Content uploaded by Erkki Soini
Author content
All content in this area was uploaded by Erkki Soini
Content may be subject to copyright.
ORIGINAL ARTICLE
Direct costs of warfarin
treatment among patients with
atrial fibrillation in a Finnish health
care setting
T. Hallinen
a
, J. A. Martikainen
a
, E. J. O. Soini
b
,
L. Suominen
c
and T. Aronkytö
d
a
Center for Pharmaceutical Policy and Economics, Department of Social
Pharmacy, University of Kuopio, Finland
b
Department of Health Policy and Management, University of Kuopio,
Finland
c
Espoo Centre for Social and Health Services, Espoo, Finland
d
City of Helsinki, Health Centre, Finland
Address for correspondence: Mrs Taru Hallinen, MSc, Center for Pharmaceutical Policy and
Economics, Department of Social Pharmacy, University of Kuopio, P. O. Box 1627, FI‑70211 Kuopio,
Finland. Tel.: +358 17 163559; Fax: +358 17 163464; email: taru.hallinen@uku.fi
Key words: Atrial fibrillation – Cost analysis – Costs – Warfarin
0300-7995
doi:10.1185/030079906X100014
All rights reserved: reproduction in whole or part not permitted
CURRENT MEDICAL RESEARCH AND OPINION®
VOL. 22, NO. 4, 2006, 683–692
© 2006 LIBRAPHARM LIMITED
Paper 3350 683
Objective: The main objective was to estimate the
mean direct costs of warfarin treatment for atrial
fibrillation (AF) patients. Secondly, the costs of
initiating warfarin treatment during a 60‑day period
and the impact of International Normalized Ratio
(INR) and co‑morbidities on costs were estimated.
Design and data: The study was performed
as a retrospective cohort study over a 12‑month
period in a Finnish communal health care setting.
All AF patients aged 65 years or older (n = 250)
with warfarin treatment were identified from the
database of the health service district of an urban
area. Patient specific information related to co‑
morbidities, INR‑control, complications and health
care resource use were collected. Cost information
was obtained from the Finnish national health care
unit cost list.
Methods: The effect of treatment balance and
other background variables on treatment costs
were evaluated using ordinary least squares
regression (OLS), log‑transformed OLS and
generalized linear model (GLM). The mean costs
were calculated on the basis of the different
models and bias corrected and accelerated
(BCa) bootstrap confidence intervals (CIs) were
calculated for the mean costs.
Results: The best fitting cost model was log‑
transformed OLS. The costs of warfarin treatment
on the basis of the log‑transformed model were
589.82 euros (BCa 95% CI: 586.68–591.99) per
patient compared to 616.00 euros (BCa 95% CI:
579.98–652.96) obtained with the OLS‑model.
For the treatment initiation period, the mean costs
were 263 euros (BCa 95% CI: 218.90–314.71).
Depending on the way that INR‑control was
defined, the mean costs were 95.27 euros or
166.92 euros higher for patients who were not in
the defined INR‑balance.
Conclusions: The INR‑control has a significant
impact on the warfarin treatment costs. The
choice of model influences the estimated mean
costs. In addition, different models identify
statistically significant effects between different
background variables and costs.
A B S T R A C T
684 Costs of warfarin treatment © 2006 LIBRAPHARM LTD – Curr Med Res Opin 2006; 22(4)
Introduction
Atrial fibrillation (AF) is the most common arrhythmic
condition in the developed countries. Every fourth
man and woman over 40 years of age will develop AF
during their remaining lifetime
1
. The prevalence of AF
is higher in the elderly as approximately 70% of AF
patients are between the ages of 65 and 85 years
2
. AF is
a notable medical and socioeconomic problem due to
the related increases in the risk of heart failure, stroke,
and both overall and cardiovascular mortality
3
.
In order to reduce the risk of stroke in AF patients,
antithrombotic therapy is often used. At the time of
the study, the only oral anticoagulant drug available in
Finland was warfarin. Warfarin use is associated with
significant reductions in all strokes (OR 0.39, 95% CI:
0.26–0.59) and death (OR 0.69, 95% CI: 0.50–0.94)
4
.
However, the administration of warfarin is problem-
atic. The treatment needs to be carefully monitored
with regular laboratory measurements of INR (Inter-
national Normalized Ratio) due to an increased risk
of bleeding when the INR-value exceeds the target
range. The therapeutic range for INR that is deemed
safe and effective is narrow (between 2 and 3)
5
. In
addition, maintaining the target range is difficult since
concomitant medications, changes in health state (for
example fever or diarrhea) and changes in the vitamin K
content of diet may interact with warfarin, destabilizing
the previously maintained INR-level. In Finland, the
warfarin treatment quality is monitored with monthly
INR-controls. The physician contacts the patient (or
vice versa) the same day or the day following the INR-
control to give dosing instructions. If the dosing needs
to be changed substantially, the patient is referred to a
new INR-control within 1–2 weeks. For patients with
very labile INR-values additional guidance is given
related to nutrition.
