Association between physician density and health care consumption: A systematic review of the evidence

Health Policy 91 (2009) 121–134

Contents lists available at ScienceDirect

Health Policy journal homepage: www.elsevier.com/locate/healthpol

Review

Association between physician density and health care consumption: A systematic review of the evidence Christian Léonard ∗ , Sabine Stordeur, Dominique Roberfroid KCE – Belgian Health Care Knowledge Centre, Boulevard du Jardin Botanique, 55, 1000 Bruxelles, Belgium

a r t i c l e

i n f o

Keywords: Supplier-induced demand Cost containment Physician density Healthcare consumption Patient and physician responsibility

a b s t r a c t Background: Supplier-induced demand (SID) for health care could be a crucial factor of rising health expenditures. However, there is thus far no consensus on the topic. Objective: To assess how physician density (physician-to-population ratio) and health care consumption correlate. Methods: A systematic review of studies retrieved through electronic databases: Medline, Econlit, PsychINFO and Embase. Search, inclusion and quality appraisal were based on standard procedures and applied independently by two researchers. Results: Twenty-five studies, generally of moderate quality, were included. Despite a substantial heterogeneity in study design and data modelling, a significant association between physician density and health care consumption was consistently observed. However, estimates varied according to a number of method parameters such as the definition of the dependent variable (physician volume or care intensity), the geographical entity or the medical specialty under consideration, and the adjustment for confounding factors. Conclusions: The exact importance of SID and the underlying motivations remain poorly understood. We discuss technical issues for better SID assessment. In the absence of more accurate information, limiting physician supply as a measure of cost containment should also be considered cautiously. © 2008 Elsevier Ireland Ltd. All rights reserved.

Contents 1. 2. 3. 4. 5.

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Materials and methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conflict of interest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

121 122 123 129 133 133 133 133

1. Introduction It is generally assumed that the utilization of health services is partly induced by the providers themselves

∗ Corresponding author. E-mail address: [email protected] (C. Léonard). 0168-8510/$ – see front matter © 2008 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.healthpol.2008.11.013

[1]. In the early 1960s, Roemer already noted the existing correlation between the density of hospital beds and the rate of hospital stays and concluded that “a bed built was a bed filled” [2,3]. Extended to medical services, the “Roemer’s Law” became the well-known ‘supplier-induced demand’ (SID). Thus the SID refers to the phenomenon of physicians deviating from their agency responsibilities

122

C. Léonard et al. / Health Policy 91 (2009) 121–134

to provide unnecessary care with the main objective of increasing their own pecuniary resources. Health expenditures are on the rise worldwide [4,5], and SID is considered a substantial contributor to it, together with economic growth, technical innovations and population ageing. This has been a strong argument to limit physician numbers in many countries, most often by restricting the number of medical trainees [6,7]. For instance, such numerus clausus was recently implemented in Belgium, a country where the very high physician density (35 medical doctors per 10 000 inhabitants [8]) was considered by the late 1990s to not longer be financially sustainable [9]. However, there is thus far no consensus on the existence and the exact importance of SID. Different factors could explain such uncertainty. First, outcomes studied have varied considerably from patient health [10–12] and quality of care [13] to economic impact [14]. Second, a variety of appraisal strategies, from ecological studies to regression modelling of individual data, have been applied. Lastly, it remains difficult to distinguish inducement by suppliers and by patients themselves [15,13,16]. The diverse, and sometimes contradictory, resulting evidence has made a global appraisal difficult. Some authors have even underscored that SID, when present, is often of small magnitude [17] and could be context-dependent, particularly in relation to the medical fees scheme [11,18,19]. Considering the importance of the topic for policymakers concerned in the fair allocation of public resources, as well as for health providers and patients, we systematically reviewed the scientific evidence on how physician density (physician-to-population ratio) and health care consumption correlate. Our review did not address specifically the link between physician density and income because income is fully correlated to activity volume when fees are pre-determined, while in other physician remuneration systems disentangling the relative contribution of those factors (i.e. modifications in fees and/or activity volume) is difficult [20]. However, it can be fairly assumed that an increased activity volume will result in higher income, regardless of the physician remuneration system. 2. Materials and methods We searched the following electronic databases: Medline (access: OVID), Econlit (access: OVID), PsychINFO (access: OVID) and Embase (access: Embase.com). We combined, with the Boolean operators OR and AND, the following keywords in the four databases: induced demand OR supplier inducement OR supplier$ induced demand OR physician$ induced demand OR physician$ created demand OR physician$ initiated demand OR demand for physician$ service$ OR (demand creation AND health care) OR inducement hypothesis OR physician density OR competition in physician$ service$ market$ OR practitioner$ behavior$ OR physician$ behavior$ OR physician$ pricing OR physician$ utilization OR information$ (asymmetr$ AND health care) OR inefficien$ in physician$ practice$ OR (financial$ incentive$ AND health care). We performed the search and the review during the first half of 2007 but used the OVID alert system to identify relevant papers subsequently published. Bibliographies of

Table 1 List of the criteria used for the critical appraisal. 1

Research question

Well explained

2

Study design

Appropriate to address the research question Cross-sectional or longitudinal Representativeness of the sample

3

Data quality

Source of data mentioned Quality check reported

4

Analysis

Methods clearly explained (management of outliers; modelling process) Appropriate statistics: cluster or multi-level accounted for; confidence intervals reported Validity of models: normality, heteroscedasticity and collinearity tested in case of regression modelling

5

Discussion

Internal validity External validity Conclusions supported by findings

retrieved papers were scrutinized for relevant references. Grey literature was also searched using the same keywords in Google and Google Scholar. The inclusion criteria were: 1. Studies addressing medical care utilization (number of medical services per physician and/or per patient) in relation to physician density (exposure variable). Studies focused on dentists, psychiatrists and physiotherapists were not considered. 2. Original studies (i.e. no opinion or methodological papers) based on individual empirical data, with analysis adjusted for the effect of at least one of the most common confounders: patient age, sex, socio-economic and health status and/or physician age and sex. Results adjustment for confounding factors is important to appraise actual size of associations in observational studies [10,16,21–25]. Studies based on data aggregated geographically (region, nation) were not included because such results adjustment on patient characteristics is difficult and because such studies are prone to ecological fallacy. An extensive illustration of problems induced by analyses of aggregated data was provided by Sorensen and Grytten [26]. 3. Published in English, French or Dutch between 1980 and today. The exclusion criteria were: 1. Studies focused on patient satisfaction. 2. Studies focused on the effect of physician density on medical fees. 3. Studies focused on the impact of medical fee scheme (copayments) on healthcare consumption. Two reviewers (CL and DR) independently screened all titles and abstracts, assessed fulfilment of inclusion criteria, and appraised study quality of included studies (Table 1). On the basis of those quality criteria, a global un-weighted score was issued for each paper (high (H), medium (M) or low (L) quality). At each step, disagreements between the two raters were solved by a consensus discussion involving the third author (SS).

