Does cognitive impairment lead to more health
care utilisation over time?
Brenda Gannona
Mark N. Harrisb
Leandro Magnussonc
June 1, 2016
Abstract
Abstract
Evaluation of health care utilization is prominent in the economics literature, generally concentrating on the impact of certain health conditions on
resource use, using varied sources of individual level self-reported panel data.
The additional impact of cognitive impairment has not yet been explored.
We now propose a model that assesses the impact of word recall, a standard
validated measure of cognition, on health care visits. Indeed, the model is
further complicated with self-reported data, as this could also in‡uence the
dependent variable. The contribution of this paper is therefore to disentangle
the additional e¤ects of recall bias introduced into the dependent variable by
the main independent variable of interest, cognitive impairment. We propose
a DOGIT model with individual random e¤ects, applied to a Poisson distribution and identify our model …rstly on functional form and secondly by
introducing novel reporting behavior variables.This models any unobserved
heterogeneity relating to digit-preferencing, reporting behaviour and the like.
By applying this model to data from the Survey of Health, Ageing and Retirement in Europe, we …nd that an initial marginal e¤ect of approximately
0.38 of word call on doctor visits, is then reduced to 0.10 after we control for
unobserved e¤ects and reporting behavior. Our model also introduces an age
speci…c random parameter, allowing us to identify the e¤ects by age.
JEL Classi…cation:
a
b
University of Manchester, UK; Curtin University, Australia; c University of Western Australia, Perth, Australia
This research is funded under the Australian Research Council Discover Grant. The paper has
been prepared for the 22nd International Panel Data Conference, Perth, Australia, June 27-29th,
2016. The usual caveats apply.
1
1
Introduction
Evaluation of health care utilization is prominent in the applied economics and
econometrics literature, generally concentrating on the impact of certain health conditions on resource use, using varied sources of individual level self-reported panel
data and econometric techniques (for example, Winklemann (2004), Bago d’Uva and
Jones (2006), Jones and O’Donnell (2002), Deb and Travedi (1997)). The additional
impact of cognitive impairment has not yet been explored. But this is an important
aspect to consider in light of global ageing and health policy developments, leading to additional econometric issues. The contribution of this paper is therefore to
disentangle the additional e¤ects of recall bias introduced into the dependent variable and any additional bias added in by the main independent variable of interest,
cognitive impairment.
It is now recognised that impaired cognitive status is a hallmark of dementia and
recent research shows that cognitive decline may begin much earlier than expected,
with onset often occurring at middle age Singh-Manoux and et al (2012). Given
that the global cost of dementia stands at US604 billion, amounting to 1-2% of
GDP in developed countries (Alzheimer (2012)), these impacts cannot be ignored
any more in health care models. Of major policy concern is any additional strain
that cognitive decline may place on the health system. Provision of information on
cognitive decline diagnosis of dementia is now encouraged to ensure people receive
adequate treatment and to try and delay the onset of dementia (e.g. in the National
Health Service, England). This is of concern to policy makers and health budget
planners who are interested to know if such new diagnoses would increase demand
for health care use.
In a theoretical sense we know that this could go either way. We use the typical
health production theoretical framework Grossman (1972) to develop our hypothesis
for our empirical models: an individual gains utility u from a stock of health, H,
and the consumption of a bundle of other commodities, C. The demand for medical
care depends on a simple process whereby an additional unit of H requires units
of medical care . We avoid further technical detail here, and recommend interested
readers to view Jack (1999) [ch4] and Grossman (1972). The aim of our analysis here
is to see how cognitive status, proxied by word recall (WR), impacts on the medical
utilisation variable . The de…nition of cognitive we use, i.e. word recall, is standard
in the medical and economics literature (Lang, Llewellyn, Langa, Wallace, Huppert,
and Melzer (2008), Guven and Sheng Lee (2012), Bonsang and Perelman (2012),
Mazzona and Peracchi (2012), Gannon, Banks, Nazroo, and Munford (2015).