The availability of published articles related to
the cost of treating AF with anticoagulation drugs is
quite limited. A recent study by Menzin et al. showed
the yearly mean direct monitoring costs of warfarin
treatment to amount to 244–383 euros (converted
into euros using the exchange rates from the European
Central Bank
6
) per patient for the year 2003 in
anticoagulation clinics located in three regions of the
United States
7
. Another study from the United States
by Anderson estimated the per-patient-per-month cost
of a decentralized outpatient pharmacy anticoagulation
service for AF patients on warfarin therapy to amount
to 47.33 euros for the year 2000 (cost per year thus 568
euros)
8
. In addition, two studies reported costs from the
United Kingdom. A monitoring cost of 158 euros per
patient related to warfarin treatment for the year 1995
can be extracted from the study by Stewart et al.
9
, and
Abdelhafiz et al.
10
reported a warfarin treatment cost of
242 euros per patient for a year during 1999–2000. At
the time of this study, no cost information related to
warfarin treatment was directly available for Finland.
The primary objective of the study was to estimate
the mean direct costs of anticoagulant treatment per
patient in Finland. In addition the cost of initiating
warfarin treatment during a 60-day period and the
impact of INR control on costs of anticoagulation
treatment were evaluated.
Methods
The study was performed as a retrospective cohort
study over a 12-month period (year 2002) in a Finnish
communal health care setting. The direct costs related
to warfarin treatment were calculated on the basis of
recorded resource use in the study population of 217
patients with continued warfarin treatment as well
as in 33 patients with newly initiated treatment. An
explorative approach was adapted in choosing the
optimal model for analyzing the cost data.
The study data were gathered from the database
of the health service district of an urban area with
over 200
000 inhabitants. The data base consisted of
information of over 400
000 patients (the figure is
larger than the actual number of inhabitants due to
changes in residence). Initially the patient database was
screened to identify patients receiving warfarin treat-
ment. Thereafter, patients eligible for the study were
manually identified by the investigators. The inclusion
criteria were the following:
age 65 years or over;
diagnosis of non-valvular atrial fibrillation;
warfarin treatment for at least 60 consecutive
days prior to the end of the study period;
data available for over a 24-month period.
After identifying the study subjects, the following
patient specific information possibly affecting the
treatment costs was collected: age, gender, duration
of follow-up, concomitant illnesses (hypertension, dia-
betes mellitus, coronary heart disease, congestive heart
failure), history of stroke, history of transient ischemic
attack (TIA), bleeding complication and INR-values
during the study period. The health care resources
(for example physician consultations, laboratory tests,
hospitalizations) used and related to warfarin treatment
were clarified for all study subjects. Unit costs for
resources were taken from the national health care unit
cost list
11
in order to increase the generalizability of
the results. The costs were transformed into monetary
values for the year 2002 using the Finnish price index
for health care from Statistics Finland.
•
•
•
•
© 2006 LIBRAPHARM LTD – Curr Med Res Opin 2006; 22(4) Costs of warfarin treatment Hallinen et al. 685
The total costs of warfarin treatment were calculated
by combining the resource use data with the cost
data from the list of Finnish health care unit costs
11
.
Similarly, the treatment costs were calculated for the
group of treatment initiators. The effect of treatment
balance, age, gender, and co-morbidities were analyzed
using three different statistical models. An explorative
approach was adopted since very little was known
about the costs and cost-drivers of warfarin treatment.
The three models chosen were ordinary least squares
(OLS) regression with and without logarithmic
transformation and generalized linear regression model
(GLM) (see Table 1).
The OLS model assumes a normal distribution of
costs. For health care costs this is often not a suitable
assumption as the distribution of costs is usually highly
skewed to the right due to the fact that a small minority
of patients is responsible for incurring very high
costs. However, since in health care decision making
the arithmetic mean costs are of importance, the
use of tests and estimates based on normality
assumption has nevertheless been recommended
12
for cost-related research. In the second model, a
logarithmic transformation of costs was performed
to account for the skewness in the data. The GLM
was chosen as a third model since it provides a
parametric method where non-normal distributions
for the outcome variable can be specified and the
way in which the explanatory variables act can be
altered
12
. In our study, the costs were assumed to follow
a gamma-distribution and the explanatory variables
were assumed to act multiplicatively on the mean
(i.e., the link function was logarithmic). The choice
of gamma-distribution and logarithmic link was done
on the basis of comparison between different GLM-
models.