C. Léonard et al. / Health Policy 91 (2009) 121–134

123

Fig. 1. Flow chart for studies selection.

Because of the important variations in settings, variables, and statistics across studies included, and the likely subsequent heterogeneity of results, a meta-analysis was not conducted. 3. Results Our search yielded 143 peer-reviewed references, and two working papers. On the basis of abstract review, 96 papers did not meet the first inclusion criterion and were excluded from further review. Complete articles for four additional references were unavailable, but two of the four abstracts included the main results. Cheng and Liu evaluated the impact of a newly opened hospital on local patient flow and found no evidence concerning physician-induced demand [27]. Tsai and Kung, in a longitudinal study conducted between 1996 and 1999, showed that the impact of increasing physician density on Chinese medical care utilization was very slight (+1.69%) [28]. The two other documents were PhD theses [29,30]. Six additional references were identified after the first literature search in February 2007. After this initial step, the 51 references appearing to meet the study eligibility criteria were reviewed thoroughly. Twenty-six of these papers were ultimately

deemed ineligible resulting in a final sample of 25 documents for review (22 peer-reviewed papers, 2 working papers [31,32] and 1 peer-reviewed scientific report [33] (Fig. 1). Studies varied by a number of characteristics (Tables 2–5). First, study designs differed. Whereas most studies used billing data to measure healthcare consumption, a specific survey to assess who, from the patient or the physician, initiated the next medical consultation was organized in four countries: Australia [34], Ireland [35,36], Norway [37] and USA [38–40]. In all these studies, patients were interviewed except in Australia, where doctors themselves were surveyed [34]. All studies, except for one [41], were cross-sectional, making it difficult to assess the time-sequence between exposure (physician density) and outcome (health care consumption). Finally, significant variation in medical specialties and geographical entities under consideration was also noted. Second, multiple measures were used to assess healthcare utilization: physician volume, i.e. the number of health services per physician per year (16 studies); individual intensity of care, i.e. the annual number of visits or medical procedures per patient per year, or the probability of getting a follow-up visit (13 studies); population intensity of care, i.e. the proportion of

124

Table 2 Physician volume and medical density. Year*

Country

Physician population

n

Analysis

Dependent variable

Independent variables

Results

Statistical test

[33]

2005

Belgium

Practising GPs 8 groups of specialists

11 626 10 922

OLS

Log total visits per GP

Log GP density per municipality Log GP density per arrondissement Log GP density per municipality

−0.80OLS −1.01OLS −0.80OLS

p < 0.0001 p < 0.05 p < 0.0001

Log GP density per arrondissement Log SP density per arrondissement

−1.54OLS

p < 0.0001

Dermatologists (n = 593) Gynaecologists (n = 1 118) Eye–Nose–Throat (n = 498) Internal medicine and neurology (n = 3 102) Ophthalmologists (n = 836) Paediatricians (n = 975) Psychiatry and neuropsychiatry (n = 1 486) All surgeons (n = 2 314)

−0.19OLS −0.17OLS −0.26OLS −0.15OLS

p < 0.05 p > 0.05 p < 0.05 p < 0.05

−0.24OLS −0.35OLS −0.16OLS

p < 0.0001 p < 0.05 p < 0.001

−0.12OLS

p > 0.05

GP density per zip-codes (zip-codes correspond to smaller geographical areas than ‘communes’) GP density/district

−0.305

p < 0.05

−0.1866OLS

p < 0.05

GP density/county

−0.6839OLS

p > 0.05

GP density per department

E

−0.201

p < 0.05

SP density per department GP density per department SP density per department

−0.125E −0.063E +0.114E

p < 0.05 p < 0.05 p < 0.05

GP density per Local Government Area (LGA) Diagnostic test ordered GP density per LGA Diagnostic test ordered GP density per LGA Diagnostic test ordered GP density per LGA Diagnostic test ordered

1.01b

p > 0.05

20.02 0.99 25.08 0.99 6.04 1.01 7.93

p < 0.01 p > 0.05 p < 0.01 p > 0.05 p < 0.01 p < 0.05 p < 0.01

Physician density

3.11LOG

p < 0.05

Log total consultations per GP Log total consultations per SP

[31]

2001

Belgium

Certified GPs

12 133

log–log OLS

[50]

2000

France

GPs

4 660

OLS

[41]

1979–1993

France

GPs and SPs of sector 1 (retrospective study)

4 500

GMM

Number of contacts (home and office visits) per GP per year Annual number of procedures per GP Average number of encounters per physician per day Quantity of care delivered

[34]a

1990

Australia

1991

GPs

495

Multi-level logit model

Selected at random

Follow-up visits for: Acute bronchitis (n = 1 328) Tonsilitis (n = 964) Sprain (n = 837) Otitis media (n = 790)

[35]c

1981

Ireland

GPs

2 874

Logistic, OLS and WLS

E (OLS)

Most recent visit is a GP’s idea Return visit

Physician density

LOG

13.29

p < 0.01

C. Léonard et al. / Health Policy 91 (2009) 121–134

Ref.

[48]

1998

Norway FFS (contract physicians) and salary systems

Contract primary care physicians

Salaried primary care physicians in 435 municipalities

[26]

[46]

[55]

1995

1995

1995

Norway

Norway

Norway

Contract primary care physicians

Contract primary care physicians

OLS and logistic

564

1 415

1 366

1 617

Proportion of consultations lasting more than 20 minutes per physician

OLS and TSLS (log–log)

OLS and TSLS

OLS and IV (log–log)

[40]f

1977–1978

USA FFS

Physicians national survey in the USA – 14 000 randomly selected households

Logit

[38]

1977

USA FFS

Physicians national survey in the USA – 40 000 individuals

[47]

1986

USA

Physicians (GP + 11 specialties)

1 700

Mean number of laboratory tests per consultation

Number of consultations per contract physician

Number of consultations per physician

Number of consultations per physician

1986

USA

Surgeons

not reported

Contract physicians Salaried physicians Physician density

0.14E (OLS) −0.08E (OLS)

p > 0.05 p > 0.05

Contract physicians Salaried physicians

0.20E (LOG) d 0.47E (LOG)

p > 0.05 p > 0.05

−0.25E (OLS)

p < 0.05

−1.00E (TSLS)

p < 0.05

−10.89OLS

p < 0.05

−10.78TSLS

p < 0.05

0.36E (OLS)

p < 0.05

0.42E (OLS) e 0.73E (IV)

p < 0.05 p < 0.05

Physician density

Physician density

Inverse of physician density

Likelihood of ambulatory physician-initiated visit

Physician density

0.12E

p < 0.05

Logit

Likelihood of surgery

Physician density

−0.004E

p > 0.05

WLS

Number of ambulatory visits

Patient initiated Doctor initiated

Physician density Physician density

−0.00 0.001

p > 0.01 p < 0.01

Number of visits for private patients

Competing physician density

−0.00006ˇ

p < 0.01

Non-competing physician density Competing physician density

0.000007ˇ −0.0006ˇ

p > 0.05 p < 0.01

Non-competing physician density

0.000083ˇ

p > 0.05

Ophthalmologists density

0.52E

p < 0.01

General surgeons density Orthopaedic surgeons density Urologists density

0.18E

p > 0.05 p < 0.01 p > 0.05

IT3SLS

Number of visits for Medicaid patients

[43]

