RESEARCH ARTICLE
The disease burden of multimorbidity and its
interaction with educational level
Yi Hsuan Chen
ID
1☯
*, Milad Karimi
1¤a‡
, Maureen P. M. H. Rutten-van Mo
¨
lken
ID
1,2¤b‡
1 Erasmus School of Health Policy and Management, Erasmus University Rotterdam, Rotterdam, The
Netherlands, 2 Institute for Medical Technology Assessment, Erasmus University Rotterdam, Rotterdam,
The Netherlands
☯ These authors contributed equally to this work.
¤a Current address: Erasmus School of Health Policy and Management, Erasmus University Rotterdam,
Rotterdam, The Netherlands
¤b Current address: Institute for Medical Technology Assessment, Erasmus University Rotterdam,
Rotterdam, The Netherlands
‡ These authors also contributed equally to this work.
* chen@eshpm.eur.nl
Abstract
Introduction
Policies to adequately respond to the rise in multimorbidity have top-priority. To understand
the actual burden of multimorbidity, this study aimed to: 1) estimate the trend in prevalence
of multimorbidity in the Netherlands, 2) study the association between multimorbidity and
physical and mental health outcomes and healthcare cost, and 3) investigate how the asso-
ciation between multimorbidity and health outcomes interacts with socio-economic status
(SES).
Methods
Prevalence estimates were obtained from a nationally representative pharmacy database
over 2007–2016. Impact on costs was estimated in a fixed effect regression model on claims
data over 2009–2015. Data on physical and mental health and SES were obtained from the
National Health Survey in 2017, in which the Katz-10 was used to measure limitations in
activities of daily living (ADL) and the Mental Health Inventory (MHI) to measure mental
health. SES was approximated by the level of education. Generalized linear models (2-part
models for ADL) were used to analyze the health data. In all models an indicator variable for
the presence or absence of multimorbidity was included or a categorical variable for the
number of chronic conditions. Interactions terms of multimorbidity and educational level
were added into the previously mentioned models.
Results
Over the past ten years, there was an increase of 1.6%-point in the percentage of people
with multimorbidity. The percentage of people with three or more conditions increased with
+2.1%-point. People with multimorbidity had considerably worse physical and mental health
outcomes than people without multimorbidity. For the ADL, the impact of multimorbidity was
PLOS ONE
PLOS ONE | https://doi.org/10.1371/journal.pone.0243275 December 3, 2020 1 / 18
a1111111111
a1111111111
a1111111111
a1111111111
a1111111111
OPEN ACCESS
Citation: Chen YH, Karimi M, Rutten-van Mo¨lken
MPMH (2020) The disease burden of
multimorbidity and its interaction with educational
level. PLoS ONE 15(12): e0243275. https://doi.org/
10.1371/journal.pone.0243275
Editor: Juan F. Orueta, Osakidetza Basque Health
Service, SPAIN
Received: January 30, 2020
Accepted: November 18, 2020
Published: December 3, 2020
Copyright: © 2020 Chen et al. This is an open
access article distributed under the terms of the
Creative Commons Attribution License, which
permits unrestricted use, distribution, and
reproduction in any medium, provided the original
author and source are credited.
Data Availability Statement: Data cannot be
shared publicly because of legal restrictions. The
databases employed in this study were provided by
Netherlands statistic (CBS). The researchers do not
own the databases but have access to them during
a given period. Researchers can obtain remote
access to those databases after their research
proposal was approved. The researchers need to
conduct their analysis under remote control
environment and make requests for analysis
results output to the CBS. Only after CBS has
ensured that the results tables do not contain any
information that might jeopardize personal
three times greater in the lowest educational level than in the highest educational level. For
the MHI, the impact of multimorbidity was two times greater in the lowest than in the highest
educational level. Each additional chronic condition was associated with a greater worsen-
ing in health outcomes. Similarly, for costs, where there was no evidence of a diminishing
impact of additional conditions either. In patients with multimorbidity total healthcare costs
were on average €874 higher than in patients with a single morbidity.
Conclusion
The impact of multimorbidity on health and costs seems to be greater in the sicker and lower
educated population.
