollié

. Bayesian image restoration, with two applications in spatial statistics. Ann Inst Stat Math 1991; 43: 1–20.

22.

MacNab

. On Gaussian Markov random fields and Bayesian disease mapping. Stat Methods Med Res 2011; 20: 49–68.

23.

Earnest A, Morgan G, Mengersen K, et al. Evaluating the effect of neighbourhood weight matrices on smoothing properties of Conditional Autoregressive (CAR) models. Int J Health Geogr 2007; 6: 54.

24.

Knaus

Draper

Wagner

, et al. APACHE II: a severity of disease classification system. Crit Care Med 1985; 13: 818–829.

25.

Harrison

Brady

Parry

, et al. Recalibration of risk prediction models in a large multicenter cohort of admissions to adult, general critical care units in the United Kingdom. Crit Care Med 2006; 34: 1378–1388.

26.

Rue

Riebler

Sorbye

, et al. Bayesian computing with INLA: a review. Annu Rev Stat Appl 2017; 4: 395–421.

27.

Duncan

White

Mengersen

. Spatial smoothing in Bayesian models: a comparison of weights matrix specifications and their impact on inference. Int J Health Geogr 2017; 16: 47.

28.

Rezaeian

Dunn

St. Leger

, et al. The production and interpretation of disease maps. Soc Psychiatry Psychiatr Epidemiol 2004; 39: 947–954.

29.

Cooke

. Risk of death influences regional variation in intensive care unit admission rates among the elderly in the United States. PLoS One 2016; 11: e0166933.

30.

Vestergaard AHS, Christiansen CF, Nielsen H, et al. Geographical variation in use of intensive care: a nationwide study. Intensive Care Med 2015; 41: 1895–1902.

31.

Waller

Gotway

. Applied spatial statistics for public health data, Hoboken, NJ: John Wiley & Sons Inc, 2004.

32.

Oliver

Mossialos

. Equity of access to health care: outlining the foundations for action. J Epidemiol Community Health 2004; 58: 655–658.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.09 MB