A priori, we expect individuals with better cognitive status to make fewer visits to
healthcare professionals, such that we expect the partial derivative . This is due to
the fact that poor cognitive status can be thought of as a marker of poor mental
health and/or dementia (Singh-Manoux and et al (2012)), and as such people with
lower levels of mental health are more likely to have a higher demand (or need)
for the services of medical doctors. However, there is also a possibility that those
individuals with lower cognition are less likely to be able to fully understand their
health conditions as a person with higher cognitive status (ceteris paribus) and hence
may not utilise healthcare as often as they should. Therefore the exact sign of the
relationship is ambiguous and hence we wish to test this hypothesis. In both cases,
demand pull and supply push factors may be operating –demand can be in‡uenced
by under-optimisation of health care by patients and supply may be in‡uenced by
…nancial incentives to health care providers.
The estimation of empirical models to test these hypotheses is not straight-forward.
The ideal data for this analysis would come from administrative sources, with little measurement error in records of health care services. In the absence of these
data, models are further complicated if applied to self-reported data, as this could
also in‡uence the dependent variable. Brusco and Watts (2015) found that in randomised control trial data, self reported primary health care visits were over-in‡ated
by 16%, compared to recorded visits from adminstrative claims data. Since most
panel analyses still utilize secondary self-reported data, we must therefore develop
econometric models to accommodate any measurement error introduced from recall
3
bias. Starting …rst with the literature on health care use, the general approach has
been to model health care use, aka doctor visits, as a count process. The double
hurdle models may be appropriate in the modelling of health care use across the
population, as many people will not have visited the doctor in the previous year.
The proportion of zeros could be up to 30%. However, in our data, we focus on
older people and the proportion of zeros is lower at 15%. However, we still allow for
a potential build-up of zero observations - in addition to what would be expected in
a standard count model - by modelling this as an additional "spike" in the within
the DOGPO framework.
We now contribute a signi…cant advancement on previous models. We estimate a
DOGPO model, i.e. DOGIT applied to a Poisson distribution Farrell and Harris
(2011). This models the ‘spikes’(multiple modes) in the data on doctor visits. The
spikes in the models are then modeled as captivity parameters (constant and word
recall1 ).We exploit the panel aspect of the data to incorporate a standard individual
level random e¤ect to control for unobserved heterogeneity. We then show how the
relationship between cognitive impairment and health care use can vary by age, by
adding an age random parameter on to the word recall variable. Our …nal model
allows for an AR(1) process on the random parameter over age. To date, no research
has included analysis of spikes but we now …ll this gap in the literature.
We apply our models to data from 3 waves of SHARE (Survey of Health, Ageing
and Retirement in Europe) for individuals age 50 and over. We …nd that an initial
marginal e¤ect of 0.40 of word call on doctor visits, is then reduced to 0.10 after
we control for unobserved e¤ects and reporting behavior. Our results are appealing
to both policy makers and econometricians, so we conclude by providing a brief
analysis of future policy and health care cost implications.
1
The next version of our paper will include reporting variables into the capitivity model, e.g.
length of time spent on the survey and number of questions each respondent refused to answer.
This will allow us to identify the e¤ect beyond functional form. Furthermore since word recall can
also impact both reporting, it is important to include this into the captivity part of our model.
4
2
Data
The main data source for our analysis is waves 1, 2, and 4 of the Survey of Health,
Ageing and Retirement in Europe (SHARE) . This dataset is a unique cross-European
interdisciplinary data set containing a rich source of information relating to the lives
of over 50 year olds. Wave 1 contains information from 2004 and each successive
wave is carried out two years apart, such that wave 4 related to 2010. We do not
use data from wave 3 (2008) as this was predominantly a retrospective life history
module, and as such did not contain su¢ cient information for our analysis.
SHARE gives a broad picture of life after age 50, measuring physical and mental
health, both objectively and subjectively; life satisfaction and well-being; and also
contains a broad range of socioeconomic information. SHARE is modelled after,
and harmonised with, the US Health and Retirement Study (HRS) and the English Longitudinal Study of Ageing (ELSA). We now outline the key variables from
SHARE that we use in our analysis.
Measure of healthcare utilisation: Within SHARE, individuals are asked about their
health care utilisation. The key question is: “How often seen or talked to medical
doctor in the last 12 months”. Individuals who report positive values to this question are then asked: “how many of these contacts were with a general practitioner
(GP)?”.
