Evidence-Based Practice Proposal – Final Paper

H I

D a

b

c

a

A R R A A

J H J I

K H A T M

1

a t t i 1 t y u i t g 1 t g

U N

h 0

Journal of Health Economics 57 (2018) 60–74

Contents lists available at ScienceDirect

Journal of Health Economics

jo u r n al homep age: www.elsev ier .com/ locate /econbase

ealth care expenditures, age, proximity to death and morbidity: mplications for an ageing population

aniel Howdona,∗, Nigel Riceb,c

Department of Economics, Econometrics and Finance, University of Groningen, Duisenberg Building, Nettelbosje 2, 9747AE Groningen, Netherlands Centre for Health Economics, University of York, York YO10 5DD, UK Department of Economics and Related Studies, University of York, York YO10 5DD, UK

r t i c l e i n f o

rticle history: eceived 18 April 2016 eceived in revised form 10 October 2017 ccepted 1 November 2017 vailable online 15 November 2017

EL classification: 51

11

a b s t r a c t

This paper uses Hospital Episode Statistics, English administrative data, to investigate the growth in admitted patient health care expenditures and the implications of an ageing population. We use two samples of around 40,000 individuals who (a) used inpatient health care in the financial year 2005/06 and died by the end of 2011/12 and (b) died in 2011/12 and had some hospital utilisation since 2005/06. We use a panel structure to follow individuals over seven years of this administrative data, containing estimates of inpatient health care expenditures (HCE), information regarding individuals’ age, time-to- death (TTD), morbidities at the time of an admission, as well as the hospital provider, year and season of admission. We show that HCE is principally determined by proximity to death rather than age, and that

19

eywords: ealth care expenditures geing ime-to-death

proximity to death is itself a proxy for morbidity. © 2017 Elsevier B.V. All rights reserved.

orbidity

. Introduction

There is concern that the demographic pressures of population geing will lead to an unprecedented rise in public expenditures o levels unsustainable under current financing arrangements. In he UK in 2013 approximately 17% of the population (11 million ndividuals) were aged 65 years or over. This represents a rise of 7.3% in this age group on a decade earlier. Projections suggest hat by 2050 this group will have increased disproportionately to ounger age groups, accounting for approximately 25% of the pop- lation (Cracknell, 2010). The growth in the proportion of older

ndividuals is partly due to increased longevity and partly due to he age structure of the population, particularly the ageing of the eneration of baby boomers of the post war period to the early

970s. Health care expenditures in the UK have also risen substan- ially over time, both in real terms and proportional to economic rowth. Close to the inception of the National Health Service (NHS),

∗ Corresponding author at: Department of Economics, Econometrics and Finance, niversity of Groningen, Duisenberg Building, Nettelbosje 2, 9747AE Groningen, etherlands.

E-mail address: [email protected] (D. Howdon).

ttps://doi.org/10.1016/j.jhealeco.2017.11.001 167-6296/© 2017 Elsevier B.V. All rights reserved.

net expenditure (net of patient charges and receipts) on the UK NHS in 1950/51 was £11.7b (GBP, in 2010/11 prices), representing 3.5% of Gross Domestic Product (GDP). This had risen to £121.3b by 2010/11, approximately 8.2% of GDP. Over the twenty-five year period from 1999/00 to 2014/15, expenditure in England almost doubled to £103.7b (2010/11 prices), with an average expenditure per head of population of £1900 (Harker, 2012). Abstracting from issues such as technological innovation, the concern is that as the share of the population at older ages rises, the economic burden of providing healthcare will become increasingly unsupportable.

Interest in the link between ageing populations and health care expenditures can be traced back 25 years, when the International Monetary Fund (IMF) asserted that ‘demographic pressures [in the UK] of an aging population will be associated with increased demand for medical services’, and presented descriptive statistics from various countries, showing that older patients, on average, had greater health care costs than younger patients (Heller et al., 1986). A report by the Organisation for Economic Co-operation and Development (OECD) predicted that across Europe population age-

ing will create a rise in age-related social expenditures from around 19% of GDP in 2000 to around 26% by 2050. Old-age pension pay- ments and expenditure on health and long-term care was deemed responsible for approximately half this increase (Dang et al., 2001).

Healt

A f f n h e p i

t ‘ t h o i 2 o l c ‘ t c I c t i t a o i d r a o t t i a m k t b m m w u v ( e

2

2

w a i F e t e

D. Howdon, N. Rice / Journal of

pproaches to predicting expenditure growth vary, but in a simple orm this consists of computing observed expenditures per head or different age-sex groups, and multiplying by projections of the umber of people expected to fall into each group. This approach, owever, fails to consider the underlying drivers of heath care xpenditures and the relative role of age, or, as has been suggested, roximity to death, or underlying levels of disability and ill-health,

n determining expenditures and its likely growth (see Gray, 2005). Additional to projections of population ageing is the poten-

ial change in the health profile of the population over time. An expansion of morbidity’ hypothesis has proposed that the ‘net con- ribution of our successes has actually been to worsen the people’s ealth’, as improvements in health care tend to lengthen the lives f those living with illness disproportionately to the effect of such mprovements on the lifespan of those living without (Gruenberg, 005). Should population ageing occur alongside a deterioration f health at older ages, then this will exacerbate impacts on pub- ic expenditures. While subsequent academic research into these laims – notably, research in the ‘compression of morbidity’ and red herring’ strands of literature – have given reason to suggest hat such concerns may have been misplaced or exaggerated, con- ern over the impact of an ageing population on HCE has persisted. ndeed, even in 2012, the UK’s then-Secretary of State for Health laimed that the fact that ‘the number of people aged over 85 in his country will double in the next 20 years’ was one of two factors n ‘costs . . . rising at an unaffordable rate’ (Lansley, 2012). He fur- her argued that ‘age is the principal determinant of health need’,1

nd that local NHS budgets should be recalibrated to be based n this, as a result (Williams, 2012). This paper uses UK admin- strative data from Hospital Episode Statistics (HES), and deaths ata from the Office for National Statistics (ONS), to consider two elated research areas. The first, in line with the ‘red herring’ thesis dvanced by Zweifel et al. (1999), is to explore the determinants f inpatient health care expenditures, with particular attention to he role played by age, time-to-death (TTD), and morbidity. We do his in a unique way by following samples of individuals who died n England, over seven years of HES data from 2005/06 to 2011/12, nd constructing a panel on individual health care expenditures and orbidity over this period. We show that TTD dominates age as a

ey driver of health care expenditures and morbidity characteris- ics dominate TTD. This finding extends the ‘red herring’ literature y showing that TTD is itself a ‘red herring’ and acts as a proxy for orbidity. This links to a second area of research by locating the odelling of health care expenditures for individuals close to death ithin the broader literature on prospective prediction of hospital se to inform resource allocation, particularly those based on indi- idual level data and which incorporate information on morbidity for example, see Iezzoni et al., 1998; van de Ven et al., 2003; Pope t al., 2004; Dixon et al., 2011).

. Literature review

.1. Compression of morbidity

The ‘compression of morbidity’ strand of literature beginning ith Fries (1980) suggests that, ‘[i]n its simplest form, “the age

t first appearance of symptoms of aging and chronic disease can ncrease more rapidly than life expectancy”’ (Fries et al., 2011).

ries (2005) identifies three separate ‘eras’ of illness and well-being xperienced during the 20th Century and beyond: an era of infec- ious disease, followed by an era of chronic disease, followed by an ra described by the author as ‘directly related to the process of

1 Emphasis ours.

h Economics 57 (2018) 60–74 61

senescence, where the aging process itself, independent of specific disease, will constitute a major burden of disease’. Senescence – the process of ageing – is characterised by the ‘decline of maximal function of [all] vital organs’, beginning before any chronic disease takes hold: deaths where this function declines below a level nec- essary to sustain life, in the absence of any disease occasioning this, may be termed ‘natural deaths’ (Fries, 2005).

The implications for HCE of an ageing population become less clear in the light of compression of morbidity, and there are two aspects to this which deserve attention. First, as the “age at first appearance of symptoms of aging and chronic disease” increases, individuals can be said to age more healthily: the implications of this for HCE are considered below. Second, the compression of mor- bidity thesis takes for granted an increase in life expectancy. The implications of this for HCE can be considered at a population level for any given year of spending. Setting aside the causal process for this health ageing (again, considered below), as the average person ages more healthily, they require lower HCE at any given age. As more people live to very old age – for instance, 90 years old – each individual requires lower health spending at that age. The overall picture for HCE is however ambiguous: a larger number of people requiring lower HCE may require greater overall costs at a popula- tion level than a smaller number of people requiring higher HCE. Similarly, an individual, who dies at age 90 and requires lower HCE at any given age than they would had they been born into an ear- lier cohort, may require greater cumulative HCE over their lifespan than they would had they aged less healthily and died at the age of 70. The implications for HCE in the presence of healthy ageing and increased lifespan may differ at an individual level to a population level.

Freedman et al. (2002), in a systematic review covering research that had been conducted between 1990 and 2002, found that many measures of disability and limitations in old age had seen declines in recent years: in particular, a change of −1.55% to −0.92% per year in those reporting any disability during the late 1980s and 1990s. Romeu Gordo (2011) observes a cohort-on-cohort fall in the num- ber of individuals with high levels of disability-related functional problems in their everyday life for those born between 1924 and 1947 in the US. Cutler et al. (2013), using Medicare records from the US, present evidence of an increase in disability-free life between 1991 and 2009. The authors conclude that ‘The major question raised by our results is why this has occurred. How much of this trend is a result of medical care versus other social and environ- mental factors?’.