Physician density

TSLS

First occurrences (initial contacts with surgeons)

0.30E 0.17E

125

GMM: generalized method of moments, OLS: ordinary least squares, GLS: generalized least squares, TSLS: two-stage least squares, IT3SLS: iterated three stages least squares, SE: standard error, SD: standard deviation, E: elasticity. * Year of data collection. a We find a very similar version in Scott and Shiell [57]. b Odds ratio were derived from regression coefficients. c The three papers about Irish GPs were based on the same source but proposed different analyses of SID. GPs density is always used as independent variable. However, other independent variables were included in each paper (respectively the ‘perceived health’, the ‘type of insurance’ and the ‘level of education’). d In the logistic regression, when the independent variable is expressed in logarithmic form, the regression coefficient can be interpreted as elasticity. It measures the change in odds ratio produced by a 1% increase in the independent variable. e This result refers to regression excluding non-significant control variables. f The three following papers were based on the same source but proposed different analyses of SID. Physicians’ density is always used as independent variable. However, other independent variables were included in each paper (respectively the ‘perceived health’, the ‘type of insurance’ and the ‘level of education’) and alternative dependent variables were used.

C. Léonard et al. / Health Policy 91 (2009) 121–134

Contract primary care physicians

1 818

126

Table 3 Individual intensity of care and medical density. Ref.

Year*

Country

Physician population

n

Analysis

Dependent variable

Independent variables

Results

Statistical test

[33]

2005

Belgium

Practising GPs

11 626

OLS

Average number of visits per patient

GP density per municipality GP density per arrondissement GP density per municipality

0.02OLS

p < 0.0001

Average number of consultations per patient

2002

Germany

Physicians (GPs and SPs together) (434 districts)

22 417

OLS 2 SLS Logistic

[50]

2000

France

GPs

4 660

OLS

[40]b

1977–1978

USA FFS

[41]

1979–1993

France

[37]

[45]

1985

1988–1989

Norway

Florida

Physicians national survey in the USA – 40 000 persons GPs and SPs of sector 1 (retrospective study)

Patients (personal interviews during 14 days in 1985)

Physical therapists in 203 facilities

Weighted Least Squares (WLS) 4 500

10 576

1 967

GMM

Logit

Regression coefficients

Number of physician visits per patient during a 3-month period Annual number of procedures per patient per GP Number of physician-initiated visits per patient Intensity of care (I)

Physician-initiated visits per patient

Physical therapy visits per patient

Minutes of physical therapist labour per patient visit

0.04

p > 0.05

−0.04OLS

p < 0.0001

GP density per arrondissement Log physician density (per district)a GP density/district

−0.11OLS

p < 0.0001

E (TSLS)

0.17

p < 0.01

0.0005OLS

p > 0.05

GP density/county

0.008OLS

p < 0.05

Physician density

0.112E

p < 0.05

GP density per department SP density per department R1 (1 818 pop/phy)

0.09E

–

−0.00355

p > 0.05

R2 (2 583 population/physician) R3 (3 796 population/physician) R4 (8 853 population/physician) R5 (40 366 population/physician) Orthopaedic surgeon density Orthopaedic surgeon density when referrals by owners Orthopaedic surgeon density

0.0015

p > 0.05

0.00031

p > 0.05

0.000096

p > 0.05

−0.000006

p > 0.05

Orthopaedic surgeon density when referrals by owners

E

0.24

␤

−0.2693

p > 0.05

126.82␤

p < 0.05

2.37␤

p > 0.05

−124.17␤

p > 0.05

C. Léonard et al. / Health Policy 91 (2009) 121–134

[32]

OLS

[43]

1975–1980

USA

Michigan

[51]

1983

Switzerland

[35]

1981

[36]

1980

Ireland (eight regional Health Board areas) Ireland

Surgeons

Surgeons, family doctors, internal medicine, paediatricians

Not reported

Not reported

TSLS

OLS (log–log)

Intensity of care

Intensity of care

Ophthalmologists density General surgeons density Orthopaedic surgeons density Urologists density

0.04E

p > 0.05

−0.12

p > 0.05

−0.19

p > 0.05

−0.04

p > 0.05 E

General surgeons density

+0.054

−0.016

p > 0.05

0.319

p < 0.001

−0.057

p < 0.01

0.258OLS

p < 0.01

0.039OLS

p < 0.01

0.08Probability 0.91Probability

p < 0.01 p < 0.01

Office-based physicians (28 districts) GPs

918

OLS

Visits per patient

Family doctors density Internal medicine doctors density Paediatric medicine doctors density Physician density

2 874

Return visits

Physician density

GPs

1 003

Logistic, OLS and WLS Logistic regression

Self referral (return visits)

GP density Min = 0.429 Max = 0.555

p > 0.05

C. Léonard et al. / Health Policy 91 (2009) 121–134

[52]

1986

*

Year of data collection. Results were reported for the full model (no distinction between statutory and private patients). The elasticity of the number of contacts with respect to the physician density was equal to 0.17 in the full model but equalled to 0.16 for statutory patients and 0.27 for private patients. b The three following papers are based on the same source but propose a different analysis of the SID. Each of them use the physician density as independent variable but complementary independent variables are specific in each of them (respectively the ‘perceived health’, the ‘type of insurance’ and the ‘level of education’) and alternative dependent variables are used. a

127

p < 0.05

p < 0.05

13.01

−7.2

Surgeon density

GP density

Rate of elective surgery per 1 000 TSLS 2 837

*

Year of data collection.

Surgeons 1969–1976 [44]

USA

0.56 0.06 0.11 0.13 Ophthalmology density General surgery density Orthopaedic surgery density Urology density

TSLS Not reported Surgeons 1986 [43]

USA 1970–1992 [53]

USA

Logistic 255 593 births

OLS (log–log) Not reported

Surgeons, family doctors, internal medicine, paediatricians Obstetricians/Gynaecolgists 1975–1980 [52]

Michigan

p < 0.01 p > 0.05 p > 0.05 p > 0.05

p < 0.01 0.23E

0.342E

Obstetrician/gynaecologists density per state Surgeon density four specialties

p < 0.01 0.269E Physician density

p < 0.01 GPs per 10 000 persons per statistical sub-division

Statistical test Dependent variable

Number of GP services per capita per year Per capita (of population) utilization of care Caesarean deliveries Surgeon’s service per Medicare enrolee in 1986 log–log OLS and TSLS

Analysis n

Not reported GPs

Physicians Population

1996 [54]

Country Year* Ref.

