<%BANNER%>

Outpatient Imaging In Primary Care

Permanent Link: http://ufdc.ufl.edu/UFE0041862/00001

Material Information

Title: Outpatient Imaging In Primary Care
Physical Description: 1 online resource (164 p.)
Language: english
Creator: Sistrom, Christopher
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2010

Subjects

Subjects / Keywords: cost, diagnostic, healthcare, hierarchical, imaging, modeling, policy, primary, utilization, variation
Health Services Research, Management, and Policy -- Dissertations, Academic -- UF
Genre: Health Services Research thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Outpatient Imaging In Primary Care Diagnostic imaging comprises a rapidly growing portion of health care dollars spent in the U.S. Additionally, imaging tests that use X-Rays (including C.T. scans) now contribute half of the entire radiation dose to the population; having increased from 15% in 1980. Like other medical services, diagnostic imaging tests are utilized in some states and cities much more frequently than in others even after controlling for factors like age, gender, and illness burden. This marked variability in how frequently and for what reasons that imaging tests are done extends down to the level of individual doctors. Understanding the causes and effects of these sorts of differences in how health care delivered (including imaging tests) is one of the core parts of health services research. The fact that patients in a stable relationship with a primary care doctor have better health outcomes and consume less health services overall has prompted efforts to increase the supply of primary care doctors and encourage patients to seek care in a so-called medical home setting. One of the key roles of the primary care doctor is as a gatekeeper for expensive tests and procedures as well as referral to specialists. This study looks at a large group of patients (about 85K) being taken care of by 148 primary care doctors to whom they have demonstrated loyalty over three years based on the pattern of office visits. This so called loyalty cohort is useful as a representative sample of other patients and doctors functioning in the ideal doctor-patient relationship. The study seeks to answer questions about how various factors in the patients (including clinical activity) relate to the amount of imaging tests that their primary care doctor orders for them over two years. This can be useful in comparing the use of imaging tests (e.g., for 100 patients in a year) between primary care doctors by helping to adjust for differences in each doctor s mixture of patients. Such risk-adjusted utilization profiles help to understand variability in practice style and resource use between doctors for purely scholarly interest as well as more practical uses by entities actually paying for the services (e.g., insurance companies, employers, government health programs). A total of 35 pieces of information about each patient (not including their name or other identifiers) and 7 characteristics of each doctor were gathered from 6 different sources of data used routinely during patient care in the practices being studied. This information included complete listing of about one quarter million diagnostic imaging tests. Of these, about 60K were ordered by the patient s own (loyal) primary care doctor and were counted by patient to form the variable of interest (outcome). Statistical relationships between all the other patient and doctor factors with the number of imaging tests were analyzed. The results demonstrated that older female patients had more imaging as did those who had many medical problems listed in the clinical record system. Also, patients who visited doctors, were admitted to the hospital, or seen in the emergency room more frequently had more imaging. Doctor factors associated with a greater tendency to order imaging tests were less experience, female gender, and having a medium size practice (500-1000 patients). A special statistical technique that accounts for all patient factors allowed creation of profiles scoring each of the 148 doctors on their general tendency to order imaging tests on the average patient and how many more tests were ordered on patients with greater comparative need for diagnostic imaging.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Christopher Sistrom.
Thesis: Thesis (Ph.D.)--University of Florida, 2010.
Local: Adviser: McKay, Niccie L.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2010
System ID: UFE0041862:00001

Permanent Link: http://ufdc.ufl.edu/UFE0041862/00001

Material Information

Title: Outpatient Imaging In Primary Care
Physical Description: 1 online resource (164 p.)
Language: english
Creator: Sistrom, Christopher
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2010

Subjects

Subjects / Keywords: cost, diagnostic, healthcare, hierarchical, imaging, modeling, policy, primary, utilization, variation
Health Services Research, Management, and Policy -- Dissertations, Academic -- UF
Genre: Health Services Research thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Outpatient Imaging In Primary Care Diagnostic imaging comprises a rapidly growing portion of health care dollars spent in the U.S. Additionally, imaging tests that use X-Rays (including C.T. scans) now contribute half of the entire radiation dose to the population; having increased from 15% in 1980. Like other medical services, diagnostic imaging tests are utilized in some states and cities much more frequently than in others even after controlling for factors like age, gender, and illness burden. This marked variability in how frequently and for what reasons that imaging tests are done extends down to the level of individual doctors. Understanding the causes and effects of these sorts of differences in how health care delivered (including imaging tests) is one of the core parts of health services research. The fact that patients in a stable relationship with a primary care doctor have better health outcomes and consume less health services overall has prompted efforts to increase the supply of primary care doctors and encourage patients to seek care in a so-called medical home setting. One of the key roles of the primary care doctor is as a gatekeeper for expensive tests and procedures as well as referral to specialists. This study looks at a large group of patients (about 85K) being taken care of by 148 primary care doctors to whom they have demonstrated loyalty over three years based on the pattern of office visits. This so called loyalty cohort is useful as a representative sample of other patients and doctors functioning in the ideal doctor-patient relationship. The study seeks to answer questions about how various factors in the patients (including clinical activity) relate to the amount of imaging tests that their primary care doctor orders for them over two years. This can be useful in comparing the use of imaging tests (e.g., for 100 patients in a year) between primary care doctors by helping to adjust for differences in each doctor s mixture of patients. Such risk-adjusted utilization profiles help to understand variability in practice style and resource use between doctors for purely scholarly interest as well as more practical uses by entities actually paying for the services (e.g., insurance companies, employers, government health programs). A total of 35 pieces of information about each patient (not including their name or other identifiers) and 7 characteristics of each doctor were gathered from 6 different sources of data used routinely during patient care in the practices being studied. This information included complete listing of about one quarter million diagnostic imaging tests. Of these, about 60K were ordered by the patient s own (loyal) primary care doctor and were counted by patient to form the variable of interest (outcome). Statistical relationships between all the other patient and doctor factors with the number of imaging tests were analyzed. The results demonstrated that older female patients had more imaging as did those who had many medical problems listed in the clinical record system. Also, patients who visited doctors, were admitted to the hospital, or seen in the emergency room more frequently had more imaging. Doctor factors associated with a greater tendency to order imaging tests were less experience, female gender, and having a medium size practice (500-1000 patients). A special statistical technique that accounts for all patient factors allowed creation of profiles scoring each of the 148 doctors on their general tendency to order imaging tests on the average patient and how many more tests were ordered on patients with greater comparative need for diagnostic imaging.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Christopher Sistrom.
Thesis: Thesis (Ph.D.)--University of Florida, 2010.
Local: Adviser: McKay, Niccie L.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2010
System ID: UFE0041862:00001


This item has the following downloads:


Full Text





OUTPATIENT IMAGING IN PRIMARY CARE


By

CHRISTOPHER L. SISTROM
















A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY

UNIVERSITY OF FLORIDA

2010
































2010 Christopher L. Sistrom





















To my wife, Brenda









ACKNOWLEDGMENTS

The research described in this dissertation was made possible through

collaboration with numerous people at the Massachusetts General Hospital Physician

Organization and Radiology Department. They include (but are not limited to) Jeffrey

Weilburg, MD, Keith Dreyer, DO, PhD, Timothy Ferris, MD, George Leehan, and

Markus Stout. I am very grateful for their help and support.

Also, this research would not have been possible without the meticulous and

consistent work by Dr. Steve Atlas and his group to identify a stable cohort of patients

and primary care doctors.

My colleagues and friends in the Department of Radiology at the University of

Florida have been very kind in allowing me to spend enough time away from routine

clinical, administrative, and educational duties to engage in a multi-year effort to learn

several new disciplines; epidemiology, public health, health services research. Without

considerable support and encouragement on their part, none of this would have been

possible. They include (but are not limited to) Anthony Mancuso, MD, Patricia Abbitt,

MD, Jonathon Williams, MD, Melinda Chitty, Anne D'Amico, and Meryll Frost.

Niccie McKay, PhD has been tireless and infinitely patient in several roles

including mentor, teacher, editor, and chair of my dissertation committee. Cyndi Garvan,

PhD gave me a crash course in hierarchical modeling and was the external member of

my committee. Jeffrey Harman, PhD, has been instrumental in bringing a

methodological rigor to my efforts and Chris Harle, PhD honored me by inaugurating his

faculty career in serving on my dissertation committee.

My wife Brenda has been a wonderful partner during what may have seemed at

times like a decade long mid-life crisis. I very much appreciate her support and









forbearance and am glad to know that she will be around as I decide 'what I want to be

when I grow up'.









TABLE OF CONTENTS

page

ACKNOWLEDGMENTS .................................... ............... 4

LIST OF TABLES ......... ............... ..................... ....... ............... 9

A B S T R A C T .............. ..... ............ ................. .................................................. 1 3

CHAPTER

1 INTRODUCTION ........................... .......... ......... 16

Background and Significance ... .... .............................................. .. .................. 16
S p e c ific A im s .............. ...... ........... ......................................................................... 1 7
Summary of Study ................ ......... ......... ......... 17
Contribution to Literature ....................................................... 19
Policy Implications ................ ........ .......... ......... 20

2 BACKGROUND ....................................................................... ............................ 22

Definition of Diagnostic Imaging .............................. ......... 22
Imaging Utilization and Costs ................ .......... ......... ............... 26
Prim ary C are Setting............................... ............... 28
Imaging as Diagnostic Testing .................................. .. ....... 29

3 LITERATURE REVIEW ....................................... ......... 32

The Andersen Model ......................................................... 32
Small Area Variation ................ ........ ......... ............... 33
C lin ic a l U n c e rta in ty ................................................................................................. 3 7
Risk Adjustment ...... .................. ............ ......... 38
Appropriateness and Supplier-Induced Demand ....... ... ............................... 40
S u m m a ry ................................................................................................... 4 2

4 CONCEPTUAL FRAMEWORK............................ ......... 44

Clinical Need.............................................. ............... 45
Context: Patient ....................................................................... ............................ 47
Context: Physician ................ ........ .......... ......... 48
Context: Malpractice ............... ........... .................... 52
Context: Practice Organization ...................................... ....... 54
Context: Payer and Prices ...................... ......................... 55
Context: Access to Imaging ........... .......... ................................... ......... 56
S u m m a ry ......... .................................. ......... ........................ .................. 5 8




6









5 SETTING, DATA SOURCES, AND VARIABLES....................... .............. 60

S e ttin g s ........... ..................... .......................................... ....... ...... 6 0
Prim ary Care Practice ......... .. .... ............................................................. 60
O outpatient R adiology .......................................... ................ .. .......... 62
D a ta S o u rce s ............... .... ...................................... .................... .......... .. 6 3
Loyalty Cohort ............ ......... ............ ....... ................................. 63
Patient D details ............ ......... ......... .......... ................................. 64
Physician Details .......... .. ...... ........... ........... ..... .. ... ...... ....... ........ 65
Im ag ing U tilization ........... ......... ......... .............................. .. .......... 65
Variables ............................ .. ................................ ..... 66
Im aging Utilization (dependent variable) ................. ................................ 66
Patient Characteristics ............... ........... ................ ............... 66
C clinical Events ................................................................. ............................ 67
Clinical Problems......................................................... 68
Outpatient Prescriptions ............................... ... ....................... 69
Other Imaging Utilization .......... .. ............... ............. ... ........ ............... 69
Physician Characteristics .............. ................ ............. ........ ..... .... 70
S ite C ha racte ristics ............... ................ ....... ................... ..... ........... 7 1
Patient, Provider, and Clinic Identifiers ......... ...... .... ........... ........ ..... 72
Data Integrity: Clinical Activity Variables......... ............. ........ ............. 72
Variable Sum m ary ............ ......... ................... ................................. 73

6 M E T H O D S ............... ............................................ ........................... ..... 84

Outcome Variable Distribution .............. ... ................... .................... 84
Correlation between Independent Variables .................. ....................... ......... 87
B ivariate R relationships ............ ......... ......... ............................. .............. 88
Variable Reduction for Modeling ...... ........................ .... ............... 88
Multivariable (logistic) Modeling: Any Imaging Use................................. 88
Multivariable (Poisson) Modeling: Imaging Intensity (non-zero)........................... 90
Preparation for Multi-Level Modeling: Imaging Propensity Scores....................... 91
M ulti-Level (H ierarchical) M odeling ..................................................... ............... 91

7 RESULTS ................................... .................................. ......... 96

O utcom e V ariable D distribution ............................................ ............................... 96
Correlation between Independent Variables .......... .......................................... 97
B ivariate relationships ................................................ .......... 99
V ariable R education for M odeling ...................................... ......................... ....... 99
Multivariable (Logistic) Modeling: Any Imaging Use ................................. 100
Multivariable (Poisson) Modeling: Imaging Intensity (non-zero)........................... 105
Comparison of Any Imaging and Imaging Intensity Results............................... 108
Preparation for Multilevel Modeling: Imaging Propensity Scores....................... 109
M ulti-Level (Hierarchical) Modeling............................... ............... 110









8 D IS C U S S IO N ............... ........................ ......... ......................... 13 7

Summary of Key Results ...................................................... 137
Discussion of Key Results ........... .......... ................................... ......... 140
Limitations of Study ................ ......... ........ ........ 142
Policy Implications .......... ................ ............... 146
C contribution to Literature .......................................... ............... ........... 148
Future Research ................ ......... ................... 149

LIST OF REFERENCES .................................. .. .................... 152

BIOGRAPHICAL SKETCH ........... ......... ......... .. ...... ...................... 164









LIST OF TABLES


Table page

5-1 All diagnostic imaging performed on study cohort during two years of study. .... 73

5-2 Outpatient diagnostic imaging performed on study cohort during two years of
stu d y .............. .. ....... .... ............ ...................................... 7 4

5-3 Outpatient diagnostic imaging ordered by patient's linked (loyal) doctor
during tw o years of study............................................................. .......... 74

5-4 Univariate statistics of the outcome variable (per patient count of outpatient
imaging tests ordered by primary care provider). ......... .......... ............. .. 75

5-5 Distribution of patient race ......... ........ .......... ......... .................. 75

5-6 Distribution of patient's payer categories ..................... ............ .................. .. 75

5-7 Patient's payer collapsed into 6 categories. .............. ....................................... 76

5-9 CPT codes and relative value units for ambulatory office visits....................... 77

5-10 Outpatient visit activity variables (per patient). ........................... ... ............... 77

5-11 Binary clinical problem variables (per patient). ............................ ... ............... 78

5-12 Four level categorization of patient active outpatient medications (per
patient). .................... .... ......... ....... ... ........ ....... ........ 78

5-13 Summary of other (non-outcome) imaging test utilization variables (per
patient). ........................................................................ ...... ........ 78

5-14 Four level categorization of the number of patients cared for by each provider
(panel size)...................................................................... ................ 78

5-15 Site (clinic) characteristics. ................ .... ..... ........ .... ... ............... 79

5-16 Description and categorization of 33 patient level independent variables. ......... 80

5-17 Description and categorization of provider (8) and clinic level (2) independent
variables. ........................................ .................. .......... .. 81

7-1 Spearman correlations between clinical activity and other imaging variables.. 116

7-2 Bivariate relationship between patient level variables and outcome (imaging
counts)............................................ .......... 117









7-3 Bivariate relationship between clinical activity variables and outcome
(imaging counts). .......... ............ ......... ............... ................ 118

7-4 Bivariate relationship between provider and clinic level variables and
outcome (imaging counts). ....................................... ... ............... 119

7-5 Patient level results from multivariable logistic model on any imaging use....... 120

7-6 Provider and clinic level results from multivariable logistic model on any
im aging use. ................... ........... .................................. 121

7-7 Patient level results from multivariable Poisson model on imaging intensity ... 122

7-8 Provider and clinic level results from multivariable Poisson model on imaging
intensity. .................... .. .............. ......... ........................ 123

7-9 Univariate statistics for raw imaging counts (IMG) and predictions from ZIP
model (IMG_PROP). .............. .......................... ..... ..... ............... 123

7-10 Dim tensions ................... ....... .... .. ................ ......... ................ ......... 123

7-11 Estim ated G correlation m atrix. ........... .. ............................ ............... 124

7-12 Covariance parameter estimates ...................................... ......................... 124

7 -1 3 F it s ta tis tic s ................ .................................................................. 1 2 4

7-14 Solution for fixed effects. ................................................................. 124

7-15 Type 3 tests of fixed effects ............... ...................... ............. 124

7-16 Results from multi-level random coefficients model .................................... 124

7-17 Quadrants in intercept versus slope relationship plot. ................................. 125

7-18 Exemplary providers in each quadrant. ...................... ................ 125

7-19 Comparison of null, and reduced model residuals with full model. ................... 125









LIST OF FIGURES
Figure page

2-1 Total number of imaging studies by year in the U.S. ............... ............... 30

2-2 Imaging shows highest cumulative growth in services per beneficiary (1999-
2 0 0 4 ) ...................... .. .. ......... .. .. ............................................ 3 0

2-3 Cumulative growth in imaging volume varies by type (1999-2004)................... 31

4-1 Summary diagram of conceptual model for outpatient imaging utilization in
primary care............................................ .......... 59

5-1 Outpatient imaging tests (per patient) ordered by the linked (loyal) primary
c a re p ro v id e r ................ ........................................................... 8 1

5-2 Number of outpatient visits by all patients in study cohort (by month) over two
years of study. .................................. .............................. ... ............ 82

5-3 Number of outpatient imaging tests performed on all patients in study cohort
(by month) over two years of study................ .............. ............. ....... ..... 82

5-4 Number of hospital encounters for all patients in study cohort over two years... 83

7-1 Comparison of imaging counts with three Poisson distributions....................... 126

7-2 Logistic regression results for any imaging utilization................................... 127

7-3 Poisson regression results for (non-zero) imaging intensity. ............................ 128

7-4 Comparison of significant variables for any imaging use and imaging intensity 129

7-5 Comparison of significant variables for any imaging use and imaging intensity
(s m a ll e ffe ct s iz e s ) ............... ...................................................... 13 0

7-6 Imaging propensity score distributions by provider ...................... ............... 131

7-7 Centered imaging propensity score distributions by provider. .......................... 131

7-8 Plot of intercept and slopes for all 148 providers obtained from multi-level
model of imaging utilization. ......................................................... ........... 132

7-9 Imaging utilization versus centered imaging propensity for a low utilizing
doctor. ................................... ................................. .......... 133

7-10 Imaging utilization versus centered imaging propensity a high utilizing doctor. 134

7-11 Provider means sorted by ascending order within each site............................. 135









7-12 Provider slopes sorted by ascending order within each site ........................ 136









Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy

OUTPATIENT IMAGING IN PRIMARY CARE


By

Christopher L. Sistrom

August 2010

Chair: Niccie McKay
Major: Health Services Research

Diagnostic imaging comprises a rapidly growing portion of health care dollars

spent in the U.S. Additionally, imaging tests that use X-Rays (including C.T. scans) now

contribute half of the entire radiation dose to the population; having increased from 15%

in 1980. Like other medical services, diagnostic imaging tests are utilized in some states

and cities much more frequently than in others, even after controlling factors like age,

gender, and illness burden. This marked variability in how frequently and for what

reasons that imaging tests are done extends down to the level of individual doctors.

Understanding the causes and effects of these sorts of differences in how health care is

delivered (including imaging tests) is one of the core parts of health services research.

The fact that patients in a stable relationship with a primary care doctor have

better health outcomes and consume less health services overall has prompted efforts

to increase the supply of primary care doctors and encourage patients to seek care in a

so-called 'medical home' setting. One of the key roles of the primary care doctor is as a

'gatekeeper' for expensive tests and procedures as well as referral to specialists. This

study looks at a large group of patients (about 85K) being taken care of by 148 primary

care doctors to whom they have demonstrated 'loyalty' over three years based on the









pattern of office visits. This so called 'loyalty cohort' is useful as a representative sample

of other patients and doctors functioning in the ideal doctor-patient relationship.

The study seeks to answer questions about how various factors in the patients

(including clinical activity) relate to the amount of imaging tests that their primary care

doctor orders for them over two years. This can be useful in comparing the use of

imaging tests (e.g., images for 100 patients in a year) between primary care doctors by

helping to adjust for differences in each doctor's mixture of patients. Such risk-adjusted

utilization profiles help to understand variability in practice style and resource use

between doctors for purely scholarly interest as well as more practical uses by entities

actually paying for the services (e.g., insurance companies, employers, government

health programs).

A total of 35 pieces of information about each patient (not including their name or

other identifiers) and 7 characteristics of each doctor were gathered from 6 different

sources of data used routinely during patient care in the practices being studied. This

information included complete listing of about 4 million diagnostic imaging tests. Of

these, about 60K were ordered by the patient's own (loyal) primary care doctor and

were counted by patient to form the variable of interest (outcome). Statistical

relationships between all the other patient and doctor factors with the number of

imaging tests were analyzed.

The results demonstrated that older female patients had more imaging as did

those who had many medical problems listed in the clinical record system. Also,

patients who visited doctors, were admitted to the hospital, or seen in the emergency

room frequently had more imaging. Doctor factors associated with a greater tendency to









order imaging tests were less experience, female gender, and having a medium size

practice (500-1000 patients). A special statistical technique that accounts for all patient

factors allowed creation of 'profiles' scoring each of the 148 doctors on their general

tendency to order imaging tests on the 'average' patient and how many more tests were

ordered on patients with greater comparative 'need' for diagnostic imaging.









CHAPTER 1
INTRODUCTION

Background and Significance

Medical technology is often cited as a major driver of health care costs in the U.S.,

with diagnostic imaging being a 'poster child' of this trend. Advanced imaging, ordered

during ambulatory care, is not only costly, but often leads to a 'cascade' of further tests

and interventions (Deyo 2002, Mold and Stein 1986, Verrilli and Welch 1996). There is

ample evidence of marked variation in utilization of imaging services in outpatient care

at multiple levels of aggregation, from international down to individual groups within a

large practice (Burkhardt and Sunshine 1996, Couchman et al. 2005, Goel et al. 1997,

Hartley et al. 1987, Katz et al. 1996, Lysdahl and Borretzen 2007). This variation in

utilization suggests that a substantial portion of diagnostic imaging may be

unnecessary. Patients receiving such studies may be needlessly exposed to radiation.

Findings of uncertain clinical meaning may prompt more imaging and costs may be

increased.

Examining the ways in which primary care physicians utilize imaging, and using

those insights to inform and enhance the consistency of their choices, can improve

health care quality and increase cost effectiveness. Thus, the event of a primary care

doctor ordering an imaging test is common and expensive enough to warrant study by

itself. Also, it represents a very fruitful paradigm for understanding medical decision

making and much of the variation in downstream discretionary health care utilization

(Parchman 1995). The interesting events that occur in primary care settings are

'upstream' from major procedures and consist of referral to specialists and diagnostic

testing. These discretionary decisions made during routine office visits to general









practitioners have substantial impact on population health and the overall cost of health

care (Sirovich et al. 2008).

Specific Aims


Quantify outpatient diagnostic imaging utilization over two years by primary care
doctors caring for a cohort of patients that regularly attend clinic (i.e., are loyal)

Collect and characterize demographic, contextual, and clinical factors for all
patients in the cohort

Collect and characterize demographic and practice factors for the primary care
physicians regularly caring for the same patients in outpatient clinics

Determine the relationships between patient, doctor, and clinic factors and the
probability that patients had at least one imaging test during the two year study
period (any use)

Determine the relationships between patient, doctor, and clinic factors and the
number of examinations performed in patients with at least one imaging test
(intensity of use)

Develop a model based risk adjustment method for imaging utilization (any use
and intensity of use) producing an 'expected' amount of imaging given known
patient and doctor factors (imaging propensity)

Estimate and partition the variability in imaging utilization using a hierarchical
method which takes into account patient's risk adjusted imaging propensity which
is in turn influenced by each doctor's tendency to use imaging in their practice



Summary of Study

The basic unit of interest for this study is a patient who regularly visits a primary

care physician and the time frame is two years. The main outcome is a measure of the

amount of outpatient imaging performed in the period of study that was ordered by the

same primary care physician on the patient in question. Specifying and quantifying the

'amount' of imaging is a non-trivial task because we wish to count all non-invasive

diagnostic studies of any modality (e.g., CT, MRI, XRAY, Ultrasound, Nuclear Medicine,









PET). Fortunately, the Radiology Department and the host institution for the study share

a longstanding commitment to a robust medical informatics infrastructure. Thus,

complete and detailed records of all imaging events for the past 15 years are readily

accessible. Also, to enhance medical management and clinical operations support, a full

spectrum of clinical data are available for health services research such as this study.

These informatics resources allow collecting a large amount of data on each patient that

will be used to form explanatory and control variables for analyzing drivers, enablers,

and inhibitors of outpatient primary care imaging utilization. Finally, using credentialing

databases from the host institution and publicly available data from state licensing

sources, provider level factors will be obtained for study.

After characterizing and quantifying the imaging utilization over two years of study,

the tests performed on each patient will be attributed to them and cross-tabulated

according to the patient's status (Inpatient, Emergency Room, Outpatient) when

performed and the type of provider ordering the study (primary care vs specialist).

Patient and provider level variables will be characterized, validated, and then

individually evaluated in bivariate fashion with the main outcome (count of outpatient

imaging tests ordered by patient's primary care doctor). Initial multivariable modeling will

be performed using logistic regression on the whole data set (outcome=any imaging

yes/no) followed by linear regression (with Poisson errors) on the non-zero

observations. The joint contribution of the various patient, provider, and clinic factors to

any use (logistic regression) and intensity of use (Poisson regression) will be inferred

from the odds ratios and coefficients respectively.









To prepare for hierarchical analysis, a zero-inflated Poisson regression will be

estimated using all observations and only patient level factors to form a single

continuous variable from the model predictions. This serves to summarize the multiple

patient level factors into a single number representing risk adjusted expected imaging or

'propensity' for each patient. A two-level hierarchical model with patients at the first level

nested within the 148 providers at the second level will be specified and fitted to the

observed imaging counts for all patients. Only two independent variables will be

included: the propensity variable and a unique ID number for each provider. By

specifying a model that defines provider level intercepts (mean imaging use) and slopes

(response to imaging propensity) two unique characteristics can be estimated for each

provider and compared with each other. In addition, variance components computed for

each level in the hierarchical structure will serve to partition variation in imaging

utilization between patient's imaging propensity, provider's general tendency to image,

and provider's response to clinical factors in their patients in the amount of imaging they

order.

Contribution to Literature

This research will contribute to the literature about primary care imaging utilization

in several ways. First of all, the study population of providers and patients is large,

comprehensive, and unique (i.e., all primary care doctors and their loyal patients at a

large metropolitan academic health center). A previously validated loyalty cohort

methodology (detailed in setting, data, and variables section) identifies patients,

doctors, and clinics in ongoing and stable relationships with each other (i.e., usual

source of care or 'medical home'). The available data about imaging utilization and

patient health status is complete and highly detailed, having come directly from primary









sources (i.e., electronic medical records and clinical radiology information systems). For

example, the requesting provider is recorded for each imaging test, which allows

stratification of utilization by who ordered the examination (patient's primary care

physician versus a specialist) and the patient care setting (outpatient, inpatient,

emergency department).

Complete data are also available for all outpatient visits, hospital stays, and

emergency room encounters. Therefore, the analytic data will contain a robust set of

clinical activity and risk adjustment variables, which will be used to estimate the risk

adjusted expectation (propensity) for imaging utilization to a high degree of accuracy

and precision. Once patient level clinical factors have been accounted for, residual

variation between providers will be quantified and partitioned in a way that should shed

considerable light on the contribution of various contextual factors. When analysis of

primary care outpatient imaging utilization is extended to provider profiling, proper

specification of patient level risk-adjustment and hierarchical analysis at the provider

level is crucial to fairly applying these medical management tools and this study will

advance knowledge of these issues.

Policy Implications

In the ongoing debate about national health care policy the most divisive and

vexing issues relate to unsustainable medical cost inflation. This is largely the result of

increasing utilization of expensive, high-technology diagnostic and therapeutic

interventions which occurs at the discretion of physicians to a substantial degree. The

magnitude of this physician discretion over utilization as opposed to evidence-based

clinical need will determine success of strategies to reduce costs targeted to physicians.

These include education, point of care intervention, clinical decision support, and post









hoc profiling for 'efficiency'. This study will provide estimates for the relative

contributions of clinical need, physician style, and non-clinical patient factors with

respect to utilization of outpatient diagnostic imaging in primary care. Assuming that

clinical need remains as a significant and substantial driver of imaging utilization, this

study provides insight into risk adjustment models which will be necessary for utilization

management efforts going forward.









CHAPTER 2
BACKGROUND

This chapter will provide background material about medical imaging as a

diagnostic test applied in routine outpatient primary care. Explanation of exactly what is

meant by adult outpatient primary care will serve to further refine the setting for the

research described in this dissertation. In addition to specifying exactly what procedures

are and are not included in a definition of diagnostic imaging, the chapter will describe

national trends toward rapidly increasing volume and cost of these services.

Definition of Diagnostic Imaging

A diagnostic imaging procedure (DIP) is defined as a discrete event with the

following attributes. The 'subject' of this event is an individual in a provider-patient

relationship with a medical practitioner. The practitioner initiates the event by means of

an order for the DIP, and this can be verbal, written, or electronic. The individual

submitting the order will be called the ordering provider (OP). Two other provider roles

are required to complete the event and these are the performing provider (PP) and the

interpreting provider (IP). The PP interacts directly with the patient using some kind of

diagnostic imaging equipment to produce images which are subsequently interpreted by

the IP who communicates the findings to the OP. Note that the while the three providers

(OP, PP, and IP) are often separate individuals, a single person may perform all three

roles. For example, an obstetrician may decide that a fetal ultrasound should be done

on her patient, may personally perform the scan, and interpret the images in real-time

from the video display all in a single step. Formal communication of the interpretation

back to the OP typically includes some permanent documentation of the findings and

interpretation, which is placed into the patient's medical record.









In addition to the people involved, other attributes serve to describe a DIP.

Although defined as a discrete event, a DIP occurs in a series of steps: ordering,

performing, interpreting, and review of results. This has been referred to as the

'radiology round trip' to emphasize the complexity of the process (Thrall 2005b, Thrall

2005a, Thrall 2005c). In some settings, the time between these steps can be lengthy

and quite variable. For most purposes, we define the procedure as having 'occurred' at

a single point in time, when the PP finishes performing or 'completes' the examination.

This is often referred to as the 'date of service' in medical record systems, billing

applications and in claims data.

Two other attributes of a DIP must be articulated and these define the sort of

equipment used to obtain the images (modality) and what part of the patient (body area)

was examined. The term 'modality' refers to the physical nature of the process used to

create images of the patient's body and is useful as a primary means of categorizing the

equipment. For example, 'radiography' uses invisible photons of high energy (X-Rays)

directed at and through the patient from a fixed generating tube and detects the photons

on a flat surface (film or digital plate). Computed Tomography (CT) also uses X-Rays

but the generating tube spins around the patient along with a detector. The varying

intensity of photons falling on the detector is combined with angular position to compute

a tomographic (planar slice) image of the patient's inner structure. Magnetic resonance

imaging (MRI) machines irradiate patients with sequences of radio waves from coils

housed in a strong magnetic field. Very sensitive antennae detect weak 'echoes' of the

radio waves emitted by hydrogen atoms within the patient's tissues and compute an

image based on strength and frequency. Nuclear medicine (including positron emission









tomography=PET) involves injecting patients with a variety of radioactive 'tracers' which

emit gamma ray photons or energetic particles as they undergo decay. These are

detected outside the patient and processed to form a map or image of activity. Finally,

ultrasound (US) produces pulses of high frequency sound from a specialized transducer

in direct contact with the patient's skin. The same transducer detects 'echoes' of those

sound waves that differ in amplitude and timing depending on tissue characteristics and

spatial location. These signals are processed into an image of anatomy immediately

beneath the transducer which can be moved around to examine an entire region.

The final attribute assigned to a particular DIP details the anatomic regions of the

patient (body area) that are exposed to the radiation, particles, or sound to produce

images for subsequent interpretation. The nomenclature of body areas typically

'imaged' is fairly straightforward and includes: head, chest, abdomen, pelvis, spine,

arms, and legs. More specialized examinations may cover particular anatomy such as

coronary arteries, lungs, aorta, gallbladder, and so forth. When a modality is combined

with a body area the result is a specific named DIP and a patient is said to have

undergone (completed it) at the date/time (of service) that they were dismissed from the

testing facility and the images became available for interpretation.

An example will help to clarify these definitions. Mary Jones (an adult) makes an

appointment with her internist (Dr. Smith) and during the visit, complains of frequent and

increasing headaches. Dr. Smith may decide to order a DIP to exclude the possibility of

a structural lesion (e.g., a mass in the brain) before treating her with drugs to relieve the

pain. In this case, Dr. Smith (OP) will order a test that evaluates the brain (body area =

head) and has several options about modality including X-Ray, CT, MRI, Nuclear









Medicine and even ultrasound. Dr. Smith's decision as to which of these modalities to

order relies on her personal assessment of the relative appropriateness for Ms. Jones.

Assuming that they opt for MRI, Dr. Smith orders an MRI of the head/brain for Ms.

Jones to be performed sometime in the future. Dr. Smith's order is transmitted in some

fashion (computer order, phone, fax, paper prescription carried by Ms. Jones) to an

imaging facility (often hospital based) that offers MRI scanning for outpatients. When

Ms. Jones keeps her appointment, she is brought into the MRI scanning suite and

asked to lie down on a movable couch which carries her into the actual MRI machine.

She is instructed to breathe quietly and hold still while the technologists (PP) execute a

pre-programmed protocol that directs the machine to acquire images of Ms. Jones'

brain over the next several minutes. Upon completion of these imaging 'sequences' and

a brief observation period, Ms. Jones is released to return home. At the same time, the

PP (technologist) executes commands to 'complete' the examination at which time the

images (computer files) are transmitted to storage media so as to be ready for

download and interpretation and/or review. This sequence of events forms a single unit

of outpatient imaging utilization and may be described by saying: Dr. Smith (OP)

requested an MRI of the head on Ms. Jones (the patient) during an outpatient visit and it

was completed on (the date of service) by (PP) to be interpreted by (IP) with the report

to be sent back to Dr. Smith (OP) for review.

The next section deals with trends in the U.S. of rapidly increasing volumes of

imaging utilization. To set the stage for this, consider the typical charges and

reimbursements that might be submitted and paid, respectively, for Ms. Jones' MRI of

the head. It is important to note that there are two separate 'billable' imaging events









arising from this sequence. The first is the encounter at the imaging facility where the

MRI images of Ms. Jones' head were made. The outpatient imaging facility will bill Ms.

Jones and/or her insurance carrier for the 'technical charge' which will likely be well over

$1000.00. The second is the bill rendered by the physician who actually interprets the

images, and this 'professional charge' will approach $500.00. The actual dollar amounts

of reimbursement received will depend on the payer, with different payers reimbursing

different amounts for the same service.

Of course, Dr. Smith will also submit a bill for the office visit during which she

ordered the scan on Ms. Jones. However, this is not directly attributed to the imaging

test and is not counted as part of its cost. It should also be noted that in an increasing

number of cases, the MRI machine might actually be owned (or leased) by Dr. Smith

and sometimes she might interpret the images herself. Under this scenario, Dr. Smith

may directly bill for, or receive through more indirect means, most or all of the revenue

generated by the technical and interpretation charges. This practice (called self-referral

of imaging) is controversial among physicians and is targeted as a driver of increasingly

burdensome costs in the U.S. by government and private payers.

Imaging Utilization and Costs

No other branch of medical technology has experienced the explosive growth in

volume and variety of available services that radiology has during the past two decades.

The medical care industry in the U.S. has purchased and installed advanced imaging

equipment at an astounding rate, outpacing all other countries. Figure 2-1 illustrates this

trend in terms of number of imaging procedures (Medicare Payment Advisory

Commission 2005). Given that the current population of the U.S. is just now reaching

300M, this translates into about 1.4 imaging tests per person year. The most dynamic









growth has occurred in CT scanning, with a steady increase in capability and indications

for use occurring over the last 30 years.

As the number of imaging procedures (many of which are CT scans) has

increased, the cumulative effective radiation dose to the average American has nearly

doubled from 3.6 milli-Sieverts (mSv) in 1980 to 6.2 mSv in 2006 (National Council on

Radiation Protection & Measurements 2009). Several high profile articles in the past

couple of years have raised concerns about a small but significant population risk for

subsequent cancers induced by ionizing radiation delivered during medical imaging

procedures (Brenner and Hall 2007, Fazel et al. 2009, Berrington de Gonzalez et al.

2009, Nyweide et al. 2009, Smith-Bindman et al. 2009).

Over the period 1985-1990, established technologies, such as CT, continued to

grow in volume for Medicare. At the same time, the new technology of MRI exploded in

terms of utilization with a 372% increase in national procedural volume for Medicare

(Boutwell and Mitchell 1993). Imaging costs to the Medicare system in the past two

decades rose much more rapidly than any other component and now comprise at least

14% of total Part B expenditures for physician services as specified in a report by the

Medical Payment Advisory Commission (MEDPAC) to the U.S. Congress (Medicare

Payment Advisory Commission 2003). Imaging costs grew by approximately 10% per

year during the period covered by the report (1999-2002) compared with average

growth of 3.3% per year for all physician services.

Testifying before Congress in 2006, Glenn Hackbarth of MEDPAC amplified and

extended prior reports and testimony (Medicare Payment Advisory Commission 2006).

He presented 1999-2004 data showing growth in Medicare claims for diagnostic









imaging as being the highest of all services at 62%. Furthermore, growth was especially

high in emerging modalities (up to 140%) with even established technologies, like head

CT, outpacing general growth at 43%. Figures 2-2 and 2-3 (exhibits in Mr. Hackbarth's

testimony) illustrate these points and have been widely reproduced.

Primary Care Setting

This research examines imaging in outpatient adult primary care, which is defined

as being rendered by doctors trained in internal medicine, family practice, and general

practice. This is the classic paradigm of clinical decision-making, in which patients

present with signs, symptoms, known diagnoses, or physical abnormalities that

generate moderate probability of one or more treatable conditions. Much of the literature

about the utilization of health services by primary care providers deals with one of two

possible responses to this situation: diagnostic testing or referral to specialists.