Cross-country international evidence on the changing pat- terns of disability rates across nine OECD countries is provide by Jacobzone et al. (2000). Consistent with the above literature, they report evidence of significant falls in severe disability rates. The importance of this issue for forecasting HCE depends upon how changes in mortality, changes in morbidity, and changes in dis- ability occur and interact with each other. If the onset of chronic conditions – those imposing large costs on health systems – can be postponed out of an individual’s lifetime, then health care costs may fall as later cohorts enjoy a longer lifespan, with a reduced level of necessary treatment for chronic conditions. Dormont et al. (2006), for instance, find that improvements in morbidity profiles in France between 1992 and 2000 have caused reductions in HCE that more than offset the rise in HCE induced by an ageing population.

The morbidity and disability profile of individuals, according to

this research, at any given age has improved over time, leading to health problems being experienced later in life and more closely to death. In the illustrated case (Figs. 1 and 22), individuals live

2 Adapted from Fries (1980) and http://www.aei.org/files/2008/06/27/20080626 WashingtonAEI.pdf.

62 D. Howdon, N. Rice / Journal of Healt

Fig. 1. Stylised change in survival curves.

u o o h a e g g i T m t t c c

2

r p t

i

Fig. 2. Stylised change in health profiles.

p to a longer observed maximum age (indicated by the shift out f the survival curve from S1 to S2 in Fig. 1), and have a higher bserved level of health at all ages (indicated by the shift out of the ealth status curve from H1 to H2 in Fig. 2). Both survival curves nd health status curves have become increasingly rectangular. The ffect on health care expenditure (HCE) is ambiguous, given that enerally more healthy ageing – a decrease in morbidity at any iven age – puts downward pressure on HCE, while an increase n life expectancy, ceteris paribus, puts upward pressure on HCE. he actual relationship between health care costs and changes in orbidity and mortality profiles at every given age depends upon

he changing shape of these two curves, and also the extent to which he changes in each are due to or caused by the healthcare that reates these HCE. The use of age per se in predicting future health are costs should be approached with caution, as a result.

.2. Age, time-to-death and healthcare expenditures

The ‘red herring’ strand of literature further gives empirical

eason to suggest that claims of steeply-rising future HCE due to opulation ageing3 may have been exaggerated, potentially owing o morbidity being concentrated in later years of life. Zweifel et al.

3 HCE may rise due to technological change brought about by new expensive nnovations in health care treatments, or due to shifting patterns of morbidity.

h Economics 57 (2018) 60–74

(1999), using Swiss sickness fund data, find that no effect of age on health care expenditures existed after controlling for TTD, i.e. the time from any given point of observation to death for an individual. Owing to the number of individuals with zero HCE, a two-step model (with a probit first stage and OLS second stage) was employed, with only deceased patients included in the model. Such work was criticised on the grounds of potential endogene- ity, with time-to-death affected by both present, previous (and, due to the nature of how TTD must be measured) future HCE. In a subsequent paper, Zweifel et al. (2004) seek to test for such prob- lems, finding that while TTD is endogenous, their results were ‘fairly robust’ to the error this induces. Werblow et al. (2007) find that age is a small (but statistically significant) determinant of HCE after controlling for TTD for patients using long-term care (LTC), such as those in care homes, and is not associated with HCE for non- LTC patients. More complicated methods, such as those employing generalised linear models, have since been used, for example by Werblow et al. (2007), in order to deal with the non-normal prop- erties (such as positive skewness) exhibited in the distribution of HCE. These papers have corroborated results obtained using probit and OLS two-step models. Felder et al. (2010), in a recent paper in this series, first predict individuals’ survival based on observed HCE and socioeconomic characteristics (in early waves), before using predicted values based on this as an instrument for TTD in explain- ing HCE in later waves. The authors find that, while TTD cannot be deemed exogenous, any effect of age on HCE becomes insignificant when TTD (or instrumented TTD) is included in the model. Further- more, results regarding the relative importance of TTD compared to age have also been corroborated in a disease-specific study carried out by Wong et al. (2011).

While use has been made of morbidity markers in models of long-term care expenditures (LTCE) (see de Meijer et al., 2011), such use has not been made in models explicitly investigating the link between HCE and population ageing. One possibility is that TTD is itself a red herring, in that it is simply a proxy for morbidity, unob- served in existing HCE models in the red herring strand of literature. Such a theory has been provisionally borne out empirically in liter- ature related to economic evaluation of healthcare, with Gheorghe et al. (2015) finding that quality of life (as measured by SF-6D scores) declines with proximity to death, and also by biological and medical literature. Indeed, Dalgaard and Strulik (2014), proposing an alternative life cycle model of ageing, note that previous work in the ‘red herring’ strand of literature is consistent with biolog- ical and medical research (see, inter alia Mitnitski et al., 2002a,b, 2005; Rockwood and Mitnitski, 2006, 2007), showing that concep- tions of ageing focusing on time-from-birth (such as that inherent in Grossman (1972)) are erroneous. This model conceptualises the human body as a system which has substantial inbuilt redundancy (that is, an ability to function at a level well over and above that required to sustain life) in youth, but redundancy which declines as ‘deficits’ (a decline in function of individual parts of the body) are accumulated. Ageing depends not upon a ‘biological clock’, but is a process of increasing frailty which is the outcome of investments in health, available health technology, the lived environment, and a physiological ‘force of aging’ parameter. According to such a model, health is predicted to decline at an increasing rate when the individ- ual’s health status is lower. The authors note that existing research in the ‘red herring’ strand of literature, in line with this, ‘suggests that health status (e.g., frailty), and not the year on the birth certifi- cate, is what matters to health investments’. The latent assumption here is that a TTD variable proxies for this health status, which declines as individuals become more morbid as they approach

death, and with ever-increasing levels of health investment to (partly) offset this decline in health status and postpone death.

This seems intuitively plausible: in the years before death, it is likely that morbidity will increase, leading to more treatment,

Healt

a b ( t t D t t i t L f t v e

t t d a t a b l a p o t o t t t f c l o v w 1 s c e t t i l

s p R w b n

e l a a b w b o o i p p

the HES dataset, which has been published for each financial year since 1989/90 and is available for admitted patient care, outpa- tient, accident and emergency and maternity cases. The admitted

D. Howdon, N. Rice / Journal of

nd that comorbidities complicating the treatment of the disease ringing about the hospital episode will also increase. Shwartz et al. 1996), in work predating the original red herring hypothesis, note hat the inclusion of variables for comorbidities increase substan- ially the explanatory power of models. It seems likely that, as algaard and Strulik (2014) suggest, variables incorporating ‘time-

o-death’ in more recent models of HCE are picking up, in large part, hese comorbidities, which are not included in existing HCE models n the red herring literature. Indeed, de Meijer et al. (2011) conclude hat time-to-death ‘largely approximates disability’ in models of TCE. Dixon et al. (2011), in proposing individual-level formulae or resource allocation in the UK’s National Health Service (often ermed ‘Person-Based Resource Allocation’, or PBRA) include indi- idual level morbidity markers, finding that these have a ‘powerful ffect. . . in predicting individual level expenditure’.

The process generating HCE is clearly not a simple function of hose explanatory variables used in existing ‘red herring’ research: he actual data-generating process behind these health care expen- itures is unlikely to be characterised accurately by a simple use of ge, historical time and time-to-death. In addition to the aforemen- ioned problems surrounding TTD and age as a proxy for morbidity, s Breyer et al. (2014) note, many existing models are likely to e characterised with substantial endogeneity problems, which

ead to potential bias in the estimation of the change in HCE as n individual ages or approaches death. The authors control for otential endogeneity introduced by differential treatment based n a physician’s view of the patient’s expected health benefits from reatment, proxied by actuarial tables of life expectancy conditional n age. If physicians expect individuals to respond differently to reatment, this may cause those who are more likely to respond to reatment to be treated more intensely than those who are not, hus increasing expected HCE for individuals who are younger, urther-from-death or with fewer comorbidities because of physi- ian selection. Conversely, HCE for older individuals – or, more ikely, individuals in the final years of life – may rise as intensity f treatment becomes stronger with heroic efforts to save an indi- idual’s life, possibly motivated by ethical ‘rule of rescue’ concerns hen faced with an identifiable, gravely sick individual (Jonsen,

986). Breyer et al. (2014) jointly estimate this possible physician election based on life expectancy alongside a model for health are expenditures, incorporating both age and time-to-death as xplanatory variables. They find that increasing survival rates for he elderly in Germany have positive impacts on HCE, arguing that his is explained by physician selection: treating patients more ntensively if they expect positive results from treatment over a onger time span.

Datasets used within the ‘red herring’ literature are, in general, ickness fund datasets, with only Seshamani and Gray (2004) using opulation-level (for users of NHS treatment) data, the Oxford ecord Linkage Study, a longitudinal dataset of all individuals ithin an area of Oxfordshire, England. We believe our paper to

e the first to use a sample of individuals from a comprehensive ational-level dataset of health care users.