Table 4 Population intensity of care and medical density.

Australia

Independent variables

0.468E

C. Léonard et al. / Health Policy 91 (2009) 121–134

Results

128

the population undergoing specific medical procedures (5 studies). Third, the relationship between physician density and the dependent variable was expressed by various statistics, i.e. mainly elasticities (19 studies) or regression coefficients (13 studies). In four studies, results were expressed as probabilities or odds ratio of a follow-up visit [34–37]. For the analysis of physician volume, a supplier-induced demand can be inferred when the elasticity or the odds ratio is significantly greater than −1.00 or +1.00 respectively. Regression coefficients present two difficulties: their size depends on the units used to measure the corresponding independent variable, and thus they are difficult to compare; when the regression coefficient is significantly smaller than 0.00, it is difficult to infer an absence of SID, because a diminution of annual services per physician does not necessarily mean that the overall number of services remains constant, given the increase in physician density. For the analysis of care intensity, at individual or population level, a SID effect can be inferred in case of an elasticity significantly greater than 0.00, a regression coefficient significantly greater than 0.00 or an odds ration significantly greater than 1.00, i.e. when there is a positive relationship between physician density and care intensity. Fourth, there was a lot of variation in modelling strategies, i.e. adjustment for confounding factors was very heterogeneous. Some studies accounted for the effect of patient characteristics such as age and sex [38,42,32,43,34,35,40], health status [32,37–40,44], insurance status [35,36,38–40,45], socio-economic status [26,32,36–38,40,43–45] and individual behaviors such as smoking and alcohol consumption [37]. Other studies accounted for the effect of physician characteristics such as age and sex [34,46,47], professional experience [39,47–50], type of practice, i.e. solo versus group practice [47,48], and in one case, perceived shortage of patients [42]. Some authors also inserted contextual variables such as density of hospital beds [26,37,39,44,47,49,51], unemployment rate [26,37,49,50], or characterization of the health, education and urbanization of the region [52]. However, only one study combined patient, physician and contextual characteristics [39]. Results regarding overall medical services per physician per year are presented in Table 2. Despite the heterogeneity discussed above, a positive association between physician density and healthcare consumption was found in the majority of the 13 studies. Most of the studies had some methodological flaws and were classified as having intermediate methodological quality. Most studies focused on general practitioners (GP), but medical specialists were also analyzed in five studies. All of the studies, except for one, reported evidence of an increased health care consumption in relation to physician density, although the size of the association was highly variable. Elasticities were all greater than −1.00 and some were even positive up to +0.14 and +0.52 [40,43], while regression coefficients were either close to 0 (no decrease of physician volume when the physician density increased) or positive (variation of independent and dependent variables in the same direction). Elasticity was −1.00 for the number of consultations per contract physician in Norway [26]. Three additional studies

C. Léonard et al. / Health Policy 91 (2009) 121–134

reported an absence of association, but only for some of the outcomes examined. In Belgium, elasticities were −1.00 for consultations and visits per GP at the arrondissement level, while at the municipality level elasticities were greater than −1.00 although remaining smaller than 0.00 [33]. Regression coefficients were also negative in Norway (coefficient = −10.9, p < 0.05) [46], and in France at the district (canton), not at the county (department), level (coefficient = −0.19, p < 0.05) [50]. With regards to the intensity of care at the individual level (Table 3), all but two studies [43,37] reported a significant increase in relation to physician density. For instance, in Ireland, the probability of patients getting a return visit was 0.91 in the area characterized by the highest physician density, whereas it was 0.08 in the area with the lowest density [36]. In some studies, the association was significant only for some of the outcomes, as in France (increase at the county, not district, level) [50] or in Michigan (increase for internal medicine doctors but not for family doctors or general surgeon) [52]. The study by Mitchell is of particular interest because the relation between physician density and the utilization of ancillary services such laboratory tests, diagnostic imaging and physical therapy was studied in a specific situation of a “joint venture”, i.e. physicians owned ancillary health care facilities to which they could refer patients. Results showed a positive relationship only for owners of such facilities, and not within independent clinics inferring that physician owners tended to overprescribe ancillary services [45]. All studies specifically addressing the question of physician-initiated visits through a questionnaire survey reported a significant positive association with physician density [34–36,38–40], although there was significant variation in results. For instance, Scott and Shiell investigated the follow-up visits initiated by the physician for four simple health conditions (i.e. acute bronchitis, tonsillitis, sprain and otitis media) [34]. They reported a significant association only with physician density for one of these conditions, i.e. otitis media. The five studies looking at the intensity of care at the population level (Table 4) consistently showed an increase in relation to physician density [44,52–54], although statistical significance was not reached for all medical specialties under study by Escarce [43]. Although present in a majority of studies, the magnitude of the relationship between healthcare consumption and physician density varied greatly between settings, even within a single country. For instance, in Norway, three studies reported evidence of no or moderate association between physician density and consumption of care [26,46,55], while a fourth showed an increased utilization of laboratory tests and lengthy consultations [48]. Although the authors of the latter study explained their results by the higher reimbursement rate for lengthy consultations, it remains unclear why this relationship was not observed for other indicators. Even within a single study, results differed substantially across various strata of analysis. A first important interacting factor seemed to be the geographical entity. In the study conducted by Béjean et al. [50], results supported the association between the number of procedures per

129

patient and the density of GPs at the county level, but not at the district level. The authors hypothesized that GPs could have a differential perception of professional competition by geographical entity considered. Such variations by geographical entity were also found in Belgium [33]. Analyzing physician density and health care activity at the smallest geographical level assumes that all patients of a GP live in the same location as their GP. Belgian data showed that cross-bordering was however non-negligible: 15% of all GPs had 30% of their patients living in another municipality. Consequently, municipality level likely misestimated the market of a GP and seemed inadequate to test the association between physician density and GP behavior. Another interacting factor was the medical specialty. Delattre and Dormont [41] illustrated that the size of the association between physician density and physician volume or care intensity was higher for specialists than for GPs. Similar findings were found by Stano [52] and Roberfroid et al. [33]. Finally, parameters other than physician density played a significant role on healthcare consumption. For instance, elasticities were also significant for patient age, patient sex, patient health status, patient insurance coverage [40], group practice [48], population age [26,48,52], population income [52], percentage of population living in urban areas [52], population education [52]. In a number of studies, those elasticities were of greater size than the one related to physician density. This was the case for patient age [31], physician age [40], patient health status [32], physician sex [26,46], patient sex [31,32], patient unemployment status [31], patient income [40,52], importance of out-of-pocket payment [32,40]. Also in the Australian study, the odds ratio for a return visit increased when a diagnostic test had been ordered during the index visit, but it was unclear if the return visit was asked for clinical follow-up or for communicating the results back to the patient [34]. 4. Discussion Twenty-five studies were included in our systematic review. Despite a substantial heterogeneity in study design and analytical modelling, the results were surprisingly consistent, whatever the outcome and the analysis used (elasticity, regression coefficient, odds ratio). Thus, the existence of SID appears quite straightforward, although estimates varied according to a number of method parameters such as the type of dependent variable (e.g. physician volume or care intensity), the geographical entity (e.g. municipality or department) or the medical specialty under consideration. Now, the crucial question for policy-makers is: to what extent can supplier inducement be interpreted as a deviant agency, and as such should be prevented? The framework proposed by Labelle et al. [10] illustrates the diversity of configurations that can be met within the broad concept of SID (see Fig. 2). An inducement beneficial to the patients can be met in two situations (cells I and II). First, the increase of healthcare consumption can be intuitively considered a ‘normal’ effect of increased care availability in regions with a previous ‘low’ physician density and unmet health needs. This is the so-called ‘availability effect’. In this case, inducement

130

C. Léonard et al. / Health Policy 91 (2009) 121–134

Table 5 Critical appraisal of the systematic review about SID. Ref.