This study will not consider imaging utilization that occurs as part of disease

screening programs, including mammography, CT colonography, cardiac calcium

detection with CT, lung cancer screening with CT, among several others. While

important for population health and public policy, fundamental differences exist between

imaging for screening and imaging for diagnostic or prognostic purposes. For example,

once a universal population screening strategy has been adopted, policy-makers are

primarily concerned with under-utilization in the target group. Conversely, in the case of

diagnostic imaging, payers and regulators (at least in the U.S.) concentrate almost

exclusively on problems of over-utilization.

Additionally, this work will not attempt to analyze utilization of imaging that occurs

during inpatient care or imaging tests that are ordered during work-up of patients in

emergency and urgent care settings due to the differing nature of medical decision-









making during inpatient and urgent care encounters compared with primary care.

Further, the study does not examine imaging utilization directed by surgeons,

oncologists or other specialists, because patients seen by doctors in these fields have

unique, disparate, and complex medical problems and co-morbidities. Finally, imaging

for children directed by pediatric providers will not be considered as it is quite rare in

primary care settings and quickly devolves into specialty-oriented utilization after

identification of congenital abnormalities, childhood cancer, or other serious problems.

Imaging as Diagnostic Testing

Imaging examinations that are interpreted by someone other than the ordering

doctor represent a hybrid of diagnostic test and specialist referral. This is because

imaging tests (especially complex ones like CT and MR) are perceived as consultations

by patients as well as by ordering physicians. In contrast, clinical laboratory tests (blood

chemistry, hematology, microbiology, and so forth) are 'reported' rather than

'interpreted' and this distinction is crucial. Non-imaging diagnostic test results are

generally reported in the form of numbers or simple fact assertions and are often

produced by automated methods with minimal analytic input by the rendering personnel.

Examples of such tests include blood chemistry, antigen/antibody assay, serum drug

levels, and urinalysis. With these, the ordering physician must synthesize an

'interpretation' about whether or not the result is normal and then decide relevance to

the clinical question. On the other hand, radiologists produce interpretative documents

that reach provisional conclusions about the probability of relevant clinical conditions (or

at least classes of disease). Sometimes these reports contain recommendations for

further follow up, clinical correlation, or even treatment and thus also function as

consultation notes.












500 471
450
430

401
41 400-
400 374
a. 349
323
C 299
300 281
I 267 268




200 -
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008


Figure 2-1. Total number of imaging studies by year in the U.S.



65- 62
60
55 Growth of all
50- physician services
45 43 44
40
0 35
30 -- 31
S30
a-
19 20
20-
15-
10
5-
0 ,
E&M Major Tests Other Imaging
procedures procedures


Figure 2-2. Imaging shows highest cumulative growth in services per beneficiary (1999-
2004).
































-2


0 o

Figure 2-3. Cumulative growth in imaging volume varies by type (1999-2004).









CHAPTER 3
LITERATURE REVIEW

This chapter will summarize relevant publications that inform a conceptual model

of imaging utilization in outpatient primary care. A general overview of theory and

empiric research about small area variation in patterns of health service utilization will

show that clinical uncertainty, physician practice style, patient preferences, and

economic factors all play roles. An important contributor to variation in use of diagnostic

imaging is clinical uncertainty. In fact, imaging tests are doubly subject to variation in

utilization related to clinical uncertainty because, by definition, their main function is to

reduce it. Thus, ambiguity about whether to observe, test, or treat in specific clinical

situations is multiplied by uncertainty about which, if any, imaging test to order. In

addition to how best to diagnose and treat their patients, doctors are also concerned

about income, leisure time, satisfaction with practice, and mitigation of malpractice risk.

These considerations influence utilization of imaging and their effect is magnified in the

presence of clinical uncertainty. Finally, patients bring a complex mixture of factors to

decisions about diagnostic imaging.

The Andersen Model

Any examination of health services utilization must consider the Andersen

behavioral model as an organizing framework. The Andersen model seeks to explain

health care utilization in an entire community, many members of whom do not seek

medical care at all. Andersen reviewed his model 25 years after it was developed and

described it mostly in terms of access to health care (Andersen 1995). The Andersen

model is primarily applied to factors that determine the seeking of health care services

in whole populations rather than what happens during encounters with providers (e.g.,









referral for diagnostic imaging). In a meta-analysis of papers using the Andersen model,

only 2 of 139 examined provider characteristics such as specialty, experience, and

gender (Phillips et al. 1998).

Thus the Andersen model is not directly applicable here because in looking at non-

screening imaging utilization in primary care, the denominator is patients who already

have an active relationship with a doctor. Like prescription drugs, one cannot undergo a

diagnostic imaging test without a prescription or referral from a health care provider. In

Folland's comprehensive review of variations in the use of medical care, this dichotomy

in overall utilization is labeled as 'first occurrence' versus 'intensity' (Folland and Stano

1989). Much of the Andersen model deals with 'first occurrence' whereas imaging

utilization falls under the 'intensity' concept. Therefore, the majority of this literature

review will focus on works that inform the intensity of outpatient imaging utilization in an

existing patient-provider relationship.

That being said, it will become evident from the empirical distribution of outpatient

imaging examined in this study, that utilization also seems to have a two-part structure.

These may be called any use and amount of--non-zero--use to account for a substantial

fraction (over half) of patients with no imaging at all during the two years of data

collection. However, in this study, even patients with no diagnostic imaging have visited

their primary care doctor regularly and a differently specified conceptual model will be

articulated below that is specific to outpatient imaging this special setting.

Small Area Variation

In an ideal world, the intensity and mixture of imaging would be appropriate to

each patient's clinical situation regardless of any contextual differences among

providers or patients. The counterfactual goes like this: Consider a situation where all









doctors and patients in a particular setting have access to identical and comprehensive

evidence about imaging appropriateness (Sistrom 2009). Further, physicians act purely

as agents and patients make rational choices based on maximizing longevity and well-

being. Finally, economic considerations are uniform across all doctors and patients

(e.g., a single payer universal insurance program). In this idealized setting, the amount

and types of imaging tests that patients underwent would be based solely on their

clinical presentation including pre-existing conditions and new symptoms, signs, and

disease trajectory. This implies that, after fully accounting for differences in clinical

presentation, the adjusted rate of imaging utilization would vary minimally over different

levels of aggregation. That being said, there is little theoretical guidance as to exactly

what this 'minimal' or 'natural' variation would entail in terms of intensity and mixture of

imaging or any other health service (Cain and Diehr 1992, Diehr et al. 1990).

Empirical investigations have identified considerable variation in the use of

medical services. The large body of literature documenting this phenomenon is referred

to as research on "small area variations". The seminal paper on this topic, which was

published by Wennberg and Gittelsohn in 1973, compared utilization and expenditures

among 13 hospital service areas in Vermont during 1969 and found large variations

among them (Wennberg and Gittelsohn 1973). For example, appendectomy rates

varied from 10 per 10,000 persons to 32 per 10,000 persons. Since then, numerous

articles, monographs and texts have examined geographic variations at the

international, state, regional, and local levels for a wide variety of types of medical

services (Health Services Research Group 1992, Folland and Stano 1990, Pekoz et al.

2003, Stano 1991, Stano 1993, Wennberg 1993). Wennberg's legacy is perhaps best









embodied in the Dartmouth Atlas of Healthcare which aggregates and analyzes

Medicare claims going back to at least 1990 through the most recent past year to

characterize variations in health service use across the entire U.S. in over 300 carefully

defined contiguous regions (Wennberg 2004b) (Wennberg 2004a). Much of the

available small area variations data deals with broad categories of services or spending.

For example, the Dartmouth Atlas provides use rates and costs per beneficiary for the

category of diagnostic services which includes laboratory tests and imaging procedures.

When more granular area utilization analyses have been undertaken, they often have

examined countable and more 'high impact' events like hospitalization, surgery, and

invasive diagnostic procedures (e.g., endoscopy and cardiac catheterization) (Leape et

al. 1990, Chassin et al. 1987).

The policy concern arising from evidence of small area variations is that they

persist even after controlling for factors such as age, gender, and race (Health Services

Research Group 1992, Blumberg 1987, Davis et al. 2000b, Wennberg 1987a,

Wennberg 2002). Moreover, research has found little evidence that the geographic

variations were correlated with differences in health status, disease-specific population

health indicators, or other measures of health outcomes (Sirovich et al. 2006, Fisher et

al. 2003b, Fisher et al. 2003a, Franks et al. 2000a, Brook and Lohr 1985). Thus, the

lack of measurable population benefit in higher use areas combined with rising health

care spending, has led policy makers to hope that health care cost reduction can be

obtained without welfare loss by targeting upper outliers (Fisher et al. 2003a, Fisher et

al. 2003b, Bodenheimer and Fernandez 2005, Stano 1993). With respect to imaging

services in particular, a Medicare Payment Advisory Commission (MedPAC) report in









2006 noted that the geographic variation in utilization of imaging services (coefficient of

variation of 28%) is exceeded only by non-imaging diagnostic tests (coefficient of

variation of 30%) (Medicare Payment Advisory Commission 2006). An earlier study in

2001 found that there was a large variation by state in the number of diagnostic imaging

studies per 1000 Medicare beneficiaries with 10th percentile being 3038 and the 90th

percentile at 4573 (Bhargavan and Sunshine 2005).

As with other medical services that display small area variation, we must consider

if inappropriate imaging accounts for the high-utilization areas and/or if poor access to

services explains the low-use areas (Leape et al. 1990). In 2008, the General

Accounting Office (GAO) published an extensive report about Medicare Part B imaging

services that noted the rapid increase in spending and commented further on the

extensive variation in utilization rates among states (Government Accountability Office

(U.S.GAO) 2008). For example, by 2006, in-office imaging spending per beneficiary

varied almost eight-fold across the states; from $62 in Vermont to $472 in Florida. The

report goes on to state that "Given the magnitude of the differences in imaging use

across geographic areas, variation is more likely due to differences in physician practice

patterns rather than patient health status. Further concerns about the appropriateness

of imaging use are raised by research on geographic variation showing that, in general,

more health care services do not necessarily lead to improved outcomes" (page 21).

Just as scientists leverage observed variability in nature to pry into its inner

workings, health policy makers target variability in utilization to better understand and

control overall cost and quality. After characterizing and quantifying medical practice

variation, focus naturally turns towards explaining it (Folland and Stano 1990, Folland









and Stano 1989). In ambulatory primary care, much of this attention is devoted to two

processes; specialist referral (Franks et al. 2000a, Franks et al. 2000b, Franks et al.

1999) and diagnostic testing (Hartley et al. 1987, Epstein and McNeil 1987, Epstein and

McNeil 1986, Epstein and McNeil 1985b, Epstein and McNeil 1985a). When making

these decisions, providers and patients labor under compound uncertainty about the

presence of diseases) as well as the efficacy and availability of referral and/or

diagnostic testing. The largest barriers to uniform medical practice are lack of evidence,

conflicting/ambiguous results, and incomplete dissemination of existing information

(Eddy 1984, Eddy 1990, Eddy 2007). Primary care has been likened to a jazz

performance where some basic structures and heuristics are in place but--owing to

uncertainty--practitioners must improvise to a large degree (Miller et al. 2001). The next

section addresses clinical uncertainty then turns to the contextual factors that play a

substantial role in determining the amount, timing, and mixture of diagnostic tests

(including imaging) performed in primary care outpatient settings.

Clinical Uncertainty

Any attempt to analyze imaging utilization must first deal with clinical uncertainty.

Economic theory provides valuable insight into understanding the role of uncertainty in

clinical decision making and makes a clear distinction between uncertainty and risk.

Given complete estimates about relevant risks of disease and treatment, patients and

doctors could 'calculate' an appropriate strategy to maximize expected clinical benefit

(Cohen 1996, Wu 1996, Eeckhoudt 1996). Uncertainty, on the other hand, can be

defined as ambiguity or imprecision attending risk estimates that renders calculation of

expected benefit difficult or impossible. Uncertainty can attach to questions about what

(if any) disease the patient has and/or the relative benefit of available treatments for the









diseases) (Eddy 1984, Wennberg et al. 1982). Given uncertainty about disease

probability and optimal treatments, practice variations can occur for several reasons

(e.g., physician practice style, patient attitudes, organizational influences) and these

lead to observed variations in health service utilization. Thus, the evidence-based

medicine (EBM) movement is predicated on helping physicians and patients to use

medical research findings to reduce their collective uncertainty wherever possible. The

hope is to decrease knowledge gaps and information asymmetry about health care

using various decision support tools to allow patients and doctors acting as their agents

to make better and more consistent clinical decisions.

Assuming that there is a scientific basis for allopathic medicine (i.e., EBM) implies

that many diagnoses can be established and optimal treatments known with at least

some certainty. Therefore, some of the variation in clinical activity must be due to

random differences in disease incidence and prevalence in the population of interest.

Routine--evidence based--care of these patients would be expected utilize different

amounts and mixtures of services to meet disparate clinical needs. The amount of

variation in health service utilization that cannot be explained and justified on the basis

of existing and incident disease burden is considered to be unwarranted (Wennberg

2004a, Wennberg 2004b). The method by which expected and unwarranted variation in

utilization are parsed from each other involves risk adjustment which seeks to account

for the effect of known disease burden on utilization.

Risk Adjustment

Case-mix and risk adjustment methods have been developed primarily to support

provider profiling of utilization or patient outcomes (Chang and McCracken 1996,

Greene et al. 1996, Salem-Schatz et al. 1994, Tucker et al. 1996, Welch et al. 1994).









Consider, for example, the subject of this dissertation: utilization of non-screening

diagnostic imaging tests by primary care providers. A payer may calculate rates of

utilization and seek to remediate doctors that are high end outliers in order to decrease

overall costs. If raw rates of images per patient year are used in this way, targeted

providers will--often correctly--object that the reason for their high use is that the

patients they see are sicker. This situation clearly calls for case mix adjustment to be

'fair' to targeted doctors, hold patients harmless, and for scientific rigor.

In addition to patient age and gender, outpatient risk adjustment should include

variables that capture relevant medical conditions and events. The prevailing method for

obtaining such clinical variables is to convert administrative data (claims and/or

prescriptions) into problem type and severity categories. The best known example in the

U.S., is Ambulatory Care Groups (ACG), a proprietary method developed at Johns

Hopkins. Outside the U.S., the most popular method for categorizing primary care case

mix is the International Classification of Primary Care (ICPC). Davis used ICPC for case

mix adjustment when studying provider variation in primary care practice activity

(prescriptions and tests) in Australia (Davis et al. 2000a, Davis et al. 2002). However,

there is little published literature that deals specifically with case mix or risk adjustment

for outpatient primary care imaging utilization. One relevant paper deals with

'ambulatory test' utilization (including chest x-ray) by general internists for patients with

hypertension (Epstein and McNeil 1985b). Another looked at 'diagnostic services',

which included radiology, among generalist and specialists caring for Medicaid patients

in community practice (Eisenberg and Nicklin 1981). Both used complex case mix/risk

adjustment schemes to account for differences in patient demographics and illness









burden. There is ample literature about case mix (or risk) adjustment in support of

performance measurement, resource use comparison, and practice profiling in primary

care settings. However, imaging is not the specific focus of these more general methods

descriptions (Chang and McCracken 1996, Greene et al. 1996, Salem-Schatz et al.

1994, Tucker et al. 1996, Majeed et al. 2001 b, Majeed et al. 2001a).

Appropriateness and Supplier-Induced Demand

There are several competing theories regarding the causes of medical practice

variation that remains after factoring out underlying patient disease burden (i.e., case

mix adjustment). An early explanatory model for clinical activity differences depended

on categorizing health care utilization events by their 'appropriateness'(Wennberg et al.

1982, Wennberg 1987b). The attractiveness of being able to correlate high-intensity

utilization with inappropriate services was enhanced by the implication that expenditures

could be reduced by targeting high use regions, practices or providers. However,

despite considerable effort and expense, no convincing evidence has emerged to show

any direct connection between the rate of various types of health service utilization and

expert consensus ratings about the appropriateness of care in both large and small

regions (Chassin et al. 1987, Leape et al. 1990, Casparie 1996, Restuccia et al. 1996).

When focus is narrowed somewhat to primary care providers and their rates of

specialist referrals, still no relationship with appropriateness has been demonstrated

(Fertig et al. 1993, Knottnerus et al. 1990).

Another theory, termed Supplier-induced Demand posits that practice variations

stem from differences in financial benefits to the supplier (i.e., physician). However,

Reinhardt is persuasive in arguing that the theory of supplier induced demand for









inappropriate services fails empirical tests and that a more nuanced explanation is

preferable (Reinhardt 1999). He argues for a preferred practice style that:

would be an amalgam of (1) what the physician has been taught to view as
best medical practice in medical school and during residency training, (2)
his or her subsequent refinement of the received doctrine on the basis of
more recent literature and continuing medical education, and (3) an
adaptation to the dominant professional norms in a given locality.
(Reinhardt 1999)

That being said, Reinhardt does not deny that financial incentives may influence

the "central tendency" of practice patterns over time (Reinhardt 1999). Davis has

studied primary care doctors in New Zealand and concludes that a 'supply hypothesis'

is not useful and that individual physician's patterns of test ordering and referral form a

'practice style' that persists over time (Davis et al. 2000b, Davis et al. 2000a, Davis et

al. 2002). The practice style theory has gained considerable traction and is commonly

cited to explain a broad range of variations in medical care delivery (Folland and Stano

1989, Grytten and Sorensen 2003, Welch et al. 1993, Wennberg et al. 1997, Sirovich et

al. 2008).

There are scholars of practice variation who argue that the simplistic picture of

clinical uncertainty allowing individual practice styles to emerge is theoretically

incomplete and does not explain geographic variations because purely personal

differences between providers should average out (Stano 1991, Stano 1993, O'Neill and

Kuder 2005). However, in some settings, variation in imaging utilization rates persists

even when there is clinical certainty. For example, there is strong consensus about

breast cancer screening intervals for women over 49 years old and considerable

controversy about younger women. In spite of this, small area variations in

mammography rates in Ontario are similar across patient age groups (Goel et al. 1997).









Westert believes that 'practice style' is, at best, shorthand for a cluster of influences

acting at the patient, community, and physician level to affect decision making (Westert

and Groenewegen 1999). In this more granular and complex model, individual provider

opportunities, incentives, and influences combine with shared clinical standards in the

group and local medical community to shape their practice. These patterns can be

broadly described as ranging between conservative and elaborate which translate into

lower and higher use rates for diagnostic tests respectively.

Summary

This dissertation will articulate and quantify factors driving outpatient primary care

diagnostic imaging utilization. Further, the study will examine the relative contributions

of case mix/clinical need versus contextual/practice style factors to the variation in

imaging utilization. Most of the literature about primary care practice variation in test

ordering and referral relies on an assumption--sometimes unstated--of proper risk

adjustment. To the extent that health service use variations are explained by differences

in patient demographics and clinical variables, they should be of less interest to health

services researchers (Diehr et al. 1990, Cain and Diehr 1992) than to epidemiologists.

This epidemiologic/contextual distinction is crucial, even in primary care settings.

However, there are relatively few empirical estimates of the relative contributions of

these to overall variance. Grytten studied use of diagnostic tests among Norwegian

primary care providers (Grytten and Sorensen 2003). The reasons for each visit as well

as patient age and gender were used for model-based risk adjustment (clinical need) of

diagnostic test expenditures on a per visit basis. The remaining variation ranged

between 47-66%, was attributed to practice style, and seemed to be a 'sticky' attribute

that followed individual providers who changed practice locations during the study.









Other authors have looked at ambulatory primary care practices and estimated the

variation in expenditure per patient from clinical need (case mix adjustment) to be about

60% (Phelps et al. 1994, Davis et al. 2000b).

The next chapter summarizes the research on drivers of resource utilization

(emphasizing diagnostic tests and referrals) by primary care providers, then presents a

conceptual framework for analyzing imaging utilization in particular. Factors related to

clinical need will be addressed first, followed by consideration of the various contextual

factors influencing imaging utilization including patient, provider, and practice factors.









CHAPTER 4
CONCEPTUAL FRAMEWORK

This dissertation examines the utilization of outpatient diagnostic imaging by

primary care doctors caring for adult patients. Because the study will not consider

imaging that occurs as part of disease screening programs, it does not include imaging

tests that might be scheduled and performed 'routinely' without an explicit decision by

the doctor. Specifically, this study examines diagnostic imaging tests ordered by a

patient's primary care doctor to address clinical issues raised either during a visit, a

phone call (or email), or another health system encounter (ER visit, hospitalization).

This chapter presents a conceptual model to account for relevant driving and modifying

factors influencing the amount of such--non-screening--imaging performed by order of

the primary care doctor on a patient.

As articulated in the literature review, drivers and modifying factors of primary care

resource utilization can be divided into two major classes. The first one is clinical need.

For a given patient, clinical need arises from existing or developing signs, symptoms,

trauma, and illness. The second class of contextual factors can be further grouped by

attribution to the agents or organizations involved. Specifically, most contextual factors

will accrue to either the patient or the doctor. Other general groupings for contextual

factors include the clinic or practice in which the doctor sees the patient and any larger

provider organization (e.g., academic faculty practice, HMO, PPO, and etc). Another

category to be considered (at least in the U.S.) relates to the payer or insurer. Other

factors relate to the structure and process of outpatient imaging facilities and methods

by which imaging tests are ordered. Finally, both patient and doctor live and work in

communities that can be represented at various levels of aggregation (e.g., city, county,









state) and the phenomenon of geographic variation implies that at least some factors

operate at the community level.

This chapter discusses the relevant driving and modifying factors of outpatient

diagnostic imaging utilization in primary care as defined above. For each individual

factor, expectations of the direction and strength of effect it should have on imaging

utilization are justified by theory and informed by existing empiric literature. The final

section summarizes and diagrams the general relationships among clinical need,

context, and imaging utilization.

Clinical Need

This set of factors is perhaps the easiest to state and comprehend and yet is the

hardest to measure and model. In essence, imaging tests are done to address clinical

uncertainty that arises about an existing condition or a clinical event. Existing conditions

are diagnoses made by a physician and therefore known to them. Uncertainty about an

existing diagnosis relates either to current stage/status of the disease or treatment

choice/response for that condition. Clinical events are defined by development or

worsening of a sign, symptom, or abnormal test results. The uncertainty engendered by

such clinical events comes from the doctor's need to determine if a new diagnosis

needs to be made or if an existing disease is responsible. In either case, the function of

the diagnostic imaging test is to better identify or exclude treatable disease to guide

therapeutic decisions. The only other clinical event not directly accounted for in the

preceding explication is trauma, for which imaging is often performed to assess severity

and type of injury. Severe trauma is treated in acute care hospitals and associated

imaging tests would not be counted as primary care outpatient utilization. On later

ambulatory care visits, the 'post traumatic' state will be identified as an existing









condition for this model and outpatient imaging tests might be done to address residual

problems.

As described previously, under ideal conditions of evidence-based practice, the

mixture and amount of imaging that a patient received in a year would depend only on

their existing diagnoses, disease status, and clinical events. However, the current state

of medical knowledge and the consistency with which it is applied in actual practice is

such that doctor's decisions about intensity and mixture of diagnostic imaging are quite

variable even when faced with identical clinical scenarios. This variability in diagnostic

decision-making 'styles' among physicians is enabled by systematic uncertainty about

what test(s)--if any--are suitable for various clinical scenarios. However, uncertainty or

ignorance about the appropriateness (expected clinical benefit) of imaging is necessary

but not sufficient for variability in utilization to occur. The contextual factors discussed

below, can influence utilization in the face of uncertainty or ignorance about the optimal

diagnostic strategy.

For the most part, clinical need variables operate exclusively at the level of

individual patients. The only exceptions to this might be in disease screening programs,

communicable illness, or environmental health issues. However, this study excludes

imaging related to disease screening programs. Furthermore, in the setting for this

study (primary care rendered in a large Northeastern metropolitan area from July 2007

through June 2009) no unusual communicable disease epidemics or environmental

health issues occurred which might distort the assumption about clinical needs working

purely at patient level.









Context: Patient

Basic patient demographics such as gender and age strongly correlate with the

amount and type of clinical need based on their complex relationships with many

diseases and health states. These two variables are often included in case-mix or risk

adjustment models. They serve as proxies for an individual's propensity to develop

health conditions and biological responses to disease, testing and treatment. However,

these same demographic factors also exert social and psychological effects on the

patient's likelihood to seek or accept diagnostic testing under various scenarios. For

example, it is conceivable that men and women of similar age might choose differently

in some systematic way about diagnostic testing for the same clinical scenario based on

level of anxiety related to attitudes about risk and uncertainty.

There is some empiric evidence about patient preferences as related to clinical

resource use in general as revealed by surveys, interviews, and responses to

hypothetical scenarios. Anthony et. al., found that elderly Medicare beneficiaries

expressed substantial differences in their preferences for seeing a doctor right away,

having tests, and for specialist care (Anthony et al. 2009). When individual Medicare

utilization was modeled with these preferences as predictors (along with demographic

control variables), those who preferred care right away and from specialists had higher

overall healthcare utilization rates. However, at larger levels of aggregation (regional

variation) differences in patient preference were uniformly distributed and did not

explain variations in cost.

Socioeconomic status (SES) factors such as level of education, income, and

ethnicity affect individual patient tendencies to seek care and comply with provider

recommendations. Empiric evidence for differences in healthcare utilization and patient









outcomes abounds in the 'disparities' literature, although separating ethnicity from

economic factors as causes of such disparities is a matter of considerable debate.

Confounded with patient's preferences based on socio-cultural characteristics are the

doctor's own biases and attitudes. Perhaps the best known (and controversial) example

of this is Schulman's survey about recommendation for cardiac catheterization

(Schulman et al. 1999). He found that, in hypothetical clinical scenarios, patient race

and gender influenced physician's tendency to recommend diagnostic work up for

otherwise identical presentations of chest pain. In one intriguing study of managed care

claims, Franks found that case mix adjusted use of diagnostic testing was actually

higher in patients from lower SES zip codes (Franks and Fiscella 2002). He

hypothesized that doctors tend to order more tests when they perceive that patients

cannot articulate their histories and current symptoms. However, in general, patients in

lower SES and minority ethnicities (in the U.S.) tend to receive lower levels of diagnostic

testing in acute and sub-acute care settings (Goldstein et al. 2006, Isaacs et al. 2004,

Pezzin et al. 2007, Quintana et al. 1997).

Context: Physician

With provider factors, it is important to remember that we are limiting the scope of

discussion to outpatient adult primary care. Primary care doctors are trained in several

distinct ways in the U.S., and they may choose different types of patients to see.

Doctors trained in family medicine often see children and pregnant women in addition to

adult patients. On the other hand, geriatrics-trained physicians tend to take care of

elderly patients. Many primary care internists have additional training after their 3 years

of internal medicine. For example, doctors with some endocrinology training after their

internal medicine residency might skew their practice towards adult diabetic patients









even though they are not rendering 'specialty' care. However, assuming that patient-

level case-mix variables are accounted for, we focus on how factors like experience,

gender, training and specialization might affect a doctor's tendency to order diagnostic

imaging in similar clinical scenarios arising in adult primary care.

Training and specialization have direct effects on test-ordering behavior in primary

care. Although, all U.S. physicians complete at least a four-year course of study leading

to an M.D. degree, specific courses about imaging use are rarely offered or required in

M.D. curricula, with the majority of training and experience about radiology gained

during residency. Since post-graduate medical education is conducted in

'apprenticeship' models, the relative intensity of imaging utilization at the training

institution strongly influences subsequent decision making about imaging during

practice. Small area variations may be relevant here because in many health care

referral regions, the academic medical centers account for and may influence much of

the measured utilization. Even within geographic regions, residency training occurs in

different types of institutions and community settings. There is wide disparity in these

with a spectrum ranging from residencies conducted in small non-affiliated rural

community health centers to the classic tertiary care safety net academic health center

owned by a medical school operating in a large city. The hypothesis is that doctors

trained in high use regions and at large academic centers will tend to order diagnostic

imaging more frequently than those trained in smaller centers and low use regions

(Chassin 1993, Eisenberg 1986a, Epstein and McNeil 1985a, Folland and Stano 1990,

Grytten and Sorensen 2003, Landon et al. 2001, O'Neill and Kuder 2005).









Experience is partly defined by length of time in practice and is also directly related

to when training took place; the two factors may be very hard to separate. Timing of

training is particularly relevant to imaging utilization because the technology has

advanced and evolved rapidly and consistently over the past 2-3 decades. The

expectation is that a recent graduate who routinely worked with advanced imaging

techniques in training will be more likely to order MR and CT in subsequent practice

than a doctor trained prior to their diffusion who may not be aware of what is available.

Length and extent of experience itself affects diagnostic decision making with doctors

having greater experience tending to order less imaging tests in identical scenarios

(Bugter-Maessen et al. 1996, Childs and Hunter 1972, Couchman et al. 2004,

Couchman et al. 2005, Eisenberg and Nicklin 1981, Sood et al. 2007, Whiting et al.

2007, Williams et al. 1982).

Physician gender also may have an effect on use of imaging independent of

experience and training, although, with the recent and substantial increase in the

fraction of women trained in and practicing medicine in the U.S., it may be difficult to

measure since women physicians tend to be younger and trained later. Empiric data

about physician gender has been mixed with studies showing both increased and

decreased tendency to order imaging and other diagnostics tests between male and

female doctors in primary care (Britt et al. 1996, Rosen et al. 1997, Sood et al. 2007).

Physician workload may influence tendency to order imaging tests in several

ways; both in the long run (months and years) and the short run (during the course of a

day or week). This effect is mediated through each physician's perception of time

pressure overall and during a particular visit. For example, an initial visit to a primary









care physician for headache occurring at the end of a busy day in clinic might be more

likely to include an imaging test than otherwise. Such a doctor, caught between time

constraint and fear about missing a diagnosis might order an MRI of the head instead of

spending an extra 15 minutes doing a detailed neurological exam. On the other hand,

the same doctor may choose to refer the patient to a specialist (neurologist in the case

of headache) rather than order any imaging if they believe that strategy would move the

patient towards a temporary disposition more rapidly. There is very little empiric

evidence about this particular factor to guide us in determining the sign of a possible

correlation between practice workload and imaging intensity.

Economic factors may influence imaging test ordering (Reinhardt 1999). Aside

from pure income maximization, physicians may seek to increase their personal utility in

other ways. Eisenberg refers to this as 'physician as self-fulfilling practitioner'

(Eisenberg 1986a, Eisenberg 1985). Also, in acting as the patient's agent, physicians

may take the patient's financial situation into account when making decisions on their

behalf (Eisenberg 1986b). However, imaging utilization also may be influenced by

physicians seeking to directly increase their income by ordering and performing imaging

tests. In such cases, the reimbursement for the imaging test itself is paid to the ordering

physician through several pathways including ownership stake in the imaging

equipment (technical fee) and/or interpretation of the examination (professional fee).

Called 'self-referral', there is a large body of literature about its practice and

ramifications. Hillman and others make a strong empiric and economic case that such

direct financial incentives have powerful positive effects on imaging utilization volume

and charges (Hillman et al. 1990, Hillman et al. 1992, Hillman 2004, Gazelle et al.









2007). Although the Starke laws have been in effect since the 1990s and have been

renewed and revised at least once, recent evidence from California shows that there is

still substantial self-referral of advanced diagnostic imaging and that various

mechanisms other than direct ownership of equipment allow this arrangement to

continue under current statute (Mitchell 2007).

Context: Malpractice

Malpractice deserves special mention because it operates at the individual

provider level in addition to the practice and community levels. The term 'defensive

medicine' is often used to describe the phenomenon of doctors' increased ordering of

imaging--and other diagnostic--tests based on fear of being sued for failure to diagnose

(Kessler and McClellan 2002, Sood et al. 2007, DeKay and Asch 1998). Personal

experience with being sued for malpractice can have profound effects on an individual

doctor's psychology and practice pattern that may persist for years or decades. By all

accounts, it is an extremely negative and unsettling event that induces a strong desire

to avoid repeat occurrences (Hermer and Brody 2010). Thus, if a doctor is sued for

missing a diagnosis, the expectation is that they will alter future practice toward more

diagnostic testing. This behavior will likely not be limited to the scenario leading to the

suit, but generalized across patients of different clinical classes and types of diagnostic

tests (including imaging). Even if a doctor is sued for malpractice unrelated to a

diagnostic error, he or she may tend to general defensiveness which may lead to

greater use of diagnostic imaging at lower levels of uncertainty than before the event.

Dekay and Asch wrote a seminal paper using expected utility theory combined

with decision analytic modeling to show the causes and consequences of malpractice

liability on diagnostic testing (DeKay and Asch 1998). They proved that consideration of









liability by physicians faced with a classic observe, treat, test choice set must widen the

zone (over disease probabilities) in which testing is the preferred strategy. They also

prove that there is an obligate utility loss to patients incurred by this extra testing. They

also assert that physicians substantially overestimate the 'protection' afforded them by

doing more testing. In retrospect, physicians generally overestimate their ability to have

made the correct diagnosis in advance. Thus, hindsight and regret bias combine with

unrealistic expectations about the efficacy of imaging and result in a near magical belief

that the right test would have 'saved the day' (DeKay and Asch 1998).

On a state by state basis, medico-legal 'climate' varies considerably based partly

on the statutory and precedent-based status of malpractice and tort laws. Baiker,

Fisher, and Chandra's paper examining trends in Medicare costs and malpractice

burden in the U.S. over the 1990s used states as the unit of measure (Baicker et al.

2007). They showed that imaging cost increases were significantly correlated with

trends in malpractice premium and payouts. Across the 50 states, a 10% increase in

malpractice premiums/payouts resulted in about two percent increase in physician

services costs. They estimated that the observed 10 year increase in malpractice of

60% resulted in more than 15B extra in spending with imaging being the largest

contributor by far (Baicker et al. 2007). It should be noted that there are diverging views

with more recent papers questioning the empiric basis for a large effect by defensive

medicine and suggesting that even comprehensive tort reform might not have much

actual effect on health costs (Hermer and Brody 2010, Sloan and Shadle 2009).

The theory of how liability concerns increase diagnostic testing at all levels of

aggregation rests on the assumption that providers perceive themselves to be at risk for









malpractice action even if they have personally not been sued before (Fenn et al. 2007,

Kessler et al. 2006). Clinicians are rather bad at assessing their own liability risk and

tend to overestimate personal probability of being sued (Holtgrave et al. 1991, Lawthers

et al. 1992, Kessler and McClellan 2002, DeKay and Asch 1998). This study will

examine a large primary care practice confined to a single institution. Therefore, the

local and state malpractice 'climate' is constant though the 'free floating' fear of medical

liability might vary by practice. In the current study, the only variable available to directly

probe the effect of 'defensive medicine' on imaging utilization is each physician's history

of being sued or not during the preceding decade.

Context: Practice Organization

After limiting consideration to adult outpatient primary care in the U.S., there are

several types of practice setting and organizational dimensions to be considered.

Perhaps the most relevant is the employment arrangement for the physicians. Health

care organizations structured as staff models, where doctors are salaried (e.g.,

traditional HMO, academic health centers, military, and VA) may have different patterns

of diagnostic imaging utilization based on individual incentives and medical

management initiatives than private practice and fee-for-service settings (Epstein and

McNeil 1985a, Kravitz and Greenfield 1995). Even if we exclude consideration of direct

financial benefit from self-referral of imaging, independent practitioners and groups are

generally less constrained in their ability to order radiology tests.

Aside from employment structures and compensation arrangements, primary care

physician practices differ in the extent and manner in which peer pressure is exerted.

For example, in a small private practice primary care group there may be minimal (if

any) formal influence on actual practice styles among members, including diagnostic









radiology utilization. At the other extreme, in some staff model practices, leadership may

routinely profile imaging utilization at the provider level and seek to control it with direct

incentives or remedial measures (Neilson et al. 2004, Axt-Adam et al. 1993, Solomon et

al. 1998).

Context: Payer and Prices

In the U.S., a patient's insurance status has a profound impact on access to

primary care services and may influence the frequency of outpatient visits. In this

conceptual model, non screening imaging tests are ordered to address issues identified

during a patient-physician encounter, thus a lower frequency of encounters provides

fewer opportunities for imaging to be ordered. Additionally, greater financial burden

(self-pay or high co-pay/deductibles) associated with imaging tests will reduce a

patient's tendency to agree to and/or undergo expensive imaging tests, even if ordered.

Physicians may be aware of a patient's financial or insurance status and in their role as

financial agents, might choose to forgo imaging tests depending on costs (Mort et al.

1996, Shen et al. 2004, Pham et al. 2007). There is a substantial literature concerning

awareness of diagnostic test costs. A recent systematic review by Allen concluded that

most doctors have a very limited understanding of diagnostic and non-drug therapeutic

costs (Allan and Lexchin 2008). Sood's more focused review of literature about multiple

contextual factors in test ordering tendencies found that cost awareness (among both

doctors and patients) was relevant (Sood et al. 2007). In general, when price

information is made available to clinicians, they tend to reduce their likelihood to order

diagnostic tests (Hoey et al. 1982, Cummings et al. 1982, Long et al. 1983). Bates

reported a 5% decrease in clinical laboratory test charges during inpatient episodes









after price information was routinely displayed during electronic order entry (Bates et al.

1997).

Increasingly, various payers (including Medicare), being aware of the rising costs

of outpatient imaging tests, have begun to employ cost-containment measures

specifically related to imaging (Government Accountability Office (U.S.GAO) 2008). One

strategy is to profile individual physicians with respect to imaging (and other resource)

utilization and place them into various 'tiers' that give preference to 'efficient' providers

in various ways. An emerging trend is for payers to contract with one of several imaging

benefits management entities (e.g., National Imaging Associates, CareCore National,

and others). These companies serve as 'gatekeepers' for outpatient diagnostic imaging

by requiring providers and/or patients to obtain pre-authorization on a case by case

basis before tests are scheduled (Otero et al. 2006, Brant-Zawadzki 1994, Bernardy et

al. 2009). In addition to simple barrier effects mediated by call center and other

administrative delays, requests for imaging tests may be denied based on proprietary

medical necessity or 'appropriateness' rules. Such arrangements can have considerable

impact on the likelihood that a given patient-doctor encounter will result in a scheduled

and completed diagnostic imaging test (Blachar et al. 2006, Levy et al. 2006, Smulowitz

et al. 2009).