The extent to which ‘red herring’ and related issues are of inter- st depends upon the intended use of such research. Much existing iterature focuses on projections of future health care costs given an geing population, with the headline results of some papers (such s Stearns and Norton (2004) and Seshamani and Gray (2004)) eing the overestimation of expected costs for a given future year hen TTD is an omitted variable. This is due to the collinearity

etween TTD and age for a given individual: an individual who gets ne year closer to death also gets one year older, and so the impact

f TTD is picked up by age in such models. The inclusion of morbid- ty markers in addition to, or replacing, TTD would allow greater recision of future estimates where reliable estimates of morbidity revalence, and the cost of treatments, conditional on age and TTD

h Economics 57 (2018) 60–74 63

were known. Certainly, if the compression of morbidity hypothesis holds, and individuals are able to postpone the onset of chronic dis- eases – with associated higher HCE – to a time period closer to their death, or even indefinitely, explicitly considering morbidity rather than proxying this by age or TTD becomes ever more important.

We build upon the compression of morbidity and red her- ring strands of existing literature, seeking to further examine the relationship between ageing, time-to-death and health care expenditures. The original red herring hypothesis is that, once time- to-death is included in models of HCE, age per se does not explain changes in HCE. While models intended for resource allocation (Dixon et al., 2011) have already included morbidity as an explana- tory variable in HCE for the general population, other applications of models of HCE have not – in particular, those focusing explicitly on ageing populations, or costs in the years approaching death.

While hospital inpatient care forms only one part of health and social care incurred later in life, it is responsible for a large pro- portion of such expenditures. While it may seem that expenditures from other categories of health-related expenditures such as that arising from general practice, prescriptions, and even long-term care should be considered in the aggregate, much richer, more uni- versal and more reliable individual-level administrative data is at our disposal for the UK for hospital expenditures than for other types of expenditure, and information regarding the relationships examined in this paper are likely to be informative for specific avenues of NHS budget-setting. Furthermore, conflicting evidence exists regarding end-of-life long-term care expenditures and their functional relationships with age and TTD (de Meijer et al., 2011; Karlsson and Klohn, 2014), and the UK’s institutional structure and its resultant incentives regarding (predominantly privately- financed) long-term care and (predominantly socially-financed) hospital care is such that different relationships for each that are particular to the UK may well be expected. Finally, hospital expen- diture is a major contributor to end-of-life health care costs: French et al. (2017) find that hospital expenditure is dominant in the final year of life with non-hospital expenditure, of which long-term care is a major component, playing a greater role in periods prior to this. Indeed, they estimate that for England 11.6% of hospital expendi- ture occurs in the last year of life.

This paper seeks to bridge the gap between the red herring strand of literature and models of resource allocation, treating mor- bidity measures as omitted variables in models of current health care expenditure, and examining what the relationship between age, TTD and HCE is once morbidity is included in these models (see, for instance, Aragon et al., 2016).

3. Data

3.1. Data sources

Information on patient-level hospital use and associated refer- ence costs for treatment are derived from the Hospital Episodes Statistics (HES) dataset, published by the Health and Social Care Information Centre (HSCIC). This is complemented with small-area data on years of potential life lost (YPLL) published by the ONS, and individual level mortality information, jointly published by the HSCIC and the ONS.

We use successive years (financial years 2005/06 to 2011/12) of 4

patient (commonly, ‘inpatient’) care HES dataset that we use pro-

4 In the UK, the financial year runs from April to March.

6 Healt

v d h

e p e g T H E p e m v D t t a F H a H s t s w fi h s t s R y

a t p r c C

c t w c h b b t l

w m r t c I c G w

w t

4 D. Howdon, N. Rice / Journal of

ides information on individual-level patient characteristics and iagnoses and procedures undergone for all patients admitted to ospitals in England.5

Information regarding inpatient spells is used to associate ref- rence costs to each spell. Reference costs are based on each NHS rovider’s estimates of their own costs for each patient spell, cat- gorised by Healthcare Resource Group (HRG, the NHS’s system of rouping clinically-similar events with comparable resource use). hese reference costs are derived from accounting costs for each RG, submitted by each organisation providing secondary care in ngland (Department of Health, 2012). The NHS Costing Manual rovides guidance to all providers to support the calculation of ref- rence costs and to enforce more uniform standards for costing ethodologies. We use the estimate provided by the hospital pro-

iding treatment as our estimated cost for the patient’s episode. The H’s Reference Cost data is submitted on a full absorption basis –

hat is, taking account of all direct and indirect costs relating to he activities in question, as well as a proportion of an estimate of ll overhead costs relating to the overall running of the provider. urther, to account for the fact that costs will vary even within RGs, hospitals are required to provide per diem costs for longer dmissions that exceed a given ‘trim point’, which differs by each RG. This trim point is defined as the upper quartile of length of

tay, plus 1.5 times the inter-quartile range for length of stay for hat HRG (Department of Health, 2012). Moreover, we augment the tandard costs incurred in each episode with the ‘unbundled’ costs here recorded for the episode. This represents one or more extra xed costs associated with the episode where additional, unusual, igh-cost treatment or procedures were involved. Even within the ame primary HRG, costs are not identical but differ according to he patient’s length of stay. An estimate of costs for each inpatient pell is obtained by matching data on costs for that provider in the eference Costs database to HRG for each episode in the relevant ear’s HES data.

HES contains diagnostic data, categorised (since 1995/96) ccording to the tenth revision of the World Health Organiza- ion’s International Classification of Diseases (ICD-10). Details of rocedures and interventions are recorded according to the fourth evision of the Office of Population, Censuses and Surveys’ Classifi- ation of Intervention and Procedures (OPCS-4) (Health and Social are Information Centre, 2013).

HES is broken down by completed “episode” – each record onsists of a continuous period of care at a single provider of reatment under the same consultant. A new record is generated hen a patient is either transferred to the care of either a new

onsultant, transferred to a new provider, or is discharged from ospital. Although individuals are not identifiable, individuals can e tracked across episodes by an anonymised identification num- er. The costing of a patient’s time in hospital and the recording of heir diagnoses and procedures undergone are made at the episode evel.

Patients can be tracked across different years of the HES dataset, hich enables the creation of a panel structure for the data. Infor- ation within the HES dataset – most commonly, information

egarding diagnosis, treatment and age of the patient – is used o apply the most appropriate Healthcare Resource Group (HRG) ategorisation to the dataset. We use the Health and Social Care nformation Centre’s Consultation ‘Grouper’ software in order to

arry out this first step. We use the most recent version of this rouper – for the 2011/12 financial year – for all seven of the years e use, to categorise patients into HRGs. HRGs are used to cate-

5 This dataset includes both day cases (patients without an overnight stay) as ell as patients who have at least one night’s stay in hospital. Our use of ‘inpatient’

hroughout this text includes both types of patient.

h Economics 57 (2018) 60–74

gorise patient spells not only by broad diagnosis, but by the type and complexity of the patient’s spell, into one of over 1400 group- ings. This allows us to apply the current best-practice methods for grouping patients into HRGs based on the information available. We apply available estimates of hospital costs for each inpatient spell, using reference costs data for the relevant financial year.

We add information regarding an individual’s death from linked HES-ONS mortality data. The latest version of this data provides information on deaths to the end of the 2012 calendar year, and therefore provides information on some individuals whose deaths are known to have occurred after the end of the final wave in our dataset. Where individuals are known to have died, they are included up to and including the final quarter of their life, and not included in the panel in following years. TTD can only be measured – for decedents – retrospectively, using information available at the time of the individual’s death. We observe individuals for a maxi- mum of seven years (from 2005/06 to 2011/12) or 28 quarters and code TTD from 1 to 28, with TTD = 1 denoting the final quarter in which death occurs.6

We adopt a strategy that employs two complementary sampling procedures, each incorporating approximately 40,000 individuals. The first draws a sample of individuals who died in 2011/12, the final year of our analysis, and who had at least one quarter of recorded positive HCE in the 28 quarters of our data. The second draws a sample of individuals who had at least one quarter of recorded positive HCE in 2005/06, and died in or before 2011/12. We believe that each of these sampling procedures has advantages and disadvantages but that, together, they can be used to establish a clear conclusion on our research question.

Our first sample for analysis consists of a random sample of 39,381 individuals (18,690 men and 20,691 women) aged 50 years and older, taken from those with at least one inpatient episode between 2005/06 and 2011/12, and whose death was recorded by the ONS in the financial year 2011/12. Our second sample consists of a random sample of 39,796 individuals (19,673 men and 20,123 women) aged 50 years and older, taken from those with at least one inpatient episode in 2005/06, and whose death was recorded by the ONS after this point, and by the end of the financial year 2011/12. Sample size was selected to enable computations not to become burdensome, and the age cut-off was selected to ensure sufficient deaths were observed in the data to make meaningful inference. We follow all sampled individuals across all quarters until their death to observe their subsequent inpatient health care use and associated morbidity characteristics.