Rate

Appraisal

[33]

H

Cross-sectional analysis, one year, all the physicians and the patients of the country, ambulatory care, two geographical levels considered (arrondissement and municipality) Different specifications of the model (linear, semi-log, log–log), clustering analysis using the GEEM, control of border-crossing effect, control of availability hypothesis, few covariates for analysis of specialists R2 : 0.22–0.28 for GPs and 0.08–0.27 for SPs (function of the specialty) Heteroscedasticity tested with White test Results concern the whole country

[50]

L

Relevancy and innovative aspect of the question unclear (3 of the 4 references cited are not peer-reviewed) Routine data from health insurance groups for year 2000 in two regions, Quality check not reported Modifying effect (interaction) among explanatory variables not tested, dummy variables issued from the cluster analysis are inserted in the regression model. But cluster variables include dependent variables. This plausibly overestimates the explanatory power of the linear modelling, but also distorts the coefficient of other explanatory variables, multi-leveling of data (GP, district, county) not accounted for R2 : 0.76–0.87. But the goodness of fit likely to be artificially boosted by inclusion of variable “practice patterns” in the regression, as reported by authors. Endogeneity of GP density not assessed Difficult to generalize the results from two regions to the whole country All conclusions should be considered with caution given the important methodological flaws

[32]

M

Cross-sectional study, no discrimination between specialists and GPs, only a 3 month-period is considered. Can capture only frequency based on visits at close intervals in a specific period. No information given on the period Survey data on individual health care utilization (SOEP 2002) + indicators at the regional level (INKAR database), no quality check reported Pseudo R2 : 0.0488 (Full Model) Endogeneity of the physician density considered, potential bias of self-reporting visit to physician by patients is not addressed, results not adjusted for the prevalence of chronic diseases, a powerful factor of repeated visits Numerous limitations reflected in the poor R2 Does not allow to disentangle demand inducement by GPs and SPs

[31]

M

Cross-sectional study, no individual information about the GP but proxy of the health and socio-economic status of the patients Routine data from the state health fund, restricted to active GPs, validity check not reported Interactions not tested, multi-levels not accounted for, endogeneity assessed R2 : 0.22 Author provides a coherent intuitive interpretation of coefficients in low density areas (availability effect) and in high density areas (SID). Nevertheless, few coefficients are really significant (i.e. p < 0.05)

[54]

M

Cross-sectional data, data aggregated on SSD (statistical sub-division) level Billing statistics from Health Insurance Commission and Australian Census (socio-economic proxy for health status for 187 statistical sub-divisions), quality check not reported, number of GPs and patients not reported Multi-leveling of data (statistical local areas, statistical divisions) not accounted for, no adjustment for health needs R2 = 0.83 No individual patient data, absence of border-crossing reported A bias due to the partial aggregation can not be excluded

[42]

M

Use of a trial making the difference between two different kinds of remuneration systems Data from a capitation trial conducted between 1993 and 1996, no quality check reported, the population of the four Municipalities is very different (8 200–114 000), the population density is not reported R2 : 0.29 Endogeneity of GP density not assessed, border-crossing not reported, Collinearity, heteroscedasticity not assessed Results and statistical tests leads to evidence of SID (reduction of services with the physician density and increase of services per patient if GPs are confronted with a patient shortage)

[41]

H

Use of a panel data. Appropriate design to highlight behavior changes in the time (5 years of observations) Patient data: Caisse Nationale d’Assurance Maladie des Travailleurs Salariés Assessing of endogeneity of density, availability effect and non observable heterogeneity Results of robustness tests reported Use of a representative sample of physicians in a long period, result can be extrapolated to the entire population of physician of the sector 1 The increase of physician density leads to SID. Authors do not discuss the potential influence of other parameters not included in the analysis

[34]

H

495 randomly selected GPs, stratified by state, disaggregated data, quality check not reported Endogeneity not addressed because use of individual data SID demonstrated only in case of otitis media. Reasons for variations across health conditions not discussed

[53]

M

Time-series study Data from the National Hospital Survey, dataset of individual birth records from 1970 to 1982, quality control of the data not reported Endogeneity of physician density not assessed, no specific test for collinearity or heteroscedasticity The magnitude of the SID attitude is small but the test about the symmetric character of the use of a new technology shows that the response to fertility is equally strong in the period when fertility is rising and falling. This suggests that the physicians are able to adapt their practices in response to financial incentives

C. Léonard et al. / Health Policy 91 (2009) 121–134

131

Table 5 (Continued ) [45]

M

Cross-sectional analysis Data from two surveys of physical therapy and rehabilitation facilities (details about the survey not reported) and of physical therapists (voluntary response via a newspaper), no quality check reported but clear explanation of data exclusion Adjusted R2 = 0.413 (for the number of visits) and 0.196 (for length of visits) Endogeneity of GP density not assessed, no control of the case-mix of the patients reported Data obtained by survey and only for one State. Difficult to assess the possibility to generalize the results

[43]

M

Cross-sectional analysis among elderly Medicare enrolees Data of Health Care Financing Administration, Medicare and Physician’s Current Procedural terminology Codes, no quality control reported Aggregation on Metropolitan area level and services level, adjustment for patients’ characteristics, use of instrumental variables (TSLS) to address the problem of endogeneity of the density. Significance level set at 0.10 R2 = 0.44 four specialties (from 0.24 to 0.29 per specialty) and R2 between 0.22 and 0.49 for first occurrence The results can not be generalized to other specialities given the substantial difference between ophthalmology and the other specialties The influence of the surgeon density on the first contact between the patient and the physician seems to support the availability hypothesis

[51]

M

Cross-section analysis of office–based physicians in one canton (Bern, Switzerland) Official sources of data (Statistical offices of Switzerland and the Canton Bern) and 1980 census (population figures) and Survey of the hospital association (hospital beds), no quality control reported Use of physicians characteristics on individual level, adjustment for the age of the patients, use of TSLS specification to assess the endogeneity of the density Results support SID but the aggregation on district level and the quasi absence of control variable for the patients are sources of bias Not possible to generalize the results to the whole country