The performance of the models was compared and
the best model for our data was chosen. The perform-
ance measures reported are the Akaike information
criteria (AIC), Bayesian information criteria (BIC) (see
for example Burnham et al.
14
) and root mean square
error (RMSE). In addition the normality and hetero-
skedasticity of residuals (deviance residuals for the
GLM) were tested using the Shapiro–Wilk W-test and
Cook–Weisberg test, respectively. A link-test was also
performed to test whether the dependent variable in
the models needed a transformation or ‘link’ function
to properly relate to the independent variables
15
. The
analyses were performed using Intercooled STATA
version 9.1
16
.
For the OLS and GLM models the expected costs
can be obtained directly from the model’s predictions.
However, for log-transformed OLS a retransformation
is needed since the unbiased and consistent quantities
on the transformed scale do not usually retransform
into unbiased and consistent quantities on the
untransformed scale
17
. Therefore, the expected costs
for the log-transformed model were calculated using
the smearing estimate by Duan
17
:
where
ˆ
β
are the OLS coefficients, n is the number of
observations and
ˆ
ε
i
are the estimated residuals from the
log-model. For the log-transformed model the effects
of explanatory variables on the treatment costs in the
untransformed scale were evaluated by calculating the
change in expected costs (∂E
(
y
|
x)) for a unit change
in each explanatory variable (∂x
i
):
where
s n
i
= ∑
−1
exp( )
ˆ
ε
The confidence intervals for the expected costs
were estimated using a bootstrapping procedure (2000
replicates) to obtain bias corrected and accelerated
confidence intervals (BCa)
18
. The bootstrapping pro-
cedure was chosen since for skewed data, the normal
theory methods can result in confidence intervals that
are not exact or accurate. These bootstrap techniques
are non-parametric and thus they provide a valid
method for constructing confidence intervals regardless
of the distribution of the statistics
19
. For log-trans-
formed data, the retransformed cost estimate (i.e., the
smearing estimate) was bootstrapped.
The OLS model:
y x x x
k
i
= + + + + +α β β β ε
1 1 2 2
...
i
i
(1)
where y is the dependent variable (costs), x
i
are the
explanatory variables,
i
are the regression coefficients and i
is the random error term.
The log‑transformed OLS model:
ln( )y
x x x
k
i
=
+ + + + +α β β β ε
1 1 2 2
...
(2)
The GLM model:
g E y x y
F
{ ( )} , = ≈β
(3)
where g
(...) is called the link function and F is the distribu-
tional family. In our study, the link function was logarithmic:
g
{
E(
y)} = ln
{
E(
y)}
Table 1. The models compared
686 Costs of warfarin treatment © 2006 LIBRAPHARM LTD – Curr Med Res Opin 2006; 22(4)
Results
Patient characteristics
The total study population consisted of 250 patients.
The characteristics of study patients are summarized
in Table 2. Most of the study patients were female
(69.2%) and elderly, over 80 years old (78%). The
most prevalent co-morbidity was hypertension
followed by coronary heart disease and congestive
heart failure. The patients in the study population
suffered from many illnesses, as many as 69% of
the patients suffered at least two other major diseases
in addition to atrial fibrillation. The mean number
of co-morbidities in addition to atrial fibrillation was
about two. Anticoagulant treatment was initiated
during the study period in 13.2% of the patients
(n = 33).
Treatment success
On average 17 INR-tests per patient were performed
during the 1 year follow-up. All INR-values in this
patient population were measured in the laboratory.
Of these tests 23.9% fell below and 12.7% above the
target range for INR-value leaving 63.4% in the target
range of two (Figure 1). In the 1 year follow-up group
good anticoagulation control was achieved by 34.6% of
the patients when the control was defined similarly to
that of the study by Menzin et al.
7
as at least 75% of
INR-values staying in the target range during the study
period. In the treatment initiator group on average
eight INR-tests per patient were performed during the
60-day initiation period. Of these tests 43.8% were in
the target range leaving 38.5% below and 17.7% above
the target range.