We collapse all inpatient episodes for each individual from HES for a given quarter into a single observation in our data. This obser- vation contains the sum of all hospital costs incurred in all episodes finishing in that quarter, as well as diagnostic information con- tained in the ICD-10 codes for those episodes in that quarter. In principle, the ICD-10 classification allows for up to 14,400 dif- ferent diagnoses. To make these more manageable for analysis, however, we collapse this information using the US Agency for Healthcare Research and Quality’s Clinical Classifications Software (CCS) method to convert ICD-10 codes to CCS codes (US Agency for Healthcare Research and Quality, 2009). This reduces the number of different groupings to a more manageable 260 mutually-exclusive, and clinically meaningful, categories.7 Where individuals do not have any episodes in a given quarter, we separately adopt two dis-

tinct methods in order to deal with such cases. In one approach, they are recorded as having zero hospital costs, and as having zero observed morbidities arising from diagnostic information. In the

6 Coding TTD in this way is akin to assuming all deaths occur at the end of a quarter.

7 A full list of these CCS groupings is provided in Appendix A.

Health Economics 57 (2018) 60–74 65

a h i a m i a c f m

i d i t n

I ( d o u o E 3 S

o u e C u a c l a L U f 2 ( f t t U a c g Y t W a

d p p o (

i p

n

Table 1 Summary statistics (quarter 1, men, first year sample).

Variable Mean Std. Dev. Min Max

HCE [missing treated as zero] 475.60 1740.26 0 82,901.09 log(HCE) [missing treated as zero] 1.57 3.00 0 11.32 log(HCE) [missing treated as missing] 7.19 1.01 3.42 11.32 Quarters to death (QTD) 9.53 7.77 0 27 log(QTD) 2.02 0.88 0 3.33 Age 75.03 10.24 50 105.66 YPLL (IMD 2007) 65.50 15.71 33.80 180.8

Table 2 Summary statistics (quarter 1, women, first year sample).

Variable Mean Std. Dev. Min Max

HCE [missing treated as zero] 504.42 1629.34 0 45,095.81 log(HCE) [missing treated as zero] 1.56 3.03 0 10.71 log(HCE) [missing treated as missing] 7.30 0.99 3.39 10.71 Quarters to death (QTD) 9.86 7.89 0 27 log(QTD) 2.05 0.89 0 3.33 Age 78.11 10.93 50 111.15 YPLL (IMD 2007) 65.85 15.54 33.30 191.5

Table 3 Summary statistics (quarter 1, men, final year sample).

Variable Mean Std. Dev. Min Max

HCE [missing treated as zero] 220.74 1339.96 0 66,770.92 log(HCE) [missing treated as zero] 0.61 2.05 0 11.11 log(HCE) [missing treated as missing] 7.28 1.08 3.85 11.11 Quarters to death (QTD) 25.57 1.13 24.00 27.00 log(QTD) 3.28 0.04 3.22 3.33 Age 72.93 9.82 50 100.83 YPLL (IMD 2007) 64.14 15.09 33.80 162.90

Table 4 Summary statistics (quarter 1, women, final year sample).

Variable Mean Std. Dev. Min Max

HCE [missing treated as zero] 213.89 1310.72 0 64,392.08 log(HCE) [missing treated as zero] 0.56 1.98 0 11.07 log(HCE) [missing treated as missing] 7.36 1.08 3.34 11.07 Quarters to death (QTD) 25.59 1.13 24.00 27.00

D. Howdon, N. Rice / Journal of

bsence of additional information on the gravity of any residual ealth problem, this assumes that such health issues are insignif-

cant relative to those leading to a hospitalisation. In a second pproach, we recognise that the recording of zero morbidities ight be unrealistic for patients observed to have hospitalisations

n recent periods and for whom there is likely to exist an underlying, lbeit less grave, health problem. Consequently, we model these ases in our second approach under the assumption that episodes or which no information is available represent non-informative,

issing data. While we include a sum of all hospital costs for episodes ending

n the quarter in question, we include only a maximum of three iagnoses for each individual, for a maximum of five episodes end-

ng in that quarter. Using the merged mortality data, we are able o add a variable for the individual’s time-to-death, measured in umber of quarters to death.

In addition, we make use of the Office for National Statistics’ ndices of Multiple Deprivation (IMD), by Lower Super Output Area LSOA) in order to construct an instrument for TTD. LSOAs are efined at the time of the UK’s decennial Census and are made up f similarly-sized small areas of the country. HES data, for the years sed in our dataset, provides information on the individual’s LSOA f residence at the time of the 2001 Census. At this time, LSOAs in ngland consisted of 32,482 areas of populations between 1000 and 000, with between 400 and 1200 households (Office for National tatistics, 2011).

Indices of Multiple Deprivation, at this LSOA level, are measures f the levels of deprivation in those small areas. Although made p of seven domains (income, employment, health and disability, ducation, housing, living environment and crime (Department for ommunities and Local Government, 2011)), we primarily make se of one of the indicators that forms part of the health and dis- bility IMD score: years of potential life lost per 1000 people. This onsists of a standardised measure of premature mortality calcu- ated using information for all individuals to have died before the ge of 75, as described in Blane and Drever (1998).8,9 Although the SOAs themselves are defined every ten years at the time of the K’s census, statistics for each domain are collected and published

or these areas more regularly: we make use of those published in 007 (produced using data from 2001 to 2005 inclusive), and 2010 produced using data from 2004 to 2008 inclusive) (Department or Communities and Local Government, 2008, 2011). For each of hese years, we use LSOAs as defined in the 2001 UK Census. While hese figures are comparable within years, the data collector (the K’s Department for Communities and Local Government) caution gainst using this data for trend analysis. These measures are highly orrelated with TTD and, by virtue of being calculated at an aggre- ate level, exogenous in a model of HCE. That is, while the level of PLL at an LSOA level is a strong predictor of an individual’s TTD, his YPLL level is not influenced by the HCE for a given individual.

e therefore include at least one wave of this measure separately s instruments.

Tables 1 and 2 present descriptive statistics for the sample of ecedents from the first wave of data, under our strategy of sam- ling from the first year of observations (2005/06). Tables 3 and 4 resent descriptive statistics from the first wave of data, under

ur strategy of sampling from the final year of observations 2011/12).

8 The Office for National Statistics, however, use 75 rather than 65 years, in their mplementation of this method, as the age at which mortality is considered to be remature (Department for Communities and Local Government, 2011). 9 Details of the method employed by the ONS were obtained in personal commu- ication with the study’s author, Chris Dibben.

log(QTD) 3.28 0.04 3.22 3.33 Age 76.80 10.02 50 105.58 YPLL (IMD 2007) 64.78 15.07 33.80 180.80

As is usual, the distribution of HCE is positively skewed, with this skewness reduced somewhat when we take a logarithmic transformation.10 As would be expected due to their longer lifes- pan, on average, the average age of women in the sample is somewhat higher than that for men. Similarly, women are observed for, on average, slightly more waves. HCE, with missing waves treated as zero-(log)-cost observations, is on average higher when sampling from the first financial year of data than when sam- pling from those who died in the final year of analysis. This is as expected: the former is drawn from those with an inpatient episode in 2005/06, whereas the latter is drawn from those with an inpa- tient episode in any of the seven financial years of analysis. Indeed, HCE is approximately similar when missing waves are treated as missing observations.

Diagrams, presented in Figs. 3 and 4, based on descriptive statis- tics from a sample of 9,957,084 individual quarterly observations

in HES, provide some illustration of the existing red herring the- sis. HCE appear to increase with age (top-left panel): this is the usual age-expenditure curve that is used to infer rising costs with

10 Due to log(0) being undefined, we add a value of one to such observations in our modelling strategies that include zero-cost quarters.

66 D. Howdon, N. Rice / Journal of Health Economics 57 (2018) 60–74

by age

p l o h e h m ( t a a t T H i a

b

w b t

Fig. 3. Healthcare expenditures

opulation ageing, with the assumption being that as the popu- ation ages, the curve continues to rise as an extrapolation of the bserved trend.11 The observation that expenditures rise with age, owever, is an artefact of a compositional effect. The naïve age- xpenditure curve is composed of individuals who are known to ave died during the period of observation (the sample used in esti- ation) – who have, on average, high expenditures for this period

top-right panel) – and individuals who are known to have survived o at least the end of the period of observation who have, on aver- ge, lower expenditures for this period (bottom-left panel).12 The verage expenditures for individuals observed to have died during he sample period are far greater than for individuals who survive. his suggests an important role for time-to-death in explaining

CE. As the proportion of the full population who are decedents

ncreases with age, the näive observed relationship between age nd expenditure displays an increasing trend. Note, however, that

11 We set aside here the drop in expenditures at very high ages, as this is likely to e due to the substantially lower sample sizes observed here. 12 While some of these survivors will be closer to death than other and therefore ould be classed as decedents over a longer observation period, such an effect would

ias us against finding a visual difference in these graphs. We therefore consider this o be strong evidence of a different age profile of HCE for decedents and survivors.

and proximity to death, males.

average expenditures for both decedents and survivors display a flatter profile than that depicted for the full population suggest- ing a less important role for age. Indeed, expenditure on decedents generally decrease, with this decrease particularly pronounced for women. Expenditure on survivors generally increase, but with a shallower gradient than observed for the full population, and at a lower average cost.

When we focus on decedents, and consider average HCE by proximity to death, we observe a large increase in costs in termi- nal quarters – particularly in the year immediately before death. Fig. 7 in Appendix A shows a similar relationship between expen- ditures and TTD for men at selected ages. In general, expenditure in quarters preceding the final three average around £500 (although there is variation). In the final three quarters, and particularly the final quarter, we observe a large increase in expenditure. With the exception of 50 year olds, there is a clear gradient of health expen- ditures rising most dramatically in the final quarter of life with average increases over the penultimate quarter ranging from £460 for 55 year olds to £1099 for 90 year olds.