Context: Access to Imaging

A final category of contextual factor relates to the facilities and processes that

underlie how diagnostic imaging examinations are ordered, authorized, scheduled,

performed, and interpreted. As described in the Background, the so-called 'radiology

round trip' is a complex chain of events that begins with a doctor-patient interaction of

some kind that raises a clinical question that might be answered by imaging. In the case









of a completed examination, the 'round trip' generally ends when an interpretative report

about the imaging test gets read by the doctor, acted upon, and relayed to the patient.

The availability of diagnostic imaging facilities in terms of proximity and capacity

influence the doctor's tendency to order and the patient's ability to obtain examinations,

even after omitting any consideration of testing facilities owned or operated by the

referring physician (self referral). Nonetheless, ready availability and ease of scheduling

for various tests will positively influence decision making about imaging by both doctor

and patient. The means by which tests are ordered and scheduled by the doctor, office

staff, and the patient can introduce barrier or enabling effects. For example, changing

from written or verbal orders to a system that requires doctors to log on to a computer

and order the test personally, may exert barrier effects if doctors believe that more of

their time and effort is required to assert the order. On the other hand, a robust

computerized point of care radiology scheduling system can allow patients to leave the

clinic with their radiology appointment in hand and will increase utilization by virtue of

convenience.

Patient experience at the diagnostic imaging facility may affect compliance with

imaging orders as well as a doctor's tendency to order in the first place. Long wait times

and other negative experiences at the testing facility will become known to the doctor

and other patients over time. If doctor and patient are contemplating a diagnostic

imaging test, expected difficulty in scheduling and/or long waiting times on the day of

examination may be perceived as 'too much trouble' and bother. Radiologist training,

skill and style will affect how they interpret any given test and this is manifest in the

report that gets sent back to the referring doctor. If reports tend to be late in arriving,









raise more questions than they answer, and frequently contain recommendations for

further testing, doctors may come to rely less on diagnostic imaging. A relatively new

development is how the resulting images are handled and distributed to patients and

referring doctors. In modern computerized radiology practices, patients are given a CD

with all the images on them which can be brought back to the referring doctor for

review. Increasingly, images can be viewed on line by the referring doctor along with

reports. Imaging providers offer these and other services to increase their market share.

Summary

The final decision of whether or not to order and undergo an imaging test thus

depends on all these factors and the complex interactions among them. As shown in

Figure 4-1, the ideal level of utilization is determined by clinical need under conditions of

certainty. Adding clinical uncertainty then allows for deviation from the ideal level of

utilization, with the variation potentially being positive or negative. The various

contextual factors add additional variation in the observed level of utilization that

persists even after case-mix and risk adjustment.

For purposes of the empirical analysis, the conceptual framework can be

summarized as: Imaging utilization = f[clinical need, patient factors, physician factors,

malpractice environment, practice organization, payer, access to imaging]

The next chapter will describe the study setting, data sources, and variables that

will be used in the analysis.









Clinical Events Necessary Imaging
Co-Morbidities [known, large benefit]
Demographics -+
3 1 7 -- -- -- -- -- -
Discretionary Imaging
Evidence Base [uncertain benefit] Measur
Evident ------------ --------------------------- Measured
Imaging
Patient Utilization
Patient
Physician
Practice
Malpractice Contingent
*Structure '2 Imaging
*Process
*Payer
*Cost
*Of Imaging

Figure 4-1. Summary diagram of conceptual model for outpatient imaging utilization in
primary care.









CHAPTER 5
SETTING, DATA SOURCES, AND VARIABLES

This chapter will describe the institutional setting for this study and then focus on

outpatient diagnostic imaging services. The general setting includes a large group of

primary care physicians practicing in 15 separate locations and a hospital based

radiology department with several service sites.

Patients in the practice who identify with a single attending physician as their

primary care provider are tracked as a 'loyalty cohort'. The majority of outpatient

imaging performed on these patients occurs at the associated radiology department.

Radiology databases provide counts of outpatient diagnostic imaging tests accruing to

patients, doctors, and clinics. Other clinical and administrative databases provide

information on patient demographics, clinical problems, medical activity (visits,

hospitalizations, etc), and physician characteristics.

Settings

Primary Care Practice

This study was conducted at Massachusetts General Hospital (MGH) and the

associated Physician Organization (MGPO). In close association with the MGH

academic medical center, MGPO is a large multi-specialty faculty group practice with a

full complement of adult primary and specialty care. The full time faculty are salaried

employees of MGH and the non-radiologists (e.g., primary care providers) have no

direct financial incentives relating to volume and/or revenues from imaging or laboratory

tests that they order. Malpractice insurance is supplied by the MGPO under a self-

insurance pool arrangement and the providers have limited personal liability. However,









malpractice actions are identified at the individual provider level for reporting to licensing

authorities and government data-banks.

This work focuses on the primary care portion of the practice and was aided by an

entity within the MGPO called the Primary Care Operations Improvement (PCOI) group.

Physician leaders and staff members in PCOI conduct analyses of various aspects of

the practice for internal quality assessment and improvement as well as for presentation

and publication in scholarly settings. In support of these efforts, Dr. Steve Atlas (a

primary care physician and health services researcher at MGH) and colleagues have

devised and validated a method for identifying a group of primary care patients, doctors,

and clinics with stable relationships to each other--the result is termed a loyalty cohort.

A cohort is identified by year and comprises a list of patients who are considered to be

'loyal' to a single primary care provider by virtue of their outpatient visits as documented

in the electronic medical record over the three years ending in the 'cohort' year. Loyalty

assertions for a given patient-doctor pair are calculated probabilistically using five

variables derived from visit data (Atlas et al. 2006, Lasko et al. 2006, Wasiak et al.

2008, Atlas et al. 2009). These are listed below.

Waiting fraction: the total number of days waited for appointments with the given
physician, divided by the total waited for all physicians combined.

Visit difference: the total number of visits that a patient has made to the given
physician minus the total to all other physicians combined.

Days since last visit: the number of days since the last visit to the given
physician.

Future difference: the total number of appointments scheduled for future visits
with the given physician, minus the total for all other physicians combined.

Idle ratio: the number of days since the last visit to the given physician, divided
by the number of days since the first visit.









A logistic model was developed and validated on subsets using loyalty assertions

as stated by the linked primary care physicians (Lasko et al. 2006). In comparison with

this stated (gold-standard) loyalty status, model predictions were over 95% accurate.

When the same technique was applied to all registered primary care patients, the study

(loyalty) cohort described below was the result.

Outpatient Radiology

The Radiology Department at MGH provides a full range of imaging services for

inpatient, emergency room, and outpatient practices. The main department is located at

the MGH campus with several ancillary outpatient sites in the Boston area. The whole

department is linked via a robust electronic infrastructure and the radiology informatics

group is widely recognized as being among the most advanced and sophisticated in the

world. The relevance for this study is two-fold. First, electronic records of all imaging

tests are housed in a data warehouse that has been created and maintained with great

care and attention to detail. This means that all imaging tests going back to at least

1995 are listed with complete and accurate information about several items relevant to

studying outpatient imaging utilization, including the identity of the doctor ordering the

test, dates of ordering/completion, modality, body area, and patient status at time of

examination (e.g., inpatient, outpatient). In total, almost 100 items of information are

stored about each test with the ones just mentioned being most relevant to the present

study. Secondly, outpatient radiology ordering by all primary care physicians at MGPO

is performed through the same web-based system. This is called Radiology Order Entry

(ROE) and it has been in full use for all modalities since 2004. After selecting from a

dynamic menu of all outpatient radiology exams, clinicians are required to input









structured information about why the test is being ordered via checkboxes for

signs/symptoms/diagnoses supplemented by a free text input field.

A popular feature of the ROE system is a patient scheduling module. This allows

an appointment for the imaging test to be made at the point of order (doctor's office)

without multiple phone calls, faxes, or other efforts. In addition, starting in 2005, a

decision support (DS) component was added to the system that is triggered with all

orders for CT, MRI, and nuclear medicine studies. The DS logic displays a 1-9 'utility'

score based on the test chosen, patient demographics, and the reasons given for the

test. The scores are grouped as follows: 1-3=Red/low, 4-6=Yellow/intermediate, and 7-

9=Green/high. A 'red' score does not preclude going ahead and ordering the test but the

clinician must provide a reason for doing so prior to proceeding with scheduling. Several

papers have been published about the ROE-DS system with the most relevant

describing the effect on total outpatient CT, MR, and ultrasound volumes at MGH

(Rosenthal et al. 2006, Sistrom et al. 2009). We found that, after correction for overall

practice activity, there were substantial reductions in growth rates, especially for CT

scans (Sistrom et al. 2009). It should be noted that during the time period (July 2007-

June 2009) covered by this study of outpatient imaging utilization in the MGPO primary

care practice, the ROE-DS system had been in use for at least two years. Further, no

substantial changes were made to the system functionality and only minor alterations

were made to the DS scores.

Data Sources

Loyalty Cohort

As described above, the loyalty cohorts are compiled by Dr. Atlas and the PCOI

staff every year and the one used in this study is for 2008. This means that MGPO









primary care patients active in 2008 were gathered and their electronic clinic scheduling

records going back through 2006 were compiled and analyzed. There were 139,609

unique patients who were candidates for 'loyalty' status. Based on the algorithm, 87,568

were flagged as being loyal to a single primary care provider in the MGPO. There were

804 patients who were loyal to 26 providers with less than 100 loyal patients in their

practices. These were excluded leaving 86,764 patients. Of these, 1483 were loyal to

four providers who had left the MGPO in late 2008 or the first or second quarters of

2009 and demographic data were not available for 4 of the remaining patients. The

analytic sample includes 85,277 patients, loyal to one of 148 primary care physicians

who will be characterized below. These physicians practice in one of 15 clinics

distributed through the greater Boston area. It should be noted that the clinics

sometimes do use residents and medical students. However, the ongoing doctor patient

relationship is with the identified primary care physician. In fact, the raison d'etre of the

loyalty cohort methodology is to unambiguously identify this relationship.

Patient Details

MGH and the MGPO have established a common Research Patient Data

Repository (RPDR), which aggregates numerous sources of information into a single

set of databases designed for use in patient-centered clinical epidemiology. These

include billing and encounter data for inpatient, emergency department, and outpatient

services as well as the contents of inpatient hospital information systems (HIS) and

outpatient electronic medical records (EMR). This study was approved by the

Institutional Review Board at MGH under an expedited protocol for analysis of existing

data. Informed consent was not required and was not obtained.









Physician Details

The MGPO credentialing database was used to obtain relevant information

pertaining to the 148 primary care doctors in the study, including gender, birth year,

medical school graduation year, and medical school state (or country for foreign medical

graduates). By definition, all doctors were licensed in the state of Massachusetts and

the publicly available web site for the Massachusetts Board of Medicine was queried to

determine whether or not each doctor had been sued for medical malpractice in the past

10 years.

Imaging Utilization

The 85,277 patient medical record numbers were queried against the radiology

data warehouse to return all diagnostic imaging tests performed during the study

interval (July 1, 2007 through June 30, 2009). The query specifically excluded

interventional procedures (e.g., biopsy, drainage, catheter angiography, embolization,

and vascular stenting among others). Also specifically excluded were mammograms, as

these are almost all related in some way to breast cancer screening. In addition to

patient medical record number, modality, body area, place of service (ER, inpatient,

outpatient), and the unique provider number of the doctor who ordered the exam was

obtained.

The 221,571 diagnostic imaging procedures performed on cohort patients over the

two-year study interval cross tabulated by place of service and modality are shown in

Table 5-1. The 157,463 outpatient diagnostic imaging procedures are cross tabulated

by the class of the ordering doctor and modality in Table 5-2. Finally, the 60,938

diagnostic imaging procedures performed in the outpatient setting and ordered by the









patient's linked (loyal) doctor are cross tabulated by body area imaged and modality in

Table 5-3.

The unit of observation for this study is the patient. The outcome variable is

constructed by aggregating and summing (by patient) the 60,983 outpatient imaging

procedures ordered by the primary care provider to whom the patient was loyal in 2008.

The remaining diagnostic imaging procedures were performed while patients were in

the emergency room (N=34,345), completed while patients were in the hospital

(29,763), or ordered as outpatient by providers other than the patient's loyal doctor

(other primary care=9,833, specialists=86,692). These other categories of imaging

utilization were also aggregated by patient and summed to produce the other patient-

level imaging utilization (independent) variables described below.

Variables

Imaging Utilization (dependent variable)

The main patient level outcome variable is called prvo_cnt and is the count of the

number of outpatient diagnostic imaging tests (CT, MR, NM, PET, X-Ray, US) ordered

by the provider to whom the patient was loyal during the study period (July 1, 2007

through June 30, 2009). Univariate statistics for this variable are summarized in Table

5-4.

Patient Characteristics

There were 35,709 men (41.9%) in the cohort whose ages in 2008 ranged from

17-100 with mean=54.5 years and standard deviation of 15.1 years. There were 49,568

women (58.1%) whose ages in 2008 ranged from 17-103 with mean=53.2 years and

standard deviation of 16.4 years. Patient race was available for all subjects and is

shown in Table 5-5.









Each patient's payer of record (in 2008) was available in the RPDR as obtained

from outpatient billing systems. These were initially categorized into 13 levels (Table 5-

6).

For modeling purposes, the 13 payer levels were collapsed into 6 levels as

follows:

Aetna into Commercial
Harvard Pilgrim Healthcare, Neighborhood Health Plan, and Tufts Health Plan
into Managed
Mass Health Net and Medicaid into State
Free Care and Self Pay into Uninsured

The resulting 6 level insurance payer categories (Table 5-7) will be used for all

subsequent analyses.

Clinical Events

The RPDR was used to obtain counts of various clinical events for each patient.

These were summed over the period from July 1, 2007 through June 30, 2009. The

events counted all occurred at MGH. Hospital activity variables (summarized in Table 5-

8) included visits to the emergency room, inpatient hospital stays, ICU days, and

inpatient observation stays. Observation (short) stays are a special category of

hospitalization where the patient remains in the hospital for less than 24 hours.

Observation stays often occur in concert with an emergency room visit and allow for

extended nursing care without a formal admission. Observation stays also are used for

minor procedures, dialysis, and administration of intravenous medications. Outpatient

visit counts for each patient during the study period (July 1, 2007 through June 30,

2009) were obtained from the RPDR and confined to the 15 CPT codes representing

outpatient office visits. The professional RVU for each of these CPT codes during 2008









was obtained from the CMS website (see Table 5-9). In addition to counting the number

of outpatient visits, the RVU of those visits were also summed to form a separate

variable. The visit counts and summed RVU for visits were stratified by type of doctor

being visited (prv=provider to whom patient was loyal, pcp=other primary care doctor,

spc=specialist). The resulting six variables are summarized in Table 5-10.

Clinical Problems

The EMR systems used by primary care providers to document outpatient care of

all patients in the study cohort allow for recording of a 'problem list' for each patient.

These problems are encoded in one of two ways depending on the clinic. The coding

systems are internal to MGH and crosswalk tables are available to parse the problem

codes into broad categories, including diabetes, hypertension, heart failure, coronary

artery disease, renal failure, cancer, trauma, obesity, and substance abuse. Active

problems not falling into one of these groups were labeled as 'other problem' for

purposes of this study. For each of the major categories, a binary variable was

constructed with value 'true/yes/1' when the patient had at least one active problem

listed in their EMR entries falling into that category. That same variable was assigned

with 'false/no/0' when the patient did not have an active problem asserted falling into the

category in question. The nine binary clinical problem variables are summarized in

Table 5-11.

In the cohort of 85,277 patients, 46,063 (54.0%) had none of the problem

categories listed above, 23,265 (27.3%) had 'yes' for a single category, 10,808 (12.7%)

had two of the problem categories asserted positively, 3,846 (4.5%) had 'yes' for three

categories, 991 (0.30%) had four positive categories, and 304 (0.36%) had 'yes' for five

or more.









The EMR systems also included additional problem assertions for many patients

that did not fall into one of the categories listed above. Examples include depression,

hepatitis, arthritis, and so forth. For each patient, the counts of unmapped problem

codes were placed into a variable of other problems (name=oth_prob). Overall summary

statistics for the count of other problems included median of 6.0, mean of 7.75, and

standard deviation of 6.94. This count of other problems was zero for 5709 patients and

of those, 4941 (5.8% of the whole cohort) had none of the problem categories

enumerated above (e.g., patient's clinical problem list was empty/null).

Outpatient Prescriptions

The EMR systems also serve prescribing and drug reconciliation functions for the

primary care practice. In the cases in which the patients do not have e-prescribing

enabled to their pharmacy and/or drug benefits program, orders for outpatient

medications are still entered into the EMR and printed prescriptions are given to

patients. To enumerate the number of outpatient medications each patient was taking

during 2008, the number of 'active' prescriptions was counted starting from the first

available entry for any given patient. The queries did not count refills of the same drug

and dose as new prescriptions. However, switches within a drug class and/or dose

changes were counted as new prescriptions which could result in over-counting and

rendering small differences in the discrete number less meaningful than the general

amount each patient was taking. Therefore, we stratified the count of active outpatient

medications into four categories (variable name=meds_cat) summarized in Table 5-12.

Other Imaging Utilization

Counts of diagnostic imaging tests performed on each patient during the study

interval (July 1, 2007 through June 30, 2009) were stratified by place of service (i.e.,









patient status/location at the time the test was performed). These include emergency

department, inpatient, and outpatient (Table 5-1). Outpatient exams were further

stratified by the category of physician ordering them; primary care physician (other than

the patient's own loyal doctor), and specialist (Table 5-2). Note that the counts of

outpatient diagnostic imaging tests ordered by the patients linked (loyal) provider is the

outcome variable which has already been described above. Summary statistics for the

four strata of (non-outcome) imaging utilization are listed in Table 5-13.

Physician Characteristics

The cohort of 85,277 patients were 'loyal to' 148 primary care physicians; 76

(51.3%) women and 72 (48.7%) men. A variable called prov_age_08 was constructed

using each doctor's birth year. The ages of the male physicians (in 2008) had

minimum=33, maximum=75, mean=49.7, and standard deviation=10.0 years

respectively. The female physicians ages (in 2008) had minimum=31, maximum=63,

mean=45.4, and standard deviation=8.4 years respectively. The year of medical school

graduation was used to construct a variable called prov_exp_08 that quantifies the

number of years between medical school graduation and 2008. This (prov_exp_08) had

minimum=5, maximum=50, mean=19.6, and standard deviation=9.6 years respectively.

Perhaps a better proxy for physician experience would have been years since

completion of residency. However, the credentials database had incomplete and

inconsistent information in this regard.

The number of (loyal and non-loyal) patients in the cohort linked to each doctor

was summed into a variable called prov_pat_count and it had minimum=172,

maximum=2394, mean=801.3, standard deviation=396.0, and median=751. This

variable (prov_pat_count) was categorized into four levels (variable









name=prov_pat_cat) and will be used as a proxy for how busy the doctor was during the

study period (see Table 5-14).

The medical school graduation state/country was used to construct a variable

called prov_fmg that was set to 'yes' (N=8, 5.4%) when the doctor had graduated from a

medical school outside the US (Argentina=1, Canada=2, Croatia=1, England=1,

Holland=1, Italy=1, Panama=1). The graduate level degrees held by each provider were

used to construct a variable called prov_md_plus that was set to 'yes' (N=14) when the

doctor had obtained a graduate degree in addition to their M.D. (MPH=6, MSC=1,

MSW=1, PHD=6). The assertions about malpractice cases in the last 10 years found on

the Massachusetts Board of Medicine web site were used to construct a variable called

mp_flag with 'yes' (N=7) when the doctor had a record of having been sued and 'no'

otherwise (N=147).

Site Characteristics

Each of the 15 sites was labeled with a unique number that will serve as

identification for subsequent analysis about inter-site variability. The only variable that

accrues to the clinics themselves is the number of primary care doctors in active

practice at each one. This variable along with the number of patients and primary care

doctors assigned to each site serve as proxies for the 'size' of the clinics. These are

summarized in Table 5-15.

Note that the sum of the number of active doctors at all sites is 168 whereas there

were only 148 included in this study (Study Doctors Column). The remaining 20 had

less than 100 loyal patients and were not included in the analytic data set.









Patient, Provider, and Clinic Identifiers

For multi-level (hierarchical) modeling it is necessary to uniquely identify the

individual primary care doctors being studied so as to be able to maintain the linking

between them and their loyal patients. Therefore, the MGH provider identifying numbers

were sorted and mapped to the corresponding rank (1, 2, 3, ... 148). Using this

mapping, each of the 85,277 patient's linked MGH provider identifier was replaced by a

unique (though now anonymous) integer. Since each observation in the final analytic

data set represents a single patient, no identifying information need be retained (e.g.,

Name, Medical Record Number, and etc) and these were all dropped. As described

above, the clinics were already identified by (uninformative) integers (1-15).

Data Integrity: Clinical Activity Variables

One way to insure that the queries of outpatient visits, inpatient stays, emergency

room visits, and outpatient imaging tests for all patients in the study were complete and

consistent is to plot them over time. This was done by counting each event type by

month for the whole study cohort and plotting as a time series. The visit counts by

month stratified by type of provider rendering the visit are plotted in Figure 5-2.

It is reassuring that the counts are relatively stable and consistent during the study

period. This implies that there are no large gaps or duplications in the data. As for any

secular patterns, this dissertation will not attempt to describe or explain them.

Similarly, the counts of outpatient diagnostic imaging examinations were

enumerated by month and stratified by the type of provider ordering the study are

plotted in Figure 5-3.

As with the counts of outpatient visits, there is apparent consistency and stability.

Further reassurance comes from the fact that relative decreases in counts during









December 2007 and August 2008 seem to match between outpatient visits and imaging

tests. Since the data came from separate and independent administrative sources,

there was an actual decrease in outpatient clinical activity during these periods. Visual

inspection of the two plots confirms that visits and imaging tests tend to rise and fall

together by month.

The other main activity variables relate to hospital events and counts of these by

month are plotted in Figure 5-4.

Again the month to month consistency and stability attests to the integrity of the

data, which came from two separate databases (one for emergency room and a second

for inpatient/observation stays).

Variable Summary

A summary of all patient-level independent variables is provided for reference in

Table 5-16. A summary of all the clinic and provider level independent variables is

provided for reference in Table 5-17.

Table 5-1. All diagnostic imaging performed on study cohort during two years of study.
Place Of
Service CT MR NM PET X-Ray US Total Percent
ER 10183 2651 489 8 18608 2406 34345 15.50
Inpatient 4768 1634 944 128 19907 2382 29763 13.43
Outpatient 26981 21999 6129 3709 76654 21991 157463 71.07
Total 41932 26284 7562 3845 115169 26779 221571 100
Percent 18.92 11.86 3.41 1.74 51.98 12.09 100
Note: CT=computed tomography, MR=magnetic resonance imaging, NM=nuclear
medicine, PET=positron emission tomography, X-Ray=radiography, US=ultrasound.
The studies performed in the outpatient setting are further stratified in Table 5-2.









Table 5-2. Outpatient diagnostic imaging performed on study cohort during two years of
study.
Who Ordered CT MR NM PET X-Ray US Total Percent
Specialist 15295 13170 3684 3294 42143 9106 86692 55.06
Other primary
Other primary 1133 820 157 17 6605 1101 9833 6.24
care doctor
Patient's own
10553 8009 2288 398 27906 11784 60938 38.70
loyal doctor
Total 26981 21999 6129 3709 76654 21991 157463 100
Percent 17.13 13.97 3.89 2.36 48.68 13.97 100-
Note: CT=computed tomography, MR=magnetic resonance imaging, NM=nuclear
medicine, PET=positron emission tomography, X-Ray=radiography, US=ultrasound.
The studies ordered by the patient's linked (loyal) doctor are further stratified in
Table 5-3.


Table 5-3. Outpatient diagnostic imaging ordered
during two years of study.
Body Area CT MR NM PET
Abdomen 4005 659 25 89
Cardiac 42 10 2003 0
Chest 4637 303 13 96
Extremity 118 1807 0 0
Head/Brain 976 2060 0 4
Maxillofacial
325 249 54 95
and/or Neck
Pelvis 256 318 0 0
Spine 193 2602 0 0
Unspecified 1 1 193 114
Total 10553 8009 2288 398
Percent 17.32 13.14 3.75 0.65


by patient's linked (loyal) doctor


X-Ray
762
0
11084
10776
244


US
4032
0
12
834
152


Total
9572
2055
16145
13535
3436


578 2055 3356


1097
3355
10
27906
45.79


4310
0
389
11784
19.34


5981
6150
708
60938
100


Percent
15.71
3.37
26.49
22.21
5.64
5.51
9.81
10.09
1.16
100


Note: CT=computed tomography, MR=magnetic resonance imaging, NM=nuclear
medicine, PET=positron emission tomography, X-Ray=radiography, US=ultrasound.









Table 5-4. Univariate statistics of the outcome variable (per patient count of outpatient
imaging tests ordered by primary care provider).
N 85277
Minimum 0 (N=53,617)
Maximum 15
Mean 0.7146
Standard Deviation 1.2563
Skewness 2.6554
Uncorrected SS 178132
Coeffient of Variation 175.8049
Sum of Observations 60938
Variance 1.5782
Kurtosis 9.8796
Corrected SS 134586
Standard Error of the Mean 0.0043
Note: SS=sum of squares.

Table 5-5. Distribution of patient race.
Race Frequency Percent
White 68432 80.25
Black 4278 5.02
Hispanic 5924 6.95
Other 6643 7.79
Total 85277 100

Table 5-6. Distribution of patient's payer categories.
Payer Group Frequency Percent
Aetna 1717 2.01
Blue Cross Blue Shield 30762 36.07
Commercial 6282 7.37
Free care 196 0.23
Harvard Pilgrim Healthcare 8574 10.05
Mass Health Net 1890 2.22
Medicaid 3646 4.28
Medicare 19555 22.93
Neighborhood Health Plan 1573 1.84
Other 3057 3.58
Self Pay 1057 1.24
Tufts Health Plan 6968 8.17
Total 85277 100.00









Table 5-7. Patient's payer collapsed into 6 categories.
Payer Group Frequency Percent
Blue Cross Blue Shield 30762 36.07
Commercial 7999 9.38
Managed 17115 20.07
Medicare 19555 22.93
Other 3057 3.58
State 5536 6.49
Uninsured 1253 1.47
Total 85277 100.00

Table 5-8. Hospital activity variables (per patient).


Description
Total hours in
Emergency room
Emergency room
visits
Inpatient admissions

total days in hospital

Days in intensive care
units
Readmitted within two
weeks of inpatient
discharge
Readmitted within one
month of inpatient
discharge
Observation (short)
stays


Variable Name

erhours

ervisits

inpt_stays

Inpt_los_total

inpt_icu_days


inpt_read_15d


inpt_read_31 d


obs_stays


Minimum
0
N=69,485
0
N=69,484
0
N=74,582
0
N=74,697
0
N=84,401

0
N=84,382

0
N=84,062
0
N=76,145


Maximum


Mean SD


483 2.71 10.15


47 0.33


1.04


22 0.20 0.67

179 0.92 4.70

49 0.05 0.76


10 0.01 0.15


17 0.02 0.21

37 0.13 0.54


Note: The number of patients with zero counts is given under Minimum (where zero).









Table 5-9. CPT codes and relative value units for ambulatory office visits.
CPT Code Visit Type Complexity RVU
99201 new patient not comprehensive not complex 0.48
99202 new patient not comprehensive mod complex 0.93
99203 new patient not comprehensive high complex 1.42
99204 new patient comprehensive moderate complexity 2.43
99205 new patient comprehensive high complexity 3.17
99211 established not comprehensive not complex 0.18
99212 established not comprehensive mod complex 0.48
99213 established not comprehensive high complex 0.97
99214 established comprehensive moderate complexity 1.50
99215 established comprehensive high complexity 2.11
99241 consultation not comprehensive not complex 0.64
99242 consultation not comprehensive mod complex 1.34
99243 consultation not comprehensive high complex 1.88
99244 consultation comprehensive moderate complexity 3.02
99245 consultation comprehensive high complexity 3.77

Table 5-10. Outpatient visit activity variables (per patient).
Description Variable Name Minimum Maximum Mean SD
Count of outpatient visits 0
to linked (loyal) provider prv_visit_count N=10,396 62 3.54 3.39
Summed RVU of visits to 0
linked (loyal) provider prv_visit_rvu N=10,396 76.22 4.95 4.83
Count of outpatient visits
to other primary care 0
doctors pcp_visit_count N=57,117 75 0.60 1.24
Summed RVU of visits to
other primary care 0
doctors pcp_visit_rvu N=57,117 59.44 0.73 1.57
Count of outpatient visits 0
to specialists spec_visit_count N=25,771 112 3.60 5.17
Summed RVU of visits to 0
loyal specialists spec_visit_rvu N=25,771 165.18 4.96 7.28
Note: The number of patients with zero counts or RVU is given under Minimum (where
zero).









Table 5-11. Binary clinical problem variables (per patient).
Problem category Variable Name Number Yes Percent Yes
Diabetes pr_dm 9485 11.12
Hypertension pr_htn 25219 29.57
Heart failure pr_chf 867 1.02
Coronary artery disease pr_cad 3940 4.62
Renal failure pr_crf 1096 1.29
Cancer pr_can 9925 11.64
Trauma pr_trm 1804 2.12
Obesity pr_obs 8855 10.38
Substance abuse pr_sub 770 0.90

Table 5-12. Four level categorization of patient active outpatient medications (per
patient).
meds_cat Frequency Percent


None
1-5
6-10
>10
Total


4475
45536
23390
11876
85277


5.25
53.40
27.43
13.93
100.00


Table 5-13. Summary of other (non-outcome)
patient).


imaging test utilization variables (per


Variable
Description Name Minimum Maximum Mean SD
Count of imaging tests
ordered during emergency 0
room visits all_e_cnt N=73,562 96 0.40 1.56
Count of imaging tests
ordered during inpatient 0
stays all i cnt N=79,118 118 0.35 2.41
Count of outpatient imaging 0
tests ordered by specialists spec o cnt N=55,105 61 1.02 2.28
Count of outpatient imaging
tests ordered by other 0
primary care doctors* pcp_o_cnt N=78,289 8 0.12 0.45
Note: The number of patients with zero counts is given under Minimum (where zero).
*Not the patient's own linked (loyal) doctor.

Table 5-14. Four level categorization of the number of patients cared for by each
provider (panel size).
prov_pat_cat Frequency Percent
<500 37 25.00
500-759 36 24.32
750-999 36 24.32
1 K+ 39 26.35









Table 5-15. Site (clinic) characteristics.


Site
(clinic) ID
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Total


Active
Doctors
7
8
20
11
15
5
7
17
12
14
9
18
5
6
14
168


% Active
Doctors
4.17
4.76
11.90
6.55
8.93
2.98
4.17
10.12
7.14
8.33
5.36
10.71
2.98
3.57
8.33
100


Study
Patients
3451
4233
11925
5058
7220
3269
3308
10381
6378
5728
5097
9109
1135
4794
4191
85277


% Study
Patients
4.05
4.96
13.98
5.93
8.47
3.83
3.88
12.17
7.48
6.72
5.98
10.68
1.33
5.62
4.91
100


Study
Doctors
5
8
16
9
13
4
6
17
12
14
7
16
5
6
10
148


% Study
Doctors
3.38
5.41
10.81
6.08
8.78
2.70
4.05
11.49
8.11
9.46
4.73
10.81
3.38
4.05
6.76
100











Table 5-16. Description and categorization of 33 patient level independent variables.


Variable Name
inpt_stays
er hours
ervisits
obs_stays
inptread_31d
inptread_15d
Inptlos_total
inpt_icu_days

all e cnt

all i cnt

speco_cnt

pcp_o_cnt

prv_visit_rvu

prv_visit_count

spec_visit_rvu
spec_visit_count

pcp_visit_rvu

pcp_visit_count
age_08
Race
Sex

payer_group
medscat
pr_cad
pr_can
pr_chf
pr_crf
pr_dm
pr_obs
pr_htn
pr_sub
pr_trm

oth prob


Level Class
Patient Activity, Hospital
Patient Activity, Hospital
Patient Activity, Hospital
Patient Activity, Hospital
Patient Activity, Hospital
Patient Activity, Hospital
Patient Activity, Hospital
Patient Activity, Hospital
Activity,
Patient Other Imaging
Activity,
Patient Other Imaging
Activity,
Patient Other Imaging
Activity,
Patient Other Imaging

Patient Activity, Visits

Patient Activity, Visits


Patient
Patient


Activity, Visits
Activity, Visits


Patient Activity, Visits


Patient
Patient
Patient
Patient


Activity, Visits
Demographics
Demographics
Demographics


Patient Insurance
Patient Medications
Patient Clinical Problem
Patient Clinical Problem
Patient Clinical Problem
Patient Clinical Problem
Patient Clinical Problem
Patient Clinical Problem
Patient Clinical Problem
Patient Clinical Problem
Patient Clinical Problem
Clinical
Patient Problems


Type
Numeric
Numeric
Numeric
Numeric
Numeric
Numeric
Numeric
Numeric

Numeric

Numeric

Numeric

Numeric

Numeric

Numeric

Numeric
Numeric

Numeric

Numeric
Numeric
Categorical
Categorical

Categorical
Categorical
Categorical
Categorical
Categorical
Categorical
Categorical
Categorical
Categorical
Categorical
Categorical

Numeric


Levels Description
count of inpatient stays
total hours spend in er
count of er visits
count of observation stays
count of readmit within 31 days
count of readmit within 15 days
total days in hospital
total days in icu

count of images done in ER

count of images done as inpatient
count of outpatient images ordered by
specialists
count of outpatient images ordered by
other primary care doctors
sum of rvu of outpatient visits to linked
(loyal) primary care physician
count of outpatient visits to linked
(loyal) physician
sum of rvu of outpatient visits to
specialists
count of outpatient visits to specialists
sum of rvu of outpatient visits to
covering pcp
count of outpatient visits to covering
pcp
Patient's age in 2008
4 patient identified race
2 patient sex
category of patient's payer of record in
6 2008
4 active outpatient prescriptions in 2008
2 coronary artery disease
2 cancer
2 congestive heart failure
2 chronic renal failure
2 diabetes
2 obesity
2 hypertension
2 substance abuse
2 trauma
count of active problems other than
those separately listed above












Table 5-17. Description and categorization of provider (8) and clinic level (2)
independent variables.


Variable Name

sitedocs
siteid

mp_flag

prov_md_plus

prov_pat_cat
prov_sex

prov_fmg
prov_age_08

prov_exp_08
prov id


Level


Class


Clinic Characteristic
Clinic Identifier

Provider Characteristic

Provider Characteristic

Provider Characteristic
Provider Characteristic

Provider Characteristic
Provider Characteristic

Provider Characteristic
Provider Identifier


Type Levels Description
number of doctors actively practicing at
Numeric the clinic in 2008
Categorical 15 anonymous clinic (site) identifier
whether provider has been sued in last
Categorical 2 10 years
whether provider has a degree beyond
Categorical 2 MD
number of patient's in provider practice
Categorical 4 in 2008
Categorical 2 provider sex
whether provider is foreign medical
Categorical 2 graduate
Numeric age in years of the provider in 2008
number of years after provider MD
Numeric graduation in 2008
Cateaorical 148 anonymous provider identifier


100000


S.
S.


2 3 4 5 6
Number Of Imaging Tests


7 8 9 >10


Figure 5-1. Outpatient imaging tests (per patient) ordered by the linked (loyal) primary
care provider. Both the percent (left Y axis, bars) and number (right Y
logarithmic axis, diamonds and dashes) of observations are shown.


60


50
0)
30
40
o
30
20






10


0


10000


U)
0
1000
(i)
-o
0


10




10




1


Si
E
z











45000
O spec
Spcp
40000- pr

35000

30000

25000

20000

15000

10000

5000







Figure 5-2. Number of outpatient visits by all patients in study cohort (by month) over
two years of study. Hatched: visits to patient's linked (loyal) doctor, Black:
visits to another (covering) PCP, White: visits to specialists.


10000 [ prv


2000


Figure 5-3. Number of outpatient imaging tests performed on all patients in study cohort
(by month) over two years of study. Outpatient imaging tests ordered by:
Hatched: patient's linked (loyal) doctor, Black: another (covering) PCP, White:
specialists.














DIN
4000- m OBs
E ER
3500

3000

2500

2000

1500

1000

500

0





Figure 5-4. Number of hospital encounters for all patients in study cohort over two
years. Hatched: emergency room visits, Black: short (observation) stays,
White: inpatient discharges.









CHAPTER 6
METHODS

Outcome Variable Distribution

Simple univariate statistics of the main outcome variable (prvo_cnt = IMG for this

chapter) have been described above in Chapter 5 (see Table 5-4 and Figure 5-1).

Based on the visual inspection, these appear to be count data with a Poisson

distribution. This is a discrete distribution, with a single parameter, Lambda (A), that

expresses the probability of a number of events occurring in a fixed period of time if

these events occur with a known average rate (A) and independently of the time since

the last event. Each instance of a Poisson random variable can be expressed as being

the result of a single Poisson 'experiment' and may take any positive integer value

(Stahl 1969). An important property of the Poisson distribution is that the variance is

equal to the mean (A). The Poisson equation may be written as Equation 6-1.

Prob(n) = (An e-) / n! (6-1)

In Equation 6-1, e is the base of the natural logarithm, n is the number of

occurrences of an event the probability of which is given by the function, and A is the

positive real number, equal to the expected number of occurrences that occur during

the given interval.

A more theoretically appealing and possibly informative way to describe the

distribution of imaging counts is that of a two stage process with the first stage

determining the occurrence of any imaging from the whole data set of 85,277 patients

and the second stage determining the count of imaging tests for patients that had at

least one (IMG > 0, N=31,660) imaging test during the study period. This can be

modeled with a logistic regression on all observations followed by Poisson regression









on the non-zero observations. For purposes of estimating patient, provider, and clinic

effects; the two step (logistic>Poisson) approach was taken.

A final method to model the distribution of outpatient imaging counts combines the

logistic portion (any use) with the Poisson assumption for intensity (non zero use) into a

single distribution; zero inflated Poisson (ZIP). The ZIP distribution is almost the same

as the standard Poisson when n>0 but has a second portion for n=0 and may be written

as Equations 6-2 and 6-3.