Introduction
Multimorbidity, defined as two or more chronic conditions occurring in one person at the
same time, is an increasing problem, particularly in countries with a rapidly ageing population
[1, 2]. The number of patients with multimorbidity in Europe is around 50 million and is
increasing every year [2]. Compared to patients without multimorbidity, those with multimor-
bidity have a higher mortality rate [3, 4], a higher rate of polypharmacy [5], worse health-
related quality of life (HRQoL) [6–8], higher healthcare costs [9–11], and more productivity
loss [12] after adjustment of age and other confounders. Moreover, they are disproportionately
affected by the fragmentation and the single-disease orientation of the health and social care
system [13]. The need for investments in potential solutions such as whole-system, person-
centred integrated care is widely recognized. To justify and plan such investments and ensure
access for all people with multimorbidity across the layers of the population, it is crucial to
have accurate data on the (change in) prevalence of multimorbidity and its impact on physical
and mental health, and costs.
Prevalence estimates of multimorbidity and their increase with age and over time vary con-
siderably due to differences in definitions of multimorbidity and sources of data used [14–18].
The most up to date prevalence estimate for the Netherlands was 31.8% in 2018 [19], which
was based on a general practitioner (GP) registry. An estimate of (change in) prevalence that is
based on a national database representing the entire Dutch population is lacking.
Several studies have investigated the association between multimorbidity and health out-
comes, both physical [6–8, 20–22], and mental [6, 8, 23, 24], but many were either small [7,
21], not representative for the entire country [8, 23], focused on a specific subgroup like elderly
[6, 7, 21, 22], or focused on a limited number of chronic diseases [24–26]. There are many
papers on socio-economic inequalities in health, which clearly demonstrate that lower SES is
associated with greater mortality [27, 28], higher prevalence of multimorbidity [29] and worse
health outcomes [30] To the best of our knowledge, none of them investigated how the associa-
tion between multimorbidity and health varies by SES.
The impact of multimorbidity on costs varies by healthcare system and methods used to
derive costs [10, 31]. In a recent review study, the ratios of multimorbidity costs to non-multi-
morbidity costs ranged from 2–16 [32]. Most of the studies in this review were cross-sectional
and half of them were based on U.S. data. There is no longitudinal study on the impact of mul-
timorbidity on healthcare costs in the Netherlands.
The aim of our study was to fill in the information gaps identified above. This study pro-
vides recent estimates of the prevalence of multimorbidity in the Netherlands and its change
PLOS ONE
The disease burden of multimorbidity and its interaction with educational level
PLOS ONE | https://doi.org/10.1371/journal.pone.0243275 December 3, 2020 2 / 18
information protection, will they send the requested
results outputs to the researchers. The contact
information for CBS: Address: Henri Faasdreef 312
2492JP The Hague Telephone: (+31) 45 570 64 00
Email: contactcenter@cbs.nl Future researchers
requesting access to the data from Netherlands
Statistics should request access through this link:
https://www.cbs.nl/en-gb/our-services/
customised-services-microdata/microdata-
conducting-your-own-research.
Funding: The author(s) received no specific
funding for this work.
Competing interests: The authors have declared
that no competing interests exist.
over time. Furthermore, we investigated the association between multimorbidity and physical
health, mental health and healthcare costs, using databases that are representative for the entire
country. In addition, we investigated how educational level interacts with the association
between multimorbidity and health.
Methods
We have used data from a nationally representative pharmacy database, claims database and a
health survey to estimate the change in prevalence of multimorbidity and the impact of multi-
morbidity on health and healthcare costs. All databases used in this study were provided by
Statistics Netherlands (CBS, Centraal Bureau voor de Statistiek), the national statistics office of
the Netherlands. The study population was the entire adult population (18 years old and over)
in the Netherlands. The prevalence estimation was based on data from 2006 to 2016, the asso-
ciation between multimorbidity and health outcomes was based on data from 2017, and the
estimated impact of multimorbidity on healthcare costs was based on panel data from 2009 to
2015.
Identification of chronic conditions and prevalence of multimorbidity
To estimate the prevalence, we used a longitudinal pharmacy database that contained data
from everyone in the Netherlands that received medication included in the basic health insur-
ance package, which is approximately 11 million adults (85% of total population) each year
[33]. In the Netherlands, every resident is mandatorily enrolled in basic health insurance. All
prescribed medicines covered by the basic insurance, including those prescribed by GP’s, to
hospital outpatients and patients in residential homes, are recorded in this dataset. The medi-
cation prescribed during in-patient hospital stays and over-the-counter medicines are not
included in the pharmacy database.