The relationship between HCE and TTD in levels is nonlinear.

Fig. 5 shows that the relationship is approximately linear on the logarithmic scale and in the modelling that follows logarithms of both HCE and TTD are used throughout.

D. Howdon, N. Rice / Journal of Health Economics 57 (2018) 60–74 67

Fig. 4. Healthcare expenditures by age and proximity to death, females.

Fig. 5. Average health care expenditures according to quarters to death (log scale for x- and y-axes).

68 D. Howdon, N. Rice / Journal of Health Economics 57 (2018) 60–74

al ind

4

s t

w s u t s b a

l

information about a patient’s morbidities at the time of their hos- pital stay. We estimate each of these models with random effects, representing unobserved heterogeneity.

Fig. 6. Change in HCE according to time-to-death and age, hypothetic

. Econometric model

We follow the general strand of the red herring literature and pecify a baseline model of HCE, including only age as an explana- ory variable.

log(HCEit) = ̨ + ˇageageit + �it + �i + εit, i = 1, . . ., N, t = 1, . . ., Ti, (1)

here �it is a vector of control variables (year and season of admis- ion, and hospital provider dummies) �i is an individual-specific nobserved effect and εit is an idiosyncratic error term. Although his model is not estimated in existing papers, it is claimed that uch a model would not adequately explain HCE. TTD is claimed to

e an omitted variable in these models, giving rise to models such s:

og(HCEit) = ̨ + ˇageageit + ˇTTD log(TTDit) + �it + �i + εit . (2)

ividual dying at 75 (top – men, bottom – women; log HCE on y-axis).

We argue that individual morbidity is an omitted variable in this type of model, where TTD functions as a proxy for such morbidity.13

Accordingly, we augment the model as follows:

log(HCEit) = ̨ + ˇageageit + ˇTTD log(TTDit)

+ 260∑

j=1 ˇCCSj CCSjit + �it + �i + εit, (3)

where CCSn represents a recorded morbidity of CCS type n (n = 1, . . ., 260). We exploit the available data in HES to include detailed

13 And, furthermore, that such a proxy relationship may change over time in the presence of a compression of morbidity.

Health Economics 57 (2018) 60–74 69

l o T H T t t t i e e u

u e i F a p T a l T u i h d s fi a e

5

l s m r T i i m a r t t m

e c d

a a o T

t w o c s t l

Table 5 Results, final wave sampling.

Model Missing observations treated as missing

(1) (2) (3) AGE ONLY AGE TTD AGE TTD MORBS

Men Age −.01459** −.01274* −.00518

(.00654) (.00652) (.00526) Age2 .00010** .00009** .00003

(.00004) (.00004) (.00003) log(TTD) −.42375*** −.14454***

(.01467) (.01206) Morbidities Included

Women Age −.00068 .00081 −.00038

(.00588) (.00585) (.00474) Age2 .00004 .00003 .00001

(.00004) (.00004) (.00003) log(TTD) −.34305*** −.13276***

(.01458) (.01218) Morbidities Included

Model Missing observations treated as zeros

(1) (2) (3)

Age .00130 .00204 .00289*

(.00180) (.00181) (.00156) Age2 0.00000 −.00001 −.00002**

(.00001) (.00001) (.00001) log(TTD) −.33712*** −.10645***

(.00679) (.00560) Morbidities Included

Women Age .00983*** .01087*** .00559***

(.00189) (.00189) (.00154) Age2 −.00005*** −.00006*** −.00003***

(.00001) (.00001) (.00000) log(TTD) −.27927*** −.09789***

(.00604) (.00520) Morbidities Included

*

D. Howdon, N. Rice / Journal of

Modelling HCE as a function of TTD suffers from potential prob- ems of endogeneity. Existing literature suggests that conditional n other covariates, being further from death – i.e. having a high TD – in time period t is likely to lead to lower levels of HCE in t. igher levels of HCEit, however, are likely to lead to high levels of TDit: if the hospital activity that generates health care expendi- ures is effective in improving health then the individual is likely o enjoy a longer remaining lifespan as a result. We therefore posit hat actual TTD at time period t has been determined in part by HCE n that time period as well as other time periods. Consequently, if ndogeneity does pose problems in this analysis, the coefficient stimate on TTD (when treated as exogenous) is likely to be an nderestimate of the true ‘effect’ of TTD.

Other models in the red herring strand of literature model HCE, sing TTD and age as explanatory variables, but highlighting this ndogeneity problem. Various attempts are made to purge TTD of ts endogeneity in HCE (Zweifel et al., 2004; Werblow et al., 2007; elder et al., 2010). We propose the use of a component of the Health nd Disability Index of Multiple Deprivation by Lower Super Out- ut Area – years of potential life lost (YPLL) – as an instrument for TD under the assumption that such measures are exogenous in

model of HCE but highly correlated with TTD. That is, while the evel of YPLL at an LSOA level is a strong predictor of an individual’s TD, this YPLL level is not influenced by the HCE for a given individ- al. Accordingly, where possible, we reestimate models (2) and (3)

nstrumenting TTD by YPLL.14 Such an instrumented approach is, owever, possible only in the case of our second sampling proce- ure, where TTD is not pre-determined by the construction of the ample. In our former sampling procedure, all individuals die in the nal four quarters (i.e., final financial year) of the sample, and thus ny relevance of variation across areas in deprivation would not be xpected.

. Results

All versions of our different sampling and modelling strategies ead to qualitatively similar results. In short, a weak (and often tatistically insignificant) relationship is observed when costs are odelled as a function of age alone. Confirming the overall red her-

ing results, a strongly significant relationship is observed between TD and HCE, when TTD is added as an explanatory variable. This s in line with our descriptive diagrams (Figs. 3 and 4), demonstrat- ng that the naïvely-estimated relationship between age and HCE is

uted when conditioning on TTD. When morbidities are included s explanatory variables, the relationship between TTD and HCE is educed (in all cases, the coefficient is reduced by approximately wo-thirds). When, where possible, instrumenting TTD, the rela- ionship between TTD and age becomes larger, with the addition of

orbidities again reducing the size of the TTD coefficient.15

Table 5 presents the results of various specification of a random

ffects panel data model of log(HCE) on age, log(TTD) and morbidity haracteristics for the sub-sample of decedents, when a sample is rawn from those who died in 2011/12. The first column of results

14 While HCE is a function of morbidity, morbidity itself will be a function of age, nd TTD is likely to be a function of morbidity. Indeed, our hypothesis is that TTD is

proxy for morbidity. Accordingly, we expect supplementing (3) with information n morbidity will temper the effect of both age (remaining after conditioning on TD) and TTD on HCE. 15 In the interests of consistency, all results presented here employ one wave of he YPLL instrument. Where both instruments appear as relevant at the first stage, e estimated the models using both YPLL waves in order to carry out a Hansen J test

f the validity of overidentifying restrictions. In all cases, we observe large p-values onsistent with failing to reject the null-hypothesis (between 0.3354 and 0.6317), uggesting evidence in favour of the exogeneity of our chosen instruments. Fur- hermore, our second stage results suggest very similar coefficients and confidence evels, such that none of our conclusions drawn below are affected.

p < 0.05. ** p < 0.01.

*** p < 1.

(model 1) shows a weak and generally non-significant relationship between age and inpatient costs. These results represent, as far as we are aware, the first reported results in the red herring strand of literature of whether hospital costs increase with age in the aggre- gate, even before control is made for other factors such as TTD and morbidities. Existing research broadly states that this is the case, but refer merely to population-level descriptive statistics. In a ran- dom effects model (2) including TTD and age, we observe a highly significant relationship with TTD. This result is in line with those in the red herring strand of existing research. As an individual gets 1% closer to death, HCE increases by between 0.34% and 0.42% for men (between 0.28% and 0.34% for women), depending on the modelling strategy adopted.16

Conditioning on morbidity markers, we find a reduced role for TTD in explaining HCE, using both sampling strategies. Our esti- mate of the TTD elasticity of HCE falls by approximately two-thirds in almost all (non-IV) cases when we condition on the individ- ual’s observed morbidity in the current time period (i.e., when

we move from model 2 to model 3). In all models, in excess of 90% of the estimated coefficients for the morbidity indicators are significant at the 1% level, yielding a p-value of 0.0000. We inter-

16 Because we aggregate costs by quarter and consequently use discrete values of TTD for each individual in each wave, this elasticity can only be considered as an approximation.

70 D. Howdon, N. Rice / Journal of Health Economics 57 (2018) 60–74

Table 6 Results, first wave sampling.