Measure of cognition: Our key measure of cognitive status is the number of words
an individual can recall (from a list of ten). The main variable of interest is the
initial recall, although the individual is also asked to recall the same list of words
after a short delay . Using word recall as a proxy for cognition is common in the
economics literature (e.g. Bonsang et al., 2012; Mazzonna and Peracchi, 2012), and
hence we use this as our key explanatory variable .
Table 1 provides summary statistics for the key variables in our analysis. The
average person makes around 5 visits to a GP per year. The average word recall (at
5
the …rst recall) is around 5/10.
The average age of our sample is 65, with age ranging between 50 and 95. 45% of
the individuals are male, and the average household comprises of 2 people.
The sample size is 73,251 (pooled observations).
3
Empirical model
The outcome measure that we have to hand is a proxy for healthcare utilisation
in the form of integer counts representing the yearly number of trips to the general
practitioner (GP). This would suggest that an appropriate starting for the statistical
model might be the Poisson regression model given by
Pr (Y = y je
x) =
exp (
)
y!
y
(1)
;
e0 e :
= exp x
with y = 0; 1; : : : ;
We immediately amend this for the availability of panel data that we have to hand,
and the fact that we have special interest on the coe¢ cient (
) on cognitive im-
pairment (CI) ; such that equation (1) becomes
Pr (Yit = yit jxit ; CIit ) =
exp (
yit
it
it )
yit !
with yit = 0; 1; : : : ;
(2)
;
=
i
where i indexes the individual; t the time period;
i
it
exp (x0it +
(3)
CIit ) ;
is the usual random e¤ect to
re‡ect the unobserved heterogeneity of individual i: The following (usual) assumption is made regarding the unobserved heterogeneity: that
i ~N
(0;
2
). However,
the basic hypothesis is that the e¤ect of cognitive impairment changes with age
of the individual. To allow for this,
must vary by age. For the most ‡exible
speci…cation possible, each age group should have a di¤erent coe¢ cient,
a;
where a
denotes the age of the individual. It would be possible to interact CI with dummy
variables for every observed age. However, this would require some additional 50
6
Figure 1: Observed Frequency of Counts; Full Sample
additional variables according to the range of observed ages in the data. A more attractive speci…cation would be to specify
a
as a random variable, where in contrast
to the usual random parameter (RP ) approach the randomness is not speci…ed at
the individual level, but at the level of the individual’s age such that
a
=
(4)
+ "a
where now "a represents (random) unobserved heterogeneity at the age-level;
age-speci…c coe¢ cient; and
to the proximity of the
a
a
the
the average e¤ect of CI on the index exp (:). Due
parameters, it is likely that neighbouring ones will be
related. To allow for a smoother evolvement of
a
as age changes equation (4) gets
amended such that
a
=
(5)
+ "a
"a = "a
1
+ ua ; ua is white-noise
Before proceeding any further, it is in intuitive to inspect the empirical distribution
of GP counts in the data. For the full sample, these are reported in Figure 1. From
this, it is obvious that most counts are relatively low, but there are obvious spikes
in this distribution. A closer inspection of a subset of counts (0
30) is given in
Figure 2 and makes these spikes even more evident. If one were to lay a Poisson
distribution over this distribution, it is clear from the raw data that there are clear
7
Figure 2: Observed Frequency of Counts; Counts 0
30
“excess”counts in, from Figure 2, counts equal to 4, 10, 12, 15, 24, 25 and 30; and
less-so, but still pronounced (Figure 1) at 35, 36, 40, 45, 48, 50, 52, 60, 70, 80, 90,
98.2 These spikes are all at observed counts where there is a strong possibility of
being adversely a¤ected by reporting behaviour. Thus we witness spikes at: all 5
and 10 intervals (digit preferencing); at those corresponding to once a quarter (4),
monthly (12), a couple of times each month (24), and “weekly”(48,52)3 ; and so on.