Prob(n) = 4 + (1-4) e- for n = 0 (6-2)

Prob(n) = (1-~ ) (An e-) / n! for n = 1, 2, ... (6-3)

In Equations 6-2 and 6-3, e is the base of the natural logarithm, n is the number of

occurrences of an event the probability of which is given by the function, A is the

positive real number, equal to the expected number of occurrences that occur during

the given interval, 4 is the real number between 0 and 1 (4 = 0 is standard Poisson)

called the zero inflation parameter.

To examine the distribution of this variable, SAS proc GENMOD was used to

estimate two null models on all 85,277 observations and a single one for the 31,660

observations with non-zero outcomes (patients who had at least one imaging test

ordered by their linked/loyal primary care doctor). Specifically the models used were

simple Poisson and Zero-Inflated Poisson (ZIP) distributions for the whole data set and

simple Poisson distribution on the non-zero observations. The intercept coefficients

from these models and the 95% confidence intervals were used as estimators of the

Poisson Lambda parameters of each proposed distribution. The dispersion for each









modeled distribution was obtained by using the PSCALE and DSCALE option in the

model statement.

To visually demonstrate the relationship between the observed distribution of

imaging test counts and the three proposed Poisson distributions, the parameter

estimates from the null models described above were submitted to a SAS provided

macro called PROBCOUNTS. This software is available for download on the public SAS

support web site at the following location (checked May 2, 2010):

http://support.sas.com/kb/26/161.html

The purpose of this program is succinctly described on the support web site as

follows:

The PROBCOUNTS macro computes the predicted count and the predicted
probabilities of specified counts for Poisson and negative binomial models
and for zero-inflated versions of these models as fit by PROC COUNTREG
in SAS/ETS software and PROC GENMOD in SAS/STAT software.

A plot of the observed counts (from 0 to 15) superimposed on line graphs of the

expected counts from each of the three distributions serves to compare them and will be

reproduced in the Chapter 7 as Figure 7-1. The purpose of this was to visually inspect

the fit of the proposed distributions with the actual imaging counts especially with

respect to the 'upper tail' and the tendency of over- or under-dispersion.

To verify the estimate of Lamda(A) for all three distributions and obtain a direct

estimate of Phi(4) for the ZIP distribution required a complementary approach. SAS

PROC NLMIXED allows exact specification of the proposed distributions) and produces

direct estimates for the parameters. Since an iterative approach is used, initial 'seed'

values for the parameters of interest (Lam and Phi in the code fragments) are supplied.

The SAS code for the two simple Poisson models is reproduced below:









proc n/mixed data=input.data_set;
parms Lam=1;
loglike=IMG*log(Lam)-Lam-lgamma(IMG+ 1);
model IMG-general(loglike);
where IMG ne 0; < run;

The SAS code to obtain the ZIP parameters is reproduced below:

proc n/mixed data =input. data_set;
parms Phi=0.5 Lam=2;
if IMG=O then prob =Phi+(1-Phi) *exp(-Lam);
if IMG=0 then loglike=log(prob);
else loglike=log(1-Phi) +1MG*log(Lam)-Lam-lgamma(IMG+ 1);
model IMG-general(loglike);
run;

Correlation between Independent Variables

Some of the clinical activity variables essentially measure the same events in

slightly different ways. This is especially true for those representing outpatient visits

where each of the three doctor classes (patient's own linked primary care doctor,

covering primary care doctors, and specialists) is quantified by simple visit counts as

well as summed RVU values. Some of the hospital activity variables also may be

correlated (e.g., inpatient stays and total inpatient length of stay). Therefore, to help

select among them for subsequent modeling, all 17 clinical activity variables were

analyzed for (Spearman) correlation with SAS PROC CORR. The relevant results were

tabulated as a correlation matrix (Table 7-1) and are discussed in the Results chapter.

Spearman correlation was also performed between the two date-derived provider level

variables; provider age in 2008 and provider experience in 2008.









Bivariate Relationships

Ordinary least squares regression with SAS PROC GLM was used to estimate the

relationship between each independent variable and the outcome (count of imaging

tests ordered by patient's loyal provider) for all observations (N=85,277). Numeric

independent variables (N=28) were entered as recorded while categorical variables

(N=20) were parameterized using SAS EFFECT method which yields L-1 design

variables where L is number of levels in the original variable. This replicates the

categorical variable coding method (reference cell) that will be used in multivariable

modeling and yields degrees of freedom that are identical. Basically, this translates into

an OLS linear regression with numeric variables and a one-way ANOVA with the

categorical variables. For each of the 48 variables, the following parameters were

obtained to measure the strength of bivariate association with the outcome: F Value, R-

Squared, Correlation (R), and p value from the single variable regression output.

Variable Reduction for Modeling

Based on the correlation analysis between independent variables and evaluating

the bivariate relationships with the outcome, redundant and/or collinear predictors were

omitted from subsequent multivariable and multi-level modeling. The selection heuristic

was to choose the one having strongest bivariate relationship with the outcome when

independent variables were strongly correlated. However, for theoretical reasons, some

independent variables were kept despite not having significant bivariate relationship

with the outcome (e.g. provider malpractice status).

Multivariable (logistic) Modeling: Any Imaging Use

Multivariate logistic regression was used to analyze a binary outcome derived from

the main outcome variable (IMG). This binary outcome variable (called ANY_IMG) is set









to 'yes' (N=31,660) when that patient had one or more outpatient diagnostic imaging

test(s) ordered by the linked (loyal) provider and 'no' (N=53,617) otherwise. This was

done with SAS PROC LOGISTIC with the modeled outcome set to 'yes'. Numeric

variables were entered as recorded. The categorical variables were parameterized

using the SAS REFERENCE encoding method. This allows explicitly setting the

reference level and results in L-1 dummy variables where L is the number of levels in

the original variable. Provider level variable reference levels were as follows: Sex-Male,

FMG-No, MD_Plus-No, Malpractice-No, Provider Patients-<500. Patient level reference

Levels were set as follows: Sex-Male, Race-White, Insurance-Uninsured, Medications-

None. The remaining patient level categorical variables were the binary clinical problem

assertions and reference for each of these was set as 'no'.

To simplify writing the model, the patient level variables may be represented by a

vector P, the provider (doctor) variables by a vector D and the clinic variable(s) by C.

The logistic model is expressed as Equation 6-4:

Logit(p[ANY_IMG=yes])i = Po + PpPip + PdDid + PcCic (6-4)

In Equation 6-4, i is the ith patient, p is the pth patient level variable, d is the dth

doctor variable, and c is the cth clinic variable. Type III p estimates, standard errors,

Chi-Squared, p-value, and odds ratios were obtained from the resulting solution output

from SAS. These were used to make inferences and comparisons about joint

significance and effect size of all included predictor variables. A Hosmer and Lemeshow

(HL) test for goodness of fit was requested as well.









Multivariable (Poisson) Modeling: Imaging Intensity (non-zero)

This portion of the analysis seeks to determine the effect of the same predictor

variables used in the logistic analysis on the intensity of imaging (given that some has

occurred). Only the 31,660 observations (patients) where IMG is between 1 and 15 are

used in this model. From preliminary analysis of the outcome (IMG) distribution, it was

determined to be reasonably represented as a Poisson count variable. To estimate

these types of models, SAS provides PROC GENMOD which allows specification of

linear models with Poisson error distribution and log link function. These may be

estimated using maximum likelihood methods. As above, to simplify writing the model,

the patient level variables may be represented by a vector P, the provider (doctor)

variables by a vector D and the clinic variable(s) by C. The Poisson model may be

written as Equation 6-5.

Log[E(IMGi I Pip, Did, Cic)] = PO + PpPip + PdDid + PcCic + ei (6-5)

In Equation 6-5, i is the ith patient, p is the pth patient level variable, d is the dth

doctor variable, and c is the cth clinic variable. The errors (ei) are distributed as Poisson.

After estimation, the solution output contains the coefficient estimate, standard error,

Chi-Squared, and p-value for each one of the numeric variables and categorical variable

levels. The model was re-estimated using DSCALE and PSCALE options to determine if

the outcome distribution was over- or under-dispersed. As will be described in the

Results chapter, the distribution turns out to be underdispersed, and standard error

would tend to be overestimated. Therefore, a correction was NOT made for dispersion

which results in somewhat conservative inferences about significance of various effect

sizes.









Preparation for Multi-Level Modeling: Imaging Propensity Scores

For multi-level modeling, we wish to answer questions about variation in imaging

utilization between providers (level 2) holding patient factors equal while accounting for

clustering of patients (level 1) within providers. For this dissertation, the possible higher

level effect of clinics as aggregations of providers will not be addressed. To simplify the

specification, estimation, and interpretation of multi-level modeling results, all patient

level factors were collapsed into a single risk adjusted expected imaging (propensity)

variable. Previously described evaluation of the overall distribution of the outcome (IMG)

determined that a zero-inflated Poisson (ZIP) distribution to be most suitable for a single

model. This was done with SAS PROC GENMOD, all 85,277 observations, and the

same patient level independent variables used in multivariable modeling described

above. The SAS PROC GENMOD instructions were constructed so as to produce

patient level predictions after the initial maximum likelihood estimation. This variable will

be called IMG_PROP for 'imaging propensity' and represents the number of outpatient

imaging tests that the 'average' patient with identical values of all independent variables

would be expected to have. Another way of describing this technique is as regression

based risk adjustment for imaging utilization. In typical provider profiling applications,

these patient level predictors would be termed the 'expected' imaging utilization and

summed across doctors to be compared with the 'observed' count of imaging tests

actually performed on those same patients.

Multi-Level (Hierarchical) Modeling

This part of the analysis will test the relationship between a summary of each

patient's imaging propensity (IMG_PROP) and imaging performed (IMG) within each of

the 148 primary care doctor's practice. This posits that each doctor has his or her 'own'









regression equation with an intercept and slope. The interpretation of the intercept for

each doctor is the 'general' tendency to obtain imaging on the average patient while the

slope corresponds to that doctor's 'response' to patients with increasing risk adjusted

expectation for imaging (IMG_PROP = imaging propensity described above). With these

interpretations in mind, the analysis will answer the following:

What is the average intercept and slope of the 148 provider regression
equations?
How much do the intercepts vary from doctor to doctor?
How much do the slopes vary from doctor to doctor?
What is the correlation between intercepts and slopes?

Generally, in two-level hierarchical modeling, where the level 2 intercepts are

allowed to vary randomly (in this study for each provider), it is important to consider

whether or not to center (offset) the level 1 variable (IMG_PROP) in any way. There are

three options; natural metric (no centering), grand mean (e.g., all patients), group mean

(e.g., patients loyal to each doctor). Consequences of these choices affect both the

value and interpretation of the intercept estimates (Raudenbush and Bryk 2002).

Further, the standard errors for estimates of both fixed and random effect intercepts

differ depending on choice of centering method (Luke 2004). In health services research

applications, centering on the (level 2) group mean (provider in this case) is

recommended (Houchens et al. 2007). For this study, the group means centering

approach will be used. One distinct advantage is that it allows interpreting the individual

provider intercepts as representing the 'average' tendency to order imaging while

holding the expected imaging propensity (IMG_PROP) of their patients equal. On the

other hand, provider intercepts estimated with the non-centered approach are clinically









less meaningful because they would have to be interpreted as the tendency to order

imaging for patients with no clinical need (which numerically could be negative).

The model used will have random coefficients for intercept and slope and at the

level of the patient (level 1), is written as Equation 6-6:

IMGij = Poj + Pij (IMG_PROPij IMG_PROP.j) + eij (6-6)

In Equation 6-6, IMGij is the Count of outpatient images for the ith patient cared for

by the jth doctor, Poj is the The intercept for the jth doctor, Plj is the The slope for the jth

doctor, IMG_PROPij is the Imaging propensity for the ith patient cared for by the jth

doctor, IMG_PROP.j is the Average imaging propensity of the patients cared for by the

jth doctor, and eij is the Error (disturbance) for the ith patient cared for by the jth doctor.

For simplicity let clMG_PROPij = IMG_PROPij IMGPROP.j, where clMG_PROPij

is the centered imaging propensity for the ith patient cared for by the jth doctor.

Therefore, our patient (level 1) model may be written as Equation 6-7.

IMGij = Poj + Pij (clMG_PROPij) + eij (6-7)

The imaging utilization of each doctor's practice is characterized by two

parameters: Poj, the intercept for the jth doctor and p3j, the slope for the jth doctor. Since

the imaging propensity for each patient is centered on the mean for their doctor, the

intercept is actually that doctor's mean imaging. These two parameters vary across

doctors in the level-2 model as a function of the grand mean and random disturbance

such that Poj = Yoo + uoj and plj = 710 + u1i, where poj is the intercept (mean imaging) for

the jth doctor, yoo is the average of the doctor means of imaging use across the

population of doctors, uoj is the individual variation from the average intercept for the jth

doctor, plj is the slope for the jth doctor, 710 is the average imaging propensity / imaging









utilization regression slope across the doctors, and uij is the individual variation from the

average slope for the jth doctor

Combining the two by substituting the level 2 random coefficient models into the

level 1 patient level model gives Equation 6-8.

IMGij = Yoo + uoj + yIo (clMG_PROPij) + u1j (clMG_PROPij) + sij (6-8)

The variance structure can be expressed as Equations 6-9 through 6-12.

uoj~ N(0, Too) (6-9)

uij~ N(0, cil) (6-10)

Cov(uoj,uij)= 1To (6-11)

eij~ N(0, C2) (6-12)

In Equations 6-9 through 6-12, coo is the Var(uoj), 'lc is the Var(uij), and coI is the

the covariance between uoj and uij.

Since the level 2 model has no predictors in either the intercept or slope equation,

it is unconditional. Therefore, we can use multi-level regression estimates for variability

in intercepts and slopes as shown in Equations 6-13 and 6-14.

Var(uoj) = Var(Poj Yoo) = Var(poj) (6-13)

Var(uij) = Var(p3j Ylo) = Var(p3j) (6-14)

For estimation with SAS PROC MIXED we divide the combined model into fixed

parts (Equation 6-15) and random parts (Equation 6-16).

Yoo + yIo clMP_PROPij (6-15)

uoj + uij clMG_PROPij + sij (6-16)

Simplified SAS Code is written:

proc mixed data =input.data_set;









class provide;
model IMG = clMG_PROP;
random intercept clMP_PROP / subject=prov_id;
run;

To quantify the fraction of variation in imaging utilization attributable to patients

and providers it is necessary to obtain patient level residuals (eij ~ 2) from reduced

models and the null model. These can be represented as follows: Patient level Imaging

Propensity only (Equation 6-17), Patient level Imaging Propensity and provider slope

only (Equation 6-18), Patient level Imaging Propensity and provider intercept only

(Equation 6-19), and Null Model (Equation 6-20):

IMGij = Yoo + Yio (IMG_PROPij) + Sij (6-17)

IMGij = Yoo + Yio (clMG_PROPij) + u1j (clMG_PROPij) + Sij (6-18)

IMGij = Yoo + uoj + y71 (clMG_PROPij) + Sij (6-19)

IMGij = Yoo + eij (6-20)









CHAPTER 7
RESULTS

Outcome Variable Distribution

The outcome variable (prv_o_cnt=IMG) was measured for each patient

(N=85,277) over two years of study and represents the count of outpatient non-invasive

diagnostic imaging tests ordered by the patient's linked (loyal) primary care provider.

Univariate statistics and the distribution were shown (Table 5-4 and Figure 5-1

respectively) in Chapter 5. To estimate the Poisson Lambda for the whole sample

(N=85,277 including the 53,617 zero observations), SAS PROC GENMOD was used

with a null model (IMG = / dist=poi). Lambda was thus estimated at 0.71459 with 95%

confidence interval of (0.70718-0.72200). The dispersion (extent to which the observed

variance exceeds the expected) was estimated by Deviance / Degrees of Freedom

(145443.64 / 85,276) at 1.7056 which means that there are more observations at higher

values of n (3-15) than expected by a Poisson distribution with Lambda=0.71459. Such

'overdispersion' can result in underestimation standard errors in subsequent modeling.

Also using SAS PROC GENMOD and a null model (IMG = / dist=poi), a second

estimation, using only the non-zero observations, of Poisson Lambda and dispersion

was made in the same manner as described above. In this case, the Poisson Lambda

was estimated at 1.92476 with 95% confidence interval of (1.90948-1.94005) and the

dispersion was 0.7796.

Beginning with Version 9.2, SAS PROG GENMOD allowed estimation of ZIP

models. In addition to specifying a model for the outcome counts, a second, so called

zeromodel, was specified (IMG = / dist=zip; zeromodel =;). Using the entire sample

(N=85,277) with null count and zeromodel statements, the parameters of this ZIP









distribution were estimated. The Poisson Lambda estimate was 1.4917 with 95%

confidence interval of (1.4760-1.5074). The dispersion was 1.2428 and the zero inflation

intercept was 0.084 with 95% confidence interval of (0.0650-0.1028). To obtain a more

meaningful value for the zero inflation parameter (4), the output of SAS PROC

NLMIXED was used to obtain a direct estimate of 0.521 (Cl: 0.516-0.526). At the same

time, all three estimates of the Poisson Lambda from NLMIXED were identical to those

obtained from GENMOD which is reassuring.

The output from PROBCOUNTS generated using parameter estimates obtained

from the null models represent the expected distributions of the simple Poisson, non-

zero Poisson and zero-inflated Poisson assumptions. They are plotted in Figure 7-1.

Clearly, the best fit appears to be with the Zero-inflated Poisson distribution which

militates for using it to generate risk adjusted expected imaging (propensity scores) for

each patient prior to multi-level modeling seeking to characterize provider variation.

However, to estimate the effect of individual patient, provider, and clinic level factors on

imaging utilization the two stage (logistic followed by Poisson) approach were used.

One reason is that odds ratios produced by the logistic analysis of any utilization have

well understood meanings that are directly interpretable. Additionally, since the Poisson

distribution of non-zero use has dispersion that is less than one, there was no need to

correct effect sizes or significance levels for intensity of use. If anything, the uncorrected

estimates of patient, provider, and clinic factor effects were somewhat conservative.

Correlation between Independent Variables

Several of the patient level clinical activity and other imaging utilization variables

are theoretically redundant in that they essentially reflect the same phenomenon. For









example, there are four variables that derive from inpatient events; number of inpatient

stays (inpt_stays), number of days in the ICU (inpt_icu_days), and the two readmission

measures (inpt_read_15d and inpt_read_31d). The most closely related variable pairs

are the counts versus summed RVU of visits to the loyal provider, other primary care

doctors, and specialists respectively. As these are highly correlated, a choice between

them (counts or summed RVU) were necessary for subsequent multivariable modeling.

Also, various classes of these clinical activity counts may be correlated even though

they are not measuring precisely the same activity. For example, emergency room visits

and imaging tests ordered from the emergency room or inpatient stays and imaging

tests ordered in the hospital are likely to be correlated. Table 7-1 lists relevant

correlations between clinical activity and other imaging utilization variables. Though

there are 18 separate variables (17 in the rows plus the first column=all_e_cnt), only 9

columns are shown. The missing columns had no correlations > 0.5 and were omitted

for brevity.

As expected, the correlations between visit count and visit RVU for specialists

(0.975), linked (loyal) provider (0.994), and covering PCP (0.978) were very high.

Therefore, only one set of these would be included in multivariable modeling (visits vs

RVU). Emergency room visits and total hours spent in the emergency room were also

highly correlated (0.996) as were the number of inpatient stays and total days spent in

hospital (0.992). The two readmission measures (15 and 31 days) were also correlated

(0.857). Not included in Table 7-1 is the correlation between the age and experience

level of patient's linked/loyal primary care doctor. As expected, this was quite high

(0.949) and justifies using only one of them for subsequent multivariable modeling.









Bivariate relationships

The results of the individual OLS regressions are listed in Table 7-2 for patient

level demographic, insurance, medications, and problem variables. For patient level

clinical activity and other imaging variables, results are summarized in Table 7-3. For

provider and clinic variables the results are given in Table 7-4. All of the patient level

variables had highly significant linear relationships with the outcome (p < 0.0001). The

only exception was substance abuse (binary problem variable named pr_sub) with

p=0.0014. The magnitude of these relationships varied considerably with correlation

coefficients ranging from 0.01 up to 0.39, with the strongest being with the summed

RVU of visits by the patient to their loyal provider (prv_visit_rvu). Other particularly

strong relationships (correlation coefficients with outcome > 0.2) were exhibited by

specialist outpatient visit variables (spec_vis_count, specvisitrvu) and the number of

active prescriptions for each patient. Finally, the unique identifier for the patient's linked

(loyal) provider (prov_id on Table 7-4) had a fairly high correlation (0.2268) with the

number of outpatient imaging tests ordered by that same doctor. Interestingly, both

measures (count and RVU) of visit intensity to the linked (loyal) provider and specialists

seemed to have stronger relationships with the outcome than the actual identity of the

patient's linked (loyal) provider with all bivariate correlations being > 0.23.

Variable Reduction for Modeling

As noted above (Table 7-1), the visit counts and RVU variables were highly

correlated with each other. The bivariate relationships with the outcome variable from

Table 7-4 were used to guide the choice between them. In each case, the RVU version

had slightly higher correlation with the outcome while all of them were significant

(p<0.0001). Accordingly, only the three visit RVU variables (prv_visit_rvu, pcp_visit_rvu,









and spec_visit_rvu) were carried forward for further analysis. Another pair of variables

with high correlation between them was the provider's age and experience. From Table

7-4, we see that the provider experience variable was significantly (p=0.0002) related to

the outcome while the provider's age was not (p=0.2542). This made choosing among

them straightforward: select the provider experience variable for subsequent

multivariable analysis. Even though the variables representing provider foreign medical

graduate (FMG) and malpractice status were not significantly related to the outcome in

bivariate fashion, they were carried forward due to theoretical considerations. For

subsequent modeling, the variables coding actual identity of providers (prov_id) and

clinics (site_id) were omitted. However, the provider identity variable was used during

the final stage of analysis: multi-level hierarchical modeling. In summary, the following

variables were dropped for purposes of multivariable logistic and Poisson regression:

er_hours (total hours in the ER)
inpt_read_15d (count of readmit within 15 days)
pcp_visit_count (count of outpatient visits to covering PCP)
prv_visit_count (count of outpatient visits to loyal doc)
spec_visit_count (count of outpatient visits to specialists)
prov_age_08 (age in years of the provider in 2008)
prov_id (provider identifier)
site_id (site (clinic) identifier)

Multivariable (Logistic) Modeling: Any Imaging Use

The logistic model with all 85,277 observations, 28 patient level, 6 provider level

and 1 clinic level independent variables was estimated using ANYIMG as the outcome

('yes' when the count of imaging tests ordered by patient's linked (loyal) doctor was

greater than zero and 'no' otherwise). This served to jointly test the effect of each of the

35 independent variables on whether or not the patient had any imaging ordered by

their linked (loyal) doctor during the two years of study. The outcome value of 'yes' was


100









set to be the event/success level. Subsequent interpretation of the resulting odds ratios

is such that when they are greater than one, that variable/level is associated with a

higher probability of imaging.

The -2 Log Likelihood was 112,501 for intercept only and 100,678 for the full

model. The 'pseudo' R-Squared was 0.13 rescaled to 0.18 and the c Statistic was

0.723. A Hosmer and Lemeshow (HL) test for lack of fit was highly significant

(p<0.0001) with Chi-Square of 588 on 8 degrees of freedom. However, it should be

noted that there is evidence that for large sample sizes (exceeding 50K as in this study)

a significant HL test does not entail that a particular logistic model is useless or even

poorly specified (Kramer and Zimmerman 2007, Bertolini et al. 2000). The only

hypothesis affirmed is that there is a high probability of at least some lack of fit and with

the R-Squared of 0.18, this is already established.

The individual independent variable results are listed in Tables 7-5 and 7-6 and

graphically depicted in Figure 7-2. In the figure, odds ratios to the right of the reference

line (1.0) imply that the variable or level was associated with an increased probability

that the patient would have any imaging test during the two years of study. Note that the

odds ratios for numeric variables (e.g., patient age, clinical activity variables, and

provider experience) represent the increase in probability of any imaging with a unit

increase in the value of that variable. For example, consider patient age. For each

additional decade (from 3rd through 9th), the probability of any imaging use increased by

16%. Thus, a 90 year old patient would be about 3 times more likely to have at least

one imaging test compared with a 20 year old (all else equal). On the other hand,

females were only 9% more likely to have imaging than males. Recall, that


101









mammography has been specifically omitted from this study. Otherwise, that number

would likely have been about an order of magnitude higher. Black and Hispanic patients

were about 20% more likely to receive imaging as compared to whites (reference level).

For the most part, insurance status was either not significant or had a small effect

size in the expected direction (greater likelihood of any imaging compared with

reference of uninsured). The notable exception was Medicare with a significant

(<0.0001) and substantial negative effect size (OR = 0.752, -25% less likely to have

any imaging than uninsured--or self-pay--patients and even greater compared with other

insurance types). The only other significant insurance type was Managed and patients

were ~18% more likely to have imaging compared with uninsured. The Blue Cross

group (BCBS) was marginally significant with a (~14%) positive effect on imaging use.

The number of medications patients were taking had no effect on imaging use.

Individual binary clinical problems were significant in six out of nine instances. There

was only one clinical problem that seemed to increase likelihood of imaging and that

was Trauma (OR=1.24). The other five problems were associated with decreased

imaging when present: Cancer (OR=0.949), Congestive Heart Failure (OR=0.765),

Diabetes (OR=0.729), Hypertension (OR=0.784), and Substance Abuse (OR=0.793).

The count variable which subsumed the remaining clinical problem list entries not

categorized above (Other Problems) showed a significant (though small) positive effect

on imaging (OR for 1 unit increase = 1.019). At the median level of 6.0, this would result

in ~12% increase in likelihood of any imaging compared with none.

Clinical activity variables tested included the summed RVU of outpatient visits. As

expected, visits to the patient's linked (loyal) doctor were strongly related to imaging


102









(prv_visit_rvu: OR=1.164 for a 1 unit increase). The maximum value for this variable is

17 RVU. Thus, patients having visits to their linked (loyal) doctor over two years totaling

17 RVU would be at least 13 times more likely to have imaging compared with those

having a single visit with fractional RVU. The variable representing visits to other

primary care doctors (pcp_visit_rvu) was not significant. However more visits to

specialists (specvisitrvu: OR=1.025 for 1 unit increase) were associated with higher

likelihood of imaging (ordered by the patient's linked/loyal doctor). Of the six clinical

activity variables that measured hospitalization, only two were significantly associated

with primary care imaging. These were 24 hour observation admissions (obs_stays:

OR=1.07 for 1 unit increase) and the total inpatient length of stay (inpt_los_total:

OR=0.987 for a 1 unit increase). The seeming discrepancy makes some sense by

speculating that short stays for observation might indicate and/or engender need for

imaging that would be performed later (as an outpatient).

Two of the additional imaging utilization variables had small effect sizes: outpatient

ordered by specialists (spec o cnt: OR=1.025) and inpatient (all_i_cnt: OR=1.020).

Emergency room imaging (all_e_cnt) was not significant. The number of outpatient

images ordered by other primary care doctors (pcp_o_cnt: OR=1.157) was positively

associated with likelihood of imaging by the patient's own (loyal) doctor. For future

applications using different data sources (risk adjustment for provider profiling of

imaging utilization) these additional imaging variables can probably be omitted with little

consequence because the remaining clinical activity variables will capture the same

phenomenon. Clearly, hospital events are, by definition, correlated with associated

imaging (e.g., ER visits and imaging performed in the ER). Likewise, visits to specialists


103









and other (covering) primary care doctors could stand in for the outpatient imaging

ordered by these same doctors (e.g., omitting pcp_o_cnt might allow pcp_visit_rvu to

become significant).

Turning to the provider and clinic level variables (Table 7-6) we see that the

amount of experience is negatively associated with likelihood to obtain imaging

(OR=0.997 for each additional year). However the effect size is rather small.

Considering that the range of experience was 5-50 years, this implies that likelihood of

any imaging decreases by only about 10% between least and most experienced

doctors. The gender of the doctor has a greater effect than experience, with women

(OR=1.14) being 14% more likely to order imaging on their patients compared with

males. Foreign medical graduates (FMG: OR=1.11) and doctors with additional

academic credentials beyond M.D. (MD_Plus: OR=1.37) tend to order tests on more of

their patients than American-trained and M.D. only primary care doctors. Malpractice

(whether or not the doctor has been sued in past 10 years) has no significant effect. It is

certainly possible that the small 'event rate' for the malpractice variable (Number

Yes=7/148) contributes to the lack of significance. However, the other two (significant)

provider variables also had small numbers of yes/true values (FMG=8/148,

MD_Plus=14/148).

The two variables measuring practice size were both positively related to likelihood

of imaging. For the (categorical) number of patients in each provider's panel, the three

levels that were greater than reference (<500) had from 10-16% more likelihood to

obtain imaging. The number of active providers practicing in each of the 15 clinics was


104









slightly positively associated with greater tendency for assigned patients to get imaging

(OR=1.014) which translates into a 20% increase over the range (5-18).

Multivariable (Poisson) Modeling: Imaging Intensity (Non-zero)

The Poisson model on the 31,660 observations with non-zero outcomes had

Deviance/df of 0.6613 and Chi-Squared/df of 0.7935. When the same model was run

with the DSCALE and PSCALE options the resulting scale parameters were 0.8132 and

0.8908 respectively. This implies that the Poisson distribution of the outcome (number

of outpatient images ordered by the linked/loyal doctor) for these non-zero patients is

underdispersed and that the standard errors for the coefficients might tend to be slightly

overestimated. Therefore, inferences about the significance of variables and levels

would, if anything, be conservative and can be discussed with some confidence. These

uncorrected coefficients and standard errors are presented in Tables 7-7 and 7-8, and

Figure 7-3 displays these same results in terms of 95% Wald confidence intervals.

When discussing the parameter coefficient estimates in terms of effect size and

direction on imaging intensity, it is important to recall that the link function for the

Poisson model was log rather than linear. This means that we can't translate the value

of the estimate into an additive number of imaging tests per patient for the variable or

level in question. However after exponentiation (far right columns in Tables 7-7 and 7-

8), the values can be interpreted as multiplicativee) incident rate ratios (IRR). To keep

these in perspective, recall that the intercept for a null Poisson model on the 31,660

non-zero observations is 0.6548 which after exponentiation is 1.925. As noted above,

this is the Poisson Lambda for the non-zero imaging counts and can be interpreted as

the expected number of images that the 'average' (non-zero) patient would have in two

years.


105









For the categorical variables, using reference cell encoding, the interpretation of

the exponentiated coefficients is straightforward. For example, females have about

1.027 times more imaging tests than males (all else equal). Therefore, for the average

female patient that had imaging, the expected count would be 1.925 times 1.027 or

1.977 images over two years. For the numerical variables the exponentiated estimate

represents a multiplier applied for each additional unit value. Thus for patient age over

the range of 17-103 years there is an 86 year difference which translates into

Exp(86*0.0053) = 1.58 times more imaging tests between the youngest and oldest

patients (all else equal). Moving on to race, we note that Black and Hispanic patients

tend to have more imaging tests than whites (the reference level). With patient

insurance category, only Medicare is significant compared with uninsured/self-pay with

IRR of ~0.92. Recalling the logistic results, we can say that patients with Medicare are

less likely to have any imaging and those that do tend to have a lower number of

imaging tests than patients in other payer categories.

As with the logistic analysis for any imaging use, number of active outpatient

medications had no effect on number of imaging tests. With three exceptions, the binary

clinical problem variables were not significant. These were chronic renal failure (CRF),

diabetes, and hypertension. All else equal, patients with these clinical problems tended

to get fewer imaging tests than those without them. The variable representing the count

of other clinical problems was significantly and positively associated with a larger

number of imaging tests as the problem count increased (IRR=1.003 per additional

problem).


106









The clinical activity variable most strongly associated with the number of imaging

tests was the summed RVU of visits to the linked (loyal) primary care doctor. Over the

range of this variable, patients with 17 RVU worth of visits to their linked primary care

doctor had about 1.6 times more imaging tests than those with a single visit (totaling

less than one RVU). Visits to specialists were also positively associated with number of

imaging tests, though much less strongly with IRR for each additional RVU of 1.006

compared with 1.029. On the other hand, when patients saw other (covering) primary

care doctors, the number of imaging tests ordered by their own (linked/loyal) doctor was

slightly lower. The only other clinical activity variable that was significantly associated

with the number of imaging tests was the total inpatient length of stay which had a

negative effect. For example, one patient spent a total of 179 days in the hospital over

the two years of study. That person would be expected to have less than 30%

(Exp(179*-0.007)=0.29) of the number of imaging tests as a patient that had not been in

hospital at all during the study. This may seem counter intuitive at first, but consider that

patients who spend many days in hospital have more imaging tests performed there.

Since the results of these tests are available to the patient's primary care doctor, he or

she would likely find the answers to their diagnostic questions in tests already

performed and not need to order new ones in the outpatient setting. Lastly, the variables

representing outpatient imaging tests ordered by other doctors (specialists and covering

primary care doctors) and while in the hospital were positively associated with the

number of outpatient imaging tests ordered by the patient's linked (loyal) doctor.

Focusing on provider and site level variables (Table 7-8) and recalling the any

imaging (logistic) results, a similar pattern emerges for the provider experience and


107









gender variables. Over the observed range of provider experience (5-50 years) the most

experienced clinicians ordered about 87% as many tests as the least experienced.

Female physicians ordered more tests than males but the difference was only about

7%. Doctors with additional training after their M.D. ordered about 10% more imaging.

As with the any imaging (logistic) analysis, the provider's malpractice status had no

effect on number of images ordered. The two middle practice size categories (500-799,

800-1000) had significant but small positive effects on number of images compared with

reference (<500). The final provider variable, foreign medical graduate status

(FMG_Yes), had a significant (p=0.023) negative effect (4%) on number of images per

patient whereas the same variable was associated with a greater likelihood (11%) of

ordering any imaging (logistic) on a given patient. Finally, the clinic size (number of

doctors) variable had a small but significant positive effect on the number of imaging

tests ordered on assigned patients. Over the range of practice sizes of 5-18 doctors,

this translates into about 4% more imaging tests.

Comparison of Any Imaging and Imaging Intensity Results

To compare the effect of the various independent variables and levels on any

imaging use (logistic) and imaging intensity (non-zero Poisson) it is useful to plot the

respective odds ratios and coefficients. Figure 7-4 shows odds ratios and coefficients

for the 32 variables / levels that were significant for either any imaging use or imaging

intensity. Figure 7-5 shows the 9 variables that had relatively small effect sizes (odds

ratio near one, coefficient near zero) with the axes scaled down to show the

relationships to better advantage. These were all numeric rather than categorical

variables which explains the small effect sizes (for a unit change in value) that are still

quite significant. Of these, only one (summed RVU of visits to covering primary care


108









doctors) was significant for imaging intensity (non-zero Poisson) but not for any imaging

use (Logistic).

In both Figures 7-4 and 7-5, note that the general tendency is for variables and/or

levels to be concordant with respect to their effect size and direction (when both are

significant) between any imaging use (logistic odds ratio, X Axis) and imaging intensity

(Poisson regression coefficient, Y Axis). There is only a single exception and this is with

the variable coding for whether or not the linked (loyal) primary care doctor is a foreign

medical graduate (FMG). It seems that FMG primary care doctors were more likely to

order some imaging but ordered fewer imaging tests when they did so.

Preparation for Multilevel Modeling: Imaging Propensity Scores

Patient level predictions from the zero inflated Poisson (ZIP) model using all 28

patient level variables for both the count and zero-model portions were calculated,

called IMG_PROP, and stored in a new data set along with the original outcome (IMG)

and the (coded/anonymous) provider ID of that patient's linked primary care doctor.

Table 7-9 compares the raw outcome (IMG) and the predictions from the ZIP model

(IMG_PPOP).

The mean of IMG_PROP for each provider was subtracted from the original value

for each patient to form a centered imaging propensity variable (clMG_PROP), which

can be expressed as Equation 7-1.

clMG_PROPij = IMG_PROPij IMG_PROP.j (7-1)

In Equation 7-1, i is the ith patient, j is the jth provider, and .j is the average for the

jth provider.

The result of the above described centering operation is illustrated by plotting the

raw imaging propensity scores (IMG_PROP) and centered imaging propensity scores


109









(clMG_PROP) against the mean outcome (images per patient) for each provider. These

are shown in Figures 7-6 and 7-7 respectively.

Multi-Level (Hierarchical) Modeling

The SAS PROC MIXED procedure on the full two level model completed in -20

seconds and converged after 4 iterations. Results from SAS are reproduced in Tables

7-10 through 7-15.

These results are summarized in Table 7-16 in terms of fixed (patient level) effects

and random (patient and provider levels) variance.

The fixed effect intercept (0.7171) is the average doctor's mean imaging. This is

nearly identical to the raw mean value of 0.7146 obtained by dividing the total number of

imaging tests ordered by the patient's linked provider (N=60,938) by the number of

patients in the whole study cohort (N=85,277). A (95%) range of plausible values for

doctor's mean imaging (intercept) around 0.7171 can be constructed using the variance

(0.0835) by 0.7171 1.96(0.0835)1/2 which gives (0.151, 1.283).

Note that this is rather wider than the 95% confidence intervals (0.6695, 0.7648)

on the estimate of the fixed intercept provided by SAS, which used the standard error.

The fixed effect of clMG_PROP (0.9919) is interpreted as the average doctor's

response (slope) in number of imaging tests ordered for a unit change in the imaging

propensity score (clMG_PROP). The fact that this is very close to 1.0 implies that the

scale of the imaging propensity score is correct (at least around the mean value). In

other words, on average, as the expected amount of imaging increases by one 'unit',

actual imaging utilization increased by one extra test per patient. As with the intercept,


110









we can calculate a (95%) range of plausible values for the slopes using the variance

(0.1567) by 0.9919 1.96(0.1567)1/2 which gives (0.216, 1.768).

As with the intercept, the plausible range of the slope is wider than the SAS

calculated 95% confidence interval on the slope parameter estimate (0.925, 1.058).

Intraclass correlations (ICC) for intercept and slope were obtained by dividing each

component variance (Too for intercept and T11 for slope) by the residual variance (2)

plus itself. For intercept, this is 0.0835 divided by (0.0835+1.1934) which gives 0.065.