Model Missing observations treated as missing

(1) (2) (3) (4) (5)

Men AGE ONLY AGE TTD AGE TTD MORBS AGE TTDIV AGE TTDIV MORBS

Age −.01800* −.00454 .00290 −.00953 −.01124 (.00932) (.00896) (.00746) (.02617) (.01087)

Age2 .00013** .00003 −.00002 .00007 .00007 (.00006) (0.00006) (.00005) (.00018) (.00008)

log(TTD) −.31565*** −.10098*** −.35626 −.13120 (.00655) (.00616) (.36994) (.30968)

Relevance F-statistic 16.38 13.24 Morbidities Included Included

Women Age −.00269 .01198 .00065 .08939 −.01118

(.00832) (.00813) (.00696) (.06773) (.01124) Age2 .00004 −.00005 −.00001 −.00059 .00007

(.00005) (.00005) (.00004) (.00047) (.00009) log(TTD) −.26423*** −.09307*** −1.82773 −.12894

(.00663) (.00612) (1.2711) (.3239383) Relevance F-statistic 66.95 11.67 Morbidities Included Included

Model Missing observations treated as zeros

(1) (2) (3) (4) (5)

Men Age −.06500*** .10191*** −.00691 .28057*** .03547

(.02465) (.02265) (.01462) (.09001) (.06062) Age2 .00049*** −.00088*** .00001 −.00363*** −.00036

(.00016) (.00015) (.00010) (.00113) (.00058) log(TTD) −1.19842*** −.18669*** −2.05176*** −.31020

(.01205) (.00745) (.38267) (.23856) Relevance F-statistic 80.44 71.30 Morbidities Included Included

Women Age .06863*** .08155*** .02635*** .61590 .05830

(.00928) (.00934) (.00806) (.40329) (.05937) Age2 −.00039*** −.00048*** −.00016*** −.00393 −.00038

(.00928) (.00006) (.00005) (.00261) (.00040) log(TTD) −.15705*** −.03723*** −6.43066 −.64920

(.00801) (.00716) (4.95179) (.8581702) Relevance F-statistic 13.09 11.81 Morbidities Included Included

p u m i o 7 r c s

d H f w t b t

a s

expected, first-stage regressions show a negative and significant relationship between YPLL and TTD and an F-test of these instru-

* p < 0.05. ** p < 0.01.

*** p < 1.

ret this as indicating that TTD does indeed serve as a proxy for nobserved morbidity. The estimated coefficients for age when orbidity markers are included see similar falls. This is illustrated

n Fig. 6 which shows the difference in log(HCE) from the quarter f death to preceding quarters for an individual who dies at age 5 for the alternative specifications of the model.17 The combined elationship of time-to-death and age is severely muted when we ondition on current morbidity markers as seen by the lines repre- enting RE AGE TTD MORBS and RE AGE TTD (Table 6).

We anticipate hospital costs to rise as individuals approach eath, and as such expect a negative relationship between TTD and CE. For the sampling strategy where this is possible – sampling

rom the first calendar year – we instrument for TTD in order to deal ith the potential endogeneity of TTD in HCE, which would mean

hat a naïve estimate of the ‘effect’ of TTD on HCE was likely to be iased towards zero (i.e. that naïve estimates would be expected o be less negative). In a further pair of models, we instrument

17 While our instrumented model including morbidity markers appears to show mildly negative relationship between age and HCE, this arises from the use of a mall and non-significant negative age coefficient in construction of this graph.

TTD with LSOA-level YPLL measures, our small-area measure of premature mortality.

When we instrument using YPLL measures – model (4) – the estimated coefficient of log(TTD) rises (in absolute terms) in all cases. While we confirm the findings of Zweifel et al. (2004) that ‘the proximity of death rather than age [being] a main determinant of HCE is fairly robust to endogeneity error,’ our results also suggest that failing to account for the endogeneity of TTD in these models may lead to a large underestimate of the true ‘effect’ of TTD in mod- els that do not include morbidity markers.18 This is also illustrated in Fig. 6, which shows the large divergence in estimated costs for these two models for an individual who dies at the age of 75. As

ments suggests their relevance as a predictor of TTD according to

18 While our point estimates rise, in most cases, however, the TTD coefficients are no longer significant when we carried out our instrumented regressions. We do not rely on these results in our conclusions regarding the relationship between TTD, morbidity, and HCE, and present these results to demonstrate its consistency with existing research.

D. Howdon, N. Rice / Journal of Health Economics 57 (2018) 60–74 71

by pr

t 1

6

p a o p t t m o l l o o t ( t m A f g

b i b ‘ t m i – i o a f a a p

Fig. 7. Healthcare expenditures

he commonly used Stock-Yogo ‘rule of thumb’ of an F-statistic of 0 in all cases.

. Conclusions

Ageing populations pose a substantial problem for public service rovision, particularly for health and social care. Estimates of how n ageing population will impact HCEs vary considerably. Devel- ping credible predictions is a core component of health systems lanning as is allocating resources efficiently and equitably to meet he health care needs of the population. Whilst it is undeniable hat health care costs will rise as the baby-boomers age, the impact

ight not be quite as large as models based on a simple extrap- lation of a crude age-expenditure curve suggests. As individuals ive longer, all other things equal, they may generate larger cumu- ative life-time costs. The extent to which this becomes a burden n the health care sector will depend on how morbidity profiles f cohorts change over time. Should a compression of morbidity hesis hold, Fries (1980), Freedman et al. (2002) and Romeu Gordo 2011), on average individuals can expect to live longer and delay he onset of morbidity into later years. This will have the effect of

oving the age-expenditure curve to the right as populations age. n expansion of morbidity would have more severe consequences

or HCEs with individuals living longer, but also experiencing a reater number of years in ill-health.

Our findings support other literature that it is not age per se, ut time-to-death (TTD), particularly the final year of life, that

s a strong driver of HCEs. Our results regarding the relationship etween age, TTD and HCE are in line with existing results in the

red herring’ strange of literature. We extend this existing literature o show that, in line with the economic implications of biological

odels of ageing as drawn out by Dalgaard and Strulik (2014), TTD n large part proxies for morbidity in explaining HCE. Our results

showing a weak relationship between HCE and age when TTD is ncluded – fall in line with existing research into the determinants f HCE for ageing populations. However, while TTD clearly plays n important role in explaining HCEs, it is unhelpful in forecasting

uture expenditure needs. At an individual level TTD is unknown nd hence to forecast future expenditure growth assumptions bout the proportions of decedents and survivors together with rojections of populations within age groups is required. By extend-

oximity to death, males by age.

ing the modelling of HCE to include morbidity characteristics we show that the impact of TTD is diminished indicating that it acts as a proxy for underlying health status. This is important to allow the planning of future resource requirements and in developing appro- priate models for budgets to be allocated equitably across providers of care in response to population health care need. Our results are robust to problems of endogeneity that exist between HCE and TTD.

Our results strengthen the need to include measures of mor- bidity in models of HCE. Merely including TTD is insufficient in predicting future HCE. To accurately forecast future expenditure needs, information on changes to profiles of morbidity are required. The existence of a compression of morbidity, along with a tendency for increased life expectancy, suggests competing and opposing pressures on HCE. While increases in life expectancy suggests that a greater number of individuals will be alive at any given age, with associated upward pressure on HCE, a compression of morbidity will tend to, on average, provide downward pressure on HCE for any given individual at any given age.

This work has focused on determinants of the demand for inpatient health care services at an individual level via age, time- to-death and morbidity characteristics. Clearly there is also a substantial role for supply-side impacts on expenditure growth notably through technological advances in health care interven- tions and the way in which health care services are organized and delivered. We do not address these issues here, but are areas that warrant further investigation at an aggregate level. Inpatient hos- pital care is one of a number of services provided by the National Health Service in England and other expenditure should also be taken into account when assessing the overall impact of an ageing population, as should costs placed on the Government by long- term care services predominantly accessed by older age groups. The increasing ability to link administrative sources of data provides a potentially valuable resource for future research in this area.

Acknowledgements

This is an independent report commissioned and funded by

the Policy Research Programme in the Department of Health from the Economics of Social and Health Care Research Unit (ESHCRU). ESHCRU is a joint collaboration between the University of York, London School of Economics and University of Kent. The views

7 Health Economics 57 (2018) 60–74

e f n ( c

A

T C

Table A1 (Continued)

CCS code Description

61 Sickle cell anemia 62 Coagulation and hemorrhagic disorders 63 Diseases of white blood cells 64 Other hematologic conditions 65 Mental retardation 66 Alcohol-related mental disorders 67 Substance-related mental disorders 68 Senility and organic mental disorders 69 Affective disorders 70 Schizophrenia and related disorders 71 Other psychoses 72 Anxiety; somatoform; dissociative; and personality disorders 73 Preadult disorders 74 Other mental conditions 75 Personal history of mental disorder; mental and behavioral

problems; observation and screening for mental condition 76 Meningitis (except that caused by tuberculosis or sexually

transmitted disease) 77 Encephalitis (except that caused by tuberculosis or sexually

transmitted disease) 78 Other CNS infection and poliomyelitis 79 Parkinson’s disease 80 Multiple sclerosis 81 Other hereditary and degenerative nervous system conditions 82 Paralysis 83 Epilepsy; convulsions 84 Headache; including migraine 85 Coma; stupor; and brain damage 86 Cataract 87 Retinal detachments; defects; vascular occlusion; and

retinopathy 88 Glaucoma 89 Blindness and vision defects 90 Inflammation; infection of eye (except that caused by

tuberculosis or sexually transmitted disease) 91 Other eye disorders 92 Otitis media and related conditions 93 Conditions associated with dizziness or vertigo 94 Other ear and sense organ disorders 95 Other nervous system disorders 96 Heart valve disorders 97 Peri-; endo-; and myocarditis; cardiomyopathy (except that

caused by tuberculosis or sexually transmitted disease) 98 Essential hypertension 99 Hypertension with complications and secondary hypertension 100 Acute myocardial infarction 101 Coronary atherosclerosis and other heart disease 102 Nonspecific chest pain 103 Pulmonary heart disease 104 Other and ill-defined heart disease 105 Conduction disorders 106 Cardiac dysrhythmias 107 Cardiac arrest and ventricular fibrillation 108 Congestive heart failure; nonhypertensive 109 Acute cerebrovascular disease 110 Occlusion or stenosis of precerebral arteries 111 Other and ill-defined cerebrovascular disease 112 Transient cerebral ischemia 113 Late effects of cerebrovascular disease 114 Peripheral and visceral atherosclerosis 115 Aortic; peripheral; and visceral artery aneurysms 116 Aortic and peripheral arterial embolism or thrombosis 117 Other circulatory disease 118 Phlebitis; thrombophlebitis and thromboembolism 119 Varicose veins of lower extremity 120 Hemorrhoids 121 Other diseases of veins and lymphatics 122 Pneumonia (except that caused by tuberculosis or sexually

transmitted disease) 123 Influenza 124 Acute and chronic tonsillitis 125 Acute bronchitis

2 D. Howdon, N. Rice / Journal of

xpressed are those of the authors and may not reflect those of the unders. Daniel Howdon acknowledges PhD funding from the Eco- omic and Social Research Council under the Large Grant Scheme RES-060-25-0045). The authors gratefully acknowledge the highly onstructive remarks of two anonymous peer reviewers.

ppendix A.