Health care resource use and costs
The health care resource use for each patient consists
of laboratory, physician and nursing staff visits, phone
Table 2. Characteristics of the study patients
Characteristic All, n = 250
Age (years), %
65–69 7.6
70–74 9.2
75–79 5.2
80–84 36.0
85–89 31.2
90
10.8
Mean age ± SD 82.5 ± 6.8
Male, % 30.8
Co-morbidities, %
Hypertension 60.0
Diabetes 23.2
Coronary heart disease 58.0
Congestive heart failure 50.4
Stroke 16.8
TIA 13.2
Bleeding 12.0
Warfarin-treatment initiated during study
period, %
13.2
Days of follow-up, mean ± SD 332.4 ± 69.5
Days spent in hospital (not related to
warfarin-treatment), mean
14.9
Figure 1. Percentage of INR tests within specified intervals
1.5
6.4
8.7
12.5
11.0
14.4
15.9
13.8
9.1
3.2
1.9
2.6
1.1
1.9
1.5
13.6
8.7
4.5
8.7
8.3
9.4
2.6
2.6
1.9
0.8
0
0.8
10.3
2.4
2.2
0.7
0.2
0.3
0.2
1.2
1.3
0.2
1.8
6.6
4.2
0.8
0.2
0
2
4
6
8
10
12
14
16
18
0.9–1.19
1.20–1.39
1.40–1.59
1.60–1.79
1.80–1.99
2.00–2.19
2.20–2.39
2.40–2.59
2.60–2.79
2.80–3.00
3.01–3.19
3.20–3.39
3.40–3.59
3.60–3.79
3.80–3.99
4.00–4.19
4.20–4.39
4.40–4.59
4.60–4.79
4.80–4.99
≥ 5
INR-intervals
Percent of INR-values
60 days
1 year
© 2006 LIBRAPHARM LTD – Curr Med Res Opin 2006; 22(4) Costs of warfarin treatment Hallinen et al. 687
consultations with either physician or nursing staff,
hospitalizations, and services provided at the home of
the patient. The mean number of visits and their costs
are shown in Table 3. The distribution of treatment
costs in our study is shown in Figure 2. The distribu-
tion is highly skewed to the right due to the fact that
only four patients required expensive hospitalizations
related to warfarin-treatment during the study period.
The distribution after log-transformation is shown in
Figure 3. The log-transformed costs are visually much
closer to normal distribution than those without
the transformation although the distribution does
still exhibit significant kurtosis (coeff. of kurtosis =
9.990925) and skewness (coeff. of skewness =
0.9067181).
Factors affecting the treatment costs
The effects of various background factors on the costs
were studied using the three models. The background
variables in the model were age (years), gender
(female = 1, male = 0), length of follow-up (days),
time spent at a hospital for reasons not related to
warfarin-treatment (days), and dummy variables
for treatment balance (at least 75% of the INR-values
in target range = 1, less than 75% of the INR-values
in the target range = 0) and concomitant illnesses
(yes = 1, no = 0). For the best fitting model
version also a model with a differently interpreted
dummy variable for treatment balance (at least one
value above target range = 1, no values above target
range = 0) was estimated. This was done to estimate
whether the effect of treatment balance was related to
higher INR-values which increase the risk of bleeding
complications. The results for the models are shown in
Table 4.
When OLS was performed as robust to account for
the heteroskedasticity detected by the Cook–Weisberg
test, the model identified no significant relationships
between any of the explanatory variables and costs.
In contrast, the log-transformed OLS model and
GLM model both detected five statistically significant
Table 3. Resource use and unit costs
Resource name Mean number of visits
(standard deviation)
Unit cost
10
(2002) €
Nursing staff 0.648 (1.734) 23.07
Phone consultations 5.928 (6.790) 16.41
Home visits by nursing staff 0.544 (2.085) 41.33
Physician visits outside normal working hours
or including operations
0.064 (0.385)
99.78
Physician visits 0.716 (1.190) 56.10
Laboratory visits 15.536 (6.250) 2.50
Prescription renewals (phone) 0.932 (2.111) 16.41
Paper consultations 10.28 (7.102) 16.41
Warfarin (3
mg) Every day 0.19/day
Hospital days related to warfarin treatment 0.184 (1.944) 334.76*
Traveling by the patient 16.964 (6.922) 5.49
*Calculated as the average cost on the basis of summed individual hospitalization costs [the cost/day
varied because treatment was given in different specialities (e.g., surgery, internal disease)]
0
20
40
60
80
Percent
0 5000 10 000
Treatment costs
0
10
20
30
40
Percent
4 6 8 10
Logarithm of treatment costs
Figure 2. The distribution of treatment costs
Figure 3. The distribution of log-transformed treatment costs
688 Costs of warfarin treatment © 2006 LIBRAPHARM LTD – Curr Med Res Opin 2006; 22(4)
Table 4. The results of the models for the patient group with continued treatment (n = 217)
OLS (robust) LOG-OLS GLM LOG-OLS 2 Mean costs
Coefficient (standard error)
Coefficient (standard error) Coefficient (standard error) Coefficient (standard error)
Age 10.40348 (11.27169) 0.0060262 (0.005287) 0.0113596 (0.0076665) 0.005672 (0.0051771)
Gender –205.7958 (242.9937) –0.0390921 (0.0735402) –0.1295258 (0.108457) –0.0521197 (0.0721244)
