University of Oslo
Institute of Health Management and
Health Economics
Working Paper 2009: 1
DECISION-MAKING IN GENERAL PRACTICE:
THE IMPORTANCE OF A NEAR PATIENT TEST
WHEN CHOOSING MEDICAL ACTIONS
Siri Fauli Munkerud,
The Norwegian Medical
Association, NOKLUS, HERO
Geir Thue,
The Norwegian Quality
Improvement of Laboratory
Services in Primary Care
© 2009 Institute of Health Management and Health Economics and the author
The report is published online at URL: www.med.uio.no/heled
Decision-making in general practice:
the importance of a near patient test when choosing medical actions
Siri Fauli Munkerud, PhD Stud1,2,
Geir Thue, MD.PhD2,
1
The Norwegian Medical Association, NOKLUS, HERO
P.O.Box 1152 Sentrum, N-0107 Oslo, Norway, +4723109115, fax 4723109100,
siri.fauli.munkerud@legeforeningen.no
2
The Norwegian Quality Improvement of Laboratory Services in Primary Care
(NOKLUS)
Abstract:
Background: Laboratory results are often important in diagnostic work-up and
monitoring of patients and may result in costly medical actions. There is a lack of
studies on the impact of near patient test results in primary care, which should be
an important prerequisite for reimbursing such tests.
Objectives: To develop a model for studying what effect the results of a near
patient laboratory analysis, may have on medical actions in primary care
Methods: A clinical vignette, describing a young woman with dyspepsia, was
surveyed to GPs in Norway, who were asked to suggest actions in response to
either a negative or a positive result of the Helicobacter pylori rapid test (HPRT).
These actions were then categorised into three management strategies for initial
handling of the patient. i.e. only symptomatic treatment, referral for upper
endoscopy, or medication to eradicate H. pylori, so-called triple therapy. Discrete
choice analysis with multinomial logit models was used to analyse the choice of
medical actions.
Results: We find that the result of the HPRT has a major influence on actions
suggested and, in particular on symptomatic treatment and triple therapy strategies.
In addition, travelling time needed to have upper endoscopy, and whether or not
the GP makes a follow-up appointment with the patient, all have significant effects
on the choice of medical actions.
1
Conclusion: Since HPRT results seem to have a major influence on the GPs’
choice of medical actions, an important prerequisite for reimbursing such near
patient tests is fulfilled. Further, the analytical quality of test results is likely to
affect patients’ health and social costs.
Keywords: Discrete choice models, Decision-making, Primary Health Care
Acknowledgements
The authors are grateful to Tor Iversen, The Health Economic Research
Programme at the University of Oslo (HERO), and John Dagsvik, Statistics
Norway for valuable guidance.
2
1.0 Introduction
Laboratory analyses are widely used in diagnostic work-up and monitoring of
patients in primary care. Getting the correct treatment at an early stage may
decrease the patient’s likelihood of developing a more severe form of the illness
that requires more complex and costly care. Thus, the economic consequences for
laboratory analyses may amount to a substantial proportion of health service costs.
In a hospital setting, laboratory services contribute to 60-70% of all critical
decision-making concerning admittance, medication, and discharge of patients
(Forsman, 1996). Interest in using laboratory analyses in the general
practitioner’s (GP) office has been increasing in Europe (DeNeef, 1986; Fischer,
1991; Hjortdahl, 1990). In Norway (Source: NOKLUS 1 ), Switzerland and the
Netherlands almost all practices have their own laboratory facilities (Leurquin et
al. 1995). It has been shown that GPs with near patient tests available in their
office laboratory tend to request it more often than GPs relying on hospital-based
tests (Fauli and Thue, 2005; Rink et al. 1993). In Norway (NIA, 2007) 24.3% of
all fees in general practices in 2006 were from the use of laboratory services.
In spite of the wide use of laboratory analyses there is a lack of studies on the
impact of laboratory results and the economic consequences of near patient2
laboratory results in primary care (Delaney et al. 1999).
In this paper, our main purpose is to develop a model to study the impact of a
test result with regard to medical actions taken in general practice. We also study
the effect of characteristics of the GP and the practice on the medical actions
chosen. Such information may be important for implementing strategies to be able
to secure rational use of laboratory tests.
Dyspepsia is a fairly common presenting symptom in general practice
consultations (Haug, 2002; Kristensen et al. 1985; Logan and Delaney, 2001;
Petersen, 2001). Sometimes dyspepsia is due to peptic ulcer, and the bacterium
1
NOKLUS (The Norwegian Quality Improvement of Laboratory Services in Primary Care) is an
organization working to ensure that laboratory tests performed outside hospitals are requested,
carried out, and interpreted correctly and in accordance with patients’ needs for investigation,
treatment, and follow-up.
2
Near patient test is defined as any test performed outside a hospital laboratory where the result is
available without the sample being sent to a laboratory for analysis, (Hobbs et al.,1997)
3
Helicobacter Pylori (HP) has been identified as the main cause of this disease. The
presence of this bacterium may be detected by the HP rapid test (HPRT), which is
a simple test kit for single use, onto which a drop of blood is applied to test for the
presence of specific antibodies. The result is read as negative or positive within a
few minutes. The HPRT is a fairly new test and may be crucial when testing
younger patients with dyspepsia in that other laboratory tests are generally not
needed. Upper endoscopy will serve as a definitive test since endoscopy will
detect peptic ulcer as well as the presence of viable bacteria with great certainty.