The individual-level prescriptions of medications are coded and categorized using the Ana-
tomical Therapeutic Chemical (ATC) classification system, up to the third level, which is the
pharmacological subgroup level [34]. We matched these ATC codes to a list of 21 chronic con-
ditions: acid-related disorders, bone diseases, cancer, cardiovascular diseases, dementia, diabe-
tes mellitus, epilepsy, glaucoma, hyperuricemia (gout), chronic viral infection, hyperlipidemia,
intestinal inflammatory diseases, iron deficiency anemia, migraines, pain, Parkinson’s disease,
psychological disorders, psychoses, respiratory illness, thyroid disorders, and tuberculosis. To
do so we used the Swiss classification by Huber et al. [35], which was an updated version of the
original Dutch classification by Lamers et al. [36], that was used in the risk equalization scheme
of the Dutch health insurance. Hence this classification was deemed suitable to be used in a
study addressing the entire adult population in the Netherlands.
In line with the WHO definition of multimorbidity, adults with two or more chronic condi-
tions in the same year were defined as having multimorbidity [37]. To avoid overestimating
the prevalence of chronic conditions, we only labeled a condition as chronic when medication
for a certain chronic disease was prescribed at least two years in a row. This continuity crite-
rium was used by previous studies [38, 39].
To estimate the prevalence of multimorbidity by gender and age, the pharmacy database
was linked to the municipal basic administration (GBA, Gemeentelijke Basis Administratie)
database from CBS, which contains demographic information (year of birth and gender) for
all residents in the Netherlands [40]. Data were linked by CBS, using a pseudonymized com-
mon ID that is unique for each individual. Information on annual size of the total population
in the Netherlands was extracted from StatLine [41].
PLOS ONE
The disease burden of multimorbidity and its interaction with educational level
PLOS ONE | https://doi.org/10.1371/journal.pone.0243275 December 3, 2020 3 / 18
Health outcomes
To estimate the association between multimorbidity and health outcomes, which include phys-
ical and mental health outcomes, we used data from the Health Survey database [42]. This
database contains data from a nation-wide, cross-sectional, annual Health Survey on health
and the need for healthcare in the Netherlands, including self-reported physical and mental
health, and demographic information such as age, gender, and highest education level.
Approximately 15,000 participants across the country are sampled for the survey each year,
with a response rate of 63% in 2017 (approximately 9,500 respondents), from which we
selected all 7,741 respondents that were 18 years and older in the year 2017. The average age of
this dataset is 52.8 years old, 50.5% of the respondents are male (S3 Table). The sample is dis-
tributed evenly over all months of the year. First, persons are asked to participate via the inter-
net. Non-responders are re-approached to participate in a face-to-face interview by way of
Computer-Assisted Personal Interviewing. A correction is applied to control for differences
between the sample and the population. For this purpose, a weighting factor is used based on
sex, age, migration background, marital status, degree of urbanisation, province, household
size, income, wealth, and survey season. The background characteristics of the Health Survey
responders were proven by CBS to be similar to the general population in the Netherlands
[42].
Physical health was measured with the Katz-10. This is a validated questionnaire [43, 44],
which consists of ten questions asking respondents whether they need help with ADL such as
eating, dressing, personal hygiene, mobility in getting up from bed, dressing, transferring, and
use of the toilet [45]. The five response options ranged from no difficulty to not even with the
help of others. Participants under 54 years only had to answer three of these ten questions, i.e.,
about their ability to get up from a chair, to get up from a bed and walk up a flight of stairs. For
both age groups, responses were linearly transformed onto a scale from 0 (best) to 100 (worst)
[46].
Mental health was measured by the Mental Health Inventory 5 (MHI-5). This is a validated
questionnaire [47], which consists of five questions asking how much of the time during the
last four weeks the respondents had felt happy, calm and peaceful, nervous, downhearted and
blue, and so down in the dumps that nothing could cheer them up [48]. The six response
options ranged from none of the time to all of the time. Responses were linearly transformed
onto a scale from 0 (worst) to 100 (best) [49].