See Fig. 7 and Table A1.

able A1 linical classifications software (CCS) groupings.

CCS code Description

1 Tuberculosis 2 Septicemia (except in labor) 3 Bacterial infection; unspecified site 4 Mycoses 5 HIV infection 6 Hepatitis 7 Viral infection 8 Other infections; including parasitic 9 Sexually transmitted infections (not HIV or hepatitis) 10 Immunizations and screening for infectious disease 11 Cancer of head and neck 12 Cancer of esophagus 13 Cancer of stomach 14 Cancer of colon 15 Cancer of rectum and anus 16 Cancer of liver and intrahepatic bile duct 17 Cancer of pancreas 18 Cancer of other GI organs; peritoneum 19 Cancer of bronchus; lung 20 Cancer; other respiratory and intrathoracic 21 Cancer of bone and connective tissue 22 Melanomas of skin 23 Other non-epithelial cancer of skin 24 Cancer of breast 25 Cancer of uterus 26 Cancer of cervix 27 Cancer of ovary 28 Cancer of other female genital organs 29 Cancer of prostate 30 Cancer of testis 31 Cancer of other male genital organs 32 Cancer of bladder 33 Cancer of kidney and renal pelvis 34 Cancer of other urinary organs 35 Cancer of brain and nervous system 36 Cancer of thyroid 37 Hodgkin’s disease 38 Non-Hodgkin’s lymphoma 39 Leukemias 40 Multiple myeloma 41 Cancer; other and unspecified primary 42 Secondary malignancies 43 Malignant neoplasm without specification of site 44 Neoplasms of unspecified nature or uncertain behavior 45 Maintenance chemotherapy; radiotherapy 46 Benign neoplasm of uterus 47 Other and unspecified benign neoplasm 48 Thyroid disorders 49 Diabetes mellitus without complication 50 Diabetes mellitus with complications 51 Other endocrine disorders 52 Nutritional deficiencies 53 Disorders of lipid metabolism 54 Gout and other crystal arthropathies 55 Fluid and electrolyte disorders 56 Cystic fibrosis 57 Immunity disorders

58 Other nutritional; endocrine; and metabolic disorders 59 Deficiency and other anemia 60 Acute posthemorrhagic anemia

126 Other upper respiratory infections 127 Chronic obstructive pulmonary disease and bronchiectasis 128 Asthma 129 Aspiration pneumonitis; food/vomitus 130 Pleurisy; pneumothorax; pulmonary collapse

D. Howdon, N. Rice / Journal of Health Economics 57 (2018) 60–74 73

Table A1 (Continued)

CCS code Description

131 Respiratory failure; insufficiency; arrest (adult) 132 Lung disease due to external agents 133 Other lower respiratory disease 134 Other upper respiratory disease 135 Intestinal infection 136 Disorders of teeth and jaw 137 Diseases of mouth; excluding dental 138 Esophageal disorders 139 Gastroduodenal ulcer (except hemorrhage) 140 Gastritis and duodenitis 141 Other disorders of stomach and duodenum 142 Appendicitis and other appendiceal conditions 143 Abdominal hernia 144 Regional enteritis and ulcerative colitis 145 Intestinal obstruction without hernia 146 Diverticulosis and diverticulitis 147 Anal and rectal conditions 148 Peritonitis and intestinal abscess 149 Biliary tract disease 150 Liver disease; alcohol-related 151 Other liver diseases 152 Pancreatic disorders (not diabetes) 153 Gastrointestinal hemorrhage 154 Noninfectious gastroenteritis 155 Other gastrointestinal disorders 156 Nephritis; nephrosis; renal sclerosis 157 Acute and unspecified renal failure 158 Chronic renal failure 159 Urinary tract infections 160 Calculus of urinary tract 161 Other diseases of kidney and ureters 162 Other diseases of bladder and urethra 163 Genitourinary symptoms and ill-defined conditions 164 Hyperplasia of prostate 165 Inflammatory conditions of male genital organs 166 Other male genital disorders 167 Nonmalignant breast conditions 168 Inflammatory diseases of female pelvic organs 169 Endometriosis 170 Prolapse of female genital organs 171 Menstrual disorders 172 Ovarian cyst 173 Menopausal disorders 174 Female infertility 175 Other female genital disorders 176 Contraceptive and procreative management 177 Spontaneous abortion 178 Induced abortion 179 Postabortion complications 180 Ectopic pregnancy 181 Other complications of pregnancy 182 Hemorrhage during pregnancy; abruptio placenta; placenta

previa 183 Hypertension complicating pregnancy; childbirth and the

puerperium 184 Early or threatened labor 185 Prolonged pregnancy 186 Diabetes or abnormal glucose tolerance complicating

pregnancy; childbirth; or the puerperium 187 Malposition; malpresentation 188 Fetopelvic disproportion; obstruction 189 Previous C-section 190 Fetal distress and abnormal forces of labor 191 Polyhydramnios and other problems of amniotic cavity 192 Umbilical cord complication 193 OB-related trauma to perineum and vulva 194 Forceps delivery 195 Other complications of birth; puerperium affecting

management of mother 196 Normal pregnancy and/or delivery 197 Skin and subcutaneous tissue infections 198 Other inflammatory condition of skin 199 Chronic ulcer of skin 200 Other skin disorders 201 Infective arthritis and osteomyelitis (except that caused by

tuberculosis or sexually transmitted disease)

Table A1 (Continued)

CCS code Description

202 Rheumatoid arthritis and related disease 203 Osteoarthritis 204 Other non-traumatic joint disorders 205 Spondylosis; intervertebral disc disorders; other back

problems 206 Osteoporosis 207 Pathological fracture 208 Acquired foot deformities 209 Other acquired deformities 210 Systemic lupus erythematosus and connective tissue disorders 211 Other connective tissue disease 212 Other bone disease and musculoskeletal deformities 213 Cardiac and circulatory congenital anomalies 214 Digestive congenital anomalies 215 Genitourinary congenital anomalies 216 Nervous system congenital anomalies 217 Other congenital anomalies 218 Liveborn 219 Short gestation; low birth weight; and fetal growth retardation 220 Intrauterine hypoxia and birth asphyxia 221 Respiratory distress syndrome 222 Hemolytic jaundice and perinatal jaundice 223 Birth trauma 224 Other perinatal conditions 225 Joint disorders and dislocations; trauma-related 226 Fracture of neck of femur (hip) 227 Spinal cord injury 228 Skull and face fractures 229 Fracture of upper limb 230 Fracture of lower limb 231 Other fractures 232 Sprains and strains 233 Intracranial injury 234 Crushing injury or internal injury 235 Open wounds of head; neck; and trunk 236 Open wounds of extremities 237 Complication of device; implant or graft 238 Complications of surgical procedures or medical care 239 Superficial injury; contusion 240 Burns 241 Poisoning by psychotropic agents 242 Poisoning by other medications and drugs 243 Poisoning by nonmedicinal substances 244 Other injuries and conditions due to external causes 245 Syncope 246 Fever of unknown origin 247 Lymphadenitis 248 Gangrene 249 Shock 250 Nausea and vomiting 251 Abdominal pain 252 Malaise and fatigue 253 Allergic reactions 254 Rehabilitation care; fitting of prostheses; and adjustment of

devices 255 Administrative/social admission 256 Medical examination/evaluation 257 Other aftercare 258 Other screening for suspected conditions (not mental

disorders or infectious disease)

259 Residual codes; unclassified 260 E Codes: All (external causes of injury and poisoning)

References

Aragon, M.J., Chalkley, M., Rice, N., 2016. Medical spending and hospital inpatient care in England: an analysis over time. Fisc. Stud.

Blane, D., Drever, F., 1998. Inequality among men in standardised years of potential life lost, 1970–93. BMJ 317 (7153), 255–256.

Breyer, F., Lorenz, N., Niebel, T., 2015. Health care expenditures and longevity: is there a Eubie Blake effect? Eur. J. Health Econ. 16 (1), 95–112.