Thus, dyspepsia and the HPRT were chosen for our impact-modelling study.
The data used in this paper are based on a case history based questionnaire
mailed in the spring of ‘99 to 739 GPs, who had the HPRT in their surgery.
In Fauli and Thue (2008), using the same clinical vignette, we have developed
a model for economic evaluation of diagnostic accuracy in general practice. But
diagnostic accuracy will only be important if the test result has a significant
influence on the choice of medical actions. Further, in Norway GPs are reimbursed
for the use of several near patient tests, and the less impact of test results, the less
motivation for the government to spend money on reimbursement.
We assume that the clinical practice of GPs for treating ulcer in Norway is
representative of the clinical practice of GPs in other developed countries, since
internationally acclaimed guidelines on dyspepsia were easily available to
Norwegian GPs, e.g. in the Norwegian textbook of general practice (Hunskår,
1997) and in an editorial in the Journal of the Norwegian Medical Association
(Petersen and Johannessen, 1998). Thus, Norway seems well suited for a study on
the impact of near patient test results.
It also seems reasonable to assume that the model may be applied generally for
clinical situations in which other diagnostic tests are used as important sources of
information e.g. some other laboratory analyses, x-ray and MRI. Hence, our
method has interest beyond the setting used in this article.
Our study has two key components:
•
First, a case history based survey of GPs’ management of a patient with
dyspepsia was performed. GPs included were those with the HPRT
4
available in their surgery and who chose to use it in the clinical situation
described in the vignette. GPs were asked to state medical actions in
response to both a negative and a positive test result. These actions were
then categorised into three management strategies for initial handling of
the patient i.e. only symptomatic treatment to ease the dyspepsia, referral
for upper endoscopy, or medication to eradicate H. pylori, so-called
triple therapy. In addition, we obtained information on the socioeconomic characteristics of the GPs.
•
Second, the chosen actions were analysed in multinomial logit models.
We assume that the GP’s medical actions depend on what he or she
thinks is best for the patients based on the best clinical evidence
available to the GP.
To the best of our knowledge, there are no other studies on the significance of
how a near patient test results and socio-economic characteristics of the GP affect
the choice of medical actions in primary health care.
Our study shows that the result of the HPRT has a significant and major
influence on the GP’s choice of medical actions.
2.0 Methods
2.1 The survey of GPs’ management of a patient with dyspepsia
The survey
The data used in this paper are based on a case history (Box 1) based questionnaire
mailed in the spring of ‘99 to 739 GPs, who, according to information from
NOKLUS, had the HPRT in their surgery.
The case history described a 31-year old woman with dyspepsia, and was used
to assess the clinical reasoning and decisions made by general practitioners. This
clinical situation was considered fairly familiar to the GPs, and in fact, with some
modifications, the case history depicts a real patient. It was chosen by one of us
5
(GT) from a number of medical record notes of consultations in which GPs had
ordered the near patient test in real life.
Anette Hansen
is 31 years old and works for 5 hours a day in the afternoon/evening as a cleaner. Married, usually happy at home, two
children aged 11 and 6 years.
During the past month she has had epigastric pain with a feeling of hunger, and some relief on eating. Experiences that
the pain increases when she under stress. Slightly loose and irregular defecation at times.
She had a similar episode just under a year ago, and then recovered rapidly with Zantac 150 mg x 2, which she took for
just over a week during her summer holidays. No other measures were taken at this consultation. She smokes 10
cigarettes a day, 2-3 cups of coffee, consumes little alcohol. No medication.
When you examine her this Tuesday she is slightly tender over her epigastrium, no other findings.
She should be at work later today.
Box 1 The case history in the questionnaire
Minor modifications were made in collaboration with several clinicians (GPs and a
gastroenterologist), and a microbiology specialist. It was an important element of
the vignette that additional tests should not be necessary.
In the accompanying questionnaire the GPs were asked to state:
-
the probability that “Mrs. Hansen's” symptoms were caused by HP before
obtaining a HPRT result (pre-test probability)
-
whether or not they would request a HPRT in this situation
-
what actions they would take based on the case history (symptoms and
clinical findings), or on the history in addition to the test result, as well as their
assessment of the relative importance of history, clinical findings and HPRT
result. Actions suggested in the questionnaire were (more than one could be
chosen):
1. lifestyle advice
2. recommend traditional antacids (Balancid etc.)
3. prescribe H2 antagonists (Zantac etc.) (i.e. more potent antacids)
4. prescribe triple therapy (antibiotics plus a potent antacid)
5. refer for breath test (detects viable H. pylori bacteria in the stomach)
6. refer for upper endoscopy (gastroscopy; enables detection of anatomical
changes such as ulcus, as well as viable bacteria)
6
7. recommend sick leave
8. set up a new appointment
9. new appointment suggested to be initiated by the patient if she did not
recover
In addition to medical questions, a section on socio-economic variables and
travelling- and waiting time for upper endoscopy was included in the
questionnaire.
The dependent variable
We reduced the alternative actions eligible for categorization into strategies (the
dependent variable) as follows:
- “lifestyle advice” was given by nearly everyone, and we therefore did not
include this alternative as a medical action
- “sick leave” (alt. 7) and “new appointment” (alts. 8 and 9) are coded as
characteristics of the GP (see section five) and therefore not included in any
strategy.
When grouping the different sets of medical actions we therefore focused on
alternatives two through six.