TIA 556.4929 (414.8848) 0.359075* (0.1032075) 0.5579731* (0.1604577) 0.3314675* (0.1003107)
Stroke –337.1925 (210.769) –0.2857673* (0.0921107) –0.3912215* (0.1356111) –0.2778644* (0.0902569)
CHD –75.48053 (93.48332) –0.0346693 (0.0710769) –0.0659079 (0.110219) –0.0360149 (0.0696318)
Congestive heart failure –5.615555 (89.30479) 0.0488102 (0.0712409) 0.0194094 (0.1091796) 0.0336718 (0.0698494)
Diabetes –14.61399 (84.05049) 0.0370718 (0.076676) 0.0018957 (0.1152095) 0.0631945 (0.0753208)
Hypertension 150.9794 (115.4604) 0.1371916 (0.0719046) 0.1419943 (0.1087414) 0.1378142* (0.0690439)
Bleeding 439.9647 (392.8306) 0.078583 (0.1025666) 0.4045855* (0.1685861) 0.1153043 (0.1003331)
Follow-up 0.2753719 (1.684179) 0.0048276* (0.0006832) 0.0038545* (0.0011316) 0.004436* (0.0006777)
Balance –121.7582 (83.80266) –0.1617108* (0.0716648) –0.1835511 (0.1072737) 0.2854725*
,
† (0.0766535)
Hospital days‡ –1.296012 (0.8348536) –0.0024703* (0.0007816) –0.0030306* (0.0012948) –0.0024271* (0.0007566)
Constant –241.5168 (502.8336) 4.060024* (0.5232739) 4.15974* (0.8100998) 3.96183* (0.5087391)
R-squared 0.1293 0.2995 0.3277
RMSE 764.27 0.47202 767.53 0.46241
AIC 3509.705 302.5905 3219.325 293.6667
BIC 3553.643 346.5292 3263.264 337.6053
Deviance/SSR 119157770 45.450828 55.50353424 43.6196273
Shapiro–Wilk W-test for normality of
residuals/deviance residuals
V = 122.952
p = 0.00000
V = 21.511
p = 0.00000
V = 53.195
p = 0.00000
V = 11.966
p = 0.00000
Cook–Weisberg test for heteroskedasticity
of prediction
chi² = 1325.95 prob > chi² = 0.0000
chi² = 1.03 prob > chi² = 0.3093
–
chi² = 3.25 prob > chi² = 0.0715
Link-test
_hatsq: p = 0.000 _hatsq: p = 0.512 _hatsq: p = 0.500 _hatsq: p = 0.403
*The effect is statistically significant at p = 0.05
†Dummy for INR-values above 3 (1 = yes, 0 = no)
‡For reasons not related to warfarin treatment
© 2006 LIBRAPHARM LTD – Curr Med Res Opin 2006; 22(4) Costs of warfarin treatment Hallinen et al. 689
coefficients for the explanatory variables in addition
to the constant in the models. Both models agreed on
the significance of the length of follow-up time, time
spent at hospital, TIA, and stroke. The effects were
also in the same direction. However, log-transformed
OLS identified a significant negative relationship
between the costs and treatment balance which was
not found by the GLM model. In contrast, GLM
detected a significant positive relationship between the
co-morbidity of bleeding and treatment cost that was
absent from the log-OLS model. With the exception
of stroke, the associated signs were as expected, since
a good treatment balance and days spent at hospital
for reasons not related to warfarin-treatment had
a negative impact on monitoring costs whereas the
length of follow-up time, bleeding, and co-morbidity
of TIA had a positive effect. However, the negative
association between stroke and monitoring costs was
not expected.
The interpretation of the results from the OLS
model is straightforward. In OLS the monitoring
costs increase or decrease by the amount shown in
the coefficients. However, none of the coefficients
were statistically significant for the heteroskedasticity
robust OLS. The interpretation of the log-cost model
and GLM with log-link is not as straightforward as
that of OLS. In GLM with log-link, the exponentiated
coefficients (exp(b) ≡ e
b
) provide a ratio of means
and they can be expressed as the percentage increase
in mean costs per unit increase in the explanatory
variable
13
(= exp(b) – 1). Thus, according to the model,
the mean costs increase by 0.4% for each extra day
of follow-up, by 74.7% if the patient has TIA and
by 49.9% if the patient has experienced bleeding. In
contrast, the mean costs decrease by 0.3% for each day
spent in hospital and by 32.4% if the patient has had a
stroke.
In the log-cost model, the coefficients reflect the
change in the logarithm of the treatment cost for a unit
change in the explanatory variable. The exponentiated
coefficients can be interpreted similarly to that of
GLM as the percentage increase in mean costs per unit
increase in the explanatory variable. Thus according to
our model, the costs increase by 43.2% if the patient
has TIA and by 0.48% for each extra day of follow-
up time. Similarly the costs decrease by 24.9% if the
patient has had a stroke, by 14.9% if the treatment
balance is good and by 0.25% for each extra day spent
in hospital. By calculating the slope of y the effects
can be evaluated in monetary terms. According to the
calculated slopes, the expected costs change by –95.27
euros if the treatment balance is good, by 211.53 euros
if the patient has TIA, by –168.35 if the patient has
suffered a stroke, by 2.84 euros for each extra day of
follow-up and by –1.46 euros for each extra day in
hospital.