Cracknell, R., 2010. The Ageing Population, Briefing Papers. House of Commons

Library, United Kingdom.

Cutler, D.M., Ghosh, K., Landrum, M.B., 2013. Evidence for Significant Compression of Morbidity in the Elderly U.S. Population, Working Paper 19268. National Bureau of Economic Research.

7 Healt

D

D

d

D

D

D D

D

F

F

F

F

F

F

G

G

G

G H

H

H

I

J

Zweifel, P., Felder, S., Meiers, M., 1999. Ageing of population and health care expenditure: a red herring? Health Econ. 8 (6), 485–496.

Zweifel, P., Felder, S., Werblow, A., 2004. Population ageing and health care expenditure: new evidence on the “Red Herring”. Geneva Pap. Risk Insur.

4 D. Howdon, N. Rice / Journal of

algaard, C.-J., Strulik, H., 2014. Optimal aging and death: understanding the Preston curve. J. Eur. Econ. Assoc. 12 (3), 672–701.

ang, T., Antolin, P., Oxley, H., 2001. Fiscal Implications of Ageing: Projections of Age-Related Spending. Working Papers No. 305. Economics Department, Organisation for Economic Co-operation and Development, Paris.

e Meijer, C., Koopmanschap, M., d’Uva, T.B., van Doorslaer, E., 2011. Determinants of long-term care spending: age, time to death or disability? J. Health Econ. 30 (2), 425–438.

epartment for Communities and Local Government, 2008. The English Indices of Deprivation 2007.

epartment for Communities and Local Government, 2011. The English Indices of Deprivation 2010.

epartment of Health, 2012. Reference Costs 2011-12. ixon, J., Smith, P., Gravelle, H., Martin, S., Bardsley, M., Rice, N., Georghiou, T.,

Dusheiko, M., Billings, J., De Lorenzo, M., Sanderson, C., 2011. A person based formula for allocating commissioning funds to general practices in England: development of a statistical model. BMJ, 343.

ormont, B., Grignon, M., Huber, H., 2006. Health expenditure growth: reassessing the threat of ageing. Health Econ. 15 (9), 947–963.

elder, S., Werblow, A., Zweifel, P., 2010. Do red herrings swim in circles? Controlling for the endogeneity of time to death. J. Health Econ. 29 (2), 205–212.

reedman, V.A., Martin, L.G., Schoeni, R.F., 2002. Recent trends in disability and functioning among older adults in the United States: a systematic review. J. Am. Med. Assoc. 288 (24), 3137–3146.

rench, E.B., McCauley, J., Aragon, M., Bakx, P., Chalkley, M., Chen, S.H., Christensen, B.J., Chuang, H., Côtó-Sergent, A., Nardi, M.D., Fan, E., Échevin, D., Geoffard, P.-Y., Gastaldi-Ménager, C., Gørtz, M., Ibuka, Y., Jones, J.B., Kallestrup-Lamb, M., Karlsson, M., Klein, T.J., Lagasnerie, G.D., Michaud, P.-C., O’Donnell, O., Rice, N., Skinner, J.S., Doorslaer, E.v., Ziebarth, N.R., Kelly, E., 2017. End-of-life medical spending in last twelve months of life is lower than previously reported. Health Aff. 36 (7), 1211–1217.

ries, J.F., 1980. Aging, natural death, and the compression of morbidity. N. Engl. J. Med. 303 (3), 130–135, PMID: 7383070.

ries, J.F., 2005. The compression of morbidity. Milbank Q. 83 (4), 801–823, PMID: 16279968; PMCID: PMC2690269.

ries, J.F., Bruce, B., Chakravarty, E., 2011. Compression of morbidity 1980–2011: a focused review of paradigms and progress. J. Aging Res. 2011, e261702.

heorghe, M., Brouwer, W.B.F., van Baal, P.H.M., 2015. Quality of life and time to death: have the health gains of preventive interventions been underestimated? Med. Decis. Mak. 35 (3), 316–327.

ray, A., 2005. Population ageing and health care expenditure. Ageing Horiz. 2, 15–20.

rossman, M., 1972. On the concept of health capital and the demand for health. J. Polit. Econ. 80 (2), 223–255, Article Type: research-article/Full publication date: Mar-Apr., 1972/Copyright 1972. The University of Chicago Press.

ruenberg, E.M., 2005. The failures of success. Milbank Q. 83 (4), 779–800. arker, R., 2012. NHS Funding and Expenditure, Standard Note. House of

Commons Library, United Kingdom. ealth & Social Care Information Centre, 2013. Hospital Episode Statistics (HES)

Analysis Guide. http://www.hscic.gov.uk/media/1592/HES-analysisguide/pdf/ HES Analysis Guide Jan 2014.pdf.

eller, P.S., Hemming, R., Kohnert, P., Feldman, R.A., 1986. Aging and Social Expenditure in the Major Industrial Countries, 1980–2025. International Monetary Fund.

ezzoni, L.I., Ayanian, J.Z., Bates, D.W., Burstin, H.R., 1998. Paying more fairly for Medicare capitated care. N. Engl. J. Med. 339 (26), 1933–1938.

acobzone, S., Cambois, E., Robine, J., 2000. Is the health of older persons in OECD countries improving fast enough to compensate for population ageing? OECD Econ. Stud., 30.

h Economics 57 (2018) 60–74

Jonsen, A.R., 1986. Bentham in a box: technology assessment and health care allocation. J. Law Med. Ethics 14 (3–4), 172–174.

Karlsson, M., Klohn, F., 2014. Testing the red herring hypothesis on an aggregated level: ageing, time-to-death and care costs for older people in Sweden. Eur. J. Health Econ. 15 (5), 533–551.

Lansley, H.A., 2012. Writing exclusively for BJN, Lansley dismisses Bill critics. Br. J. Nurs. 21 (6), 333.

Mitnitski, A.B., Mogilner, A.J., MacKnight, C., Rockwood, K., 2002a. The accumulation of deficits with age and possible invariants of aging. Sci. World J. 2, 1816–1822.

Mitnitski, A.B., Mogilner, A.J., MacKnight, C., Rockwood, K., 2002b. The mortality rate as a function of accumulated deficits in a frailty index. Mech. Ageing Dev. 123 (11), 1457–1460.

Mitnitski, A., Song, X., Skoog, I., Broe, G.A., Cox, J.L., Grunfeld, E., Rockwood, K., 2005. Relative fitness and frailty of elderly men and women in developed countries and their relationship with mortality. J. Am. Geriatr. Soc. 53 (12), 2184–2189.

Office for National Statistics, 2011. Super Output Areas (SOAs). In: Super Output Areas are a Geography Designed to Improve the Reporting of Small Area Statistics. http://www.ons.gov.uk/ons/guide-method/geography/beginner-s- guide/census/superoutput-areas-soas-/index.html.

Pope, G.C., Kautter, J., Ellis, R.P., Ash, A.S., Ayanian, J.Z., Lezzoni, L.I., Ingber, M.J., Levy, J.M., Robst, J., 2004. Risk adjustment of Medicare capitation payments using the CMS-HCC model. Health Care Financ. Rev. 25 (4), 119–141.

Rockwood, K., Mitnitski, A., 2006. Limits to deficit accumulation in elderly people. Mech. Ageing Dev. 127 (5), 494–496.

Rockwood, K., Mitnitski, A., 2007. Frailty in relation to the accumulation of deficits. J. Gerontol. Ser. A: Biol. Sci. Med. Sci. 62 (7), 722–727.

Romeu Gordo, L., 2011. Compression of morbidity and the labour supply of older people. Appl. Econ. 43 (4), 503–513.

Seshamani, M., Gray, A., 2004. Ageing and health-care expenditure: the red herring argument revisited. Health Econ. 13 (4), 303–314.

Shwartz, M., Iezzoni, L.I., Moskowitz, M.A., Ash, A.S., Sawitz, E., 1996. The importance of comorbidities in explaining differences in patient costs. Med. Care 34 (8), 767–782.

Stearns, S.C., Norton, E.C., 2004. Time to include time to death? The future of health care expenditure predictions. Health Econ. 13 (4), 315–327.

US Agency for Healthcare Research and Quality, 2009. Clinical Classifications Software (CCS) for ICD-10. http://www.hcup-us.ahrq.gov/toolssoftware/icd 10/ccs icd 10.jsp.

van de Ven, W.P.M.M., Beck, K., Buchner, F., Chernichovsky, D., Gardiol, L., Holly, A., Lamers, L.M., Schokkaert, E., Shmueli, A., Spycher, S., Van de Voorde, C., van Vliet, R.C.J.A., Wasem, J., Zmora, I., 2003. Risk adjustment and risk selection on the sickness fund insurance market in five European countries. Health Policy (Amst. Neth.) 65 (1), 75–98.

Werblow, A., Felder, S., Zweifel, P., 2007. Population ageing and health care expenditure: a school of ‘red herrings’? Health Econ. 16 (10), 1109–1126.

Williams, D., 2012. Lansley: CCG allocations should be based on age, not poverty. Health Serv. J.

Wong, A., van Baal, P.H.M., Boshuizen, H.C., Polder, J.J., 2011. Exploring the influence of proximity to death on disease-specific hospital expenditures: a Carpaccio of red herrings. Health Econ. 20 (4), 379–400.

Issues Pract. 29 (4), 652–666.