Clearly some of these observations will be valid (for example, some individuals may
have regularly-spaced check-up appointments), but clearly some will be related to
reporting behaviour (rounding to the nearest 5/10 integer) and recall bias. Laying
an unadulterated count-based density over this distribution of observed counts with
such pronounced spiking behaviour is unlikely to …t the data well, and moreover
could give biased results of key covariates of interest (Farrell and Harris (2011)).
Modelling count data with such observed spikes has, however, been considered by
Farrell and Harris (2011). The approach adopted in that paper was to combine the
“captivity” elements of particular choice outcomes of the DOGIT model (Gaudry
and Dagenis (1979)) with a standard Poisson process (and in doing such, breaks
the mean equal variance restriction of the standard Poisson model). So here, such
captivity would account for all of the unobserved heterogeneity relating to digit2
The latter somewhat due to the fact that the sparse counts of 98 and above counts were collated
into a single group.
3
48 would correspond to once a week over four weeks in the month over twelve months.
8
preferencing, reporting behaviour and the like. That is, akin to zero-in‡ated models,
these counts gain an additional component in addition to that as determined by
the underlying count (Poisson) process. For the current paper, this is achieved by
…rstly assigning captivity parameters,
j
; to all of the j in‡ated outcomes from the
full j = 0; 1; : : : choice set, the observed spiked-outcome choices listed above. The
captive probability of the j count outcomes will be similar to that in Gaudry and
Dagenis (1979):
captive
Pr (Y = j ) =
1+
and for “free-choice”
f ree
Pr (Y = j) =
By enforcing
model as
j
1+
j
with yit = 0; 1; : : : ;
a
1
P
j
(6)
j
j
(7)
:
j
= 0 for all j 2
= f4; 10; : : : ; 98g we can write the full probabilistic
P
Pr (Yit = yit jxit ; CIit ) =
1+ j
"
1
P
1+ j
and
Pj
+
#
(8)
j
j
it
=
exp (
it )
yit
it
yit !
i
exp (x0it +
;
a CIit ) ;
is driven by equation (5). Equation (8) will have no simple analytical closed
form due to the presence of the two unobserved e¤ects
i
and "a ; both assumed be
normally distributed with mean zero and respective variances,
2
and
2
a.
Therefore
to evaluate this likelihood function these random elements are integrated out of the
likelihood using simulation techniques, with 100 Halton draws for each sequence
Train (2003).
4
Results
To put the additional contribution of our models into perspective, we …rst present
results from the simple Poisson model (Model 1). Table 2 shows that the marginal
9
e¤ect of remembering one more word in the word test would lead to 0.38 additional
visits to the doctor. The marginal e¤ects on other conditions are interesting in
their own right – for example, those with heart conditions, diabetes, stroke, lung
conditions, arthritis or high blood pressure are much more likely to be high users of
services.
In Table 3, we present results from our full speci…cation (Model 2). We now include
the full random e¤ect at the individual level, the random parameter on CI varying
by age (
a ),
and the captivity parameters are modelled with CI entered as a control
variable. The interesting result from this parsimonious model is that the impact of
CI is now disentangled into a count process and captivity process. This shows that
overall marginal e¤ect of CI is in fact 0.11 and most of the full e¤ect (of -0.479) is
captured in the captivity process where the partial e¤ect is shown as 0.37.
This
means that the overall e¤ect of remembering one additional word is 0.11 more visits
on average.
Table 4 presents the captivity probabilities and shows the partial e¤ect of CI.
5
Concluding comments
The main conclusion is that cogntive impairment does impact on health care use,
over and above the visits made for standard health conditions (comorbidities of
people). The marginal e¤ect is around 0.11, meaning that one less word recalled
would lead to 0.11 more visits per year, on average. So if the prevalence rate of
mild cognitive impairment is approx. 5 per cent Sachdev et al. 2015), then there
could be 457,605 visits overall, at a direct cost of £ 22.9 million annually in the
EU28.This is not expensive in the long term and if it means many dementia cases
can be detected early, the longer-term returns on investment may be massive. A
number of questions emerge from this …nding. Could these (relatively low) additional
costs be even lower if people were looked after outside primary care for non-medical
10
needs? On the other hand, the little additional cost indicates further investment
for early cognitive impairment diagnosis would not be overly expensive to primary
care. While we …nd that there is very little additional use of resources over and
above other health conditions, this can imply people with cognitive problems are
(a) already getting su¢ cient health care, along with their comorbidities OR (b)
These people are not optimising resource use either due to their own decision or due
to little uptake so far in any supply incentives.