For slope, the ICC is 0.1567 divided by (0.1567+1.1934) which gives 0.116. This implies

that about 6% of the variance in intercepts is between doctors and about 12% of the

variance in slopes is between doctors. It is helpful to recall that an ICC of 0 would mean

that all doctors exhibit the same IMG/clMG_PROP relationship and clustering of

patients by doctor had no effect (i.e., the hierarchical modeling not informative). On the

other hand, an ICC approaching 1 would mean that any given doctor's patients have

nearly identical adjusted imaging utilization and very small variation between them.

Model based estimates (including standard errors and 95% confidence intervals)

of individual provider intercept and slope were obtained by requesting the solution for

the random portion of the model. Reliability for the individual provider intercept

estimates can be calculated as the overall provider intercept variance (0.1567) divided

by the provider's own variance (standard error squared) and the overall reliability for

provider intercept estimates is the average of our 148 doctors which is 0.965. Similarly,

the individual reliability for each provider's slope estimate is the overall variance

(0.0835) divided by the individual variance (standard error squared) with the aggregate

reliability being the average of these for the 148 doctors which is 0.939. The high


111









reliability of both the intercept and slope estimates is reassuring and supports

interpreting them as representing each provider's mean tendency to order imaging tests

(intercept) and their response (slope) to patient level imaging propensity represented by

the clMG_PROP (risk adjusted expected imaging) variable.

The correlation between each doctor's general tendency to order imaging

(intercept) and his or her response to patient imaging propensity (slope) can be

expressed as p(poj Plj) which is estimated by ToI/(Too cl)1/2. Substituting from Table

7-16 gives 0.0810 / (0.0835 x 0.1567)1/2 which turns out to be 0.7081. This implies a

substantial correlation between the average tendency to use imagining, and the

increase in the number of images providers order on their patients with higher imaging

propensity (i.e., sicker). A scatter plot of the intercepts (X axis) and slopes (Y axis) for

all 148 providers is shown in Figure 7-8 and serves to visualize the relationship between

them. The quadrants in this slope versus intercept plot are labeled A-D and are further

detailed in Tables 7-17 and 7-18.

Less than 20% of the providers have discordance between their slopes and

intercepts (quadrants A and D) with the remaining 80% (quadrants B and C) being

concordant with respect to their tendency to image and their response to patient

imaging 'need' as represented by the imaging propensity variable. Therefore, in general,

if a primary care doctor tends to order imaging less than average, odds are 4.5:1 that he

or she will also increase the amount of images that they obtain on 'sicker' patients less

than average. For example, provider C.C. obtained about 29 images per 100 patients

over the course of the study and increased their image utilization by about 15 images

per 100 for each unit increase in imaging propensity. It is useful to plot imaging


112









propensity versus actual imaging utilization for all the patients cared for by Doctor C.C.

and this is shown in Figure 7-9.

Likewise, if a doctor tends to order imaging more than average, odds are 3.7:1 that

he or she will increase the amount of images that they order on 'sicker' patients more

than average. For example, provider B.B. obtained about 199 images per 100 patients

over the course of the study and increased their image utilization by about 264 images

per 100 for each unit increase in imaging propensity. The observed imaging versus

imaging propensity for Dr. B.B. is plotted in Figure 7-10.

The discordant doctors (quadrants A and D in Figure 7-8) are not only few in

number but tend to cluster near the mean value of the slope (with standard errors

overlapping the average slope of zero) such that only three doctors are significantly

discordant with one having low intercept / high slope and two having high intercept / low

slope. For example, the provider labeled A.A. in the scatterplot obtained about 50

images per 100 patients over the two years of study and increased imaging by about

119 images per 100 for every unit increase in imaging propensity. Another way of

saying this is that Doctor A.A. has a somewhat higher threshold for obtaining imaging

on any given patient but tends to order more images as patient need for imaging

increases.

In contrast, provider D.D. obtained about 85 images per 100 patients and

increased their imaging by about 58 images per 100. One might speculate that provider

D.D. is generous with imaging in general but less discriminating in terms of increasing

utilization according to patient need. On the other hand, the two quadrant D providers

(D.D. and the one just above on the scatter plot) may be relatively liberal in terms of


113









both imaging and referral to specialists. This might mean that their sicker patients get

less images ordered by the primary care provider because they tend to refer at a lower

threshold and at least some images would be ordered by the specialists rather than

themselves.

Another interesting visualization is to plot the only the intercepts (provider's mean

imaging) and 95% confidence intervals sorted according to practice site, and this shown

in Figure 7-11. Similarly, a plot of the slopes (provider's change in imaging as imaging

propensity increases) with 95% confidence intervals is shown as Figure 7-12. One

important observation from the provider slopes plot (Figure 7-12) is that all providers

increase their diagnostic image ordering in response to additional patient need (none of

the scaled imaging vs. propensity slopes are below zero). The alternative is that some

doctor's slopes could be negative such that the amount of imaging actually decreased

for sicker patients. One explanation for this (counterfactual) would be that those with

negative slopes, refer sicker patients to specialists who themselves obtain the needed

imaging tests. The fact that this did NOT occur at all in the current study implies that

even sick patients who see many specialists continue to have at least some of their care

rendered by the linked (loyal) primary care provider. This should come as no surprise

given that demonstration of this 'loyalty' relationship was the main inclusion criteria for

both patients and providers.

The reduced level 1 model (IMG = IMG_PROP) yielded a patient level residuals

(error variance) of 1.3061 while the null model (IMG=;) gave patient level residuals

(error variance) of 1.5782. These are combined with the error variance from the full two

level model (1.1934) in Table 7-19.


114









The overall amount of explained variation in outpatient imaging utilization after

accounting for provider ID and all the patient level variables (risk-adjustment) is 24.4%

(0.3848/1.5782). Of that, roughly 70% (0.2721/0.3848) is attributable to patient level

factors as captured in the imaging propensity variable and the remaining 30% is

attributable to provider variation. One implication is that about three quarters of the

variation in the number of outpatient imaging tests ordered by a primary care doctor on

loyal patients is unexplained. This is despite taking into account a robust and large set

of patient factors as well as all between doctor differences in imaging utilization habits

(by directly modeling unique provider identity). Of the roughly 25% of variation in

primary care outpatient imaging utilization that can be explained, the majority (70%) is

attributed to factors that mostly relate to each patient's clinical 'need' for imaging,

regardless of who their primary care doctor is. The remaining 30% arises from

differences in the tendency for primary care doctors to order imaging which may be

partitioned into intercept (~10%) and slope (~20%) components. The next chapter will

summarize and discuss these results in terms of advances in knowledge, applications in

imaging utilization management and provider profiling. Directions for future research

with these (and similar) data sources as well as some policy implications will be covered

as well.


115










Table 7-1. Spearman correlations between clinical activity and other imaging variables.


Variable Name




all i cnt
spec_o_cnt
pcp_o_cnt
spec_visit_count
pcp_visit_count
prv_visit_count
spec_visit_rvu
pcp_visit_rvu
prv_visit_rvu
ervisits
erhours
obs_stays
inpt_stays
inpt_los_total
inpt_icu_days
inpt_read_l 5d
inpt_read_31d


all e cnt
all i cnt
spec_o_cnt
pcp_o_cnt
spec_visit_count
pcp_visit_count
prv_visit_count
spec_visit_rvu
pcp_visit_rvu
prv_visit_rvu
ervisits
erhours
obs_stays
inpt_stays
inpt_los_total
inpt_icu_days
inpt_read_l 5d
inpt_read_31d


CD







0.450
0.226
0.095
0.260
0.111
0.253
0.260
0.115
0.254
0.843*
0.857*
0.295
0.469
0.474
0.203
0.243
0.283


0.273
0.056
0.278
0.048
0.230
0.279
0.053
0.239
0.423
0.429
0.148
0.726*
0.742*
0.366
0.343
0.398


0.107
0.574*
0.132
0.241
0.575*
0.135
0.249
0.228
0.229
0.250
0.303
0.299
0.101
0.106
0.126


0.195
0.342
0.975*
0.199
0.351
0.272
0.273
0.298
0.331
0.327
0.110
0.121
0.141


_0

I,-
0-

















0.131
0.130
0.076
0.066
0.066
0.011
0.028
0.030
0.030


0.128
0.978*
0.263
0.267
0.136
0.263



0.231
0.232
0.091
0.099
0.116


CD
,.
$-


















0.996*
0.290
0.455
0.459
0.184
0.231
0.267


0.992*
0.283
0.302
0.352


0.857*


Key To Variable Names Is Below
count of images done in ER
count of images done as inpatient
count of outpatient images ordered by specialists
count of outpatient images ordered by covering PCP
count of outpatient visits to specialists
count of outpatient visits to covering PCP
count of outpatient visits to loyal doc
sum of RVU of outpatient visits to specialists
sum of RVU of outpatient visits to covering PCP
sum of RVU of outpatient visits to loyal doc
count of ER visits
total hours in the ER
count of observation stays
count of inpatient stays
total days in hospital
total days in ICU
count of readmit within 15 days
count of readmit within 31 days


NOTE: For brevity, columns where ALL correlations were < 0.5 are omitted.
* Correlations above 0.5.


116











Table 7-2. Bivariate relationship between patient level variables and outcome (imaging


counts).
Type Description
Categorical patient identified race
Categorical patient sex
Numeric patient age in 2008
Categorical patient's payer of record
in 2008
Active prescriptions in
Categorical 2008
Binary coronary artery disease
Binary cancer
Binary congestive heart failure
Binary chronic renal failure
Binary diabetes
Binary obesity
Binary hypertension
Binary substance abuse
Binary trauma
count of active problems
Numericr tn t
other than those listed


Variable Name
Race
Sex
age_08
PayerGroup

meds cat
pr_cad
pr_can
pr_chf
pr_crf
pr_dm
pr_obs
pr_htn
pr_sub
pr_trm
oth_prb


F Value
29.98
258.44
3087.75
343.33

1313.04
445.63
476.77
123.35
132.79
436.64
150.85
628.14
10.19
48.85
3115.43


R-Squared
0.0011
0.0030
0.0349
0.0236

0.0442
0.0052
0.0056
0.0014
0.0016
0.0051
0.0018
0.0073
0.0001
0.0006
0.0353


Correlation
0.0324
0.0550
0.1869


p value
<0.0001
<0.0001
<0.0001


0.1536 <0.0001

0.2101 <0.0001


0.0721
0.0746
0.0379
0.0394
0.0713
0.0421
0.0855
0.0110
0.0239


<0.0001
<0.0001
<0.0001
<0.0001
<0.0001
<0.0001
<0.0001
0.0014
<0.0001


0.1877 <0.0001


117










Table 7-3. Bivariate relationship between clinical activity variables and outcome


(imaging counts).
Type Description
Numeric total hours in the ER
Numeric count of ER visits
Numeric total days in ICU
Numeric total days in hospital
eric count of readmit
Numeric
within 15 days
eric count of readmit
Numeric
within 31 days
Numeric count of inpatient
Numeric
stays
Numeric count of observation
Numeric
stays
Numc count of images done
Numeric
in ER
Numeric count of images done
Numeric
as inpatient
count of outpatient
Numeric images ordered by
covering PCP
count of outpatient
Numeric images ordered by
specialists
Numeric count of outpatient
Numeric
visits to covering PCP
sum of RVU of
Numeric outpatient visits to
covering PCP
Numeric count of outpatient
Numeric
visits to loyal doc
sum of RVU of
Numeric outpatient visits to
loyal doc
Numeric count of outpatient
Numeric
visits to specialists
sum of RVU of
Numeric outpatient visits to
specialists


Variable Name
erhours*
ervisits
inpt_icu_days
inpt_lostotal
inpt_readl 5d*

inpt_read_31d

inpt_stays

obs_stays

all e cnt

all i cnt


pcp_o_cnt


spec o cnt

pcp_visitcount*

pcp_visit_rvu

prv_visitcount*

prv_visit_rvu

specvisitcount*

spec visit rvu


F Value
1323.13
1412.2
78.36
753.11
214.52


R-
Squared
0.0153
0.0163
0.0009
0.0088
0.0025


273.09 0.0032

1562.36 0.0180

647.26 0.0075

1281.08 0.0148

663.8 0.0077

581.73 0.0068


2729.8 0.0310

436.63 0.0051

536.26 0.0063

14540.96 0.1457

15228.91 0.1515

4786.98 0.0532

5683.23 0.0625


Correlation
0.1236
0.1276
0.0303
0.0935
0.0501


p value
<0.0001
<0.0001
<0.0001
<0.0001
<0.0001


0.0565 <0.0001

0.1341 <0.0001

0.0868 <0.0001

0.1217 <0.0001

0.0879 <0.0001

0.0823 <0.0001


0.1761 <0.0001

0.0713 <0.0001

0.0791 <0.0001

0.3817 <0.0001

0.3893 <0.0001

0.2305 <0.0001

0.2500 <0.0001


* Variables NOT carried forward to multivariable analysis.


118










Table 7-4. Bivariate relationship between provider and clinic level variables and


outcome (imaging coui

Type Description
whether provider has
Categorical been sued in last 10
years
whether provider is
Categorical foreign medical
graduate
whether provider has
Categorical
a degree beyond MD
number of patient's in
Categorical provider practice in
2008
Categorical provider sex
Numeric age in years of the
provider in 2008
number of years after
Numeric provider MD
graduation in 2008
Identifier anonymous
provider identifier
number of doctors
Numeric actively practicing at
the clinic in 2008
Identifier site (clinic) identifier
* Variables NOT carried forward


Variable
Name

mp_flag


prov_fmg

prov_md_plus

prov_pat_cat

prov_sex
prov_age_08*

prov_exp_08

provide*

sitedocs


site id*


F R-
Value Squared

1.05 0.0000


0.17 0.0000

32.71 0.0004

79.4 0.0028

66.9 0.0008
1.3 0.0000

14.37 0.0002

31.43 0.0514

189.38 0.0022

100.16 0.0162


Correlation


p value


0.0032 0.3054


0.0000 0.6761

0.0195 <0.0001

0.0528 <0.0001

0.0279 <0.0001
0.0045 0.2542

0.0130 0.0002

0.2268 <0.0001

0.0471 <0.0001

0.1272 <0.0001


to multivariable analysis.


119









Table 7-5. Patient level results from multivariable logistic model on


Type
Demographics

Race


Insurance





Medications


Problems


Visits


Other Imaging



Hospital


Variable /
Level
Age
Sex F
Black
Hispanic
Other
BCBS
Commercial
Managed
Medicare
State
Other
1-5
6-10
>10
CAD
Cancer
CHF
CRF
Diabetes
Obesity
Hypertension
Substance
Abuse
Trauma
Other (count)
prv_visit_rvu
pcp_visit_rvu
specvisitrvu
pcp_o_cnt
spec_o_cnt
all e cnt
all i cnt
ervisits
obs_stays
inpt_stays
inpt_read_31d
inpt_los_total
inpt_icu_days


Standard
Estimate t r
Error
0.016 0.001
0.083 0.018
0.212 0.035
0.232 0.032
0.091 0.029
0.133 0.067
0.092 0.071
0.168 0.068
-0.286 0.070
0.028 0.072
0.016 0.078
0.044 0.037
0.044 0.040
-0.017 0.046
-0.026 0.039
-0.052 0.025
-0.268 0.079
-0.113 0.069
-0.316 0.027
0.013 0.026
-0.243 0.019
-0.232 0.085


0.215
0.018
0.152
-0.006
0.025
0.146
0.024
0.005
0.020
-0.012
0.067
-0.022
-0.077
-0.013
-0.013


0.053
0.001
0.002
0.006
0.002
0.019
0.004
0.009
0.007
0.014
0.016
0.024
0.053
0.004
0.015


Chi-
Square
483.106
22.011
36.472
52.404
9.695
3.933
1.712
6.065
16.570
0.150
0.041
1.381
1.234
0.134
0.463
4.491
11.453
2.650
141.985
0.267
160.556
7.440
16.426
193.493
4253.497
1.115
276.583
62.486
32.949
0.282
8.978
0.730
17.880
0.868
2.070
10.709
0.736


any imaging use.
Pr > Odds
ChiSq Ratio
<0.0001 1.016
<0.0001 1.087
<0.0001 1.236
<0.0001 1.261
0.0018 1.095
0.0473 1.143
0.1908 1.097
0.0138 1.183
<0.0001 0.752
0.6982 1.028
0.8390 1.016
0.2400 1.045
0.2666 1.045
0.7146 0.983
0.4961 0.974
0.0341 0.949
0.0007 0.765
0.1036 0.893
<0.0001 0.729
0.6056 1.013
<0.0001 0.784
0.0064 0.793


<0.0001
<0.0001
<0.0001
0.2910
<0.0001
<0.0001
<0.0001
0.5952
0.0027
0.3929
<0.0001
0.3515
0.1503
0.0011
0.3910


1.240
1.019
1.164
0.994
1.025
1.157
1.025
1.005
1.020
0.989
1.070
0.978
0.926
0.987
0.987


Reference Levels: Sex-Male, Race-White, Insurance-Uninsured, Medications-None


120









Table 7-6. Provider and clinic level results from multivariable logistic model on any


imaging use.
Variable /
Level
Experience
sex F
FMG Yes
MD Plus Yes
Malpractice
Yes


500-759
750-999
1K+


Estimate
-0.003
0.133
0.108
0.318
0.020


0.123
0.150
0.098


Standard
Error
0.001
0.018
0.034
0.029
0.040


0.028
0.027
0.025


Chi-
Square
10.993
55.406
10.412
117.859
0.262


18.621
31.511
14.974


Odds
Pr > ChiSq R
Ratio
0.0009 0.997
<0.0001 1.142
0.0013 1.114
<0.0001 1.374
0.6085 1.021

<0.0001 1.130


<0.0001
0.0001


1.162
1.103


Active
Clinic size ve 0.014 0.002 69.051 <0.0001 1.014
Providers
Reference levels: Sex-Male, FMG-No, MD_Plus-No, Malpractice-No, Provider Patients-
<500


121


Type
Provider


Provider
Patients









Table 7-7. Patient level results from multivariable Poisson model on


Type


Demographics

Race


Insurance





Medications


Problems


Visits


Other Imaging



Hospital


Variable /
Level


Age
Sex F
Black
Hispanic
Other
BCBS
Commercial
Managed
Medicare
State
Other
1-5
6-10
>10
CAD
Cancer
CHF
CRF
Diabetes
Obesity
Hypertension
Substance
Abuse
Trauma
Other (count)
prv_visit_rvu
pcp_visit_rvu
spec_visit_rvu
pcp_o_cnt
spec_o_cnt
all e cnt
all i cnt
ervisits
obs_stays
inpt_stays
inptread_31d
inpt_los_total
inpt_icu_days


Standard
Estimate
Error


0.0053
0.027
0.034
0.073
-0.013
-0.047
-0.052
-0.031
-0.085
-0.048
0.000
0.001
0.010
0.028
-0.028
0.013
-0.047
-0.078
-0.058
0.018
-0.058
0.063
0.013
0.003
0.029
-0.009
0.006
0.026
0.007
-0.004
0.012
-0.005
0.006
0.010
0.023
-0.007
-0.010


0.0004
0.010
0.018
0.016
0.017
0.039
0.041
0.040
0.040
0.041
0.045
0.024
0.025
0.027
0.017
0.012
0.033
0.030
0.013
0.013
0.010
0.040
0.026
0.001
0.001
0.003
0.001
0.008
0.002
0.003
0.002
0.005
0.007
0.010
0.021
0.002
0.005


Chi-
Square
183.800
8.220
3.480
19.500
0.650
1.440
1.590
0.620
4.520
1.360
0.000
0.000
0.170
1.060
2.730
1.200
2.040
6.860
21.250
2.060
36.160
2.520
0.270
32.560
1479.330
11.390
90.200
10.740
18.360
1.640
26.340
0.790
0.870
1.030
1.200
17.580
3.660


imaging intensity.
Pr> (RR)
Exp
ChiSq Exp
Estimate
<0.0001 1.005
0.0042 1.027
0.062 1.034
<0.0001 1.075
0.4198 0.987
0.2305 0.954
0.2071 0.949
0.4305 0.969
0.0335 0.918
0.2435 0.953
0.9932 1.000
0.959 1.001
0.6787 1.010
0.3028 1.028
0.0982 0.972
0.2725 1.013
0.153 0.954
0.0088 0.925
<0.0001 0.944
0.1508 1.019
<0.0001 0.944
0.1124 1.065


0.6011
<0.0001
<0.0001
0.0007
<0.0001
0.001
<0.0001
0.200
<0.0001
0.3754
0.3503
0.3101
0.274
<0.0001
0.0558


1.013
1.003
1.029
0.991
1.006
1.026
1.007
0.996
1.012
0.996
1.006
1.010
1.023
0.993
0.990


Reference Levels: Sex-Male, Race-White, Insurance-Uninsured, Medications-None


122









Table 7-8. Provider and clinic level results from multivariable Poisson model on imaging
intensity.
Variable / Estimate Standard Chi- Pr> (RR) Exp
TypeLevel EstimatError Square ChiSq Estimate
Level Error Square ChiSq Estimate


Experience
sex F
FMG Yes
MD Plus Yes
Malpractice
Yes


500-759
750-999
1K+


-0.003
0.063
-0.041
0.097
-0.008

0.044
0.084
0.026


Active
Clinic size ve 0.003
Providers
Reference levels: Sex-Male, FMG-No, MD
<500


0.001
0.009
0.018
0.015
0.021


28.060
45.860
5.150
40.980


<0.0001
<0.0001
0.0232
<0.0001


0.140 0.7091


0.015 8.540 0.0035


0.014
0.014


35.880
3.630


<0.0001
0.0567


0.001 12.460 0.0004


0.997
1.065
0.960
1.102
0.992


1.044
1.087
1.026
1.003


_Plus-No, Malpractice-No, Provider Patients-


Table 7-9. Univariate statistics for raw imaging counts (IMG) and predictions from ZIP
model (IMG_PROP).
Statistic IMG IMGPROP


85277


85277


Minimum 0 (N=53,617) 0.0446
Maximum 15 *14.8, 15,2, 20.1
Mean 0.7146 0.7164
Standard Deviation 1.256 0.5651
Skewness 2.655 3.844
Coefficient of Variation 176 78
Sum of Observations 60938 61088
Variance 1.578 0.319
Kurtosis 10 46
Standard Error of the 0
0.004 0.002
Mean
NOTE: The highest 3 observations (*) are shown for IMG_PROP Maximum.

Table 7-10. Dimensions.
Covariance Parameters
Columns in X
Columns in Z Per Subject
Subjects
Max Obs Per Subject


4
2
2
148
2101


123


Provider


Provider
Patients









Table
Row
1
2


7-11. Estimated G correlation matrix.
Effect provide Coll
Intercept 10562 1.0000
clMG PROP 10562 0.7081


Col2
0.7081
1.0000


Table 7-12. Covariance parameter estimates.
Cov Parm Subject Estimate Standard Error
UN(1,1) prov_id 0.08345 0.01003
UN(2,1) prov_id 0.08098 0.01195
UN(2,2) prov_id 0.1567 0.01953
Residual 1.1934 0.005789
NOTE: UN(1,1) = Too, UN(2,2) = Tl, UN(2,1) = TOI


Table 7-13. Fit statistics.
-2 Res Log Likelihood
AIC (smaller is better)
AICC (smaller is better)
BIC (smaller is better)


Z Value
8.32
6.78
8.02
206.13


257971.1
257979.1
257979.1
257991.1


Table 7-14. Solution for fixed effects.


Effect
Intercept
clMG PROP


Estimate Standard Error
0.7171 0.02412
0.9919 0.03364


DF
147
147


t Value
29.74
29.48


Pr > Itl
<0.0001
<0.0001


Table 7-15. Type 3 tests of fixed effects.
Effect Num DF Den DF
clMG PROP 1 147


F Value
869.19


Table 7-16. Results from multi-level random coefficients model.
Standard
Fixed Effect Symbol Coefficient Standa t-
Error


Intercept
clMGPROP


Random Effect
Patient
Residual
Provider
Intercept
Provider Slope
Covariance:
Intercept, Slope


0.7171
0.9919


Symbol


eij (&2)

Uoj (Too)
ulj (z 1)
Cov(uoj,u1j)
(T01)


Variance
Component
1.1934

0.0835
0.1567
0.0810


0.02412
0.03364


Standard
Error
0.00579

0.01003
0.01953
0.01195


value

29.74
29.48


z-value

206.13

8.32
8.02
6.78


p-value
<0.0001
<0.0001


p-value

<0.0001

<0.0001
<0.0001
<0.0001


124


PrZ
<0.0001
<0.0001
<0.0001
<0.0001


Pr> F
<0.0001









Table 7-17.
Quadrant


A
B
C
D


Quadrants in intercept versus slope relationship plot.
Tendency To Response To Number Of
Order Imaging: Imaging Propensity: Providers
Intercept Slope
Low (<0) High (>0) 1
High (>0) High (>0) 5:
Low (<0) Low (<0) 6'
High (>0) Low (<0) 1l


Percent Of
Providers


9.46
35.14
45.27
10.14


Table 7-18. Exemplary providers in each quadrant.
Provider Tendency To Order Imaging:
Intercept


Response To Imaging Propensity:
Slope


A.A. -0.2176 0.1988
(0.4995) (1.1907)
B.B. 1.2708 1.6451
(1.9879) (2.637)
C.C. -0.4268 -0.8387
(0.2903) (0.1532)
D.D. 0.1310 -0.4087
(0.8481) (0.5832)
NOTE: Numbers in parentheses for intercept are adjusted to the fixed effect mean by
adding 0.7171 and numbers in parenthesis for slope are adjusted to the fixed effect
mean by adding 0.9919.

Table 7-19. Comparison of null, and reduced model residuals with full model.
Patient Provider Provider Residual Absolute Fraction
Model Name
IMGPROP Intercept Slope (error) Reduction Reduction
Null 1.5782


Provider
Intercept Only
Imaging
Propensity
Only*
Two Level:
Provider Slope
Two Level:
Provider
Intercept
Two Level:
Provider
Intercept and
Slope


1.4996 0.0786 0.0498


1.3061


0.2721


0.1724


1.2727 0.3055 0.1936

1.2389 0.3393 0.2150


1.1934 0.3848 0.2438


NOTE: Fraction reduction in the residual is equivalent to R-Squared for that model.
*NON Centered. If centered, error=1.3180.


125












60000


Observed Counts
- -Simple Poisson
S..... Nnn-7prn Pnissnn


50000
5 0 Zero-inflated F


40000 -



0 30000 \
O \


20000



10000



0 .-...... *-*
0 1 2 3 4 5 6 7 8
Number Of Imaging Tests

Figure 7-1. Comparison of imaging counts with three Poisson distributions.


'oisson


9 >10


126














DEMOGRAPHICS [Age]
[Sex F]
RACE [Black]
[Hispanic]
[Other]
INSURANCE [BCBS]
[Commercial]
[Managed]
[Medicare]
[State]
[Other]
MEDICATIONS [1-5]
[6-10]
[>10]
PROBLEMS [CAD]
[Cancer]
[CHF]
[CRF]
[Diabetes]
[Obesity]
[Hypertension]
[Substance Abuse]
[Trauma]
PROBLEMS [Other (count)]
VISITS [prv_visit_rvu ]
[pcp_visit_rvu ]
[spc_visit_rvu ]
OTHER IMAGING [pcp_o_cnt ]
[spc_o_cnt ]
[all_e_cnt ]
[all i cnt ]
HOSPITAL [er_visits]
[obs_stays ]
[inpt_stays ]
[inpt_read_31d ]
[inpt_los_total ]
[inpt_icu_days ]
PROVIDER [Experience ]
[sex F]
[FMG Yes]
[MD Plus Yes]
[Malpractice Yes]
PROVIDER PATIENTS [500-759]
[750-999]
[1K+]
CLINIC SIZE [Active Providers ]


1.016
1.087
1.236
1.261
1.095
1.143
1.097
1.183
0.752
1.028
1.016
1.045
1.045
0.983
0.974
0.949
0.765
0.893
0.729
1.013
0.784
0.793
1.240
1.019
1.164
0.994
1.025
1.157
1.025
1.005
1.020
0.989
1.070
0.978
0.926
0.987
0.987
0.997
1.142
1.114
1.374
1.021
1.130
1.162
1.103
1.014


IrII


I I II


I I I I


II


I I


II


Ii
I:


0.5 0.6 0.7 0.8 0.9 1.0

Odds Ratio


1.1 1.2 1.3 1.4 1.5


Figure 7-2. Logistic regression results for any imaging utilization. Horizontal bars

represent 95% confidence intervals on odds ratio for each variable/level.

Patient variable reference levels: Sex-Male, Race-White, Insurance-

Uninsured, Medications-None, Problems-No. Provider variable reference

levels: Sex-Male, FMG-No, MD_Plus-No, Malpractice-No, Provider Patients-

<500.


127


I I I I


i' "-


I I I I I














DEMOGRAPHICS [Age]
[Sex F]
RACE [Black]
[Hispanic]
[Other]
INSURANCE [BCBS]
[Commercial]
[Managed]
[Medicare]
[State]
[Other]
MEDICATIONS [1-5]
[6-10]
[>10]
PROBLEMS [CAD]
[Cancer]
[CHF]
[CRF]
[Diabetes]
[Obesity]
[Hypertension]
[Substance Abuse]
[Trauma]
PROBLEMS [Other (count)]
VISITS [prv_visit_rvu]
[pcp_visit_rvu ]
[spc_visit_rvu ]
OTHER IMAGING [pcp_o_cnt ]
[spc_o_cnt ]
[all_e_cnt ]
[all i cnt ]
HOSPITAL [er_visits]
[obs_stays ]
[inpt_stays ]
[inpt_read_31d ]
[inpt_los_total ]
[inpt_icu_days ]
PROVIDER [Experience ]
[sex F]
[FMG Yes]
[MD Plus Yes]
[Malpractice Yes]
PROVIDER PATIENTS [500-759]
[750-999]
[1K+]
CLINIC SIZE [Active Providers ]


I I II


i I I


I


II I I


II


I


-0.2 -0.1


0.0 0.1
Beta Estimate


Figure 7-3. Poisson regression results for (non-zero) imaging intensity. Horizontal bars

represent 95% confidence intervals on estimated coefficient for each

variable/level. Patient variable reference levels: Sex-Male, Race-White,

Insurance-Uninsured, Medications-None, Problems-No. Provider variable

reference levels: Sex-Male, FMG-No, MD_Plus-No, Malpractice-No, Provider
Patients-<500.


128


s I I


I


I


I I



















0 10




005
c

0o

0 000
E
(0


0 -005




-0 10


06 07 08 09 10 11
Odds Ratios Any Imaging


12 13 14 15


Figure 7-4. Comparison of significant variables for any imaging use and imaging

intensity. Unless underlined, all variable/levels were significant for both any

imaging use and imaging intensity. The underlined variable/levels were not

significant for imaging intensity, except for chronic renal failure (CRF), which

not significant for any imaging use. The variables indicated by open circles

(near the origin) are shown again in Figure 13 with appropriate axis scaling.


129


diabetes CHF
hypertension

CRF
medicare


-0 15 -
05


provider has
extra degree
0
pract size=750-000
substance race: hispanic
bus i in y provider is female
abuse imaging by
covering pcp
c r pract size=500-799
pat female ,S krae:black
cancer Oract size=lK+ tr a linked (loyal)
-% trauma doc visits
observationdays

race: other


provider managed
is FMG *
BCBS


I





























provider experience



total LOS

covering pcp visits


* inpatient imaging
imaging by
specialists

specialist visits
* other problems
(count)


1 00
Odds Ratios Any Imaging


Figure 7-5. Comparison of significant variables for any imaging use and imaging
intensity (small effect sizes). Unless underlined, all variable/levels were
significant for both any imaging use and imaging intensity. The summed RVU
of visits to covering primary care doctors (underlined) was not significant for
any imaging use.


130


0.02 7


0.01






0.00
E


E


-0


patient age

clinic size docss) *


-0.02 --
0 96















2.0



0. 1.5
0


1.0



0.5



0.0
0.0 0.5 1.0 1.5 2.0
Observed Images/ patient


Figure 7-6. Imaging propensity score distributions by provider. Each provider's (N=148)
mean imaging propensity (diamonds) along with 10th and 90th percentiles
(error bars) are plotted on the Y axis against the observed mean number of
images per patient for that provider (X axis). The grand mean of imaging
propensity (0.73) is indicated by the horizontal line.



1.5


1.0

aI
0
O
.. 0.5

00

w 0.0
oD


0.5 1.0 1.5 2.0
Observed Images / Patient


Figure 7-7. Centered imaging propensity score distributions by provider. Each provider's
(N=148) centered imaging propensity (diamonds) along with 10th and 90th
percentiles (error bars) are plotted on the Y axis against the observed mean
number of images per patient for that provider (X asis). The overall centered
mean of imaging propensity (zero) is indicated by the horizontal line.


131











B.B.

1.5 -+


1.0

A.A.
o 0.5 +*+-

A0 B
0.0
-PF -'-

-0.5
C.c


Intercept

Figure 7-8. Plot of intercept and slopes for all 148 providers obtained from multi-level
model of imaging utilization. Error bars represent + standard error for each
provider's intercept (horizontal) and slope (vertical).
w \ -d^^-^^Intercept^
Fiur 78.Plt f ntrcptan sops oral 148A prov4idrobandfmmut-el
modelofimain utilizaion.Erorbas epesnt_+tadad rorfo ec
prviers nerep (orzntal nsoe(eria)


132
















12

"0
0 10
E

5 8-


2
(D 6-
0)
4-

*") 4 .

2 *** ** *** y= 0.1532x + 0.2903

--* -



-2 ----
-2 -1 0 1 2 3 4 5
Centered Imaging Propensity


Figure 7-9. Imaging utilization versus centered imaging propensity for a low utilizing
doctor (C.C. in Figure 7-8). Each diamond represents a single patient and the
dashed line is the linear equation defined by the multi-level model intercept
and slope for provider C.C.


133













y = 2.637x+ 1.9879 ,


*




**...*...
** 4. 4 ^*
4m.4..4 4 4* '* *


il. **111 ..
- i inu iiiiii.nf
1
jpriii ** *
^ _


Centered Imaging Propensity


Figure 7-10. Imaging utilization versus centered imaging propensity a high utilizing
doctor (B.B. in Figure 7-8). Each diamond represents a single patient and the
dashed line is the linear equation defined by the multi-level model intercept
and slope for provider B.B.


134


0 10
E
L
a, 8
a

6

r-
I) 4
E


-
loo
ol


.









B.B.


rr i


I.C. A
I || A A.A |


r


8 9 10 11 12 13 14 15


Figure 7-11. Provider means sorted by ascending order within each site. Site (clinic)
numbers shown along the bottom. Each provider's intercept is scaled by
adding the model coefficient for each provider to the fixed intercept (0.7171)
and multiplied by 100 to represent adjusted images per 100 patients over the
two year study interval. The solid horizontal line is at 71.7 images per 100
patients which is the grand mean number of images per 100 patients. Error
bars are 95% confidence intervals scaled up the same way. The individual
providers (A.A. D.D.) as discussed in the text are labeled).


135


200.0-


150.0


100.0


f

V


DD.


i


1 2 3 4 5 6: 7


I


ri


:a


T- T: ;1:: : ":::


I II I









300.0


Figure 7-12. Provider slopes sorted by ascending order within each site. Site (clinic)
numbers shown along the bottom. Each provider's sloped is scaled by adding
the model coefficient for each provider to the fixed imaging propensity effect
(0.9919) and multiplied by 100 to represent adjusted images per 100 patients
over the two year study interval. Error bars are 95% confidence intervals
scaled up the same way. The individual providers (A.A. D.D.) as discussed
in the text are labeled).


136









CHAPTER 8
DISCUSSION

Summary of Key Results

Utilization of outpatient diagnostic imaging in a cohort of 85,277 patients was

evaluated over a two year period extending through June of 2009. The adult patients in

this study were cared for in a stable 'medical home' as defined by regular visits to a

primary care doctor practicing in one of 15 clinics. The institutional setting is an

academic health center located in Boston. In general, the study revealed that older

female patients had more imaging as did those who had many medical problems listed

in the clinical record system. Also, patients who visited doctors, were admitted to the

hospital, or seen in the emergency room more frequently had more imaging. Doctor

factors associated with a greater tendency to order imaging tests were less experience,

female gender, and having a medium size practice (500-1000 patients). A special

statistical technique (hierarchical modeling) that accounts for all patient factors allowed

creation of 'profiles' scoring each of the 148 doctors on their general tendency to order

imaging tests on the 'average' patient and how many more tests were ordered on

patients with greater comparative 'need' for diagnostic imaging.

On a per patient basis, the amount of imaging ordered by their own (linked / loyal)

primary care doctor ranged from 0-15 examinations, with average of 0.7146 (60,938 /

85,277). This translates into 35.7 images per 100 patient years. For comparison, during

the same period, the cohort had 50.8 outpatient images per 100 patient years ordered

by specialists caring for them and 37.6 images per 100 patient years performed while

they were in the emergency department or hospitalized. This study concentrated

exclusively on variability in and factors contributing to the outpatient imaging ordered by


137









the linked (loyal) primary care doctor, which accounts for about 27.5% of the total

imaging received by the entire patient cohort.

With the patient experience over 2 years as the unit of analysis and the count of

primary care ordered outpatient images as the outcome, fitting a simple Poisson

distribution yields mean (X) of 0.715 (Cl: 0.707-0.722). The outcome distribution also fits

quite well to a Zero-Inflated Poisson (ZIP) distribution with mean (X) of 1.492 (CI: 0.389-

0.411) and zero-inflation parameter (4) of 0.521 (CI: 0.516-0.526). This (ZIP) distribution

was used for model based creation of patient level imaging propensity expected number

of images using available risk adjustment variables.

An alternate way to model the same phenomenon is in two stages; any imaging

versus none (logistic process) followed by imaging intensity (Poisson process) for

patients with at least one imaging test. For the logistic process, the overall 'success'

rate (patient had some imaging) was 37.13% (31,660 / 85,277). The Poisson mean (X)

for the non-zero imaging (N=31,660 patients) was 1.925 (CI: 1.910-1.940). This two

stage approach was used to test joint effect of 28 patient, 6 provider, and 1 clinic

variable(s) on any imaging utilization (logistic regression) followed by imaging intensity

(Poisson regression).