In our study, when comparing the OLS regression
model with and without log-transformation, it can be
seen that the log-transformation improves the ability of
the model to explain the variation of costs. The better
performance of log-transformed OLS is reflected in the
higher R-squared value and the lower values for the
RMSE and the information criteria. Also the link-test
implies that the dependent variable in the OLS-model
needs to be transformed in order to properly relate
to the independent variables. When comparing the
performance of log-transformed OLS and GLM, it can
be seen that the values for the information criteria and
RMSE are lower for the log-transformed OLS. These
support the log-transformed OLS as the better fitting
model.
When the treatment balance was interpreted more
strictly, as the presence or absence of any INR-values
above 3 (76% of patients had at least one value above
3), the results were somewhat different as shown by
the last column in Table 4 (in a similar model for
INR-values below the target range, the effect of the
balance-dummy on costs was not significant). The
statistically significant effects (
p < 0.05) found by
this model were TIA, stroke, balance, hypertension,
follow-up time, and length of hospital stays for reasons
not related to warfarin treatment. With the exception
of stroke, the associated signs were in the expected
direction. Thus the expected treatment costs increased
by 39.3% if the patient had TIA, by 14.8% if the
patient had hypertension, by 33% if the patient had
any INR-values above the target range and by 0.4%
for each extra day of follow-up. Conversely the costs
decreased by 24.3% if the patient had had a stroke and
by 0.2% for each extra day spent in hospital for reasons
not related to warfarin treatment. In monetary values,
the changes were 193.8 euros for TIA, 80.58 euros for
hypertension, 166.92 euros if the patient had INR-
values above the target range, 2.59 euros for each extra
day of follow-up, –162.47 euros for stroke and, –1.42
euros for hospital days.
As the diseases and co-morbidities in this study are
non-curable, the only variables in the models that can
be influenced by treatment practices are bleeding and
treatment balance. As the effects of bleeding were
not statistically significant, it can be concluded that
the INR-control is the most significant cost driver
in this study population. The cost savings per 100
patients would be approximately 12
690 euros (95%
CI: 5969.80–19402.04) if their INR-values stayed
below 3 (or 6230 euros (95% CI: 786.10–11674.55)
if treatment balance is defined according to
Menzin et al.
7
).
690 Costs of warfarin treatment © 2006 LIBRAPHARM LTD – Curr Med Res Opin 2006; 22(4)
Mean costs
The mean costs of warfarin treatment for the OLS model
were 616 euros (normal based 95% CI: 577.92–654.08
and BCa 95% CI: 579.98–652.96). However, the mean
costs are somewhat lower when they are calculated on
the basis of the OLS with logarithmic transformation
and re-transformation being then equivalent to 589.82
euros (95% BCa CI: 586.68–591.99). As was to be
expected due to heavy tail in our data, the mean costs
were lower for the better fitting log-cost model than
they were for the OLS model. Also the confidence
intervals were narrower for the log-transformed and
re-transformed costs. The difference between expected
costs from the GLM model and OLS were not large
as the expected costs from the GLM were 611.53
euros (BCa 95% CI: 584.55–642.67). However, the
confidence interval for GLM was narrower than that
achieved with OLS.
The mean costs for the 60-day initiation period
were estimated on the basis of 33 patients, whose
treatment was initiated in the year 2002. Since the
treatment costs for the 60-day initiation period were
normally distributed, the mean costs were calculated
directly from the data. The mean costs for the 60-day
initiation period were 263.05 euros (normal based 95%
CI: 212.02–314.09/BCa CI: 218.90–314.71). When
costs were compared for those who had achieved
treatment balance and those who had not, the related
costs were 151.67 (95% CI: 46.42–256.93) and 278.42
euros (95% CI: 224.25–332.59), respectively. The
difference was statistically significant according to the
performed two sample t-test with equal variances (the
equality of variance was tested using a variance ratio
test).
Discussion
The mean monitoring cost of warfarin for the year
2002 in our study amounted to 589.82 euros. This
is somewhat higher than the cost of 242–383 euros
reported in the studies by Menzin et al.
7
and Abdel-
hafiz et al.