Each GP responded first to when the test result was taken to be negative, and
then to when it was positive. In these two instances, medical actions chosen were
categorised into one of three strategies: symptomatic treatment (only alts. 2 and 3
were chosen), referral (at least alts. 5 or 6 were included) and triple therapy (at
least alt. 4 was included).
2.2. Hypotheses regarding independent variables
Table 2 gives an overview of the variables included in the impact model.
The pre-test probability is the GP’s assumption that the patient in our case
history had an HP infection as the cause of her dyspepsia before the HPRT was
taken and medical actions chosen. We assume that if the GP states a low pre-test
probability, he will be inclined to choose symptomatic treatment, and if the GP
states a high pre-test probability he will be inclined to choose triple therapy.
7
GPs were asked to distribute ten points between the case history, clinical
findings described, and the laboratory result (negative or positive), by giving most
points to the factor he/she considered most important. We assume that the GPs
who allotted a relatively high score to the importance of the HPRT will tend to
choose triple therapy if the test is positive, and symptomatic treatment if the test is
negative.
Upper endoscopy is the preferred diagnostic procedure for detecting peptic ulcer
and is done either in hospitals or by primary care internists. Norway is a scarcely
populated country, and many inhabitants live hours away from the nearest
gastroscopy facility. The GPs were asked to state both the travelling time (hours)
and the waiting period (weeks) to get an upper endoscopy. Referral to upper
endoscopy will be more inconvenient for the patient if the waiting time or the
travelling time is long, and thus we assume that the probability of referral
decreases with increased travelling or waiting time.
The GPs who prefer to make a follow-up appointment with the patient may
choose medical actions that demand more follow-up. Only prescribing
symptomatic treatment may induce more follow-up by the GP than referring the
patient for upper endoscopy. We assume, therefore, that the GPs who make a new
appointment or ask the patient to make a new appointment tend to give
symptomatic treatment rather than referral or triple therapy.
General practices in urban areas may face competition for patients (Iversen,
2005), and one way of getting a competitive advantage is to give quicker service to
the patients, including referral to specialist care. We assume, therefore, that,
compared with GPs located in rural and semi-urban areas, GPs in urban areas have
a higher probability of choosing referral or prescribing triple therapy versus
resorting to symptomatic treatment.
In Norway, primary health care is the responsibility of the municipalities. In
principle, either GPs are paid fee-for-service, or they are on a fixed salary as
employees. We include whether the GP is on fee-for-service (in private practice)
as a control variable in the empirical analysis, without having any particular
8
hypothesis regarding the effect of private practice on clinical decision-making in
this particular case.
2.3 Data
Responses were received from 425 GPs (57%), of whom 210 decided to use the
HPRT. Of these, nine GPs who had obviously misunderstood the questionnaire, or
had very deviant characteristics (e.g. less than 10 working hours per week) were
excluded. The focus of our study, therefore, is the medical actions taken by the
201 eligible GPs who decided to use HPRT in this situation, responding first to a
negative and then to a positive HPRT result “received” after the vignette
consultation. We compared age, sex, and type of payment in our sample of GPs
using HPRT with the total Norwegian population of GPs from a register kept by
the Norwegian Medical Association. The sample had the same mean value
regarding age, but had a slightly higher percentage of men (81% versus 73.6%),
and fewer on fixed salary (8% versus 28%).
2.3.1 The dependent variable
The GPs chose many different sets of actions, which were categorized according
to main medical strategies. Table 1 shows that the medical actions strongly depend
on the result of the HPRT.
Table 1 Medical strategies (dependent variable) in response to a negative or a
positive HPRT, as suggested by GPs (n=201)
Independent variable
HPRT neg HPRT pos
Medical strategies
1. Symptomatic treatment 112 (56%)
8 ( 4%)
2. Referral
85 (42%) 100 (53%)
3. Triple therapy
4 ( 2%)
83 (43%)
2.3.2 The independent variables
Table 2 gives an overview of our data for the 201 GPs as well as responses to
either a positive or a negative HPRT result.
9
Table 2 Overview of independent variables (n=201 GPs)
Variables
Definition
Mean. Std.dev
Data on the GPs and practice characteristics
Sex
Age
Need of info on
HPRT
Type of info on
HPRT
Group practice
Urban
Semi-Urban
Rural
Consultations
Working hours.
Private practice
Specialist
Wait.upper endo.
Trav.upper endo.
Binary variable:1 if male, 0 if female
Number of years
Need for information about the use of HPRT; Binary: 1 if
some or a lot, 0 if none or only modest
The two most important information sources of HPRT.
Binary variable; 1 if only suppliers info, 0 if other
Type of practice. Binary variable: 1 if group practice, 0 if
solo practice
Reference category for location of practice:
Binary variable: 1 if inhab.>15000, 0 if other
Category for location of practice. Binary variable: 1 if
5000≤inhab.≤15000, 0 if other
Category for location of practice
Binary variable: 1 if inhab.<5000, 0 if other
Number of consultations per week
Number of working hours per week
Binary:0=fixed salary, 1=are reimbursed
The GP’s education. A number of courses are required to
have a specialist certificateBinary variable: 1 if specialist certificate, 0 if other
Waiting time in weeks for upper endoscopy assumed by
the GP
Travelling time in hours for the patient (one way) for upper
endoscopy assumed by the GP
0.806
46.200
0.642
7.299
0.517
0.776
0.632
0.209
0.159
89.139
35.129
0.92
0.721
27.316
8.088
4.900
3.700
1.015
3.495
The pre-test probability that Mrs Hansen's symptoms are
due to a HP infection
The post-test probability that Mrs Hansen's symptoms are
due to a HP infection
The post-test probability that Mrs Hansen's symptoms are
due to a HP infection
The relative importance of the HPRT-test result on a scale
from 1 to 10 (see text)
The relative importance of the HPRT-test result on a scale
from 1 to 10 (see text)
Binary:0= no sick leave, 1=sick leave suggested
49.459
20.515
15.652
16.312
76.375
18.985
2.797
1.447
3.923
1.8
Binary:0= no sick leave, 1=sick leave suggested
0.318
Binary:0= no appointm.1=new appointm suggested by GP
0.438
Binary:0= no appointm.1=new appointm.