There are two main limitations in our research to date that are worth noting. Firstly
we will introduce reporting type variables into the model, to impact on reporting
behaviour but not on health care utilisation, e.g. the type of reporting being frequently ticking the same response, or duration of interview. We will also produce
the RP e¤ects by age.
The main contribution to the literature in applied panel econometric methodology
is the application of the DOGIT to health care use for the …rst time, combined with
the novel introduction of random e¤ects at both the individual and age aspects of
the panel - a pseudo panel model approach.
References
Alzheimer, W. (2012): “World Alzheimer Report,”Discussion paper, Alzheimer’s
Disease International.
Bago d’Uva, T., and A. Jones (2006): “Latent class models for utilisation of
health care,”Health Economics, 15(4), 329–343.
Bonsang, E., A. S., and S. Perelman (2012): “Does retirement a¤ect cognitive
functioning?,”Journal of Health Economics, 31(3), 490–501.
Deb, P., and P. Travedi (1997): “Demand for medical care by the elderly: A
…nite mixture approach,”Journal of Applied Econometrics, 12(3), 316–36.
11
Farrell, L. Fry, T., and M. Harris (2011): “A pack a day for 20 years: Smoking
and cigarette pack sizes.,”Applied Economics, 43(21), 2833–2842.
Gannon, B., J. Banks, J. Nazroo, and L. Munford (2015): “Cognitive impairment and health care utilsation among the ageing population: An analysis of
SHARE data,”Presented at HESG UK Conference, She¢ eld.
Gaudry, M., and M. Dagenis (1979): “The Dogit Model,” Transportation
Research-Part B, 13B, 105–112.
Grossman, M. (1972): “On the Concept of Health Capital and the Demand for
Health,”Journal of Political Economy, 80(2), 223–255.
Guven, C., and W. Sheng Lee (2012): “Height and congitive function at older
ages: is height a useful summary measure of early childhood experiences?,”Health
Economics Letter.
Jack, W. (1999): Principles of Health Economics for Developing Countries. World
Bank Publications.
Jones, A., and O. O’Donnell (eds.) (2002): Econometric Analysis of Health
Data. John Wiley and Sons.
Lang, I., D. Llewellyn, K. Langa, R. Wallace, F. Huppert, and
D. Melzer (2008): “Neighbourhood deprivation, individual socioeconomic status, and cogntive function in older people: Analyses from the English Longitudinal
Study of Ageing,”JAGS, 56, 191–198.
Mazzona, F., and F. Peracchi (2012): “Ageing, cognitive abilities and retirement,”European Economic Review, 56(4), 691–710.
Singh-Manoux, and et al (2012): “Timing of onset of cognitive decline:Results
from Whitehall II Prospective Cohort Study,”BMJ: British Medical Journal, 344.
Train, K. (2003): Discrete Choice Methods with Simulation. Cambridge University
Press.
12
Winklemann, R. (2004): “Health care reform and the number of doctor visits. An
Econometric Analysis,”Journal of Applied Econometrics, 19(4), 455–472.
13
Table 1: Descriptive statistics
Variable
Word Recall
GP Visits
Age
Male
Household Size
Heart
High Blood Pressure
Cholestorol
Diabetes
Stroke
Lung
Asthma
Arthritis
Osteoporosis
Cancer
Ulcer
Cateract
Hip
Parkinson’s Disease
Other
N=73251
14
Mean
5:037
4:844
65:199
0:453
2:136
0:109
0:333
0:221
0:104
0:033
0:059
0:030
0:229
0:055
0:049
0:045
0:074
0:019
0:006
0:195
Standard
Deviation
(1:88)
(7:40)
(10:05)
(0:498)
(0:966)
(0:312)
(0:471)
(0:415)
(0:306)
(0:179)
(0:236)
(0:172)
(0:420)
(0:228)
(0:218)
(0:207)
(0:262)
(0:137)
(0:079)
(0:396)
Table 2: Model 1:Pooled Poisson
Variable
Constant
Cognitive Impairment (CI)
Ln Age
Male
Household Size
Heart
High Blood Pressure
Cholestorol
Stroke
Diabetes
Lung
Asthma
Arthritis
Osteoporosis
Cancer
Ulcer
Parkinson’s Disease
Cateract
Hip
Other
N=73251
Marginal E¤ect
0:817
0:378
1:800
0:363
0:037
1:574
1:287
0:367
1:157
1:420
1:273
0:885
1:334
0:811
1:006
0:878
1:814
0:142
0:840
1:091
Notes: *p<0.10,**p<0.05,***p<0.01.