10
but similar to the 568 euros described by
Anderson
8
. The figures of Menzin et al. and Abdelhafiz
et al. seem to be rather low compared to our results,
especially since the costs of treatment initiators are
included in their costs but they are absent from our
figures. However, these cost differences were not totally
unexpected since the settings in which the services
were provided differed greatly in the two studies from
the United States and one from the United Kingdom
and in our study. In the studies by Menzin et al. and
Abdelhafiz et al., the services were provided by clinics
specializing in the monitoring of warfarin treatment
which might lead to more efficient treatment practices
compared to the corresponding services in Finland.
Many of the responsibilities carried out by physicians
in the Finnish setting (such as discussing the laboratory
results with patients over the phone and adjusting
the warfarin dose) were carried out by pharmacists
or nurses in the settings described in the three
other studies. Also the study populations were
different as the patients in our study were notably
older (83.9% aged 75 years or older compared to
45.5 and 32% in the two US studies and 49.5% in
the UK study), had more co-morbidities and the
share of female (69.6 compared to 46.5, 29, and 44.3%)
was also higher. This reflects the fact that in Finland
there are more females in the studied age-groups
(i.e., 71% of people over 80 years were female in
2004).
The mean costs (adjusted for age and gender) for the
initiation of warfarin treatment in a 60-day period was
263.05 euros. The monthly initiation costs are thus
larger than the implied monthly continuation costs
(131.5 euros vs. 49.15 euros) due to more intensive
monitoring needed to achieve the initial treatment
balance. No other studies related to the initiation costs
of warfarin treatment could be found.
The quality of warfarin treatment measured as the
share of INR-values in the target range was comparable
to those in studies by Menzin et al.
7
and Anderson
8
despite the differences in health care settings and
study populations. The INR-values in our study were
in target range for 63.4% of the time while the figures
for the aforementioned studies were 62% and 60.4%,
respectively. The patients in all three studies spent
more time below the target range than above it, and
these figures were also similar in magnitude (approx-
imately 25% below and 13–14% above).
The effect of background variables on the treatment
costs were analyzed using three models, OLS,
log-transformed OLS and GLM with gamma
distribution and log-link, of which the best fitting
model was chosen. Compared to the log-transformed
OLS, the performances of OLS and GLM were
poorer on the basis of RMSE and information criteria
values. The poorer performance of GLM is probably
due to the fact that our data are heavy-tailed. In
Manning et al.
20
the efficiency losses of GLM compared
to OLS-based estimates were shown to be substantial
and increasing in the coefficient of kurtosis of the log-
scale error. The standard errors of the coefficients in
their study were seven times larger for the gamma
model when kurtosis was 5.0. In our data, the kurtosis
of the log-scale error was even more substantial
(kurtosis = 10.3).
© 2006 LIBRAPHARM LTD – Curr Med Res Opin 2006; 22(4) Costs of warfarin treatment Hallinen et al. 691
According to the log-transformed OLS, the impact
of treatment balance, follow-up time, hospital days,
having a stroke, and having TIA had a statistically
significant effect on treatment costs. The expected
warfarin treatment costs decreased by 95.27 euros if
the treatment balance was good (at least 75% of INR-
values in the target range) and by 168.35 euros if the
patient had had a stroke and by 1.46 euros for each
extra day spent in hospital for reasons not related to
warfarin treatment. Conversely, the expected costs
rose by 211.53 euros if the patient had TIA and by
2.84 euros for each extra day of follow-up time.
In our second log-transformed OLS version, where
the treatment balance variable was replaced by a
variable describing whether the patient had any INR-
values above the target range, the costs for patients with
these excessive values were 166.92 euros higher than
the costs of patients with no above range INR-values.
Also the effect of hypertension reached statistical
significance as the costs were 80.58 euros higher in
patients with hypertension. The effect of above range
INR values on costs might be explained by the need
to adjust the dosage of warfarin and the subsequent
laboratory tests to be certain that the INR had returned
to the target range. The effects on costs in these models
were thus in the expected direction with the exception
of those due to stroke. However, the decrease in costs
of stroke patients might be explained either by the
reduced ability of the non-hospitalized stroke patients
to utilize health care resources or that the stroke
prior to the study period had been mild enough
not to cause excess use of health care but instead
acted as an incentive to comply with the warfarin
treatment.
The retrospective study design and the age structure
of the study population create some limitations in the
generalizability of the study results. Firstly, there was
a significant number of hospitalizations in the study
population during the study period. However, in
retrospect, only four hospitalizations could be unam-
biguously linked to warfarin treatment. Nevertheless,
the long stays in hospital for reasons not related to
warfarin treatment reduced the use of those resources
that were measured in our study (physician visits, visits
of nursing staff, phone consultations etc.). Therefore,
the costs of warfarin treatment may be undervalued
in our study. Secondly, the study patients’ health
care resource use within the private health care sector
and the resource use by relatives in taking care of the
patients were not available for analysis which may also
underestimate the true costs of warfarin treatment.