Patient demographic and clinical factors significantly associated with a greater

likelihood of any imaging and greater imaging intensity included: increasing patient age,

female sex, Hispanic race (compared with white), and more clinical problems. Patient

level clinical activity variables associated with greater likelihood of any imaging and

greater imaging intensity included: office visits to the linked (loyal) primary care doctor,

office visits to specialists, inpatient imaging tests, outpatient imaging ordered by other


138









(covering) primary care doctors, and outpatient imaging ordered by specialists. Provider

and clinic level factors associated with a greater likelihood of any imaging and greater

imaging intensity included: an extra degree (after M.D.) held by the patient's linked

(loyal) primary care doctor, patient's linked (loyal) doctor was female, mid-level practice

size (500-799 patients), and increasing size of the clinic (number of practicing doctors).

Patient level factors associated with decreased likelihood of any imaging and

decreased imaging intensity included: insurance with Medicare (compared to self-pay),

diabetes, and hypertension. Patients with longer total length of stay in the hospital

actually had lower likelihood of any imaging and lower imaging intensity. As the primary

care provider's experience increased, both likelihood and intensity of imaging for their

linked (loyal) patients decreased. The provider's place of M.D. training (foreign medical

graduate=FMG) had a discordant effect on imaging utilization. When the doctor had

FMG status of 'yes', they tended to be more likely to order at least some imaging on

their patients but the amount of imaging tests was less than American trained

(FMG='no') counterparts.

One factor that did not have any significant effect during multivariable modeling of

imaging utilization (any imaging or imaging intensity) was the amount of outpatient

prescription medications (in 4 ordinal categories) each patient was taking. This is

notable because there was a strong bivariate relationship between this medication

variable and the outcome (outpatient imaging ordered by primary care doctor). This will

be discussed below.

When all patient level factors were combined in a single ZIP model, the predictions

for each patient were used as an (expected) imaging propensity score for subsequent


139









multi-level hierarchical modeling. This imaging 'risk adjustment' model had R-Squared

of 0.17 at the patient level. Adding the provider's unique (anonymous) identity as a

predictor in a two level hierarchical model brought the full model R-Squared up to 0.24.

Since the patient level imaging propensity scores were centered on each provider's

mean, the provider intercept can be interpreted as the average imaging utilization for

that doctor (all else equal). The plausible range of these intercepts was 0.151 to 1.283

images per patient over two years with mean of 0.7171. This translates into 35.9 images

per 100 patient years with plausible range of 7.6 to 64.2. At the same time, the

individual provider slopes from the hierarchical model can be interpreted as the extent

to which each doctor responds to a unit increase in patient imaging propensity (which

also ranged from 0 to 15 with a single higher score of -20). The mean was 0.9919 and

the plausible range was 0.216 to 1.768. Scaling these up to images per 100 patient

years gives mean of ~50 and plausible range of 10.8 to 88.4 for every unit increase in

patient imaging 'need'. These estimates of provider imaging utilization parameters

(average/intercept and slope) are quite precise with calculated reliability of 0.965 and

0.939 respectively.

Discussion of Key Results

Perhaps the most vexing and interesting question arising from this study has to do

with the fact that no more than 25% of the variation in the number of primary care

imaging tests per patient is explainable using a quite robust and complete set of patient,

doctor, and clinic variables. This holds true for even the most complete models tried.

For example, the ZIP model with all 28 patient level variables that produced the imaging

propensity scores for multi-level modeling can be modified to include all variation due to

providers and clinics by placing the unique identity of each patient's doctor and clinic


140









into the model (as class variables). The R-Squared of this 'fullest' model is still just over

0.24 and about 30% of the explanatory 'power' comes from knowing who the doctors

and clinics are. Further, the empirically determined variation in the 148 providers

tendency to order imaging on the average patient (intercepts and slopes from multi-level

modeling) is quite substantial. This conundrum goes to the very heart of philosophical

considerations concerning causes and consequences of variations in medical resource

utilization. These questions have been posed occasionally in the health services

research literature but remain unanswered (Cain and Diehr 1992, Diehr et al. 1990).

With the preceding in mind, consider that the practice setting (MGH) is among the

most sophisticated and consistent with respect to the processes of outpatient imaging

ordering, scheduling, and provider feedback. As mentioned in the Chapter 5, virtually all

outpatient imaging was ordered and scheduled via a web-based radiology order entry

(ROE) system. During the entire study period, the ROE system had fully functional and

complete real-time appropriateness decision support (DS) feedback for all CT, MRI, and

nuclear medicine tests. Additionally, the primary care doctors included in this study were

all given periodic (bi-yearly) feedback about their utilization of outpatient imaging

compared with peers. It can be persuasively argued that the practice examined herein is

'as good as it gets' with respect to outpatient imaging decision support and utilization

management. This implies that the amount of variability in primary care outpatient

imaging utilization accruing to doctors (~30%) is a lower bound. Therefore, if similar

studies were to be conducted elsewhere, the absolute amount of variability between

doctors would be greater and the fraction of total variation also larger.


141









Limitations of Study

In terms of generalizability, the primary care practice being studied is not widely

representative of other private, or even other academic settings. As mentioned above,

near complete use of electronic order entry with imaging-specific decision support

coupled with active utilization management means that most other practice settings and

locales will differ on several axes. The absolute number (and distribution over modality)

of images obtained per patient in other primary care settings may be substantially

higher or lower. Other primary care doctors/groups may actually obtain advanced

imaging less often without easy access to electronic ordering, scheduling, and clinical

decision support. For, example they may more often refer complex patients to

specialists and defer imaging to them resulting in lower apparent utilization by the

primary care providerss. On the other hand, without 'barrier' effects of a formal order

entry system and decision support, which sometimes recommends against imaging,

overall utilization could be much higher. In either event, variability of utilization between

patients (after risk adjustment) and providers will almost certainly be greater in other

settings. This would manifest in a lower fraction of overall explained variation (25% in

this setting), less effective risk adjustment at the patient level, and a greater fraction of

variation attributable to providers. These observations should not discourage others

from using risk adjusted benchmarking of imaging utilization. Quite to the contrary, such

provider profiles under conditions of greater variation in imaging utilization will have

potentially greater impact.

The main study outcome, outpatient imaging utilization, was quantified by counting

imaging tests. Clearly, not all imaging tests are equal and some cost more others. For

utilization management efforts that seek to understand and control expenditures,


142









resource use (cost) of each imaging procedure is important. The relative value unit

(RVU) of each imaging procedure is the obvious choice as a proxy for cost and is in fact

often used to calculate reimbursement. Thus, the summed RVU of outpatient imaging

tests performed on each patient has a potential advantage over simple counts if cost

were to be the main focus of analysis. For example, an 'old school' doctor who ordered

chest X-Ray on most patients having respiratory symptoms might seem to have the

same level of utilization (by simple counting of procedures) as a doctor that ordered

chest CT scans much more often. In comparing relative contribution of primary care and

specialist ordered tests to outpatient imaging expenditures, summed RVU (as opposed

to simple counts) would account for differences in the type of tests (modalities) that get

ordered. Also, the summed RVU as outcome approach would allow model based

prediction/speculation about potential cost savings that could be realized by reducing

provider variation and/or curtailing utilization among the 'high outliers'. This study only

examined outpatient imaging ordered by each patient's linked primary care doctor which

accounts for less than half (~40%) of all outpatient imaging. The vast majority of the rest

is ordered by the (sometimes many) specialists caring for the same patient. However,

the relationship between primary care and specialist ordered imaging was partly

addressed by including the amount specialist ordered imaging as a patient level

predictor.

The assumption about distribution of errors in modeling counts of imaging was that

they were Poisson or ZIP. In nature, true Poisson data generating processes have

identical probability of events during time t+1 independent of the cumulative number of

events through time t. It can be argued that in actual patient care, this independence


143









assumption may not hold because 'sicker' patients tend to have more imaging tests in

the future. At the same time, patients who have been ill may already had had more

imaging tests in the past than otherwise healthy peers such that prior imaging is a

'marker' for having been sick. However, the distribution of imaging test counts over

patients has an empirical shape that is quite well fitted by Poisson or ZIP distributions.

At the same time, one of the main reasons for selecting non normal error distribution for

count modeling is to avoid deflation of standard errors of estimates and resulting

mistakes in hypothesis tests about them. Thus, even if the data generating

phenomenon is not a perfect 'natural' Poisson or ZIP process, these distributions may

still be most suitable for error fitting.

The multivariable modeling of any imaging use (logistic) and imaging intensity

(Poisson on non-zero observations) did not account for the nested structure of the data

so that the doctor and clinic factors were repeated over all patients in each respective

unit. This may result in biased estimation of the effect size and somewhat lower

standard errors for these coefficients. The subsequent hierarchical modeling helps to

address this shortcoming. Despite selecting one from each group of highly correlated

variables, there may be additional problems with multi-colinearity as evidenced by the

behavior of the variable measuring patient's outpatient medications (significant at

bivariate analysis but not significant when analyzed jointly with other variables). Also,

additional work will better characterize the residuals from the logistic, Poisson, ZIP, and

multi-level models used at various stages of the analysis. Alternate error distribution

assumptions for the count of imaging tests per patient might also be worthy of

exploration, including Negative Binomial.


144









There were 10,396 (12.2%) patients that had no recorded visits to the linked

(loyal) provider during the two year study period. These were retained in the analytic

data set. This raises the concern that the linked (loyal) provider visit RVU variable

(which was strongly associated with the number of imaging tests ordered by that doctor)

was confounded in some way or that the loyalty attribution methodology was flawed. By

definition of the (2008) loyalty cohort, all patients had at least one visit to their linked

provider from 2006-2008 and this was confirmed. All 10,396 patients had at least one

visit to the linked provider occurring between January 1, 2006 (start of the loyalty cohort

definition period) and July 1, 2007 (start of the study period). The patients with no visits

during the study period were distributed across 143 of the providers. That is, only 5

providers saw all their loyal patients at least once during the study. A by provider

distribution of the percent patients with no visit during the study period had mean=11.8,

median=11.3, and standard deviation=8.25 in a nearly normal distribution (skew=1.11)

of loyal patients who did not visit them during the study period. This is reassuring in that

it would seem to reflect actual practice variation rather than a substantial data integrity

problem (e.g., provider identifier mismatch).

There were 804 patients excluded from the analytic data set because they were

cared for by one of 26 physicians that had less than 100 loyal patients. In general, these

were doctors who had mixed practices, worked part time, or left practice during the

study period. For example, an endocrinologist (cardiologist, gastroenterologist) that

worked out of a primary care clinic might have a few patients identified as 'loyal' to them

by the Atlas methodology. Alternatively, a doctor with administrative, teaching, or

research commitments taking up most of their time might attend in one of the primary


145









care clinics a few days per month. Of the 804 excluded patients 593 (74%) had no

images ordered by the linked (loyal) provider, 764 (95%) had two or less, and 40 (5%)

had from 3-7 images. Exclusion of patients (and doctors) in these small and/or mixed

practices may reduce generalizability. However, the specific aim of this research was to

evaluate imaging utilization by actively practicing primary care doctors and including this

small number of patients and their doctors might have biased the results without adding

any additional useful information.

The hierarchical modeling was only carried out with two levels (patients and

doctors) which discounts the effect of having patients nested within doctors that are in

turn nested in clinics. Perhaps a single grand three level (patient, doctor, and clinic)

model that incorporates all relevant predictors individually and handles error distribution

robustly (e.g. ZIP) might provide greater insight into the phenomenon and produce

superior estimates of parameters. This would be a very complex undertaking and might

well involve a full ZIP specification (primary and zero-model) at each of 3 hierarchical

levels (patient, doctor, and clinic).

Policy Implications

Two overarching policy concerns attach to outpatient diagnostic imaging;

substantial and rapidly rising costs as well as increasing population radiation burden

from medical imaging (especially due to CT scans) with attendant risk of cancer

induction. In addition, the situation of primary care doctors deciding between test

(imaging), treat, observe, or refer when faced with their patient's varying clinical

presentations is a classic paradigm of medical decision making. Any attempts to curtail

imaging utilization growth in general or to target 'high users' for remediation or sanction

must be informed by proper modeling of drivers of variation at patient, provider, and


146









perhaps higher levels. The empiric findings of this study will help to understand the

manner and extent to which risk adjustment and variation analysis methods can help

with outpatient imaging utilization management. For example, there is a growing body of

literature reflecting a vigorous debate about the reliability, utility, and fairness of provider

'efficiency' profiles promulgated by payers (Adams et al. 2010). Unlike typical 'observed

/ expected' metrics, the multi-level modeling described herein, directly produces highly

reliable (>95%) and much more meaningful provider level measures about average

utilization (intercept) and response to patient clinical need variables (slope). Further,

each provider intercept and slope has its own standard error which allows much more

meaningful comparison with individual peers and the overall average utilization.

As stated above, the clinical leadership for the large group practice serving as the

setting for this study has engaged in quite robust and longstanding utilization

management efforts specific to outpatient imaging utilization and appropriateness. Also,

the doctors studied were all salaried employees of the group practice and none had any

financial incentives (or disincentives) associated with diagnostic imaging. For example,

there was almost no growth in the use of CT scans by this practice for 4 years which

include the period of study (Sistrom et al. 2009). At the same time, double digit rates of

growth in CT volumes have occurred in many other settings in the U.S. Thus, the

average yearly number of imaging tests ordered by these primary care doctors (~36 per

100 patients) would seem to be a lower bound estimate of what occurs nationally. At the

same time, there was substantial variation between doctors in their average use of

imaging (intercept from multi-level model) and their response to patient propensity

(clinical need) for imaging (slopes from multi-level analysis).


147










There are policy implications from this study that relate to high technology in

medical care more generally. Advanced imaging is a 'poster child' for a broad array of

diagnostic and therapeutic interventions made possible by advances in basic sciences,

engineering, and informatics. These devices and techniques are attractive and

compelling to health care providers, patients, and lay public such that hospitals now

actively compete to obtain and advertise extensively about the 'latest and greatest'

advances. Further, the regulation of medical devices by the Food and Drug

Administration and other agencies is much less stringent than for drugs. Specifically,

there is little or no requirement that developers and vendors demonstrate clinical

effectiveness; only safety and functionality. Comparative effectiveness evaluations to

determine appropriateness of various devices and technologies for different clinical

purposes will be needed to guide reimbursement determinations. These have two

separate stages: first, whether or not to allow claims for a new device/technology at all,

and second, what clinical situations warrant reimbursement on a case by case basis. As

with imaging, there are many contextual factors surrounding utilization of high

technology medical interventions that operate in concert (or opposition) with clinical

need (appropriateness).

Contribution to Literature

The use of the MGH/Atlas loyalty cohort methodology provides a unique

population of patients and doctors participating in a stable 'medical home' type of

primary care practice. As described above, the robust medical and imaging informatics

infrastructure at MGH provides an optimal situation for standardizing the

appropriateness and intensity of outpatient imaging utilization. Thus the practice under


148









observation is quite likely at or near optimum with respect to variation between primary

care providers with respect to diagnostic imaging. This same rich informatics

environment also means that the available empirical data about patients, doctors, and

the clinical activity they engage in (including but not limited to imaging tests) is

unparalleled in fidelity and completeness. The estimates of the effect size and direction

of many patient and provider level factors on imaging utilization should be useful in

themselves.

Despite increasing popularity and application of hierarchical techniques to health

outcomes, cost analyses, and risk-adjustment, this is the first study in which hierarchical

modeling has been used to study outpatient imaging utilization in primary care.

Preparatory risk-adjustment for 'imaging propensity' at the patient level using ZIP

modeling is also unique. The combination of these methods yields important and

interesting insights into how doctors differ in both their general tendency to use imaging

on the 'average' patient but how they respond to changing 'need' for imaging in their

own panel of patients.

Future Research

The current data source includes individual records for every outpatient visit with

CPT code indicating type and intensity of visit, ICD-9 codes for the visit reasonss,

patient ID, rendering provider ID, and date of service. The CPT codes were already

used to create summed RVU of visits by provider type for the completed study.

Combining these, it should be possible to use the visit data in much more robust ways.

For example, the date, provider, and patient information common to both visit records

and outpatient imaging events can be submitted to 'attribution logic' which matches

imaging tests to visits. A simple set of rules serves to do this. Also, it is important to


149









have at least two contiguous years worth of visits and three years of imaging data on

the same population (which is available for this cohort).

This 'visit-based' method has been validated on a separate set of neurology

outpatient visits and associated imaging tests with successful attribution of more than

90% of imaging tests to a visit. The analysis becomes a visit-based rather than panel-

based and the 'measure' becomes images per visit rather than images per patient year.

By modeling visits and grouping by provider, comparison between providers concerning

their relative tendency to order imaging can be performed. Using the same set of

patients and primary care doctors over the same time frame would allow comparing

visit-based and panel-based imaging utilization profiles to see if the visit-based method

gives similar results (e.g., ranks providers in the same order in terms of imaging

utilization intensity). If a visit-based method is acceptable, it can be generalized to data

sets that are less granular and robust (e.g., Medicare claims). The other advantage of

visit-based provider profiles is that they will work with specialists who are much less

likely--than primary care doctors--to have a stable 'panel' of patients.

Another interesting set of questions arising from the current data source has to do

with the relationship between outpatient imaging ordered by the patient's linked (loyal)

primary care doctor and other doctors, mostly specialists. Specifically, in this cohort of

patients the majority (about 60%) of all outpatient imaging was ordered by specialists.

As described above, primary care doctors faced with clinical uncertainty have a limited

set of options: observe, treat, (imaging) test, or refer to specialist. The current study

lumps three of the choices; observe, treat, or refer into a 'no imaging' category and

evaluates factors relating to the single alternate choice: order an imaging test. Parallel


150









and/or simultaneous modeling (perhaps with multivariate techniques) of the imaging

performed by specialists could add an additional dimension to our characterization of

primary care doctor behavior. It may turn out that some providers who seem to be

'conservative' with respect to imaging are actually 'liberal' in terms of referring to

specialists and this behavior may costlier overall compared with providers who order

more of their own imaging tests.


151









LIST OF REFERENCES


Adams, J. L., A. Mehrotra, J. W. Thomas, et al. 2010. Physician Cost Profiling-
Reliability and Risk of Misclassification. New England Journal of Medicine
362:1014-1021.

Allan, G. M., and J. Lexchin. 2008. Physician Awareness of Diagnostic and Nondrug
Therapeutic Costs: A Systematic Review. International Journal of Technology
Assessment in Health Care 24:158-165.

Andersen, R. M. 1995. Revisiting the Behavioral Model and Access to Medical Care:
Does it Matter? Journal of Health and Social Behavior 36:1-10.

Anthony, D. L., M. B. Herndon, P. M. Gallager, et al. 2009. How Much Do Patients'
Preferences Contribute to Resource Use? Health Affairs (Millwood) 28:864-873.

Atlas S. J., Y. Chang, T. A. Lasko, et al. 2006. Is this "My" Patient? Development and
Validation of a Predictive Model to Link Patients to Primary Care Providers.
Journal of General Internal Medicine 21:973-978.

Atlas, S. J., R. W. Grant, T. G. Ferris, et al. 2009. Patient-Physician Connectedness and
Quality of Primary Care. Annals of Internal Medicine 150: 325-335.

Axt-Adam, P., J. C. van der Wouden, and E. van der Does. 1993. Influencing Behavior
of Physicians Ordering Laboratory Tests: A Literature Study. Medical Care
31:784-794.

Baicker, K., E. S. Fisher, and A. Chandra. 2007. Malpractice Liability Costs and the
Practice of Medicine in the Medicare Program. Health Affairs (Millwood) 26:841-
852.

Bates, D. W., G. J. Kuperman, A. Jha, et al. 1997. Does the Computerized Display of
Charges affect Inpatient Ancillary Test Utilization? Archives of Internal Medicine
157:2501-2508.

Bernardy, M., C. G. Ullrich, J. V. Rawson, et al. 2009. Strategies for Managing Imaging
Utilization. Journal of the American College of Radiology 6:844-850.

Berrington de Gonzalez, A., M. Mahesh, K. P. Kim, et al. 2009. Projected Cancer Risks
from Computed Tomographic Scans Performed in the United States in 2007.
Archives of Internal Medicine 169:2071-2077.

Bertolini, G., R. D'Amico, D. Nardi, et al. 2000. One Model, Several Results: The
Paradox of the Hosmer-Lemeshow Goodness-of-Fit Test for the Logistic
Regression Model. Journal of Epidemiology and Biostatistics 5:251-253.


152









Bhargavan, M., and J. H. Sunshine. 2005. Utilization and Radiology Services in the
United States: Levels and Trends in Modalities, Regions, and Populations.
Radiology 234:824-832.

Blachar, A., S. Tal, A. Mandel, et al. 2006. Preauthorization of CT and MRI
Examinations: Assessment of a Managed Care Preauthorization Program based
on the ACR Appropriateness Criteria and the Royal College of Radiology
Guideliness. Journal of the American College of Radiology 3:851-859.

Blumberg, M. S. 1987. Inter-area Variations in Age-adjusted Health Status. Medical
Care 25:340-353.

Bodenheimer, T., and A. Fernandez. 2005. High and Rising Health Care Costs. Part 4:
Can Costs be Controlled While Preserving Quality? Annals of Internal Medicine
143:26-31.

Boutwell, R. C. and J. B. Mitchell. 1993. Diffusion of New Technologies in the Treatment
of the Medicare Population. Implications for Patient Access and Program
Expenditures. International Journal of Technology Assessment in Health Care
9:62-75.

Brant-Zawadzki, M. 1994. Cost Containment Drives Utilization Management. Diagnostic
Imaging (San Francisco) 16:37-43.

Brenner, D. J., and E. J. Hall. 2007. Computerized Tomography-An Increasing Source
of Radiation Exposure. The New England Journal of Medicine 357:2277-2284.

Britt, H., A Bhasale, D. A. Miles, et al. 1996. The Sex of the General Practitioner: A
Comparison of Characteristics, Patients, and Medical Conditions Managed.
Managed Care 34:403-415.

Brook, R. H., and K. N. Lohr. 1985. Efficacy, Effectiveness, Variations, and Quality.
Boundary-crossing Research. Medical Care 23:710-722.

Bugter-Maessen, A. M., R.A. Winkens, R. P. Grol, et al. 1996. Factors Predicting
Differences among General Practitioner in Test Ordering Behaviour and in the
Response to Feedback on Test Requests. Family Practice 13:254-258.

Burkhardt, J. H., and J. H. Sunshine. 1996. Utilization of Radiologic Services in Different
Payment Systems and Patient Populations. Radiology 200:201-207.

Cain, K. C., and P. Diehr. 1992. Testing the Null Hypothesis in Small Area Analysis.
Health Services Research 27:267-294.

Casparie, A. F. 1996. The Ambiguous Relationship between Practice Variation and
Appropriateness of Care: An Agenda for Further Research. Health Policy 35:247-
265.


153









Chang, W. and S. B. McCracken. 1996. Applying Case Mix Adjustment in Profiling
Primary Care Physician Performance. Journal of Health Care Finance 22:1-9.

Chassin, M. R. 1993. Explaining Geographic Variation. The Enthusiasm Hypothesis.
Medical Care 31 :YS37-YS44.

Chassin, M. R., J. Kosecoff, R. E. Park, et al. 1987. Does Inappropriate Use Explain
Geographic Variations in the Use of Health Care Services? A Study of Three
Procedures. Journal of the American Medical Association 258:2533-2537.

Childs, A. W., and E. D. Hunter. 1972. Non-medical Factors Influencing Use of
Diagnostic X-ray by Physicians. Medical Care 10:323-335.

Cohen, B. J. 1996. Is Expected Utility Theory Normative for Medical Decision Making?
Medical Decision Making 16:1-6.

Couchman, G. R., S. N. Forjuoh, M. H. Rajab, et al. 2004. Nonclinical Factors
Associated with Primary Care Physicians' Ordering Patterns of Magnetic
Resonance Imaging/Computed Tomography for Headache. Academic Radiology
11:735-740.

Couchman, G. R., S. N. Forjuoh, M. D. Reis, et al. 2005. Variation in MRI/CT Utilization
among Family Physicians and General Internists in a Multi-specialty Group
Practice. Medical Science Monitor 11 :MT19-MT25.

Cummings, K. M., K. B. Frisof, M. J. Long, et al. 1982. The Effects of Price Information
on Physicians' Test-ordering Behavior. Ordering of Diagnostic Tests. Medical Care
20:293-301.

Davis, P., B. Gribben, R. Lay-Yee, et al. 2002. How Much Variation in Clinical Activity is
There Between General Practitioners? A Multi-Level Analysis of Decision-Making
in Primary Care. Journal of Health Services Research & Policy 7:202-208.

Davis, P., B. Gribben, A. Scott, et al. 2000a. Do Physician Practice Styles Persist Over
Time? Continuities in Patterns of Clinical Decision-Making Among General
Practitioners. Journal of Health Services Research & Policy 5:200-207.

Davis, P., B. Gribben, A. Scott, et al. 2000b. The "Supply Hypothesis" and Medical
Practice Variation in Primary Care: Testing Economic and Clinical Models of Inter-
Practitioner Variation. Social Science & Medicine 50:407-418.

DeKay, M. L. and D. A. Asch 1998. Is the Defensive Use of Diagnostic Tests Good for
Patients, or Bad? Medical Decision Making 18:19-28.

Deyo, R. A. 2002. Cascade Effects of Medical Technology. Annual Review of Public
Health 23:23-44.


154









Diehr, P., K. Cain, F. Connell, et al. 1990. What is too Much Variation? The Null
Hypothesis in Small-Area Analysis. Health Services Research 24:741-771.

Eddy, D. 2007. Reflections on Science, Judgment, and Value in Evidence-Based
Decision Making: A Conversation with David Eddy by Sean R. Tunis. Health
Affairs (Millwood) 26:w500-w515.

Eddy, D. M. 1984. Variations in Physician Practice: The Role of Uncertainty. Health
Affairs (Millwood) 3:74-89.

Eddy, D. M. 1990. Clinical Decision Making: From Theory to Practice. Guidelines for
Policy Statements: The Explicit Approach. Journal of the American Medical
Association 263:2239-40, 2243.

Eeckhoudt, L. 1996. Expected Utility Theory--Is it Normative or Simply "Practical"?
Medical Decision Making 16:12-13.

Eisenberg, J. M. 1985. Physician Utilization. The State of Research about Physicians'
Practice Patterns. Medical Care 23:461-483.

Eisenberg, J. M. 1986a. The Physician as Self-Fulfilling Practitioner. in Doctor's
Decisions and the Cost of Medical Care, Health Administration Press
Perspectives, Ann Arbor, MI, pp. 29-56.

Eisenberg, J. M. 1986b. The Physician as the Patient's Agent. in Doctor's Decisions and
the Cost of Medical Care, Health Administration Press Perspectives, Ann Arbor,
MI, pp. 57-78.

Eisenberg, J. M., and D. Nicklin. 1981. Use of Diagnostic Services by Physicians in
Community Practice. Medical Care 19:297-309.

Epstein, A. M., and B. J. McNeil. 1985a. Physician Characteristics and Organizational
Factors Influencing Use of Ambulatory Tests. Medical Decision Making 5:401-415.

Epstein, A. M., and B. J. McNeil. 1985b. The Effects of Patient Characteristics on
Ambulatory Test Ordering. Social Science & Medicine 21:1071-1075.

Epstein, A. M., and B. J. McNeil. 1986. Relationship of Beliefs and Behavior in Test
Ordering. The American Journal of Medicine 80:865-870.

Epstein, A. M., and B. J. McNeil. 1987. Variations in Ambulatory Test Use: What Do
They Mean? Medical Clinics of North America 71:705-717.

Fazel, R., H. M. Krumholz, Y. Wang, et al. 2009. Exposure to Low-Dose Ionizing
Radiation from Medical Imaging Procedures. The New England Journal of
Medicine 361:849-857.


155









Fenn, P., A. Gray, and N. Rickman. 2007. Liability, Insurance and Medical Practice.
Journal of Health Economics 26:1057-1070.

Fertig, A., M. Roland, H. King, et al. 1993. Understanding Variation in Rates of Referral
Among General Practitioners: Are Inappropriate Referrals Important and Would
Guidelines Help to Reduce Rates? British Medical Journal 307:1467-1470.

Fisher, E. S., D. E. Wennberg, T. A. Stukel, et al. 2003a. The Implications of Regional
Variations in Medicare Spending. Part 1: The Content, Quality, and Accessibility of
Care. Annals of Internal Medicine 138:273-287.

Fisher, E. S., D. E. Wennberg, T. A. Stukel, et al. 2003b. The Implications of Regional
Variations in Medicare Spending. Part 2: Health Outcomes and Satisfaction with
Care. Annals of Internal Medicine 138:288-298.

Folland, S., and M. Stano. 1989. Sources of Small Area Variations in the Use of Medical
Care. Journal of Health Economics 8:85-107.

Folland, S., and M. Stano. 1990. Small Area Variations: A Critical Review of
Propositions, Methods, and Evidence. Medical Care Review 47:419-465.

Franks, P., and K. Fiscella. 2002. Effect of Patient Socioeconomic Status on Physician
Profiles for Prevention, Disease Management, and Diagnostic Testing Costs.
Medical Care 40:717-724.

Franks, P., C. Mooney, and M. Sorbero. 2000a. Physician Referral Rates: Style Without
Much Substance? Medical Care 38:836-846.

Franks, P., G. C. Williams, J. Zwanziger, et al. 2000b. Why do Physicians Vary so
Widely in Their Referral Rates? Journal of General Internal Medicine 15:163-168.

Franks, P., J. Zwanziger, C. Mooney, et al.1999. Variations in Primary Care Physician
Referral Rates. Health Services Research 34:323-329.

Gazelle, G. S., E. F. Halpern, H. S. Ryan, et al. 2007. Utilization of Diagnostic Medical
Imaging: Comparison of Radiologist Referral Versus Same-Specialty Referral.
Radiology 245:517-522.

Goel V., K. Iron, and J. I. Williams. 1997. Enthusiasm or Uncertainty: Small Area
Variations in the use of Mammography Services in Ontario, Canada. Journal of
Epidemiology & Community Health 51:378-382.

Goldstein J. N., C. A. Camargo, Jr., A. J. Pelletier, et al. 2006. Headache in United
States Emergency Departments: Demographics, Work-up and Frequency of
Pathological Diagnoses. Cephalalgia 26:684-690.


156









Government Accountability Office (U.S.GAO). 2008. Medicare Part B Imaging Services:
Rapid Spending Growth and Shift to Physician Offices Indicate Need for CMS to
Consider Additional Management Practices GAO-08-462.

Greene, B. R., J. Barlow, and C. Newman. 1996. Ambulatory Care Groups and the
Profiling of Primary Care Physician Resource Use: Examining the Application of
Case Mix Adjustments. Journal of Ambulatory Care Management 19:86-89.

Grytten, J., and R. Sorensen. 2003. Practice Variation and Physician-Specific Effects.
Journal of Health Economics 22:403-418.

Hartley, R. M., A. M. Epstein, C. M. Harris, et al. 1987. Differences in Ambulatory Test
Ordering in England and America. Role Of Doctors' Beliefs and Attitudes. The
American Journal of Medicine 82:513-517.

Health Services Research Group. 1992. Small-Area Variations: What are They and
What do They Mean? Health Services Research Group. Canadian Medical
Association Journal 146:467-470.

Hermer, L. D., and H. Brody 2010. Defensive Medicine, Cost Containment, and Reform.
Journal of General Internal Medicine

Hillman, B. J. 2004. Self-Referral: The Time for Whining is Over. Journal of the
American College of Radiology 1:795-796.

Hillman, B. J., C. A. Joseph, M. R. Mabry, et al. 1990. Frequency and Costs of
Diagnostic Imaging in Office Practice--A Comparison of Self-Referring and
Radiologist-Referring Physicians. The New England Journal of Medicine
323:1604-1608.

Hillman, B. J., G. T. Olson, P. E. Griffith, et al. 1992. Physicians' Utilization and Charges
for Outpatient Diagnostic Imaging in a Medicare Population. Journal of the
American Medical Association 268:2050-2054.

Hoey, J., J. M. Eisenberg, W. O. Spitzer, et al. 1982. Physician Sensitivity to the Price of
Diagnostic Tests: A U.S.-Canadian analysis. Medical Care 20:302-307.

Holtgrave, D. R., F. Lawler, and S. J. Spann. 1991. Physicians' Risk Attitudes,
Laboratory Usage, and Referral Decisions: The Case of an Academic Family
Practice Center. Medical Decision Making 11:125-130.

Houchens, R., B. Chu, and C. Steiner. 2007. Hierarchical Modeling Using HCUP Data.,
2007-01 edn.

Isaacs, D. M., J. Marinac, and C. Sun. 2004. Radiograph Use in Low Back Pain: A
United States Emergency Department Database Analysis. The Journal of
Emergency Medicine 26:37-45.


157









Katz, S. J., L. F. McMahon, and W. G. Manning. 1996. Comparing the Use of Diagnostic
Tests in Canadian and US Hospitals. Medical Care 34:117-125.

Kessler, D. P., and M. B. McClellan. 2002. How Liability Law Affects Medical
Productivity. Journal of Health Economics 21:931-955.

Kessler, D. P., N. Summerton, and J. R. Graham. 2006. Effects of the Medical Liability
System in Australia, the UK, and the USA. Lancet 368:240-246.

Knottnerus, J. A., J. Joosten, and J. Daams. 1990. Comparing the Quality of Referrals
of General Practitioners with High and Average Referral Rates: An Independent
Panel Review. British Journal of General Practice 40:178-181.

Kramer, A. A., and J. E. Zimmerman. 2007. Assessing the Calibration of Mortality
Benchmarks in Critical Care: The Hosmer-Lemeshow Test Revisited. Crit Care
Med 35:2052-2056.

Kravitz, R. L., and S. Greenfield. 1995. Variations in Resource Utilization Among
Medical Specialties and Systems of Care. Annual Review of Public Health 16:431-
445.

Landon, B. E., J. Reschovsky, M. Reed, et al. 2001. Personal, Organizational, and
Market Level Influences on Physicians' Practice Patterns: Results of a National
Survey of Primary Care Physicians. Medical Care 39:889-905.

Lasko, T. A., S. J. Atlas, M. J. Barry, et al. 2006. Automated Identification of a
Physician's Primary Patients. Journal of the American Medical Informatics
Association 13:74-79.

Lawthers, A. G., A. R. Localio, N. M. Laird, et al. 1992. Physicians' Perceptions of the
Risk of Being Sued. Journal of Health Politics, Policy and Law 17:463-482.

Leape, L. L., R. E. Park, D. H. Solomon, et al. 1990. Does Inappropriate Use Explain
Small-Area Variations in the Use of Health Care Services? Journal of the
American Medical Association 263:669-672.

Levy, G., A. Blachar, L. Goldstein, et al. 2006. Nonradiologist Utilization of American
College of Radiology Appropriateness Criteria in a Preauthorization Center for MRI
Requests: Applicability and Effects. American Journal of Roentgenology 187:855-
858.

Long, M. J., K. M. Cummings, and K. B. Frisof. 1983. The Role of Perceived Price in
Physicians' Demand for Diagnostic Tests. Medical Care 21:243-250.

Luke, D. A. 2004. Multilevel Modeling, 07-143 edn. Sage Publications, Thousand Oaks,
Calif.


158









Lysdahl, K. B., and I. Borretzen. 2007. Geographical Variation in Radiological Services:
A Nationwide Survey. BMC.Health Services Research 7:21.

Majeed, A., A. B. Bindman, and J. P. Weiner. 2001a. Use of Risk Adjustment in Setting
Budgets and Measuring Performance in Primary Care I: How it Works. British
Medical Journal 323:604-607.

Majeed, A., A. B. Bindman, and J. P. Weiner. 2001b. Use of Risk Adjustment in Setting
Budgets and Measuring Performance in Primary Care II: Advantages,
Disadvantages, and Practicalities. British Medical Journal 323:607-610.

Medicare Payment Advisory Commission. 2003. MedPAC: Variation and Inovation in
Medicine. Report to congress, June.

Medicare Payment Advisory Commission. 2005. MedPAC Recommendations on
Imaging Services. Testimony Before Congress, March 17.

Medicare Payment Advisory Commission. 2006. MedPAC Recommendations on
Imaging Services. Testimony Before Congress, July 18.

Miller, W. L., R. R. McDaniel, Jr., B. F. Crabtree, et al. 2001. Practice Jazz:
Understanding Variation in Family Practices Using Complexity Science. The
Journal of Family Practice 50:872-878.

Mitchell, J. M. 2007. The Prevalence of Physician Self-Referral Arrangements After
Stark II: Evidence from Advanced Diagnostic Imaging. Health Affairs (Millwood)
26:w415-w424.

Mold, J. W., and H. F. Stein. 1986. The Cascade Effect in the Clinical Care of Patients.
The New England Journal of Medicine 314:512-514.

Mort, E. A., J. N. Edwards, D. W. Emmons, et al. 1996. Physician Response to Patient
Insurance Status in Ambulatory Care Clinical Decision-Making. Implications for
Quality of Care. Medical Care 34:783-797.

National Council on Radiation Protection and Measurements. 2009. NCRP Report No.
160: Ionizing Radiation Exposure of the Population of the United States. Journal of
Radiological Protection 29:465.

Neilson, E. G., K. B. Johnson, S. T. Rosenbloom, et al. 2004. The Impact of Peer
Management on Test-Ordering Behavior. Annals of Internal Medicine 141:196-
204.

Nyweide, D. J., W. B. Weeks, D. J. Gottlieb, et al. 2009. Relationship of Primary Care
Physicians' Patient Caseload with Measurement of Quality and Cost Performance.
The Journal of the American Medical Association 302:2444-2450.


159









O'Neill, L., and J. Kuder. 2005. Explaining Variation in Physician Practice Patterns and
Their Propensities to Recommend Services. Medical Care Research and Review
62:339-357.

Otero, H.J., S. Ondategui-Parra, E. M. Nathanson, et al. 2006. Utilization Management
in Radiology: Basic Concepts and Applications. Journal of the American College of
Radiology 3:351-357.

Parchman, M. L. 1995. Small Area Variation Analysis: A Tool for Primary Care
Research. Family Medicine Journal 27:272-276.

Pekoz, E. A., M. Shwartz, L. I. Lezzoni, et al. 2003. Comparing the Importance of
Disease Rate Versus Practice Style Variations in Explaining Differences in Small
Area Hospitalization Rates for Two Respiratory Conditions. Statistics in Medicine
22:1775-1786.