[36]

M

Cross-sectional survey Source: Nationwide sample survey of households in Ireland. One observation randomly taken in each household of the sample, no quality control specified Age and gender of the patients used as proxies for health status R2 adjusted for degree of freedom = 0.136 (Full Model) and 0.123 (Model including only significant variables) Endogeneity of GP density not assessed or discussed Results support the SID hypothesis, more return visits in high density area

[44]

M/H

Cross-sectional study with control for patient characteristics, estimation of the modelling for three sub-samples, estimation (OLS) making the difference between areas with shortage and surplus of surgeons Source: Health Interview Surveys, sample of all US households in 349 Primary Sampling Units, no quality control reported TSLS modelling R2 between 0.05 and 0.09 Endogeneity of physician density assessed, control for border-crossing effect Surgeon density plays a significant role on the elective surgery rate and the surgical fees

[39]

M

Cross-sectional study Source: National Medical Care Expenditure Survey, no quality control reported Adjustment for health status of the patient, data clustering not accounted R2 = 0.13 Endogeneity of insurance status addressed, no control for border-crossing effect Other explanatory variables play a substantial role in the determination of the physician’s initiated expenditures

[40]

M

Cross-sectional data 14 000 randomly selected households during 18 months, due to interview techniques Patients characteristics, insurance status and physician characteristics to control the estimations Endogeneity of physician density not assessed but individual data, no control for border-crossing effect Values of elasticities can be considered as a support for the SID hypothesis

[38]

M

Interview of 40 000 individuals representative of the US population, possible bias due to interview techniques R2 = 0.07 for patient-initiated model and 0.12 for physician-initiated model Endogeneity of physician density not assessed but individual data, no control for border-crossing effect Weak evidence for SID, other independent variables play a more important explanatory role (health status)

[35]

M

Cross-sectional study based individual data collected by interview Nationwide interview of 1 069 households, details about the interview reported, validity of the answers reported Health and age as proxies of the health status R2 = 0.103–0.118 depending on the dependent variable and the modelling used Endogeneity of GP density not discussed, author considers as a potential weakness of the study, the lack of control for medical factors influencing both physician density and the return visits, such as regional prevalence of chronic conditions Results support the SID hypothesis

[48]

L

Cross-sectional design, sample selection not reported Results are poorly reported, independent variable representing the physician density not clearly defined Four variables to measure inducement, physician characteristics and patient health status as control variables R2 = 0.02 for model explaining the number of laboratory tests (very low) Endogeneity of GP density assessed The magnitude and sign of elasticities should be interpreted with caution, coefficients for contract and salaried physicians are different

132

C. Léonard et al. / Health Policy 91 (2009) 121–134

Table 5 (Continued ) [46]

M

Cross-sectional data Data from national Insurance Administration and from fax forms, individual’s data for physicians and area’s data for patients, no explicit quality control reported R2 : 0.15 for consultations as dependent variable and = 0.34 for revenue as dependent variable Endogeneity for non practice incomes tested but not for GP density Nationwide sample. The results, if robust, should be generalized to the whole country The interpretation of the magnitude of the regression coefficient is not correct. They are of small magnitude but have to be multiplied by the value of the variable to assess the real effect of the independent variable. In the sensitivity analysis, coefficients are significant with TSLS and not with OLS and R2 are not reported. The coefficients of the ‘Non practice’ income are considered as evidence against SID. Nevertheless, these coefficients are negative and significant, this means that the number of consultations decreases if the non-practice income increases. This relation can be considered as a proof of the ability of the physician to modify the demand of care. SID can not be excluded. On the other hand, the coefficient for the physician density is very strong (−10.9), multiplied by the average number of consultations, this leads to a dramatic reduction in the number of consultations in areas where the concurrence is ‘high’

[26]

M

Cross-sectional data of volume of care collected per individual physician, and density per administrative area Data from national Insurance Administration and from fax forms, individual’s data for physicians and area’s data for patients, no explicit quality control reported Patient (municipality characteristics) and physician (individual) characteristics as control variables R2 : 0.17 for consultations as dependent variable, 0.32 for revenue as dependent variable and 0.15 for net income as dependent variable Endogeneity of GP density assessed Nationwide sample. Results, if robust, should be generalized to the whole country Absence of evidence of SID congruent with results

[55]

M

Cross-sectional data (volume of care per individual physicians), and density per administrative area Data from national Insurance Administration and from fax forms, individual’s data for physicians and area’s data for patients, no explicit quality control reported R2 = 0.167–0.174 Endogeneity of GP density assessed Nationwide sample. The results, if robust, should be generalized to the whole country Results show that the number of consultations and the revenue from laboratory tests increase when physician density increases. These results are not a proof of absence of SID but the physicians’ ability to adapt their practices according to exogenous variations

[47]

M

Cross-sectional data (volume of care per individual physicians) and density per speciality Physicians Practice Cost and Income Survey–personal physician characteristics, no quality control reported R2 = not reported, endogeneity of physician density not assessed (individual data) but reported Conclusion supported by the results

[37]

L

The part of the study concerning the laboratory tests is not taken into account because data were aggregated Central Bureau of Statistics, no quality control reported Gamma: 0.65 Results can be generalized to the whole country No indication for inducement effect (OR between 0.996 and 1.001). Few results are significant

[52]

M

Cross-sectional data concerning volume of care given by GP’s and surgeons on the level of individual patient, use of aggregated variable (utilization per capita) and individual variable (intensity of care per patient) Data from 15 markets in Michigan and for 424 procedures, cross-sectional micro-data, no quality control reported R2 = 0.952 for dependent variable “Quantity per capita” (only 0.099 and 0.102 for intensity of care given by GP’s and surgeons respectively) Border-crossing assessed, endogeneity not assessed and not reported (individual data) The elasticity of the utilization per capita supports SID. Author argues that the absence of modification in intensity of care corresponds to an ‘availability effect’. However, the values of elasticities of intensity of care can also be considered as an increase of return visits, initiated by the physician

results from improved accessibility to health services. Correlation coefficients in areas of low physician density in studies conducted in Norway [55] and in Belgium [33,49] can be interpreted as such. However, in such studies areas

Fig. 2. Conceptual framework for SID [10].

of low physician density were defined relatively to areas of higher density, i.e. the concept of unmet health needs was also relative and not defined by clear-cut criteria. The availability hypothesis could only be ascertained through an analysis of unmet health needs in relation to physician density. This type of study design was not present in our review. Second, the physician could induce a care activity that he considered beneficial to his patients, i.e. perform an altruistic inducement. Grytten and Sorensen reported that the health status of the patient was a major predictor of a return visit, i.e. that the care intensity was correlated to the patient needs, not to the physician density [37]. This could also explain the results of the four studies demonstrating an association between follow-up visits initiated by the physician and the physician density [34,35,39,40]. Scott demonstrated that the prescription of a diagnosis test during the index visit was also a strong determinant of the