0.477
Binary:0= no appointm.1=new appointm. (patient makes a
new appointment as needed)
Binary:0= no appointm.1=new appointm.
0.468
GPs’ suggestions in response to the case history
Pre-testprobability
Post- test prob of
negative test
Post -test prob of
positive test
Importance of
negative test
Importance of
positive test
Sick leave – neg
test
Sick leave – pos
test
New appointment
negative test
New appointment
positive test
Patient initiated
appointm. – neg
Patient initiated
appointm. – pos.
0.269
0.248
On the average, the pre-test-probability is 49.5%, indicating that the GP needs
further information in order to choose a medical strategy, i.e. a test result. The
10
post-test probability measure changes in pre-test probability when the HPRT result
is known.
Table 2 shows that a negative test result decreases the average probability by
33% while a positive test result increases the probability by 27%.
3.0 Empirical models
The theoretical framework is based on the theory of discrete choice analysis; see
Greene (2000) and Malchow-Moeller and Svarer (2003).
We want to study variables influencing the GPs’ choice of medical actions when
using a laboratory test in a specific situation, and thus to predict the probability of
a GP choosing a particular medical strategy. We have data on GPs hypothetical
choices among three alternatives (symptomatic treatment, referral, and triple
therapy) that are mutually exclusive, and will use the framework of multinomial
logit models to analyse these data. In addition, GPs stated the attributes of the
choices and socio-economic characteristics.
All the GPs evaluated the same patient vignette – so the focus here is on
responses which may reflect the GP’s own objectives and preferences, knowledge,
experience and uncertainty, since the patient’s preferences are not included in the
vignette.
We assume that the impact of income in the GPs utility function is small due to
the small net surplus of using the HPRT (Fauli and Thue, 2005), and thus that the
GP’s medical actions depend on what he or she thinks is best for the patient, based
on the best clinical evidence available. The usefulness of a laboratory analysis to
detect an HP infection will depend on the pre-test probability stated by the GP, and
that the test is performed correctly and the result properly interpreted by the GP.
The GP’s choice setting can be viewed as a case of choice between lotteries
because of the uncertainty of the initial health status of the patient described as
well as of the laboratory analysis. The uncertainty of the laboratory analysis occurs
because the HPRT measures antibodies to the HP bacteria and not the disease as
such, and because healthy carriers of the HP bacteria exist. Because some patients
are carriers of HP without being ill, the use of the test result will depend on the
11
GP’s knowledge of HP and its role in dyspepsia. However, our model is a reduced
form version in the sense that the treatment of uncertainty from the individual
GP’s point of view is not modelled.
The GP may also have unstable preferences, meaning that he or she may have
problems in evaluating the expected utility of the different alternatives. This means
that the GP may make different choices in seemingly identical choice settings. The
variation in responses across seemingly identical choice settings may change
because colleagues, medical journals, and experience gained from treating other
patients have a continuous influence on the GP. Further, there will be nonsystematic variation in the choices that cannot be explained by the variables
available to the researcher, especially those stemming from the patient-physician
interaction.
Recall that we have two observations per GP, one set of medical actions when
the HPRT is negative and one set of medical actions when the HPRT is positive.
Encounters by the same GP may be correlated and then standard regression
techniques may not be suitable. To take this into account, we use a multinomial
logit model with random effects or a so-called mixed multinomial logit model with
normal mixing distribution, which is a method used for panel data.
Let Vijt be the utility for GPi, as evaluated by the GPi, with respect to the
events mentioned above, given alternatives j = 0,1,2. We assume that the utility
Vijt is stochastic and has the representation
(1)
Vijt= αij + Xitβj + Zijtθj + εijt
where βj and θj are alternative specific parameter vectors, t = 1 the test result is
negative and when t = 2 the test result is positive, Xit is a vector of individual
characteristics of the GPs [sex, age, need of info on HPRT, type of info on HPRT,
type of practice (fixed salary/fee for service), group/solo practice, location of
practice, number of consultations per week, working hours per week, has a
specialist certificate or not, pre-test probability, relative importance of test result,
suggestions with regard to sick leave, a new appointment, a patient initiated
12
appointment], in addition to the test result. The attribute vector Zijt is individual
and alternative-specific, stated by the GP (travelling time and waiting time for
endoscopy). The travelling time and waiting time will vary among GPs and are
only relevant for referrals, that is, travelling times and waiting times are zero for
the other alternatives. The terms αij and εijt are unobserved random components.
The terms αij represents a time constant, and individual specific levels of the
respective utilities across seemingly identical choice settings accounts for the GP’s
practice style. The term εijt is the stochastic term that is supposed to account for
unobserved variables of the GP that affect his preferences and vary randomly over
observations per GP and across alternatives.