Thirdly, the results are based on the Finnish communal
health care setting and may not be generalized to other
health care settings.
Conclusions
The mean costs of warfarin treatment for an elderly
population in Finland were 589.82 euros in 2002.
The costs of initiating warfarin treatment were propor-
tionally higher, as the mean costs for a 60-day initiation
period were 263.05 euros. The treatment success,
defined either as INR-values staying in the target range
of 2–3 for 75% of the time or as INR-values staying
below the upper limit of 3, had a statistically signif-
icant impact on costs. When the defined treatment
success was not reached, the mean yearly costs of
treatment increased by 95.27 and 166.92 euros,
respectively. The choice of the model influenced the
estimated mean costs. In addition, different models
identified statistically significant effects between differ-
ent background variables and costs. Thus the testing
diagnostics and comparative choice of model is crucial
in cost analyses related to health care.
Acknowledgments
Declaration of interest: The research funding was
provided by AstraZeneca Oy, Finland.
We thank Ismo Linnosmaa, PhD, for comments on
an earlier draft of this paper.
References
1. Lloyd-Jones D, Wang T, Leip E, et al. Lifetime risk for develop-
ment of atrial fibrillation: the Framingham Heart Study.
Circulation 2004;110:1042-6
2. Feinberg W, Blackshear J, Laupacis A, et al. Prevalence, age
distribution, and gender of patients with atrial fibrillation.
Analysis and implications. Arch Intern Med 1995;155:469-73
3. Chatap G, Giraud K, Vincent J. Atrial fibrillation in the elderly:
facts and management. Drugs Aging 2002;19:819-46
4. Aguilar M, Hart R. Oral anticoagulants for preventing stroke
in patients with non-valvular atrial fibrillation and no previous
history of stroke or transient ischemic attacks. Cochrane
Database Syst Rev 2005, Issue 3. Art. No.: CD001927 [DOI:
10.1002/14651858.CD001927.pub2]
5. Hylek E, Skates S, Sheehan M, et al. An analysis of the
lowest effective intensity of prophylactic anticoagulation for
patients with nonrheumatic atrial fibrillation. New Engl J Med
1996;335:540-6
6. ECB/EU. ECU/EUR exchange rates versus national currencies:
Eurostat 2.11.2005
7. Menzin J, Boulanger L, Hauch O, et al. Quality of anticoagula-
tion control and costs of monitoring warfarin therapy among
patients with atrial fibrillation in clinic settings: a multi-site
managed care study. Ann Pharmacother 2005;39:446-51
8. Anderson R. Cost analysis of a managed care decentralized
outpatient pharmacy anticoagulation service. J Manag Care
Pharm 2004;10:159-65
9. Stewart S, Murphy N, Walker A, et al. Cost of an emerging
epidemic: an economic analysis of atrial fibrillation in the UK.
Heart 2004;90:286-92
10. Abdelhafiz AH, Wheeldon NM. Use of resources and cost
implications of stroke prophylaxis with warfarin for patients
with nonvalvular atrial fibrillation. Am J Geriatr Pharmacother
2003;1:53-60
692 Costs of warfarin treatment © 2006 LIBRAPHARM LTD – Curr Med Res Opin 2006; 22(4)
11. Hujanen T. Terveydenhuollon yksikkökustannukset Suomessa
vuonna 2001. Stakes Aiheita 2003(1)
12. Thompson SG, Barber JA. How should cost data in pragmatic
randomised trials be analysed? Br Med J 2000;320:1197-200
13. Barber JA, Thomson SG. Multiple regression of cost data: use of
generalised linear models. J Health Serv Res Policy 2004;9:197-
204
14. Burnham KP, Anderson DR. Multimodel inference, under-
standing AIC and BIC in model selection. Sociol Methods Res
2004;33:261-304
15. Stata base reference manual, vol. 2, K-Q, release 9. College
Station (TX): Stata Press; 2005
16. Stata statistical software: release 9 [program]. College Station
(TX): StataCorp LP; 2005
17. Duan N. Smearing estimate: a nonparametric retransformation
method. J Am Statistical Assoc 1983;78:605-10
18. Efron B. Better bootstrap confidence intervals. J Am Statistical
Assoc 1987;82:171-85
19. Drummond M, McGuire A. Statistical considerations in analys-
ing health care resource utilization and cost data. Economic
evaluation in health care, merging theory with practice. Oxford:
Oxford University Press; 2001
20. Manning WG, Mullahy J. Estimating log models: to transform
or not to transform? J Health Econ 2001;20:461-94
CrossRef links are available in the online published version of this paper:
http://www.cmrojournal.com
Paper CMRO-3350_3, Accepted for publication: 16 February 2006
Published Online: 07 March 2006
doi:10.1185/030079906X100014