The Health Survey database also contains age, gender, and highest education level achieved,
which was classified into five levels, i.e. primary education; pre-vocational training; high school
or vocational training; higher education until Bachelor; and master/doctorate [50]. The multi-
morbidity status of the Health Survey respondents was obtained by linking to the prevalence
data (2016). In all the databases provided by CBS, each individual has one unique, encrypted
RIN code. This RIN code enable us to link several databases on individual level.
Healthcare costs
To estimate the impact of multimorbidity on healthcare expenditure, we used a database of
health insurance data, containing annual individual-level claims data from 2009 to 2015 [51].
Sixteen categories of costs are included, that cover all expenditures reimbursed by the compul-
sory basic health insurance in the Netherlands, including GP care, hospital care, dental care,
pharmaceuticals, other costs (e.g. paramedic, assistive device and birth care), and total costs.
Costs of GP care include the annual per capita registration fee, the consultation fees and other
costs incurred by the GP. Costs of hospital care include both inpatient and outpatient hospital
care. Pharmacy costs include all the costs of medicines, except for the ones administered
PLOS ONE
The disease burden of multimorbidity and its interaction with educational level
PLOS ONE | https://doi.org/10.1371/journal.pone.0243275 December 3, 2020 4 / 18
during a hospital admission. From the health insurance database, we included only the adult
population (18 or older), which was around 2,450,000 each year in the period from 2009 to
2015 (17,048,049 observations for adults across these years in total).
As healthcare costs increase substantially in the year of death [52, 53], it was essential to
have mortality data to know whether an individual had died in a certain year. Mortality data
for all deceased Dutch residents were obtained from the National Mortality Database, which
includes the date and cause of death of all residents in the Netherlands. End-of-life costs were
defined as the healthcare expenditures during the year of death.
Statistical analyses
Prevalence. For each year from 2007 to 2016, the prevalence of multimorbidity was calcu-
lated as the number of adults with multimorbidity (in a certain age class) divided by the num-
ber of adults (in that age class) living in the Netherlands that year. We further classified
patients according to their number of chronic conditions (0, 1, 2, 3, 4, and 5 or more) and esti-
mated the prevalence of each class.
Health outcomes. To analyse the association between multimorbidity and ADL, we esti-
mated two-part models (2PM), because a large proportion of the people has a score of 0, indi-
cating no problems with their ADL [45]. This makes the conventional ordinary least squares
(OLS) inappropriate. In the first part of the 2PM, we used a logit model to estimate the proba-
bility of having an ADL score higher than 0, including morbidity status (0 for no morbidity, 1
for single morbidity, and 2 for multimorbidity), age, gender and educational level:
P
r
ADL
i
> 0jX
i
ð Þ ¼
e
aþbx
i
1 þ e
aþbx
i
ð1Þ
Where P
r
(ADL
i
> 0|X
i
) is the probability of ADL score to be positive with given age, gender,
multimorbidity status, and education level, α is the constant, and β
i
are parameters in the
model, which are age, gender, multimorbidity status and education level.
The second part of the 2PM estimated the ADL score among those with a non-zero score,
using a generalized linear model with a gamma distribution and log-link function, because the
data were skewed to the right. The choice of the model was driven by the Manning and Mul-
lahy algorithm [54]. Like in the first part, the model included morbidity status, age, gender and
educational level:
lnðEðADL
i
ÞÞ ¼ a þ b x
i
þ ε ð2Þ
Where ε stands for the error term. The results of the 2PM combined the probability estimated
in the first part and the ADL score estimated in the second part:
EðADL
i
jx
i
Þ ¼ P
r
ðADL
i
> 0jX
i
Þ EðADL
i
jx
i
; ADL
i
> 0Þ ð3Þ
To analyse the association between multimorbidity and MHI score, we estimated a general-
ized linear model with the same independent variables as above. Based on the Manning and
Mullahy algorithm [54], the identity link function and normal distribution were chosen in this
analysis.
MHI
i
¼ a þ b x
i
þ ε ð4Þ
To investigate whether the association between multimorbidity and ADL or MHI scores
differed by level of education, interaction variables of morbidity status and level of education
were added to the models above.
PLOS ONE
The disease burden of multimorbidity and its interaction with educational level
PLOS ONE | https://doi.org/10.1371/journal.pone.0243275 December 3, 2020 5 / 18