Since the relation in (1) is an expected utility relation it also depends on the
probabilities of the possible outcomes given strategy (lottery) j. Since these
probabilities are not known, this relation must be interpreted as a reduced form
one.
In our model symptomatic treatment is the reference, and the models predict
the probability of referral versus symptomatic treatment and the probability of
choosing triple therapy versus symptomatic treatment.
Assuming that the εijt are all independently distributed according to a type I
extreme value distribution, it follows from well known results (Greene, 2000), that
P(Yij = 1⏐Vi,αi ) =
(2)
exp(αij + Xitβj + Zijtθj )
2
∑
exp(αik + Xitβk + Ziktθk)
k =0
for j = 0, 1, 2, where Vi is a vector that contains all explanatory variables for GP i
and αi = (αi0, αi1, αi2). Furthermore, it is assumed that the random effect
parameters are independent across alternatives and across individuals, and
normally distributed. The mean of one of the random effects (for symptomatic
treatment) is normalized to zero. Consequently, the model for the observed choices
is given by
13
Pij=P(Yij = 1|Vi) = E P(Yij = 1⏐Vi,αi ),
(3)
where the expectation is taken with respect to αi. Thus, the final model in (3) is a
so-called mixed multinomial logit model with normal mixing distribution or
alternatively a multinomial logit model with normally distributed random effects
(Cameron and Trivedi, 2005).
The maximum likelihood estimation procedure for multinomial logit models
with random effects is described in Malchow-Moeller and Svarer (2003). The
model is estimated in NLOGIT 4.0 (Greene, 2007).
4.0 Results
We examine the effect of the variables included in Table 2 on the probability for
choosing different medical actions by estimating a mixed multinomial logit model,
specified in (2) and (3).
Since four GPs chose no medical action, the number of observations was
reduced from 402 to 394, and further reduced to 367 for the standard model due to
missing values. For the mixed model we needed balanced data and had to exclude
GPs who had only chosen one set of medical actions, and thus included 350
observations.
Table 3 shows the results from both the standard multinomial logit model and
the mixed multinomial logit model for variables that were significant as well as all
laboratory-related variables. The full table is in Appendix A and shows the values
and the t-ratios on the estimated parameters in the models. By using the LR-test,
we found that the standard multinomial model tested against the mixed
multinomial logit model is rejected (LR-stat 9.805, 5% significance level), and we
will, therefore, focus on the mixed multinomial logit model when interpreting
results. The estimated parameters in the mixed model are slightly less significant,
but have a larger effect on the probability of choice of medical strategy. The pretest probability and the relative importance of the HPRT result, versus history and
clinical findings, are not significant in the mixed model.
14
Table 3 shows that if the HPRT-result is positive versus negative the GP is more
likely to choose referral or triple therapy versus symptomatic treatment. This
seems reasonable because if the HPRT-test is positive, there are reasons for further
investigations to find out whether this patient has an HP-infection, or to adapt a
test-and-treat strategy with no further investigations. The result is consistent with
Table 1.
Table 3 Results of estimations from the standard- and mixed multinomial logit models.
Reference: Symptomatic treatment
Independent variables
Medical action
Standard Model
Mixed Model
Parameter
t-ratio Parameter
t-ratio
Constant
Referral
1.212
0.724
2.059
0.847
Triple therapy
-4.743
-1.921
- 13.477
-1.453
Referral
HPRT: positive vs. negative
3.000
6.188
3.895
4.247
Triple therapy
6.069
7.957
11.178
2.506
Relative importance of HPRT,
Referral
-0.075
-0.684
-0.122
-0.784
result vs. history/findings
Triple therapy
1.027
1.522
0.311
2.194
Pre-test-probability
Referral
0.007
0.877
0.011
0.895
Triple therapy
0.016
1.374
0.030
1.136
Travelling time
Referral
-0.655
-2.072
-1.291
-2.004
Triple therapy
New appointment vs. not a new
Referral
-1.791
-3.837
-2.456
-3.207
appointment
Triple therapy
-0.957
-1.669
-0.361
-0.287
New appointment initiated by
Referral
-2.236
-4.724
-3.049
-3.651
patient vs. not a new appointm.by Triple therapy
-1.203
-1.937
-0.994
-0.729
patient
Sick leave vs. no sick leave
Referral
1.129*
1.941
0.853
2.225
Triple therapy
3.471*
1.868
1.670
3.355
Variance of the random effect
Referral
1.587
2.385
Triple therapy
3.786
1.590
Log-L
-344.2526
-338.4920
Restricted Log-L
-485.2030
McFaddens R2
0.285
0.302
McFaddens adjusted R2
0.256
0.275
Bold figures indicate that the effect is significant at 5% level, *close to significant i.e. p=0.053 and
p=0.061).
If the travelling time to get endoscopy increases, the GP is more likely to prefer
symptomatic treatment than referral.
GPs that make a new appointment or ask the patient to make a new
appointment if she does not recover will tend to prefer symptomatic treatment
compared to the GPs who do not arrange for a follow up. This makes sense, since
only prescribing symptomatic treatment demands more follow up from the GP in
case a referral should be necessary later.
15
GPs who recommend sick leave are more likely to choose referral than
symptomatic treatment, compared with the GPs who do not recommend sick leave.
This is probably because GPs choosing referral versus symptomatic treatment
assume the patient to have a more serious dyspepsia.