Pezzin, L. E., P. M. Keyl, and G. B. Green. 2007. Disparities in the Emergency
Department Evaluation of Chest Pain Patients. Academic Emergency Medicine
14:149-156.

Pham, H. H., G. C. Alexander, and A. S. O'Malley. 2007. Physician Consideration of
Patients' Out-of-Pocket Costs in Making Common Clinical Decisions. Archives of
Internal Medicine 167:663-668.

Phelps, C. E., C. Mooney, A.I. Mushlin, et al. 1994. Doctors Have Styles-- and They
Matter! Working paper.

Phillips, K. A., K. R. Morrison, R. Andersen, et al. 1998. Understanding the Context of
Healthcare Utilization: Assessing Environmental and Provider-Related Variables in
the Behavioral Model of Utilization. Health Services Research 33:571-596.

Quintana, J. M., D. Goldmann, and C. Homer. 1997. Social Disparities in the use of
Diagnostic Tests for Children with Gastroenteritis. International Journal for Quality
in Health Care 9:419-425.

Raudenbush, S. W., and A. S. Bryk. 2002. Hierarchical Linear Models
Applications And Data Analysis Methods, 2nd ed edn. Sage Publications,
Thousand Oaks.

Reinhardt, U. E. 1999. The Economist's Model of Physician Behavior. Journal of the
American Medical Association 281:462-465.

Restuccia, J., M. Shwartz, A. Ash, et al. 1996. High Hospital Admission Rates and
Inappropriate Care. Health Affairs (Millwood) 15:156-163.

Rosen, M. P., R. B. Davis, and L. G. Lesky 1997. Utilization of Outpatient Diagnostic
Imaging. Does the Physician's Gender Play A Role? Journal of General Internal
Medicine 12:407-411.


160









Rosenthal, D. I., J. B. Weilburg, T. Schultz, et al. 2006. Radiology Order Entry with
Decision Support: Initial Clinical Experience. Journal of the American College of
Radiology 3:799-806.

Salem-Schatz, S., G. Moore, M. Rucker, et al. 1994. The Case for Case-Mix Adjustment
in Practice Profiling. When Good Apples Look Bad. Journal of the American
Medical Association 272:871-874.

Schulman, K. A., J. A. Berlin, W. Harless, et al. 1999. The Effect of Race and Sex on
Physicians' Recommendations for Cardiac Catheterization. The New England
Journal of Medicine 340:618-626.

Shen, J., R. Andersen, R. Brook, et al. 2004. The Effects of Payment Method on Clinical
Decision-Making: Physician Responses To Clinical Scenarios. Medical Care
42:297-302.

Sirovich, B., P. M. Gallagher, D. E. Wennberg, et al. 2008. Discretionary Decision
Making by Primary Care Physicians and the Cost of U.S. Health Care. Health
Affairs (Millwood) 27:813-823.

Sirovich, B. E., D. J. Gottlieb, H. G. Welch, et al. 2006. Regional Variations in Health
Care Intensity and Physician Perceptions of Quality of Care. Annals of Internal
Medicine 144:641-649.

Sistrom, C.L. 2009. The Appropriateness of Imaging: A Comprehensive Conceptual
Framework. Radiology 251:637-649.

Sistrom, C.L., P. A. Dang, J. B. Weilburg, et al. 2009. Effect of Computerized Order
Entry with Integrated Decision Support on the Growth of Outpatient Procedure
Volumes: Seven-Year Time Series Analysis. Radiology 251:147-155.

Sloan, F.A., and J. J. Shadle. 2009. Is there Empirical Evidence for "Defensive
Medicine"? A reassessment. Journal of Health Economics 28:481-491.

Smith-Bindman, R. J. Lipson, R. Marcus, et al. 2009. Radiation Dose Associated with
Common Computed Tomography Examinations and the Associated Lifetime
Attributable Risk of Cancer. Archives of Internal Medicine 169:2078-2086.

Smulowitz, P. B., L. Ngo, and S. K. Epstein. 2009. The Effect of a CT and MRI
Preauthorization Program on ED Utilization. The American Journal of Emergency
Medicine 27:328-332.

Solomon, D. H., H. Hashimoto, L. Daltroy, et al. 1998. Techniques to Improve
Physicians' Use of Diagnostic Tests: A New Conceptual Framework. Journal of the
American Medical Association 280:2020-2027.


161









Sood, R., A. Sood, and A. K. Ghosh. 2007. Non-evidence-based Variables Affecting
Physicians' Test-ordering Tendencies: A Systematic Review. The Netherlands
Journal of Medicine 65:167-177.

Stahl, F. W. 1969. The Mechanics of Inheritance, 2d edn. Prentice-Hall, Englewood
Cliffs, N.J.

Stano, M. 1991. Further Issues in Small Area Variations Analysis. Journal of Health
Politics, Policy and Law 16:573-588.

Stano, M. 1993. Evaluating the Policy Role of the Small Area Variations and Physician
Practice Style Hypotheses. Health Policy 24:9-17.

Thrall, J. H. 2005a. Reinventing Radiology in the Digital Age. Part II. New Directions
and New Stakeholder Value. Radiology 237:15-18.

Thrall, J. H. 2005b. Reinventing Radiology in the Digital Age. Part III. Facilities, Work
Processes, and Job Responsibilities. Radiology 237:790-793.

Thrall, J. H. 2005. Reinventing Radiology in the Digital Age. Part I. The All-digital
Department. Radiology 236:382-385.

Tucker, A. M., J. P. Weiner, S. Honigfeld, et al. 1996. Profiling Primary Care Physician
Resource Use: Examining the Application of Case Mix Adjustment. Journal of
Ambulatory Care Management 19:60-80.

Verrilli, D., and H. G. Welch. 1996. The Impact of Diagnostic Testing on Therapeutic
Interventions. Journal of the American Medical Association 275:1189-1191.

Wasiak, R., G. S. Pransky, and S. J. Atlas. 2008. Who's in Charge? Challenges in
Evaluating Quality of Primary Care Treatment for Low Back Pain. Journal of
Evaluation in Clinical Practice 14:961-968.

Welch, H. G., M.E. Miller, and W. P. Welch. 1994. Physician Profiling. An Analysis of
Inpatient Practice Patterns in Florida and Oregon. The New England Journal of
Medicine 330:607-612.

Welch, W. P., M.E. Miller, H.G. Welch, et al. 1993. Geographic Variation in
Expenditures for Physicians' Services in the United States. The New England
Journal of Medicine 328:621-627.

Wennberg, D. E., J. D. Dickens, Jr., L. Biener, et al. 1997. Do Physicians Do What They
Say? The Inclination to Test and its Association with Coronary Angiography Rates.
Journal of General Internal Medicine 12:172-176.

Wennberg, J., and A. Gittelsohn. 1973. Small Area Variations in Health Care Delivery.
Science 182:1102-1108.


162









Wennberg, J. E. 1987a. Population Illness Rates Do Not Explain Population
Hospitalization Rates. A Comment on Mark Blumberg's Thesis that Morbidity
Adjusters are Needed to Interpret Small Area Variations. Medical Care 25:354-
359.

Wennberg, J. E. 1987b. The Paradox of Appropriate Care. Journal of the American
Medical Association 258:2568-2569.

Wennberg, J. E. 1993. Future Directions for Small Area Variations. Medical Care
31 :YS75-YS80.

Wennberg, J. E. 2002. Unwarranted Variations in Healthcare Delivery: Implications for
Academic Medical Centres. British Medical Journal 325:961-964.

Wennberg, J. E. 2004a. Practice Variation: Implications for Our Health Care System.
Managed Care 13:3-7.

Wennberg, J. E. 2004b. Practice Variations and Health Care Reform: Connecting the
Dots. Health Affairs (Millwood) Suppl Web Exclusives, VAR140-VAR144.

Wennberg, J. E., B. A. Barnes, and M. Zubkoff. 1982. Professional Uncertainty and the
Problem of Supplier-Induced Demand. Social Science & Medicine 16:811-824.

Westert, G. P., and P. P. Groenewegen. 1999. Medical Practice Variations: Changing
the Theoretical Approach. Scandinavian Journal of Public Health 27:173-180.

Whiting, P., M. Toerien, I. de Salis, et al. 2007. A Review Identifies and Classifies
Reasons for Ordering Diagnostic Tests. Journal of Clinical Epidemiology 60:981-
989.

Williams, S. V., J. M. Eisenberg, L. A. Pascale, et al. 1982. Physicians' Perceptions
about Unnecessary Diagnostic Testing. Inquiry 19:363-370.

Wu, G. 1996. The Strengths and Limitations of Expected Utility Theory. Medical
Decision Making 16:9-10.


163









BIOGRAPHICAL SKETCH

After serving in the U.S. Army Signal Corps as a cryptographer from 1973-1977,

Chris Sistrom, MD, MPH obtained his undergraduate degree in Computer Science

(1980) from the University of Oregon in Eugene. He attended at Oregon Health

Sciences University (MD in 1984) and completed radiology residency at the University

of Virginia in 1988. He is now Associate Chairman of Radiology, Chief Information

Officer for Radiology, and Associate Professor at the University of Florida, College of

Medicine. Dr. Sistrom obtained an MPH degree in epidemiology and health policy in

2003 from the University of Florida, and is in final stages of a PhD in Health Services

Research there. The topic of his dissertation is Imaging Utilization in Primary Care. The

research goal is to quantify and model various factors that affect the intensity and

mixture of outpatient imaging performed on primary care patients. The resulting models

will be useful in practitioner profiling at institutional and regional levels. The eventual

goal is to produce a 'Map of Imaging' along the lines of the Dartmouth Atlas of

Healthcare and to create risk adjusted population based estimates of optimal and

appropriate imaging utilization.


164





PAGE 1

1 OUTPATIENT IMAGING IN PRIMARY CARE By CHRISTOPHER L. SISTROM A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORID A IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2010

PAGE 2

2 2010 Christopher L. Sistrom

PAGE 3

3 To my wife, Brenda

PAGE 4

4 ACKNOWLEDGMENTS The research described in this dissertation was made possible through collaboration with numer ous people at the Massachusetts General Hospital Physician Organization and Radiology Department. They include (but are not limited to) Jeffrey Weilburg, MD, Keith Dreyer, DO, PhD, Timothy Ferris, MD, George Leehan, and Markus Stout. I am very gratef ul for their help and support. Also, this research would not have been possible without the meticulous and consistent work by Dr. Steve Atlas and his group to identify a stabl e cohort of patients and primary care doctors. My colleagues and friends in the Department of Radiology at the University of Florida have been very kind in allowing me to spend enough time away from routine clinical, administrative, and educational duties to engage in a multi-year effort to learn several new disciplines; epidemiology, public health, health services research. Without considerable support and encouragement on t heir part, none of this would have been possible. They include (but are not limited to) Anthony Mancuso, MD, Patricia Abbitt, MD, Jonathon Williams, MD, Melinda Chitty Anne DAmico, and Meryll Frost. Niccie McKay, PhD has been tireless and infinitely patient in several roles including mentor, teacher, edito r, and chair of my dissertat ion committee. Cyndi Garvan, PhD gave me a crash course in hierarchical modeling and was the external member of my committee. Jeffrey Harman, PhD, has been instrumental in bringing a methodological rigor to my efforts and Chri s Harle, PhD honored me by inaugurating his faculty career in serving on my dissertation committee. My wife Brenda has been a wonderful partner during what may have seemed at times like a decade long mid-life crisis. I very much appreciate her support and

PAGE 5

5 forbearance and am glad to know that she will be around as I decide what I want to be when I grow up.

PAGE 6

6 TABLE OF CONTENTS page ACKNOWLEDG MENTS .................................................................................................. 4LIST OF TABLES ............................................................................................................ 9ABSTRACT ................................................................................................................... 13CHAPTER 1 INTRODUC TION .................................................................................................... 16Background and Si gnificance ................................................................................. 16Specific Aims .......................................................................................................... 17Summary of Study .................................................................................................. 17Contribution to Literatu re ........................................................................................ 19Policy Implic ations .................................................................................................. 202 BACKGRO UND ...................................................................................................... 22Definition of Diagnostic Im aging ............................................................................. 22Imaging Utilizatio n and Costs ................................................................................. 26Primary Care Setting............................................................................................... 28Imaging as Diagnostic Testing ................................................................................ 293 LITERATURE REVIEW .......................................................................................... 32The Anders en Model .............................................................................................. 32Small Area Variation ............................................................................................... 33Clinical Unce rtainty ................................................................................................. 37Risk Adjust ment ...................................................................................................... 38Appropriateness and Supp lier-Induced De mand .................................................... 40Summary ................................................................................................................ 424 CONCEPTUAL FR AMEWORK ............................................................................... 44Clinical Need ........................................................................................................... 45Context: Pa tient ...................................................................................................... 47Context: Physician .................................................................................................. 48Context: Malpractice ............................................................................................... 52Context: Practice Organiza tion ............................................................................... 54Context: Payer and Pric es ...................................................................................... 55Context: Access to Imagi ng .................................................................................... 56Summary ................................................................................................................ 58

PAGE 7

7 5 SETTING, DATA SOURCES, AND VAR IABLES .................................................... 60Settings ................................................................................................................... 60Primary Care Practice ...................................................................................... 60Outpatient Radiology ........................................................................................ 62Data S ources .......................................................................................................... 63Loyalty Cohort .................................................................................................. 63Patient De tails .................................................................................................. 64Physician De tails .............................................................................................. 65Imaging Utiliz ation ............................................................................................ 65Variables ................................................................................................................. 66Imaging Utilization ( dependent vari able) .......................................................... 66Patient Characteristics ..................................................................................... 66Clinical Events .................................................................................................. 67Clinical Pr oblems .............................................................................................. 68Outpatient Pr escripti ons ................................................................................... 69Other Imaging Utilizat ion .................................................................................. 69Physician Charac teristics ................................................................................. 70Site Charac teristics .......................................................................................... 71Patient, Provider, and Clinic Ident ifiers ................................................................... 72Data Integrity: Clinical Activity Variables ................................................................. 72Variable Summary .................................................................................................. 736 METHODS .............................................................................................................. 84Outcome Variable Distribution ................................................................................ 84Correlation between Independent Vari ables ........................................................... 87Bivariate Rela tionships ........................................................................................... 88Variable Reducti on for Modeling ............................................................................. 88Multivariable (logistic) Modeling: Any Imaging Use ................................................. 88Multivariable (Poisson) Modeling: Imaging Intensit y (non-zero) .............................. 90Preparation for Multi-Level Mode ling: Imaging Propensity Scores .......................... 91Multi-Level (Hierarc hical) M odelin g ......................................................................... 917 RESULTS ............................................................................................................... 96Outcome Variable Distribution ................................................................................ 96Correlation between Independent Vari ables ........................................................... 97Bivariate rela tionships ............................................................................................. 99Variable Reducti on for Modeling ............................................................................. 99Multivariable (Logistic) M odeling: Any Imaging Use ............................................. 100Multivariable (Poisson) Modeling: Imaging Intensit y (non-zero) ............................ 105Comparison of Any Imaging an d Imaging Intens ity Resu lts .................................. 108Preparation for Multilevel Modeli ng: Imaging Propensity Scor es .......................... 109Multi-Level (Hierarc hical) M odelin g ....................................................................... 110

PAGE 8

8 8 DISCUSSI ON ....................................................................................................... 137Summary of Ke y Results ...................................................................................... 137Discussion of Key Resu lts .................................................................................... 140Limitations of Study .............................................................................................. 142Policy Implic ations ................................................................................................ 146Contribution to Literatu re ...................................................................................... 148Future Re search ................................................................................................... 149LIST OF RE FERENCES ............................................................................................. 152BIOGRAPHICAL SKETCH .......................................................................................... 164

PAGE 9

9 LIST OF TABLES Table page 5-1 All diagnostic imaging per formed on study cohort during two years of study. .... 735-2 Outpatient diagnostic imaging perform ed on study cohort during two years of study. .................................................................................................................. 745-3 Outpatient diagnostic imaging ordered by patients linked (loyal) doctor during two year s of study. ................................................................................... 745-4 Univariate statistics of the outcome variable (per patient count of outpatient imaging tests ordered by prim ary care pr ovider). ............................................... 755-5 Distribution of patient race. ................................................................................. 755-6 Distribution of patients payer ca tegories. ........................................................... 755-7 Patients payer collaps ed into 6 ca tegories. ....................................................... 765-9 CPT codes and relative value units for ambulatory o ffice visits .......................... 775-10 Outpatient visit activity variables (per patient). ................................................... 775-11 Binary clinical problem variables (per patient ). ................................................... 785-12 Four level categorizat ion of patient active outpatient medications (per patient ). .............................................................................................................. 785-13 Summary of other (non-outcome) im aging test utilization variables (per patient ). .............................................................................................................. 785-14 Four level categorization of the number of patients ca red for by each provider (panel si ze). ........................................................................................................ 785-15 Site (clinic) c haracterist ics. ................................................................................. 795-16 Description and categorization of 33 patient level independent variables. ......... 805-17 Description and categorization of provider (8) and clinic level (2) independent variables. ............................................................................................................ 817-1 Spearman correlations between clinical activity and other imaging variables. 1167-2 Bivariate relationship between patient level variables and outcome (imaging counts) ............................................................................................................. 117

PAGE 10

10 7-3 Bivariate relationship between clinic al activity variables and outcome (imaging c ounts). .............................................................................................. 1187-4 Bivariate relationship between provider and clinic level variables and outcome (imagi ng count s). ............................................................................... 1197-5 Patient level results from multivariabl e logistic model on any imaging use. ...... 1207-6 Provider and clinic level results from multivariable logistic model on any imaging us e. ..................................................................................................... 1217-7 Patient level results from multivariable Poisson model on im aging intensity. ... 1227-8 Provider and clinic level results from multivariable Poisson model on imaging intensit y. ........................................................................................................... 1237-9 Univariate statistics for raw imagi ng counts (IMG) and predictions from ZIP model (IMG _PROP). ........................................................................................ 1237-10 Dimens ions. ...................................................................................................... 1237-11 Estimated G correlation ma trix. ........................................................................ 1247-12 Covariance paramet er estimates. ..................................................................... 1247-13 Fit stat istics. ...................................................................................................... 1247-14 Solution for fixed effe cts. .................................................................................. 1247-15 Type 3 tests of fixed effe cts. ............................................................................. 1247-16 Results from multi-leve l random coefficients model. ......................................... 1247-17 Quadrants in intercept vers us slope relation ship plot. ...................................... 1257-18 Exemplary provider s in each quadrant ............................................................ 1257-19 Comparison of null, and reduced model residuals with full m odel. ................... 125

PAGE 11

11 LIST OF FIGURES Figure page 2-1 Total number of imaging studies by year in the U. S. .......................................... 302-2 Imaging shows highest cumulative growth in services per beneficiary (19992004) .................................................................................................................. 302-3 Cumulative growth in imaging volume varies by type (1999-2004) ..................... 314-1 Summary diagram of conceptual model for outpatient imaging utilization in primary care. ....................................................................................................... 595-1 Outpatient imaging tests (per patient ) ordered by the linked (loyal) primary care provider. ...................................................................................................... 815-2 Number of outpatient visi ts by all patients in study cohort (by month) over two years of study. .................................................................................................... 825-3 Number of outpatient imaging tests performed on all patients in study cohort (by month) over two years of study. .................................................................... 825-4 Number of hospital encount ers for all patients in study cohort over two years. .. 837-1 Comparison of imaging counts wit h three Poisson distribut ions. ...................... 1267-2 Logistic regression results for any imaging utilization ....................................... 1277-3 Poisson regression results fo r (non-zero) imagi ng intensity. ............................ 1287-4 Comparison of significant variables for any imaging use and imaging intensity 1297-5 Comparison of significant variables for any imaging use and imaging intensity (small effect sizes). ........................................................................................... 1307-6 Imaging propensity score dist ributions by provider. .......................................... 1317-7 Centered imaging propensity score distributions by provider. .......................... 1317-8 Plot of intercept and slopes for all 148 providers obtained from multi-level model of imaging utilizat ion.. ............................................................................ 1327-9 Imaging utilization versus center ed imaging propensity for a low utilizing doctor. .............................................................................................................. 1337-10 Imaging utilization versus centered imaging propensity a high utilizing doctor. 1347-11 Provider means sorted by a scending order withi n each si te. ............................ 135

PAGE 12

12 7-12 Provider slopes sorted by a scending order withi n each si te. ............................ 136

PAGE 13

13 Abstract of Dissertation Pr esented to the Graduate School of the University of Florida in Partial Fulf illment of the Requirements for t he Degree of Doctor of Philosophy OUTPATIENT IMAGING IN PRIMARY CARE By Christopher L. Sistrom August 2010 Chair: Niccie McKay Major: Health Services Research Diagnostic imaging comprises a rapidly gr owing portion of health care dollars spent in the U.S. Additionally, imaging tests that use X-Rays (including C.T. scans) now contribute half of the entire r adiation dose to the population; having increased from 15% in 1980. Like other medical services, diagnosti c imaging tests are utilized in some states and cities much more frequently than in other s, even after controlling factors like age, gender, and illness burden. This marked variabi lity in how frequently and for what reasons that imaging tests are done extends down to the level of individual doctors. Understanding the causes and effect s of these sorts of differenc es in how health care is delivered (including imaging tests) is one of t he core parts of health services research. The fact that patients in a stable rela tionship with a primar y care doctor have better health outcomes and consume less health services overall has prompted efforts to increase the supply of primary care docto rs and encourage patients to seek care in a so-called medical home setting. One of the key roles of the primary care doctor is as a gatekeeper for expensive tests and procedures as well as referral to specialists. This study looks at a large group of patients (about 85K) being taken care of by 148 primary care doctors to whom they have demonstrated loyalty over three years based on the

PAGE 14

14 pattern of office visits This so called loyalty cohort is useful as a representative sample of other patients and doctors functioning in the ideal doctor-patient relationship. The study seeks to answer questions about how various factors in the patients (including clinical activity) re late to the amount of imaging tests that their primary care doctor orders for them over two years. This can be useful in comparing the use of imaging tests (e.g., images for 100 patients in a year) between primary care doctors by helping to adjust for differences in each docto rs mixture of patients. Such risk-adjusted utilization profiles help to understand variability in practice style and resource use between doctors for purely scholarly interest as well as more practical uses by entities actually paying for the services (e.g., insurance companies, employers, government health programs). A total of 35 pieces of information about each patient (not incl uding their name or other identifiers) and 7 characteristics of each doctor were gathered from 6 different sources of data used routinely during patient care in the practices being studied. This information included complete listing of about million diagnostic imaging tests. Of these, about 60K were ordered by the patient s own (loyal) primary care doctor and were counted by patient to form the vari able of interest (outcome). Statistical relationships between all the other pati ent and doctor factors with the number of imaging tests were analyzed. The results demonstrated that older fema le patients had more imaging as did those who had many medical problems listed in the clinical record system. Also, patients who visited do ctors, were admitted to the hospi tal, or seen in the emergency room frequently had more imaging. Doctor factors associated with a greater tendency to

PAGE 15

15 order imaging tests were less experience, female gender, and having a medium size practice (500-1000 patients). A s pecial statistical technique that accounts for all patient factors allowed creation of profiles sco ring each of the 148 doctors on their general tendency to order imaging tests on the average patient and how many more tests were ordered on patients with great er comparative need for diagnostic imaging.

PAGE 16

16 CHAPTER 1 INTRODUCTION Background and Significance Medical technology is often cited as a major driver of health care costs in the U.S., with diagnostic imaging being a poster child of this trend. Advanced imaging, ordered during ambulatory care, is not only costly, but often leads to a cascade of further tests and interventions (Deyo 2002, Mold and St ein 1986, Verrilli and Welch 1996). There is ample evidence of marked variation in utilizati on of imaging services in outpatient care at multiple levels of aggregation, from intern ational down to individual groups within a large practice (Burkhardt and Sunshine 1996, Couchman et al. 2005, Goel et al. 1997, Hartley et al. 1987, Katz et al. 1996, Lysdahl and Borretzen 2007). This variation in utilization suggests that a substantial portion of diagnostic imaging may be unnecessary. Patients receiving such studies may be needlessly exposed to radiation. Findings of uncertain clinical meaning may prompt more imaging and costs may be increased. Examining the ways in which primary ca re physicians utilize imaging, and using those insights to inform and enhance the consistency of their choices, can improve health care quality and increase cost effectiv eness. Thus, the event of a primary care doctor ordering an imaging test is common and expensive enough to warrant study by itself. Also, it represents a very frui tful paradigm for underst anding medical decision making and much of the variation in downs tream discretionary health care utilization (Parchman 1995). The interesting events that occur in primary care settings are upstream from major procedures and consist of referral to specialists and diagnostic testing. These discretionary decisions made during routine office visits to general

PAGE 17

17 practitioners have substantial impact on populati on health and the overall cost of health care (Sirovich et al. 2008). Specific Aims Quantify outpatient diagnostic imaging utilization over tw o years by primary care doctors caring for a cohort of patients that regularly attend clinic (i.e., are loyal) Collect and characterize demographic, c ontextual, and clinical factors for all patients in the cohort Collect and characterize demographic and pr actice factors for the primary care physicians regularly caring for the same patients in outpatient clinics Determine the relationships between pati ent, doctor, and clinic factors and the probability that patients had at least one imaging test during the two year study period (any use) Determine the relationships between pati ent, doctor, and clinic factors and the number of examinations performed in patients with at least one imaging test (intensity of use) Develop a model based risk adjustment me thod for imaging utilization (any use and intensity of use) produc ing an expect ed amount of imaging given known patient and doctor factors (imaging propensity) Estimate and partition the variability in imaging utilization using a hierarchical method which takes into account patients risk adjusted imaging propensity which is in turn influenced by each doctors tendency to use imaging in their practice Summary of Study The basic unit of interest for this study is a patient who regularly visits a primary care physician and the time frame is two year s. The main outcome is a measure of the amount of outpatient im aging performed in the period of st udy that was ordered by the same primary care physician on the patient in question. Specifying and quantifying the amount of imaging is a non-trivial task bec ause we wish to count all non-invasive diagnostic studies of any modality (e.g., CT, MRI, XRAY, Ultrasound, Nuclear Medicine,

PAGE 18

18 PET). Fortunately, the Radiology Department and the host institution for the study share a longstanding commitment to a robust m edical informatics infrastructure. Thus, complete and detailed records of all imagi ng events for the past 15 years are readily accessible. Also, to enhance medical management and clinical operations support, a full spectrum of clinical data are available for health services research such as this study. These informatics resources allow collecting a large amount of data on each patient that will be used to form explanatory and control va riables for analyzing drivers, enablers, and inhibitors of outpatient prim ary care imaging utilization. Finally, using credentialing databases from the host inst itution and publicly availabl e data from state licensing sources, provider level fact ors will be obtained for study. After characterizing and quantifying the im aging utilization over two years of study, the tests performed on each patient will be attributed to them and cross-tabulated according to the patients status (I npatient, Emergency Room, Outpatient) when performed and the type of provi der ordering the study (primary care vs specialist). Patient and provider level variables will be characterized, validated, and then individually evaluated in bivariate fashion wit h the main outcome (c ount of outpatient imaging tests ordered by patients primary care doctor). Initial multivariable modeling will be performed using logistic regression on t he whole data set (outcome=any imaging yes/no) followed by linear regression (with Poisson errors) on the non-zero observations. The joint contribution of the vari ous patient, provider, and clinic factors to any use (logistic regression) and intensity of use (Poisson regression) will be inferred from the odds ratios and coefficients respectively.

PAGE 19

19 To prepare for hierarchical analysis, a ze ro-inflated Poisson regression will be estimated using all observations and only patient level factors to form a single continuous variable from the m odel predictions. This serves to summarize the multiple patient level factors into a single number r epresenting risk adjust ed expected imaging or propensity for each patient. A two-level hierar chical model with patients at the first level nested within the 148 providers at the second level will be specified and fitted to the observed imaging counts for a ll patients. Only two ind ependent variables will be included: the propensity variable and a uni que ID number for each provider. By specifying a model that defines provider level intercepts (mean imaging use) and slopes (response to imaging propensity) two unique c haracteristics can be estimated for each provider and compared with each other. In addition, va riance components computed for each level in the hierarchical structure will serve to partition variation in imaging utilization between patients im aging propensity, providers general tendency to image, and providers response to clinical factors in their patients in the amount of imaging they order. Contribution to Literature This research will contribute to the literat ure about primary care imaging utilization in several ways. First of all, the study population of providers and patients is large, comprehensive, and unique (i.e., all primary care doctors and their loyal patients at a large metropolitan academic health center). A previously validated loyalty cohort methodology (detailed in setti ng, data, and variables secti on) identifies patients, doctors, and clinics in ongoing and stable relationships with each other (i.e., usual source of care or medical home). T he available data about imaging utilization and patient health status is comple te and highly detailed, having come directly from primary

PAGE 20

20 sources (i.e., electronic medical records and clinical radiology information systems). For example, the requesting provider is recorded for each imaging test, which allows stratification of utilization by who ordered the examination (patients primary care physician versus a specialist) and the pati ent care setting (outpatient, inpatient, emergency department). Complete data are also available for a ll outpatient visits, hospital stays, and emergency room encounters. T herefore, the analyt ic data will contain a robust set of clinical activity and risk adjustment variabl es, which will be used to estimate the risk adjusted expectation (propensity) for imagi ng utilization to a high degree of accuracy and precision. Once patient level clinical factors have been accounted for, residual variation between providers will be quantified and partitioned in a way that should shed considerable light on the contribution of va rious contextual factors. When analysis of primary care outpatient imagi ng utilization is extended to provider profiling, proper specification of patient leve l risk-adjustment and hierarchical analysis at the provider level is crucial to fairly applying thes e medical management tools and this study will advance knowledge of these issues. Policy Implications In the ongoing debate about natio nal health care policy the most divisive and vexing issues relate to unsustain able medical cost inflation. Th is is largely the result of increasing utilization of expensive, hi gh-technology diagnostic and therapeutic interventions which occurs at the discretion of physicians to a substantial degree. The magnitude of this physician discretion over utilization as opposed to evidence-based clinical need will determine succe ss of strategies to reduce costs targeted to physicians. These include education, point of care inte rvention, clinical decision support, and post

PAGE 21

21 hoc profiling for efficiency. This study will provide estimates for the relative contributions of clinical need, physician style, and non-clinical patient factors with respect to utilization of outpatient diagnosti c imaging in primary care. Assuming that clinical need remains as a significant and substantial driver of imaging utilization, this study provides insight into risk adjustment models which will be necessary for utilization management efforts going forward.

PAGE 22

22 CHAPTER 2 BACKGROUND This chapter will provide background material about medical imaging as a diagnostic test applied in routi ne outpatient primary care. Explanation of exactly what is meant by adult outpatient prim ary care will serve to further refine the setting for the research described in this dissertation. In addition to specifying exactly what procedures are and are not included in a definition of diagnostic imaging, the chapter will describe national trends toward rapidly increasing volume and cost of these services. Definition of Diagnostic Imaging A diagnostic imaging procedure (DIP) is defined as a discrete event with the following attributes. The subject of this ev ent is an individual in a provider-patient relationship with a medical practitioner. The pr actitioner initiates the event by means of an order for the DIP, and this can be verbal written, or electronic. The individual submitting the order will be call ed the ordering provi der (OP). Two other provider roles are required to complete the event and these are the performing pr ovider (PP) and the interpreting provider (IP). The PP interacts directly with the patient using some kind of diagnostic imaging equipment to produce images which are subsequently interpreted by the IP who communicates the fi ndings to the OP. Note that the while the three providers (OP, PP, and IP) are often s eparate individuals, a single person may perform all three roles. For example, an obstetrician may dec ide that a fetal ultrasound should be done on her patient, may personally perform the sc an, and interpret the images in real-time from the video display all in a single step. Formal communication of the interpretation back to the OP typically includes some permanent documentation of the findings and interpretation, which is placed into the patients medical record.

PAGE 23

23 In addition to the people involved, other attributes serve to describe a DIP. Although defined as a discrete event, a DIP occurs in a series of steps: ordering, performing, interpreting, and review of re sults. This has been referred to as the radiology round trip to emphasize the comple xity of the process (Thrall 2005b, Thrall 2005a, Thrall 2005c). In some settings, the time between these steps can be lengthy and quite variable. For most purposes, we def ine the procedure as having occurred at a single point in time, when t he PP finishes performing or c ompletes the examination. This is often referred to as the date of service in medical record systems, billing applications and in claims data. Two other attributes of a DIP must be articulated and these define the sort of equipment used to obtain the im ages (modality) and what part of the patient (body area) was examined. The term modality refers to the physical nature of the process used to create images of the pati ents body and is useful as a pr imary means of categorizing the equipment. For example, radiogr aphy uses invisible photons of high energy (X-Rays) directed at and through the pati ent from a fixed generating tube and detects the photons on a flat surface (film or digital plate). Computed Tomography (CT) also uses X-Rays but the generating tube spins around the patient along with a detector. The varying intensity of photons falling on the detector is combined with angular position to compute a tomographic (planar slice) im age of the patients inner structure. Magnetic resonance imaging (MRI) machines irradiate patients with sequences of radio waves from coils housed in a strong magnetic field. Very sens itive antennae detect weak echoes of the radio waves emitted by hydr ogen atoms within the patient s tissues and compute an image based on strength and fre quency. Nuclear medicine (including positron emission

PAGE 24

24 tomography=PET) involves injecting patients wit h a variety of radioactive tracers which emit gamma ray photons or energetic parti cles as they undergo decay. These are detected outside the patient and processed to form a map or image of activity. Finally, ultrasound (US) produces pulses of high frequen cy sound from a specialized transducer in direct contact with the patients skin. T he same transducer detects echoes of those sound waves that differ in amplitude and ti ming depending on tissue characteristics and spatial location. These signals are proce ssed into an image of anatomy immediately beneath the transducer which can be moved around to examine an entire region. The final attribute assigned to a particular DIP details the anatom ic regions of the patient (body area) that are exposed to t he radiation, particles or sound to produce images for subsequent interpretation. The nomenclature of body areas typically imaged is fairly straightforward and in cludes: head, chest, abdomen, pelvis, spine, arms, and legs. More specialized examinatio ns may cover particular anatomy such as coronary arteries, lungs, aorta, gallbladder, and so forth. When a modality is combined with a body area the result is a specific named DIP and a patient is said to have undergone (completed it) at the date/time (of serv ice) that they were dismissed from the testing facility and the images became available for interpretation. An example will help to clar ify these definitions. Mary Jones (an adult) makes an appointment with her internist (D r. Smith) and during the visi t, complains of frequent and increasing headaches. Dr. Smith may decide to order a DIP to excl ude the possibility of a structural lesion (e.g., a mass in the brain) before treating her with drugs to relieve the pain. In this case, Dr. Smith (OP) will order a test that eval uates the brain (body area = head) and has several options about modality including X-Ray, CT, MRI, Nuclear

PAGE 25

25 Medicine and even ultrasound. Dr. Smiths decision as to wh ich of these modalities to order relies on her personal a ssessment of the relative appr opriateness for Ms. Jones. Assuming that they opt for MRI, Dr. Sm ith orders an MRI of the head/brain for Ms. Jones to be performed sometime in the future Dr. Smiths order is transmitted in some fashion (computer order, phone, fax, paper prescription carried by Ms. Jones) to an imaging facility (often hospital based) that o ffers MRI scanning for outpatients. When Ms. Jones keeps her appointment, she is brought into the MRI scanning suite and asked to lie down on a movable couch which ca rries her into the actual MRI machine. She is instructed to breathe quietly and hold st ill while the technologists (PP) execute a pre-programmed protocol that directs the machine to ac quire images of Ms. Jones brain over the next several minutes. Upon completion of these imaging sequences and a brief observation period, Ms. Jones is released to return home. At the same time, the PP (technologist) executes commands to compl ete the examination at which time the images (computer files) ar e transmitted to storage medi a so as to be ready for download and interpretation and/or review. This sequence of events forms a single unit of outpatient imaging utilization and may be described by saying: Dr. Smith (OP) requested an MRI of the head on Ms. Jones (t he patient) during an outpatient visit and it was completed on (the date of service) by (PP) to be interpreted by (IP) with the report to be sent back to Dr. Smith (OP) for review. The next section deals with trends in the U.S. of rapidly increasing volumes of imaging utilization. To set the stage for this, consider the typical charges and reimbursements that might be submitted and paid, respectively, for Ms. Jones MRI of the head. It is important to note that there are two separat e billable imaging events

PAGE 26

26 arising from this sequence. The first is t he encounter at the imaging facility where the MRI images of Ms. Jones head were made. The outpatient imaging facility will bill Ms. Jones and/or her insurance carrier for the technical charge which will likely be well over $1000.00. The second is the bill rendered by t he physician who actually interprets the images, and this professional charge will a pproach $500.00. The ac tual dollar amounts of reimbursement received will depend on the payer, with di fferent payers reimbursing different amounts for the same service. Of course, Dr. Smith will also submit a bill for the office visit during which she ordered the scan on Ms. Jones. However, this is not directly attributed to the imaging test and is not counted as part of its cost. It should also be noted that in an increasing number of cases, the MRI ma chine might actually be owned (or leased) by Dr. Smith and sometimes she might interpret the image s herself. Under this scenario, Dr. Smith may directly bill for, or receive through more indirect means, most or all of the revenue generated by the technical and interpretation c harges. This practice (called self-referral of imaging) is controversial among physicians and is targeted as a driver of increasingly burdensome costs in the U.S. by government and pr ivate payers. Imaging Utilization and Costs No other branch of medical technology has experienced the explosive growth in volume and variety of available services that radiology has during the past two decades. The medical care industry in the U.S. has purchased and installed advanced imaging equipment at an astounding rate, outpacing all other countries. Figure 2-1 illustrates this trend in terms of number of imaging pr ocedures (Medicare Payment Advisory Commission 2005). Given that the current population of the U.S. is just now reaching 300M, this translates into about 1.4 imaging tests per person year. The most dynamic

PAGE 27

27 growth has occurred in CT scanning, with a steady increase in capability and indications for use occurring over the last 30 years. As the number of imagi ng procedures (many of which are CT scans) has increased, the cumulative effective radiati on dose to the average American has nearly doubled from 3.6 milli-Sieverts (mSv) in 1980 to 6.2 mSv in 2006 (National Council on Radiation Protection & Measur ements 2009). Several high profile articles in the past couple of years have raised concerns about a small but significant population risk for subsequent cancers induced by ionizing radiation delivered during medical imaging procedures (Brenner and Hall 2007, Fazel et al. 2009, Berrington de Gonzalez et al. 2009, Nyweide et al. 2009, Smith-Bindman et al. 2009). Over the period 1985-1990, established tec hnologies, such as CT, continued to grow in volume for Me dicare. At the same time, the ne w technology of MRI exploded in terms of utilization with a 372% increase in national procedural volume for Medicare (Boutwell and Mitchell 1993). Imaging costs to the Medicare system in the past two decades rose much more rapidly than any ot her component and now comprise at least 14% of total Part B expenditures for physician services as specified in a report by the Medical Payment Advisory Commission (MEDPAC) to the U.S. Congress (Medicare Payment Advisory Commission 2003). Imaging costs grew by approximately 10% per year during the period covered by t he report (1999-2002) co mpared with average growth of 3.3% per year for all physician services. Testifying before Congress in 2006, Glenn Hackbarth of MEDPAC amplified and extended prior reports and te stimony (Medicare Payment Advisory Commission 2006). He presented 1999-2004 data s howing growth in Medicare claims for diagnostic

PAGE 28

28 imaging as being the highest of all services at 62%. Furthermore, growth was especially high in emerging modalities (up to 140%) with even established technologies, like head CT, outpacing general growth at 43%. Figures 2-2 and 2-3 (exhibits in Mr. Hackbarths testimony) illustrate these poi nts and have been wid ely reproduced. Primary Care Setting This research examines imaging in outpat ient adult primary care, which is defined as being rendered by doctors trained in internal medicine, family practice, and general practice. This is the classic paradigm of clinical decision-making, in which patients present with signs, symptoms, known diagnoses, or physical abnormalities that generate moderate probability of one or more treatable conditi ons. Much of the literature about the utilization of health services by pr imary care providers deals with one of two possible responses to this situation: diagnos tic testing or referral to specialists. This study will not consider imaging utilization that o ccurs as part of disease screening programs, includ ing mammography, CT colonography, cardiac calcium detection with CT, lung cancer screening with CT, among several others. While important for population health and public policy, fundamental differences exist between imaging for screening and imaging for diagnostic or prognostic purposes. For example, once a universal population screening strat egy has been adopted, policy-makers are primarily concerned with under-utilization in the ta rget group. Conversely, in the case of diagnostic imaging, payers and regulators (at least in the U.S.) concentrate almost exclusively on problems of over-utilization. Additionally, this work will not attempt to analyze utilization of imaging that occurs during inpatient care or imaging tests that are ordered during work-up of patients in emergency and urgent care settings due to t he differing nature of medical decision-

PAGE 29

29 making during inpatient and urgent care encounters compared with primary care. Further, the study does not examine imaging utilization directed by surgeons, oncologists or other specialists, because patients seen by doctors in these fields have unique, disparate, and comp lex medical problems and co-mo rbidities. Finally, imaging for children directed by pediatric providers will not be considered as it is quite rare in primary care settings and quickly devolves in to specialty-oriented utilization after identification of congenital abnor malities, childhood cancer, or other serious problems. Imaging as Diagnostic Testing Imaging examinations that are interpre ted by someone other than the ordering doctor represent a hybrid of diagnostic test and specialist referral. This is because imaging tests (especially complex ones like CT and MR) are perceived as consultations by patients as well as by ordering physicians. In contrast, clinical laboratory tests (blood chemistry, hematology, mi crobiology, and so forth) ar e reported rather than interpreted and this distinction is crucia l. Non-imaging diagnosti c test results are generally reported in the form of numbers or simple fact assertions and are often produced by automated methods with minimal analytic input by the rendering personnel. Examples of such tests include blood chem istry, antigen/antibody assay, serum drug levels, and urinalysis. With these, t he ordering physician must synthesize an interpretation about whether or not the re sult is normal and then decide relevance to the clinical question. On the other hand, r adiologists produce interpretative documents that reach provisional conclusions about the probability of relev ant clinical conditions (or at least classes of disease). Sometimes these reports contain recommendations for further follow up, clinical correlation, or even treatment and thus also function as consultation notes.