C. Léonard et al. / Health Policy 91 (2009) 121–134

follow-up visit, whatever the physician density [34]. Thus follow-up visits might be considered as ‘good practice’ by physicians, but affordable only if the workload is manageable, which is more likely in areas with higher physician density. More objective data concerning the patient health status and the clinical guidelines is necessary to evaluate the legitimacy of such a follow-up visit. Sorting out the underlying motivations would require studies at the microlevel with a qualitative component. There is also good evidence that SID can result from a will to maximize income (cells IVa and IVb of Labelle’s frame), i.e. physicians inducing health services that the patient would have refused if he had been provided with the correct information. In the USA, Mitchell and Sass [45] showed that the inducement attitude could result from pure monetary interests. They found that a physical therapy clinic that received all of its referrals from physician owning the clinic treated a patient with approximately 50% more visits than a clinic receiving no referral from owners. Authors also reported that this difference was directly related to the increase in physician density, the decrease in population density and the increase in the total number of owners. Iversen [42] measured the physician’s perception of ‘patient shortage’ and related it to activity levels. Patient shortage was defined as the difference between the desired and the actual number of patients for a given physician. Authors observed that ‘rationed’ physicians had about 15% more income per listed patient than their un-rationed colleagues. In a paper focused on economic motives and professional norms, Iversen and Luras [56] found that GPs who experienced a shortage of patients had higher income, longer and more frequent consultations as well as more laboratory tests per listed patient than their unconstrained colleagues. The GPs who experienced a light patient shortage had 9% higher income per patient than their colleagues without shortage, while GPs who experienced a severe shortage of patients were expected to have 17% higher income per patient. This latter group of GPs prescribed 29% more laboratory tests per listed patient than the other group of unconstrained GPs [56]. In France, Delattre and Dormont referred to a similar concept, although not having collected direct data to ascertain ‘shortage’ [41]. It should be noted that a deviant agency is not necessarily income-related. Physicians can conduct procedures in order to protect themselves against litigation, when they are practising a defensive medicine [1]. Our review also yielded a number of methodological issues regarding SID. The first one concerns the means used to appraise its existence. All but one study used cross-sectional designs. In such design, it is impossible to ascertain a time-sequence, and reverse causality cannot be ruled out. Also, the goodness of fit of statistical models as measured by the R2 statistics was generally very low. The effect of physician density on healthcare use if other important parameters, such as patient health status, were included in the model remains unknown. Moreover, the analysis, in the vast majority of the studies, did not account for the multi-level character of the data. Applying ordinary regression technique to such data set is likely to underestimate the standard error. Finally, physician density was

133

often computed for one medical speciality at a time. However, Cromwell and Mitchell [44] showed that the density of GPs plays a reduction role on the rate of elective surgery. This could have resulted from a gate-keeping role of the GPs. To study SID, a global appraisal considering ‘physician supply’ as a whole should be carried out, rather than analyzing separately specialties without considering their interactions. The second technical issue relates to the magnitude of SID. In many studies, the coefficient was statistically significant but as sample sizes were commonly great, such statistical significance was evidenced already for modest differences. Moreover, as confidence intervals were often not reported, it was impossible to figure out the actual effect size. There were also a lot of variations in effect sizes across studies, plausibly due to the heterogeneity in assessment methods. So, if the existence of SID is ascertained, its exact importance remains unknown and is likely to vary in different health systems. In addition, none of the studies included a cost-efficiency analysis. 5. Conclusion An increase in the physician density tends to lead to physician-induced demand. However, the exact importance of the phenomenon and the underlying motivations are still poorly understood. It is necessary to employ more rigorous study designs and to collect more qualitative data to estimate the utilization inducement and to differentiate such an effect from an availability effect or an altruistic inducement. Surveys on the enhancement process of follow-up visits or about the patient’s pathology are interesting examples of data that should be collected in the future. The choice of the geographical entity of analysis should also be based on an analysis of patient flow. In the absence of more accurate information, measures encouraging physician supply limitation to contain costs should also be considered cautiously. Conflict of interest Authors declare no conflicts of interest. The study was commissioned by the Belgian government. Acknowledgements We would like to thank Eric Schokkaert, Professor at the Catholic University of Leuven (Belgium) for his stimulating comments on a preliminary version of this paper, and Dr Batya Elul, Assistant Professor of Clinical Epidemiology at the Mailman School of Public Health, Columbia University (USA), for her careful editing of our manuscript. References [1] McGuire TG. Physician agency. In: Culyer AJ, Newhouse JP, editors. Handbook of health economics, vol. 1A. North Holland: Elsevier; 2000. p. 461–536. [2] Shain M, Roemer MI. Hospital costs relate to the supply of beds. Modern Hospital 1959;92:71–3. [3] Roemer MI. Bed supply and hospital utilisation: a national experiment. Hospitals Journal of American Health Affairs 1961;35:988–93.