We have conditional choice probabilities, and the probability of choosing
referral or triple therapy versus symptomatic treatment depends on whether the
GPs have recommended sick leave or not. This aspect is further discussed in
section five.
Using NLOGIT 4.0 (Greene, 2007) the model allows for calculating detailed
strategy probabilities for a particular GP32 including the effect of changes in
significant attribute or characteristics. Table 4 shows the probability that “our” GP
chooses the different medical strategies, given that he does not recommend sick
leave for Mrs. Hansen and the HPRT result is positive: 2.5% for symptomatic
treatment, 74.7% for referral and 22.8% for triple therapy. The HPRT has a major
influence on the GP’s decision on strategy, and in particular on symptomatic
treatment and triple therapy which is consistent with Table 1. Whether or not the
GP makes a new appointment or the patient is asked to make a new appointment,
has the most influence on referral. If the HPRT is taken to be negative, the
probability for choosing symptomatic treatment will increase by 42 percentage
points to 44.5% and the probability of using triple therapy will decrease by 22.4
percentage points to 0.4%. If the GP asks the patient to make a new appointment,
the probability of using referral will decrease by 27.5 percentage points and the
probability of using symptomatic treatment and triple therapy will increase by 15.7
and 11.8 percentage points respectively.
3
male, 46 years old, needs information on the use of HPRT, the HPRT is positive
does not use supplier's information only, works 35 hours with 89 consultations per week in a group
practice, is paid fee-for-service, lives in an urban area, estimates a pre-test-probability of 30% ,
does not have a specialist licence in general practice/family medicine, gives the relative importance
of the HPRT 3 points on a scale from 1 to 10, does not schedule a follow-up appointment, has one
hour’s travelling time for upper endoscopy, and 5 weeks waiting time
16
Table 4. Impact on average choice probabilities from changes in selected
explanatory variables for standard and mixed models respectively
Independent variables
Strategy probabilities for “our
GP”, HPRT positive
Result of the HP-analysis
Positive result → negative result
Travelling time 1→ 0
Not a new appointment → new
appointment
The patient is not asked to make a
new appointment → is asked
Does not → does recommend
sick leave
Symptomatic
Referral
Triple therapy
Std
Mixed Std
Mixed Std
Mixed
2.7
2.5
72.3
74.7
25.0
22.8
39.6
42.0
-15.5
-19.6
-24.1
-22.4
-1.1
8.3
-1.6
10
11.1
-22.8
9.1
-17.7
-10
14.5
-7.5
13.8
12.3
15.7
-29.2
-27.5
16.9
11.8
-1.8
-2.0
-16.7
-16.7
18.5
18.7
5.0 Discussion
The main purpose of our study was to construct a model for the impact of a near
patient test. There are several limitations regarding our findings using a case
history based approach and the method chosen.
First, the results are conditional based on the GP having already chosen
whether or not to carry out the test in private practice, and whether or not to
recommend sick leave. In Fauli and Thue (2005) we found that there was a
selection effect because the variable “private practice” depended on characteristics
of the GP. We also assume that whether or not the GP recommends sick leave
depends on the characteristics of the GP. If so, the variables “sick leave” and
“private practice” become endogenous variables and will correlate with the error
term. Hence, our estimates may be biased due to the selection effect, and this
aspect needs further investigation in another study.
Second, we do not have any information about the sensitivity and the
specificity of the HPRT actually used in the GP’s office. Several studies show that
test results are perceived to be accurate in various clinical settings (Aakre et al.
2008; Fraser and Fogarty, 1989; Skeie et al. 2005), implying that the GPs usually
do not consider sources of error involved in laboratory work when interpreting
17
results, i.e. a single result is taken at “face value” and acted upon. We, therefore
assume that the GP consider the test-result to be correct without considering
diagnostic accuracy. The difference between the post-test and pre-test probability
(ref. Table 2) somewhat supports this assumption.
Third, we wanted to measure the effect of the test result and there were several
options: including focusing on both the pre-test probability and the test result,
focusing on post-test probability, or focusing on the difference between the posttest probability and the pre-test probability. We chose to include the pre-test
probability and the test result variables because this improved the model, and
because we are particularly interested in identifying the effect of the test result in
the impact model.
Fourth, we have revealed preference data, as our data is based on a
questionnaire where the GP is assumed to have enough information to establish a
preliminary diagnosis and conduct reasonable medical actions. Thus, when we
composed the clinical vignette it was important to describe a realistic situation. In
the literature there have been discussions about the validity of written case
scenarios in medical decision-making. One might argue that by using a clinical
vignette we measure competence (what a physician is capable of doing), and not
performance (what a physician actually does in his day-to-day practice). However,
Kuyvenhoven et al. (1983) conclude that written simulations give a realistic
impression of a GP’s diagnostic and therapeutic approach to patients with vague
symptoms as was the case in our clinical vignette. In a review of 74 published
studies using written simulations, the validity issue was addressed in only 11
studies, and the conclusions were conflicting (Jones et al. 1990). Peabody et al.
2000 have validated clinical vignettes as a method for measuring the competence
of physicians and the quality of their actual practice, and conclude that the quality
of care can be measured by using clinical vignettes. Bias is more likely if the
respondents feel obliged to display some kind of expected behaviour or if the
written scenario differs from a typical situation. However, our case history depicts
a real patient with some minor modifications, in order to make the situation as
18
realistic as possible, and Norwegian GPs are used to responding to clinical
scenarios like these, thus making a less biased response plausible.