15
Standard error
(0:728)
(0:014)
(0:160)
(0:047)
(0:027)
(0:067)
(0:048)
(0:053)
(0:124)
(0:069)
(0:089)
(0:122)
(0:052)
(0:093)
(0:109)
(0:104)
(0:265)
(0:083)
(0:150)
(0:057)
Table 3: Model 2:Poisson with RP by age and captivity parameterised. Marginal
E¤ects
Variable
Constant
Cognitive Impairment (CI)
Ln Age
Male
Household Size
Heart
High Blood Pressure
Cholesterol
Stroke
Diabetes
Lung
Asthma
Arthritis
Osteoporosis
Cancer
Ulcer
Parkinson’s Disease
Cateract
Hip
Other
b*sigma_t
sigma_i
N=
Overall E¤ect
0:594
(0:614)
0:479
(0:013)
0:948
(0:139)
0:162
(0:032)
0:035
(0:026)
0:966
(0:071)
0:833
(0:031)
0:357
(0:033)
0:512
(0:099)
0:715
(0:039)
0:653
(0:048)
0:389
(0:065)
0:851
(0:043)
0:525
(0:066)
0:597
(0:112)
0:379
(0:067)
1:587
(0:428)
0:036
(0:055)
0:325
(0:087)
0:571
(0:042)
Count Process
2:576
(0:614)
0:112
(0:014)
Captivity Process
3:366
(0:110)
0:366
(0:011)
16
Notes: All captivity probabilities are signi…cant and CI impacts on each particularly for 12 visits.
17
Table 4: Model 2:Captive Probabilities
Spike
0
4
10
12
15
20
24
25
30
35
36
40
45
48
50
52
60
70
80
90
98
b*sigma_t
sigma_i
N=73251
Captive Probabilities
0:005
(0:002)
0:062
(0:002)
0:014
(0:001)
0:059
(0:001)
0:003
(0:000)
0:004
(0:000)
0:004
(0:000)
0:001
(0:000)
0:002
(0:000)
0:0002
(0:000)
0:0001
(0:000)
0:001
(0:000)
0:0001
(0:000)
0:001
(0:000)
0:001
(0:000)
0:001
(0:000)
0:0002
(0:000)
0:0002
(0:000)
0:0001
(0:000)
0:0001
(0:000)
0:001
(0:000)
0:000
0:829
Partial e¤ect of Z variables:constant
0:009
(0:003)
0:1216
(0:003)
0:035
(0:002)
0:049
(0:002)
0:012
(0:001)
0:013
(0:001)
0:014
(0:001)
0:004
(0:001)
0:006
(0:001)
0:002
(0:001)
0:001
(0:000)
0:005
(0:001)
0:001
(0:000)
0:003
(0:005)
0:005
(0:001)
0:004
(0:001)
0:001
(0:000)
0:001
(0:000)
0:001
(0:000)
0:001
(0:000)
0:004
(0:001)
18
Partial e¤ect of Z variables: CI
0:003
(0:001)
0:002
(0:001)
0:003
(0:000)
0:0153
(0:000)
0:001
(0:0000)
0:001
(0:000)
0:001
(0:000)
0:000
(0:000)
0:000
(0:000)
0:000
(0:000)
0:000
(0:000)
0:000
(0:000)
0:000
(0:000)
0:000
(0:000)
0:000
(0:000)
0:000
(0:000)
0:000
(0:000)
0:000
(0:000)
0:000
(0:000)
0:000
(0:000)
0:000
(0:000)