134

C. Léonard et al. / Health Policy 91 (2009) 121–134

[4] Kotlikoff L, Hagist C. Who’s going Broke? Comparing growth in healthcare costs in ten OECD Countries. NBER Working Papers. Massachusetts: National Bureau of Economic Research, Inc.; 2005. [5] Kotlikoff L, Hagist C. Health care spending: what the future will look like. Dallas: National Center for Policy Analysis; 2006. p. 15. [6] Cooper RA, Aiken LH. Human inputs: the health care workforce and medical markets. Journal of Health Politics, Policy & Law 2001;26:925–38. [7] Dussault G, Dubois CA. Human resources for health policies: a critical component in health policies. Human Resources for Health 2003;1:1. [8] Corens D. Health system review. Brussels, Belgium: European Observatory on Health Systems and Policies; 2007. [9] Schokkaert E, Van de Voorde C. Health care reform in Belgium. Health Economics 2005;14:S25–39. [10] Labelle R, Stoddart G, Rice T. A re-examination of the meaning and importance of supplier-induced demand. Journal of Health Economics 1994;13:347–68. [11] Rochaix L, Jacobzone S. L’hypothese de demande induite: Un bilan economique (The Induced Demand Hypothesis: An Economic Assessment. With English summary). Economie & Prevision 1997:25–36. [12] Labelle R, Stoddart G, Rice T. Response to Pauly on a re-examination of the meaning and importance of supplier-induced demand [comment]. Journal of Health Economics 1994;13:491–4 [discussion 5–6]. [13] Couderc N, Touze V, Ventelou B. Politiques de sante dans un modele macroeconomique: Un exercice de simulation. (The Macroeconomics of Health Care. With English summary.). Revue de L’Ofce 2001:209–44. [14] Pauly MV. Editorial: a re-examination of the meaning and importance of supplier-induced demand: reply. Journal of Health Economics 1994;13:495–6. [15] Lambert ML. Young general practitioners in an area with oversupply of doctors: the case of Brussels. Cahiers de sociologie et de démographie médicales 1998;38:271–96. [16] Dormont B. Microeconometrie de la sante: Remarques sur l’experience francaise (Health Econometrics Using Micro Data: Some Comments on the French Experience. With English summary.). Revue Economique 2000;51:583–94. [17] Bickerdyke I, Dolamore R, Monday I, Preston R. Supplier-induced demand for medical services. Productivity Commission, Staff Working Paper; 2002. [18] Evans RG. Supplier-induced demand: some empirical evidence and implications. Economic theory and the welfare state. vol. 3. Benefits in kind Barr, Nicholas, editor, Elgar reference collection international library of critical writings in economics. vol. 132. Cheltenham, UK and Northampton, MA: Elgar; distributed by American International Distribution Corporation, Williston, VT; 2001. p. 147–58. [19] Tabbush V, Swanson G. Changing paradigms in medical payment. Archives of Internal Medicine 1996;156:357–60. [20] Liu X, Mills A. Supplier-induced demand and unnecessary care. In: Preker AS, Xingzhu L, Velenyi EV, Baris E, editors. Public ends, private means strategic purchasing of health services. Washington: The World Bank; 2007. p. 279–310. [21] Trinquard S. Demande induite par l’offre ambulatoire:un survol de la littérature théorique et empirique. Cahiers du LASER: Université Montpellier 1; 2006. p. 34. [22] Muller C. Review of twenty years of research on medical care utilization. Health Services Research 1986;21:129–44. [23] Rice TH, Labelle RJ. Do physicians induce demand for medical services? Journal of Health Politics, Policy & Law 1989;14:587–600. [24] Rossiter LF, Wilensky GR. Health economist-induced demand for theories of physician-induced demand. Journal of Human Resources 1987;22:624–7. [25] Phelps CE. Induced demand—can we ever know its extent? Editorial. Journal of Health Economics 1986;5:355–65. [26] Sorensen RJ, Grytten J. Competition and supplier-induced demand in a health care system with fixed fees. Health Economics 1999;8:497–508. [27] Cheng S-H, Liu L-I. The impact of a new hospital on the operation of local clinics. Part II: Patient flow and physician-induced demand. Taiwan Journal of Public Health 2001;20:61–8. [28] Tsai W-C, Kung P-T. Relationship between Chinese medical utilization and growth of physicians. Taiwan Journal of Public Health 2001;20:463–74. [29] Bunn D. Physician choice: a fresh approach to testing the induced demand hypothesis. Washington State University; 1996. [30] Tareque M. Physician-induced demand and the effect of organized buyers on hospital behavior. Essays on health markets. Boston University; 1990.

[31] Schaumans C. Supplier inducement in the Belgian primary care market. Leuven: KUL; November, 2007. p. 37. [32] Jürges H. Health insurance status and physician-induced demand for medical services in Germany: new evidence from combined district and individual level data. Discussion paper of the Mannheim Research Institute for Economic of Aging; 2007. [33] Roberfroid D, Stordeur S, Camberlin C, Van de Voorde C, Vrijens F, Léonard C. Physician workforce supply in Belgium: current situation and challenges. In: (D/2006/10.273/) KrC, editor. Health Services Research (HSR). Brussels: Belgian Health Care Knowledge Centre (KCE); 2008. [34] Scott A, Shiell A. The use of multilevel analysis in health economics: an application to examining the effect of competition on general practitioners’ behaviour. Health, the medical profession, and regulation, Zweifel, Peter, editor, Developments in health economics and public policy. vol. 6. Boston, Dordrecht and London: Kluwer Academic; 1998. p. 159–68. [35] Tussing AD. Physician-induced demand for medical care: Irish general practitioners. Economic & Social Review 1983;14:225–47. [36] Tussing AD, Wojtowycz MA. Physician-induced demand by Irish GPs. Social Science & Medicine 1986;23:851–60. [37] Grytten J, Carlsen F, Sorensen R. Supplier inducement in a public health care system. Journal of Health Economics 1995;14:207–29. [38] Rossiter LF, Wilensky GR. A reexamination of the use of physician services: the role of physician-initiated demand. Inquiry 1983;20:162–72. [39] Rossiter LF, Wilensky GR. Identification of physician-induced demand. Journal of Human Resources 1984;19:231–44. [40] Wilensky GR, Rossiter LF. The relative importance of physicianinduced demand in the demand for medical care. Milbank Memorial Fund Quarterly: Health & Society 1983;61:252–77. [41] Delattre E, Dormont B. Fixed fees and physician-induced demand: a panel data study on French physicians. Health Economics 2003;12:741–54. [42] Iversen T. The effects of a patient shortage on general practitioners’ future income and list of patients. Journal of Health Economics 2004;23:673–94. [43] Escarce JJ. Explaining the association between surgeon supply and utilization. Inquiry 1992;29:403–15. [44] Cromwell J, Mitchell JB. Physician-induced demand for surgery. Journal of Health Economics 1986;5:293–313. [45] Mitchell JM, Sass TR. Physician ownership of ancillary services: indirect demand inducement or quality assurance? Journal of Health Economics 1995;14:263–89. [46] Grytten J, Carlsen F, Skau I. The income effect and supplier induced demand: evidence from primary physician services in Norway. Applied Economics 2001;33:1455–67. [47] Bradford WD, Martin RE. Supplier-induced demand and quality competition: an empirical investigation. Eastern Economic Journal 1995;21:491–503. [48] Grytten J, Sorensen R. Type of contract and supplier-induced demand for primary physicians in Norway. Journal of Health Economics 2001;20:379–93. [49] Schaumans C. Supplier inducement in the Belgian primary care market. Leuven: K.U. leuven; 2007. [50] Béjean S, Peyron C, Urbinelli R. Variations in activity and practice patterns: a French study for GPs. The European Journal of Health Economics 2007;1:1–12. [51] Kraft K, van der Schulenburg JMG. Co-insurance and supplier-induced demand in medical care: what do we have to expect as the physician’s response to increased out-of-pocket payments? Journal of Institutional & Theoretical Economics 1986;142:360–79. [52] Stano M. An analysis of the evidence on competition in the physician services markets. Journal of Health Economics 1985;4:197–211. [53] Gruber J, Owings M. Physician financial incentives and cesarean section delivery. Rand Journal of Economics 1996;27:99–123. [54] Peacock SJ, Richardson JRJ. Supplier-induced demand: re-examining identification and misspecification in cross-sectional analysis. European Journal of Health Economics 2007;8(3):267–77. [55] Carlsen F, Grytten J. More physicians: improved availability or induced demand? Health Economics 1998;7:495–508. [56] Iversen T, Luras H. Economic motives and professional norms: the case of general medical practice. Journal of Economic Behavior and Organization 2000;43:447–71. [57] Scott A, Shiell A. Analysing the effect of competition on general practitioners’ behaviour using a multilevel modelling framework. Health Economics 1997;6:577–88.