With regard to our use of clinical scenarios, two studies (Redelmeier and
Tversky, 1990; Sandvik, 1995) imply in our setting that travelling and waiting
time for upper endoscopy in a real consultation may have a higher impact on the
choice of medical strategies than with a case history. Redelmeier and Tversky
(1990) noted that physicians give more weight to the personal concerns of patients
when considering them as individuals as compared with the aggregate perspective,
and more weight to general criteria of effectiveness when considering them as a
group. Sandvik (1995) studied the validity of responses to patient vignettes in a
situation related to the management of female urinary incontinence, and found that
when cueing items were provided, the physician claimed more actions with
vignettes than were actually performed. This is of minor importance for the
construction of the impact model, but may imply that the GP in a real situation
might resort more to symptomatic treatment or a test-and-treat strategy.
Finally, the response rate was 57% which may imply selection bias in results.
Still, participants were similar to the total population of Norwegian GPs regarding
age and sex, but fewer were on a fixed salary (8% (see Table 2) vs. 28%) since the
H. pylori test was more abundant among GPs on fee-for-service. If selection bias
is present, we believe it would be sensible to assume that participants are more
knowledgeable of dyspepsia than non-responders, and therefore, more likely to
adhere to medical guidelines in this field.
6.0 Conclusion
We have developed a model that can be used to evaluate the effect of a specific
laboratory analysis, or other “crucial” diagnostic measures, on choosing medical
actions. By using discrete choice analysis, we have seen that the result of the
HPRT has a very important influence on GPs’ choice of medical actions, thereby
fulfilling an important prerequisite for reimbursing near patient tests. In addition,
the accuracy of test results is important and cost-effectiveness analyses of
measures to achieve sufficient analytical accuracy should be done. Still, further
19
studies of different types of laboratory analyses and other “crucial” diagnostic
measures, e.g., MRI or x-ray, are needed before we can draw general conclusions.
20
References
Aakre KM, Thue G, Subramaniam-Haavik S, Bukve T, Morris H, Müller M,
Lovrencic MV, Plum I, Kallion K, Aab A, Kutt M, Gillery P, Schneider N,
Horvath AR, Onody R, Oosterhuis W, Ricos C, Perich C, Nordin G, Sandberg S.
Postanalytical external quality assessment of urine albumin in primary health care:
an international survey. Clinical Chemistry 2008;54;1630-6.
Cameron C, Trivedi PK. Microeconometrics: Methods and Applications.
Cambridge University Press, New York, 2005.
Delaney BC, Hyde CJ, McManus RJ, Wilson S, Fitzmaurke DA, Jowest S, Ros
T, Thorpe GH, Hobbs R. Systematic review of near patient test evaluations in
primary care. British Medical Journal 1999, 319; 824–827.
DeNeef P. The expanding role of the office laboratory. Journal of Family Practice
1986; 22; 215-216.
Fauli S, Thue G. Decision making in general practice: the availability and use of a
specific laboratory analysis. International Journal of Technology Assessment in
Health Care 2005;21;386-92.
Fauli S. Thue G. Economic consequences of near-patient test results: the case of
tests for the Helicobacter Pylori bacterium in dyspepsia. European Journal of
Health Economics 2008;9;221-8.
Fischer PM. Laboratory testing in the 1990s. Journal of Family Practice 1991;
33;453-454.
Forsman RW. Why is the laboratory an afterthought for managed care
oganizations? Clinical Chemistry 1996;42;813-6.
21
Fraser CG, Fogarty Y. Interpreting laboratory results. British Medical Journal
1989: 298;1659-1680.
Greene WH. Econometric Analysis, Prentice-Hall International, Englewood Cliffs,
New Jersey, 2000.
Greene WH. NLOGIT version 4.0, Users manual, Econometric Software,
Australia, 2007.
Haug TT, Functional dyspepsia-a psychosomatic disease. The Journal of the
Norwegian Medical Association 2002;122;1218-22.
Hjortdahl P. The silent revolution. Scandinavian Journal of Primary Health Care
1990; 8;188-190.
Hobbs FDR, Delaney BC, Fitzmaurice DA, Wilson S, Hyde CJ, Thorpe GH, et al.
A review of near patient testing in primary care. Health Technology Assessment
1997;5.
Hunskår S (ed). Allmennmedisin – Klinisk arbeid (Norwegian textbook of general
practice). Ad Notam Gyldendal; 1997.
Iversen T: A study of income-motivated behaviour among general practitioners in
the Norwegian list patient system. University of Oslo. Health research programme.
Working paper 2005; 8.
Jones TV, Gerrity MS, Earp J. Written case simulations: do they predict
physicians’ behaviour? Journal of Clinical Epidemiology 1990;43; 805-15.
22
Kristensen P, Sandbakken P, Johannessen T, Løge I, Hafstad PE. Petersen H,
Fjøsne U, Kleveland PM. Fordøyelsessykdommer I almenpraksis. The Journal of
Norwegian Medical Association 1985;105;728-31.
Kuyvenhoven MM, Jacobs HM, Touw-Otten FWMM, Van Es JC. Written
simulation of patient-doctor encounters. 1: research instrument for registration of
the performance of general practitioners, Journal of Family Practice 1983;1;14-19.
Leurquin P, Van Casteren V, De Maeseneer J. Use of blood tests in general
practice: a collaborative study in eight European countries. Eurosentinel Study
Group. British Journal of General Practice 1995;45;21–25.
Logan R, Delaney B. ABC of the upper gastrointestinal tract. Implications of
dyspepsia for the NHS. British Medical Journal 2001; 323:675-677.
Malchov-Møller N. and Svarer M. Estimation of the Multinomial Logit Model
with Random Effects, Applied Economics Letters 10, 2003;389-392.
National Insurance Administration (NAV). What’s going on in general practice?
2007. Report nr. 4.
Peabody JW, LuckJ. Glassman P, Dresselhaus TR, Lee M .Comparison of
vignettes, standardized patients and chart abstraction, Journal of the American
Medical Association 2000;283;1715-1722.
Petersen H. Legemiddelbehandling av funksjonell dyspepsi. (Medicinal treatment
of functional dyspepsia). The Journal of Norwegian Medical Association
2001;121;2839-40.
23
Petersen H, Johannessen T. Helicobacter pylori – vet vi hva vi gjør (Helicobacter
pylori – do we know what we are doing?). The Journal of Norwegian Medical
Association 1998;118;1540-1.
Redelmeier DA, Tversky A. Discrepancy between medical decisions for individual
patients and for groups. The New England Journal of Medicine 1990; 322; 116264.
Rink E, Hilton S, Szczepura A, Fletcher J, Sibbald B, Davies C, Freeling P,
Stilwell J. Impact of introducing near patient testing for standard investigations in
eneral practice. British Medical Journal 1993; 307; 775–778.
Sandvik H. Criterion validity responses to patient vignettes; an analysis based on
management of female urinary incontinence, Family Medicine 1995;27;388-92.
Skeie S, Perich C, Ricos C, Araczki A, Horvath AR, Oosterhuis WP, Bubner T,
Nordin G, Delport R, Thue G, Sandberg S. Postanalytical External Quality
Assessment of Blood Glucose and Hemoglobin A1c: An International Survey.
Clinical Chemistry 2005;51;1145-1153.
24
Appendix A. Estimation results – full models.
Table A.1 Results of the estimation of the model for GPs with HPRT- Reference: Balancid/Zantac
Independent variables
Medical action
Standard Model
Mixed Model
Parameter
t-ratio Parameter
t-ratio
Constant
Referral
1.212
0.724
2.059
0.847
Triple therapy
13.477
-1.453
-4.743
-1.921
Referral
Result of HPRT:
3.000
6.188
3.895
4.247
positive vs. negative result
Triple therapy
6.069
7.957
11.178
2.506
Relative importance of HPRT
Referral
-0.075
-0.684
-0.122
-0.784
result vs. history/findings
Triple therapy
1.027
1.522
0.311
2.194
Pre-test-probability
Referral
0.007
0.877
0.011
0.895
Triple therapy
0.016
1.374
0.030
1.136
Age
Referral
0.005
0.206
0.008
0.223
Triple therapy
-0.001
-0.025
-0.025
-0.340
Sex
Referral
0.010
0.022
0.165
0.253
Triple therapy
0.086
0.140
0.338
0.278
Need of information
Referral
-0.325
-0.907
-0.364
-0.699
Triple therapy
-0.271
-0.553
-0.447
-0.443
Type of information
Referral
0.614
1.850
0.825
1.642
Triple therapy
0.741
1.638
1.292
1.271
Group practice:group vs. solo
Referral
-0.337
-0.816
-0.436
-0.719
Triple therapy
-0.636
-1.133
-1.175
-0.993
Semi-Urban vs. Urban
Referral
-0.365
-0.875
-0.517
-0.852
Triple therapy
-0.215
-0.383
0.177
0.144
Rural vs. Urban
Referral
0.286
0.562
0.547
0.747
Triple therapy
-0.368
-0.533
-0.761
-0.523
Private practice vs. fixed salary
Referral
-0.888
-1.315
-1.098
-1.105
Triple therapy
0.283
0.305
1.808
0.777
Travelling time
Referral
-0.655
-2.072
-1.291
-2.004
Triple therapy
Waiting time
Referral
0.013
0.377
0.008
0.131
Triple therapy
Working hours
Referral
0.001
0.042
-0.001
-0.030
Triple therapy
0.005
0.151
0.010
0.140
Consultations
Referral
0.003
0.357
0.003
0.243
Triple therapy
-0.001
-0.108
-0.005
-0.228
Specialist vs. not a specialist
Referral
0.064
0.157
0.077
0.129
Triple therapy
-0.300
-0.546
-0.912
-0.768
New appointment vs. not a new
Referral
-1.791
-3.837
-2.456
-3.207
appointment
Triple therapy
-0.957
-1.669
-0.361
-0.287
New appointment initiated by
Referral
-2.236
-4.724
-3.049
-3.651
patient vs. not a new appointment. Triple therapy
-1.203
-1.937
-0.994
-0.729
initiated by patient
Sick leave vs. not a sick leave
Referral
1.129*
1.941
0.853
2.225
Triple therapy
3.471*
1.868
1.670
3.355
Variance of the random effect
Referral
1.587
2.385
Triple therapy
3.786
1.590
Log-L
-344.2526
-338.4920
Restricted Log-L
-485.2030
McFaddens R2
0.285
0.302
McFaddens adjusted R2
0.256
0.275
Bold figures indicate that the effect is significant at 5% level, *close to significant at 5% level
(p=0.053 and p=0.061).
25