PAGE 30

30 Figure 2-1. Total number of imaging studies by year in the U.S. Figure 2-2. Imaging shows highest cumulative growth in services per beneficiary (19992004).

PAGE 31

31 Figure 2-3. Cumulative growth in imaging volume varies by type (1999-2004).

PAGE 32

32 CHAPTER 3 LITERATURE REVIEW This chapter will summarize relevant publications that inform a conceptual model of imaging utilization in out patient primary care. A gener al overview of theory and empiric research about small area variation in patterns of health service utilization will show that clinical uncertainty, physician practice style, patient preferences, and economic factors all play roles. An important c ontributor to variation in use of diagnostic imaging is clinical uncertainty In fact, imaging tests are doubly subject to variation in utilization related to clinical uncertainty becau se, by definition, their main function is to reduce it. Thus, ambiguity about whether to obser ve, test, or treat in specific clinical situations is multiplied by uncertainty about which, if any, imaging test to order. In addition to how best to diagnos e and treat their patients, doc tors are also concerned about income, leisure time, satisfaction with pr actice, and mitigation of malpractice risk. These considerations influence utilization of imaging and their effect is magnified in the presence of clinical uncertainty. Finally, pat ients bring a complex mi xture of factors to decisions about diagnostic imaging. The Andersen Model Any examination of health services ut ilization must consider the Andersen behavioral model as an organizing framework. The Andersen model seeks to explain health care utilization in an entire community many members of whom do not seek medical care at all. Andersen reviewed hi s model 25 years after it was developed and described it mostly in terms of access to health care (Andersen 1995). The Andersen model is primarily applied to factors that det ermine the seeking of health care services in whole populations rather than what happens during encount ers with providers (e.g.,

PAGE 33

33 referral for diagnostic imaging). In a meta-analysis of papers using the Andersen model, only 2 of 139 examined provider characteristics such as specialty, experience, and gender (Phillips et al. 1998). Thus the Andersen model is not directly applicable here because in looking at nonscreening imaging utilization in primary care the denominator is patients who already have an active relationship with a doctor. Like prescription drugs, one cannot undergo a diagnostic imaging test without a prescription or referral from a health care provider. In Follands comprehensive review of variations in the use of medical care, this dichotomy in overall utilization is labeled as first o ccurrence versus intensity (Folland and Stano 1989). Much of the Andersen model deals wi th first occurrence whereas imaging utilization falls under the intensity concept. Therefore, the majority of this literature review will focus on works that inform the int ensity of outpatient imagi ng utilization in an existing patient-provider relationship. That being said, it will become evident from the empirical distribution of outpatient imaging examined in this study, that utilizati on also seems to have a two-part structure. These may be called any use and amount of--non-zero--use to account for a substantial fraction (over half) of patients with no imagi ng at all during the two years of data collection. However, in this study, even pat ients with no diagnostic imaging have visited their primary care doctor regularly and a diffe rently specified conceptual model will be articulated below that is specific to outpatient imaging this special setting. Small Area Variation In an ideal world, the intensity and mixture of imaging would be appropriate to each patients clinical sit uation regardless of any contextual differences among providers or patients. The c ounterfactual goes like this: Consider a situation where all

PAGE 34

34 doctors and patients in a particular setting have access to identical and comprehensive evidence about imaging appropriateness (Sistrom 2009). Further, physicians act purely as agents and patients make rational choices based on maximizing longevity and wellbeing. Finally, economic considerations are uniform across all doctors and patients (e.g., a single payer universal insurance progr am). In this idealized setting, the amount and types of imaging tests that patients underwent would be based solely on their clinical presentation includi ng pre-existing conditions and new symptoms, signs, and disease trajectory. This implies that, after fully accounting for differences in clinical presentation, the adjusted rate of imaging utilizatio n would vary minimally over different levels of aggregation. That bei ng said, there is little theoret ical guidance as to exactly what this minimal or natural variation woul d entail in terms of intensity and mixture of imaging or any other health servic e (Cain and Diehr 1992, Diehr et al. 1990). Empirical investigations have identified considerable variation in the use of medical services. The large body of literature documenting this phenomenon is referred to as research on small area variations. The seminal paper on this topic, which was published by Wennberg and Gittelsohn in 1973, compared utilization and expenditures among 13 hospital service areas in Vermont during 1969 and found large variations among them (Wennberg and Gi ttelsohn 1973). For example, appendectomy rates varied from 10 per 10,000 persons to 32 per 10,000 persons. Since then, numerous articles, monographs and texts have exam ined geographic variations at the international, state, regional, and local levels for a wide variety of types of medical services (Health Services Research Group 1992, Folland and Stano 1990, Pekoz et al. 2003, Stano 1991, Stano 1993, Wennberg 1993). Wennbergs legacy is perhaps best

PAGE 35

35 embodied in the Dartmouth Atlas of Healthcare which aggregates and analyzes Medicare claims going back to at least 1990 through the most recent past year to characterize variations in health service use across the entire U.S. in over 300 carefully defined contiguous regions (Wennberg 2004b) (Wennberg 2004a). Much of the available small area variations data deals with broad categories of services or spending. For example, the Dartmouth At las provides use rates and costs per beneficiary for the category of diagnostic services which includes laboratory tests and imaging procedures. When more granular area utilization analyse s have been undertaken, they often have examined countable and more high impact events like hospitalization, surgery, and invasive diagnostic procedures (e.g., endoscopy and cardiac catheterization) (Leape et al. 1990, Chassin et al. 1987). The policy concern arising from evidence of small area variations is that they persist even after controlling for factors such as age, gender, and race (Health Services Research Group 1992, Blumberg 1987, Davis et al. 2000b, Wennberg 1987a, Wennberg 2002). Moreover, research has found little evidence that the geographic variations were correlated with differences in health status, disease-specific population health indicators, or other measur es of health outcomes (Sirovich et al. 2006, Fisher et al. 2003b, Fisher et al. 2003a, Franks et al. 2000a, Brook and Lohr 1985). Thus, the lack of measurable population be nefit in higher use areas combined with rising health care spending, has led policy makers to hope that health care cost reduction can be obtained without welfare loss by ta rgeting upper outliers (Fisher et al. 2003a, Fisher et al. 2003b, Bodenheimer and Fernandez 2005, St ano 1993). With respect to imaging services in particular, a Medicare Paym ent Advisory Commission (MedPAC) report in

PAGE 36

36 2006 noted that the geographic variation in utilization of imaging services (coefficient of variation of 28%) is exceeded only by nonimaging diagnostic tests (coefficient of variation of 30%) (Medicare Payment Adviso ry Commission 2006). An earlier study in 2001 found that ther e was a large variation by state in the number of diagnostic imaging studies per 1000 Medicare beneficiaries with 10th percentile being 3038 and the 90th percentile at 4573 (Bhargav an and Sunshine 2005). As with other medical services that displa y small area variation, we must consider if inappropriate imaging accounts for the hi gh-utilization areas and/or if poor access to services explains the low-use areas (Leape et al. 1990). In 2008, the General Accounting Office (GAO) published an extens ive report about Medicare Part B imaging services that noted the rapid increase in spending and commented further on the extensive variation in utiliz ation rates among states (Gov ernment Accountability Office (U.S.GAO) 2008). For ex ample, by 2006, in-office imaging spending per beneficiary varied almost eight-fold across the states; from $62 in Vermont to $472 in Florida. The report goes on to state that Given the magnitude of the di fferences in imaging use across geographic areas, variation is more likel y due to differences in physician practice patterns rather than patient health status. Further concerns about the appropriateness of imaging use are raised by research on geogr aphic variation showing that, in general, more health care services do not necessar ily lead to improved outcomes (page 21). Just as scientists leverage observed variability in nature to pry into its inner workings, health policy makers target vari ability in utilization to better understand and control overall cost and quality. After char acterizing and quantifying medical practice variation, focus naturally turns towards ex plaining it (Folland and Stano 1990, Folland

PAGE 37

37 and Stano 1989). In ambulatory primary care, much of this attention is devoted to two processes; specialist referral (Franks et al. 2000a, Franks et al. 2000b, Franks et al. 1999) and diagnostic testing (Hartley et al. 1987, Epstein and McNeil 1987, Epstein and McNeil 1986, Epstein and McNeil 1985b, Epstein and McNeil 1985a). When making these decisions, providers and patients l abor under compound uncertainty about the presence of disease(s) as well as the efficacy and availability of referral and/or diagnostic testing. The largest barriers to uni form medical practice are lack of evidence, conflicting/ambiguous results, and incomple te dissemination of existing information (Eddy 1984, Eddy 1990, Eddy 2007). Prim ary care has been likened to a jazz performance where some basic structures and heuristics are in place but--owing to uncertainty--practitioners must im provise to a large degree (Miller et al. 2001). The next section addresses clinical uncertainty then turns to the contextual factors that play a substantial role in determi ning the amount, timing, and mi xture of diagnostic tests (including imaging) performed in pr imary care outpatient settings. Clinical Uncertainty Any attempt to analyze imaging utilization must first deal with clinical uncertainty. Economic theory provides valuable insight in to understanding the role of uncertainty in clinical decision making and makes a clear distinction between uncertainty and risk. Given complete estimates about relevant risks of disease and treatment, patients and doctors could calculate an appropriate strat egy to maximize expected clinical benefit (Cohen 1996, Wu 1996, Eeckhoudt 1996). Un certainty, on the other hand, can be defined as ambiguity or imprecision attending risk estimates that renders calculation of expected benefit difficult or impossible. Unce rtainty can attach to questions about what (if any) disease the patient has and/or the rela tive benefit of available treatments for the

PAGE 38

38 disease(s) (Eddy 1984, Wennberg et al. 1982). Given uncertainty about disease probability and optimal treatments, practice variations can occur for several reasons (e.g., physician practice style, patient attitudes, organizational influences) and these lead to observed variations in health service utilization. Thus, the evidence-based medicine (EBM) movement is predicated on helping physicians and patients to use medical research findings to reduce their co llective uncertainty wherever possible. The hope is to decrease knowledge gaps and in formation asymmetry about health care using various decision support tools to allo w patients and doctors acting as their agents to make better and more cons istent clinical decisions. Assuming that there is a scientific basis for allopathic medicine (i.e., EBM) implies that many diagnoses can be established and optimal treatm ents known with at least some certainty. Therefore, so me of the variation in clinic al activity must be due to random differences in disease incidence and prevalence in the population of interest. Routine--evidence based--care of these pat ients would be expected utilize different amounts and mixtures of services to meet disparate clinical needs. The amount of variation in health service utilization that cannot be explained and justified on the basis of existing and incident disease burden is considered to be unwarranted (Wennberg 2004a, Wennberg 2004b). The method by which expected and unwarranted variation in utilization are parsed from each other involv es risk adjustment which seeks to account for the effect of known di sease burden on utilization. Risk Adjustment Case-mix and risk adjustment methods have been developed primarily to support provider profiling of utiliz ation or patient outcomes (Chang and McCracken 1996, Greene et al. 1996, Salem-Schatz et al. 1994, Tucker et al. 1996, Welch et al. 1994).

PAGE 39

39 Consider, for example, the subject of th is dissertation: utiliz ation of non-screening diagnostic imaging tests by primary care pr oviders. A payer may calculate rates of utilization and seek to remedi ate doctors that are high end outliers in order to decrease overall costs. If raw rates of images per patient year are used in this way, targeted providers will--often correctly--object that the reason for their hi gh use is that the patients they see are sicker. Th is situation clearly calls fo r case mix adjustment to be fair to targeted doctors, hold pati ents harmless, and for scientific rigor. In addition to patient age and gender, outpat ient risk adjustment should include variables that capture relevant medical conditions and events. The prevailing method for obtaining such clinical variables is to convert administrative data (claims and/or prescriptions) into problem type and severity categories. The best known example in the U.S., is Ambulatory Care Groups (ACG), a proprietary method developed at Johns Hopkins. Outside the U.S., the most popular method for categorizing primary care case mix is the International Classif ication of Primary Care (ICPC). Davis used ICPC for case mix adjustment when studying provider variat ion in primary care practice activity (prescriptions and tests) in Australia (Davis et al. 2000a, Davis et al. 2002). However, there is little published literature that deals specifically with case mix or risk adjustment for outpatient primary care imaging utilization. One relevant paper deals with ambulatory test utilization (including chest x -ray) by general internists for patients with hypertension (Epstein and McNeil 1985b). Anot her looked at diagnostic services, which included radiology, am ong generalist and specialists caring for Medicaid patients in community practice (Eisenberg and Nicklin 1981). Both used complex case mix/risk adjustment schemes to account for diffe rences in patient demographics and illness

PAGE 40

40 burden. There is ample literat ure about case mix (or risk) adjustment in support of performance measurement, resource use com parison, and practice profiling in primary care settings. However, imaging is not the spec ific focus of these more general methods descriptions (Chang and McCracken 1996, Greene et al. 1996, Salem-Schatz et al. 1994, Tucker et al. 1996, Majeed et al. 2001b, Majeed et al. 2001a). Appropriateness and Supplier-Induced Demand There are several competing theories r egarding the causes of medical practice variation that remains after factoring out underlying patient disease burden (i.e., case mix adjustment). An early expl anatory model for clinical activity differences depended on categorizing health care utilization ev ents by their appropriateness(Wennberg et al. 1982, Wennberg 1987b). The attractiveness of being able to correlate high-intensity utilization with inappropriate services was enhanced by the implicatio n that expenditures could be reduced by targeting high use regi ons, practices or providers. However, despite considerable effort and expense, no convincing evidence has emerged to show any direct connection between the rate of vari ous types of health service utilization and expert consensus ratings about the appropria teness of care in both large and small regions (Chassin et al. 1987, Leape et al. 1990, Casparie 1996, Restuccia et al. 1996). When focus is narrowed somewhat to primary care providers and their rates of specialist referrals, still no relationship with appropriateness has been demonstrated (Fertig et al. 1993, Knottnerus et al. 1990). Another theory, termed S upplier-induced Demand posits that practice variations stem from differences in fi nancial benefits to the supplier (i.e., physician). However, Reinhardt is persuasive in arguing that the theory of supplier induced demand for

PAGE 41

41 inappropriate services fails empirical test s and that a more nuanc ed explanation is preferable (Reinhardt 1999). He argues fo r a preferred practice style that: would be an amalgam of (1) what the physician has been taught to view as best medical practice in medical sc hool and during residency training, (2) his or her subsequent re finement of the received doctrine on the basis of more recent literature and contin uing medical education, and (3) an adaptation to the dominant professi onal norms in a given locality. (Reinhardt 1999) That being said, Reinhardt does not deny t hat financial incentives may influence the central tendency of practice pattern s over time (Reinhardt 1999). Davis has studied primary care doctors in New Zeala nd and concludes that a supply hypothesis is not useful and that individual physicians pa tterns of test ordering and referral form a practice style that per sists over time (Davis et al. 2000b, Davis et al. 2000a, Davis et al. 2002). The practice style theory has gained considerable traction and is commonly cited to explain a broad range of variations in medical care delivery (Folland and Stano 1989, Grytten and Sorensen 2003, Welch et al. 1993, Wennberg et al. 1997, Sirovich et al. 2008). There are scholars of practice variation who argue that the si mplistic picture of clinical uncertainty allowing individual practice styles to emerge is theoretically incomplete and does not explain geographi c variations because purely personal differences between providers should aver age out (Stano 1991, St ano 1993, O'Neill and Kuder 2005). However, in some settings, vari ation in imaging utiliz ation rates persists even when there is clinical certainty. For example, there is st rong consensus about breast cancer screening intervals for wome n over 49 years old and considerable controversy about younger women. In spit e of this, small area variations in mammography rates in Ontario are similar across patient age groups (Goel et al. 1997).

PAGE 42

42 Westert believes that practice style is, at best, shorthand for a cluster of influences acting at the patient, community, and physician level to affect decision making (Westert and Groenewegen 1999). In this mo re granular and complex model, individual provider opportunities, incentives, and influences combi ne with shared clinical standards in the group and local medical community to shape their practice. These patterns can be broadly described as ranging between conserva tive and elaborate which translate into lower and higher use rates for diagnostic tests respectively. Summary This dissertation will articulate and quantif y factors driving outpatient primary care diagnostic imaging utilization. Further, the st udy will examine the relative contributions of case mix/clinical need versus contextual /practice style factors to the variation in imaging utilization. Most of t he literature about prim ary care practice variation in test ordering and referral relies on an assumpti on--sometimes unsta ted--of proper risk adjustment. To the extent that health service use variations are explained by differences in patient demographics and clinical variables, t hey should be of less interest to health services researchers (Diehr et al. 1990, Cain and Diehr 1992) than to epidemiologists. This epidemiologic/contextual distinction is crucial, even in primary care settings. However, there are relatively few empirical estimates of the relative contributions of these to overall variance. Grytten studied use of diagnostic tests among Norwegian primary care providers (Grytten and Sorensen 2003). The reasons for each visit as well as patient age and gender were used for model -based risk adjustment (clinical need) of diagnostic test expenditures on a per visit basis. The remaining variation ranged between 47-66%, was attributed to practice style, and seemed to be a sticky attribute that followed individual providers who changed practice locations during the study.

PAGE 43

43 Other authors have looked at ambulatory pr imary care practices and estimated the variation in expendit ure per patient from clinical need (case mix adjustment) to be about 60% (Phelps et al. 1994, Davis et al. 2000b). The next chapter summarizes the research on drivers of resource utilization (emphasizing diagnostic tests and referrals) by primary care provi ders, then presents a conceptual framework for analyzi ng imaging utilization in parti cular. Factors related to clinical need will be addr essed first, followed by considerat ion of the various contextual factors influencing imaging utilization includi ng patient, provider, and practice factors.

PAGE 44

44 CHAPTER 4 CONCEPTUAL FRAMEWORK This dissertation examines the utiliza tion of outpatient diagnostic imaging by primary care doctors caring for adult pati ents. Because the study will not consider imaging that occurs as part of disease scr eening programs, it does not include imaging tests that might be scheduled an d performed routinely without an explicit decision by the doctor. Specifically, th is study examines diagnostic imaging tests ordered by a patients primary care doctor to address clinical issues raised either during a visit, a phone call (or email), or another health syst em encounter (ER visit, hospitalization). This chapter presents a conceptual model to account for relevant driving and modifying factors influencing the amount of such--non-screening--imagi ng performed by order of the primary care doctor on a patient. As articulated in the literature review, dr ivers and modifying fact ors of primary care resource utilization can be divided into two major classes. The first one is clinical need. For a given patient, clinical need arises fr om existing or devel oping signs, symptoms, trauma, and illness. The second class of contextual factors can be further grouped by attribution to the agents or organizations involv ed. Specifically, most contextual factors will accrue to either the patient or the doctor. Other general groupings for contextual factors include the clinic or practice in wh ich the doctor sees the patient and any larger provider organization (e.g., academic facu lty practice, HMO, PPO, and etc). Another category to be considered (at least in the U.S. ) relates to the payer or insurer. Other factors relate to the stru cture and process of outpatient imaging facilities and methods by which imaging tests are ordered. Finally both patient and doctor live and work in communities that can be represented at various levels of aggregation (e.g., city, county,

PAGE 45

45 state) and the phenomenon of geographic variation implies that at least some factors operate at the community level. This chapter discusses the relevant driv ing and modifying fact ors of outpatient diagnostic imaging utilization in primary ca re as defined above. For each individual factor, expectations of the direction and strength of effect it should have on imaging utilization are justified by t heory and informed by existing empiric literature. The final section summarizes and diagrams the general relationships among clinical need, context, and imagi ng utilization. Clinical Need This set of factors is perhaps the easie st to state and comprehend and yet is the hardest to measure and model. In essence, imaging tests are done to address clinical uncertainty that arises about an existing condition or a clinical event. Existing conditions are diagnoses made by a physician and theref ore known to them. Uncertainty about an existing diagnosis relates either to current stage/status of the disease or treatment choice/response for that condition. Clinical events are defined by development or worsening of a sign, symptom, or abnormal test results. The uncertainty engendered by such clinical events comes from the doctors need to determine if a new diagnosis needs to be made or if an existing disease is responsible. In either case, the function of the diagnostic imaging test is to better i dentify or exclude treatable disease to guide therapeutic decisions. The only other clinical event not directly accounted for in the preceding explication is trauma, for which imaging is often performed to assess severity and type of injury. Severe trauma is treat ed in acute care hospitals and associated imaging tests would not be counted as primar y care outpatient utilization. On later ambulatory care visits, the post traumatic state will be identified as an existing

PAGE 46

46 condition for this model and outpatient imagi ng tests might be done to address residual problems. As described previously, under ideal condit ions of evidence-based practice, the mixture and amount of imaging that a patient received in a year would depend only on their existing diagnoses, disease status, and c linical events. However, the current state of medical knowledge and the cons istency with which it is appli ed in actual practice is such that doctors decisions about intens ity and mixture of diagnostic imaging are quite variable even when faced with identical clinical scenarios. This variability in diagnostic decision-making styles among physicians is enabled by systematic uncertainty about what test(s)--if any--are suitable for various c linical scenarios. Howe ver, uncertainty or ignorance about the appropriateness (expected c linical benefit) of imaging is necessary but not sufficient for variability in utilizatio n to occur. The contextual factors discussed below, can influence utilization in the face of uncertainty or ignor ance about the optimal diagnostic strategy. For the most part, clinical need variables operate exclusively at the level of individual patients. The only exceptions to this might be in disease screening programs, communicable illness, or env ironmental health issues. However, this study excludes imaging related to disease screening programs. Furthermore, in th e setting for this study (primary care rendered in a large Northeastern metropolitan area from July 2007 through June 2009) no unusual communicable disease epidemics or environmental health issues occurred which might distort the assumption about clinical needs working purely at patient level.

PAGE 47

47 Context: Patient Basic patient demographics such as gender and age strongly correlate with the amount and type of clinical need based on t heir complex relationships with many diseases and health states. These two variables are often included in case-mix or risk adjustment models. They serve as proxies for an individuals propensity to develop health conditions and biological responses to disease, testing and treatment. However, these same demographic factors also exert social and psychological effects on the patients likelihood to seek or accept di agnostic testing under various scenarios. For example, it is conceivable that men and wom en of similar age might choose differently in some systematic way about diagnostic testing for the same clinical scenario based on level of anxiety related to attitudes about risk and uncertainty. There is some empiric evidence about patie nt preferences as related to clinical resource use in general as revealed by surveys, interviews, and responses to hypothetical scenarios. Anthony et. al., found that elderly Medicare beneficiaries expressed substantial differenc es in their preferences for seeing a doctor right away, having tests, and for specialist care (Anthony et al. 2009). When individual Medicare utilization was modeled with t hese preferences as predict ors (along with demographic control variables), those who preferred care right away and from specialists had higher overall healthcare utilization rates. However, at larger levels of aggregation (regional variation) differences in patient prefer ence were uniformly distributed and did not explain variations in cost. Socioeconomic status (SES) factors such as level of education, income, and ethnicity affect individual patient tendencie s to seek care and comply with provider recommendations. Empiric evidence for differenc es in healthcare utilization and patient

PAGE 48

48 outcomes abounds in the disparities literat ure, although separatin g ethnicity from economic factors as causes of such dispar ities is a matter of considerable debate. Confounded with patients pref erences based on socio-cultur al characteristics are the doctors own biases and attitudes. Perhaps the best known (and controversial) example of this is Schulmans survey about re commendation for cardiac catheterization (Schulman et al. 1999). He found that, in hypothetical clinical scenarios, patient race and gender influenced physicians tendency to recommend diagnostic work up for otherwise identical presentations of chest pain. In one intriguing study of managed care claims, Franks found that case mix adjust ed use of diagnostic testing was actually higher in patients from lower SES zip codes (Franks and Fiscella 2002). He hypothesized that doctors tend to order more tests when they perceive that patients cannot articulate their histories and current symptoms. However, in general, patients in lower SES and minority ethnicities (in the U.S.) tend to receive lower levels of diagnostic testing in acute and sub-acute care settings (Goldstein et al. 2006, Isaacs et al. 2004, Pezzin et al. 2007, Quintana et al. 1997). Context: Physician With provider factors, it is important to remember that we ar e limiting the scope of discussion to outpatient adult pr imary care. Primary care doc tors are trained in several distinct ways in the U.S., and they may c hoose different types of patients to see. Doctors trained in family medicine often s ee children and pregnant women in addition to adult patients. On the other hand, geriatrics-trained physicians tend to take care of elderly patients. Many primar y care internists have additional training after their 3 years of internal medicine. For example, doctors with some endocrinology training after their internal medicine residency might skew t heir practice towards adult diabetic patients

PAGE 49

49 even though they are not render ing specialty care. However, assuming that patientlevel case-mix variables are accounted for, we focus on how factors like experience, gender, training and specialization might affect a doctors tendency to order diagnostic imaging in similar clinical scenario s arising in adult primary care. Training and specialization have direct effects on test-ordering behavior in primary care. Although, all U.S. physicians complete at least a four-year course of study leading to an M.D. degree, specific c ourses about imaging use are rare ly offered or required in M.D. curricula, with the majority of training and experience about radiology gained during residency. Since post-graduate medical education is conducted in apprenticeship models, the relative int ensity of imaging utilizat ion at the training institution strongly influences subsequent decision making about imaging during practice. Small area variations may be relevant here because in many health care referral regions, the academic medical cent ers account for and may influence much of the measured utilization. Even within geographic regions, residency training occurs in different types of institutions and community settings. There is wide disparity in these with a spectrum ranging from residencies conducted in small non-affiliated rural community health centers to the classic tertiary care safety net academic health center owned by a medical school operating in a lar ge city. The hypothesis is that doctors trained in high use regions and at large ac ademic centers will tend to order diagnostic imaging more frequently than those trained in smaller centers and low use regions (Chassin 1993, Eisenberg 1986a, Epstein and McNeil 1985a, Folland and Stano 1990, Grytten and Sorensen 2003, Landon et al. 2001, O'Neill and Kuder 2005).

PAGE 50

50 Experience is partly defined by length of time in practice and is also directly related to when training took place; the two fact ors may be very hard to separate. Timing of training is particularly relevant to imagi ng utilization because the technology has advanced and evolved rapidly and consiste ntly over the past 2-3 decades. The expectation is that a recent graduate w ho routinely worked with advanced imaging techniques in training will be more likely to order MR and CT in subsequent practice than a doctor trained prior to their diffusion w ho may not be aware of what is available. Length and extent of experience itself affects diagnostic de cision making with doctors having greater experience tending to order less imaging tests in identical scenarios (Bugter-Maessen et al. 1996, Childs and Hunter 1972, Couchman et al. 2004, Couchman et al. 2005, Eisenberg and Nicklin 1981, Sood et al. 2007, Whiting et al. 2007, Williams et al. 1982). Physician gender also may have an effe ct on use of imaging independent of experience and training, although, with the recent and substantia l increase in the fraction of women trained in and practicing medicine in the U.S., it may be difficult to measure since women physicians tend to be younger and trained la ter. Empiric data about physician gender has been mixed with studies showing both increased and decreased tendency to order imaging and ot her diagnostics tests between male and female doctors in primary care (Britt et al. 1996, Rosen et al. 1997, Sood et al. 2007). Physician workload may influence tendency to order imaging tests in several ways; both in the long run (months and years) and the short run (durin g the course of a day or week). This effect is mediated through each physicians perception of time pressure overall and during a pa rticular visit. For example, an initial visit to a primary

PAGE 51

51 care physician for headache occurring at the end of a busy day in clinic might be more likely to include an imaging test than otherwis e. Such a doctor, caught between time constraint and fear about missing a diagnosis mi ght order an MRI of the head instead of spending an extra 15 minutes doing a detail ed neurological exam. On the other hand, the same doctor may choose to refer the patient to a specialist (neur ologist in the case of headache) rather than order any imaging if t hey believe that strategy would move the patient towards a temporary disposition more rapidly. There is very little empiric evidence about this particular factor to gui de us in determining the sign of a possible correlation between practice workload and imaging intensity. Economic factors may influence imaging te st ordering (Rein hardt 1999). Aside from pure income maximization, physicians may seek to increase their personal utility in other ways. Eisenberg refers to this as physician as self-fulfilling practitioner (Eisenberg 1986a, Eisenberg 1985). Also, in ac ting as the patients agent, physicians may take the patients financial situation in to account when making decisions on their behalf (Eisenberg 1986b). However, imaging ut ilization also may be influenced by physicians seeking to directly increase their income by ordering and performing imaging tests. In such cases, the reim bursement for the imaging test it self is paid to the ordering physician through several pathways includ ing ownership stake in the imaging equipment (technical fee) and/or interpretati on of the examination (professional fee). Called self-referral, there is a large body of literature about its practice and ramifications. Hillman and others make a strong empiric and economic case that such direct financial incentives have powerful pos itive effects on imaging utilization volume and charges (Hillman et al. 1990, Hillman et al. 1992, Hillman 2004, Gazelle et al.

PAGE 52

52 2007). Although the Starke laws have been in effect since the 1990s and have been renewed and revised at least once, recent evidence from California shows that there is still substantial self-referral of advanced diagnostic imaging and that various mechanisms other than direct ownership of equipment allow this arrangement to continue under current statute (Mitchell 2007). Context: Malpractice Malpractice deserves special mention because it operates at the individual provider level in addition to the practice and community levels. The term defensive medicine is often used to describe t he phenomenon of doctors increased ordering of imaging--and other diagnostic--tests based on fear of being sued for failure to diagnose (Kessler and McClellan 2002, Sood et al. 2007, DeKay and Asch 1998). Personal experience with being sued for malpractice c an have profound effects on an individual doctors psychology and practice pattern that may persist for years or decades. By all accounts, it is an extremely negative and unsettling event that induces a strong desire to avoid repeat occurrences (Hermer and Brody 2010). Thus, if a doctor is sued for missing a diagnosis, the expectation is that they will alter future pr actice toward more diagnostic testing. This behavior will likely not be limited to the sc enario leading to the suit, but generalized across patients of differ ent clinical classes and types of diagnostic tests (including imaging). Ev en if a doctor is sued for malpractice unrelated to a diagnostic error, he or she may tend to general defensiveness which may lead to greater use of diagnostic imaging at lower leve ls of uncertainty than before the event. Dekay and Asch wrote a seminal paper us ing expected utility theory combined with decision analytic m odeling to show the causes and consequences of malpractice liability on diagnostic testing (DeKay and Asch 1998). They proved that consideration of

PAGE 53

53 liability by physicians faced with a classic ob serve, treat, test choice set must widen the zone (over disease probabilities) in which test ing is the preferred strategy. They also prove that there is an obligate utility loss to patients incurred by this extra testing. They also assert that physicians substantially overestimate the protecti on afforded them by doing more testing. In retrospect, physicians generally overestimate their ability to have made the correct diagnosis in advance. Thus hindsight and regret bias combine with unrealistic expectations about the efficacy of imaging and result in a near magical belief that the right test would have saved the day (DeKay and Asch 1998). On a state by state basis, medico-legal climate varies considerably based partly on the statutory and precedentbased status of malpractice and tort laws. Baiker, Fisher, and Chandras paper examining trends in Medicare costs and malpractice burden in the U.S. over the 1990s used states as the unit of measure (Baicker et al. 2007). They showed that imaging cost increases were significantly correlated with trends in malpractice premium and payouts. Across the 50 states, a 10% increase in malpractice premiums/payouts resulted in about two percent increase in physician services costs. They estimated that the observed 10 year increase in malpractice of 60% resulted in more than 15B extra in spending with imaging being the largest contributor by far (Baicker et al. 2007). It should be noted that there are diverging views with more recent papers questioning the empiri c basis for a large effect by defensive medicine and suggesting that even comprehens ive tort reform might not have much actual effect on health costs (Herme r and Brody 2010, Sloan and Shadle 2009). The theory of how liability concerns increase diagnostic testing at all levels of aggregation rests on the assumpti on that providers perceive themselves to be at risk for

PAGE 54

54 malpractice action even if they have personally not been sued before (Fenn et al. 2007, Kessler et al. 2006). Clinicians are rather bad at a ssessing their own liability risk and tend to overestimate personal pr obability of being sued (Holtgrave et al. 1991, Lawthers et al. 1992, Kessler and McClellan 2002, DeKay and Asch 1998). This study will examine a large primary care practice confined to a single institution. Therefore, the local and state malpractice climate is constant though the free floati ng fear of medical liability might vary by practice In the current study, the only va riable available to directly probe the effect of defensive medicine on imaging utilization is each physicians history of being sued or not during the preceding decade. Context: Practice Organization After limiting consideration to adult outpati ent primary care in the U.S., there are several types of practice setting and organi zational dimensions to be considered. Perhaps the most relevant is the employment arrangement for the physicians. Health care organizations structured as staff m odels, where doctors are salaried (e.g., traditional HMO, academic health centers, m ilitary, and VA) may have different patterns of diagnostic imaging utilization based on individual incentives and medical management initiatives than private practice and fee-for-service settings (Epstein and McNeil 1985a, Kravitz and Greenfield 1995). Even if we ex clude consideration of direct financial benefit from self-re ferral of imaging, independent practitioners and groups are generally less constrained in their ability to order radiology tests. Aside from employment st ructures and compensation arrangements, primary care physician practices differ in the extent and manner in which peer pressure is exerted. For example, in a small private practice primary care group there may be minimal (if any) formal influence on actual practice styles among members, including diagnostic

PAGE 55

55 radiology utilization. At the ot her extreme, in some staff m odel practices, leadership may routinely profile imaging utilization at the provi der level and seek to control it with direct incentives or remedial measures (Neilson et al. 2004, Axt-Adam et al. 1993, Solomon et al. 1998). Context: Payer and Prices In the U.S., a patients insurance status has a profound impact on access to primary care services and may influence t he frequency of outpatient visits. In this conceptual model, non screening imaging tests are ordered to address issues identified during a patient-physician enc ounter, thus a lower frequen cy of encounters provides fewer opportunities for imaging to be ordered. Additionally, greater financial burden (self-pay or high co-pay/deductibles) a ssociated with imaging tests will reduce a patients tendency to agree to and/or undergo expensive imaging tests, even if ordered. Physicians may be aware of a patients financial or insurance status and in their role as financial agents, might choose to for go imaging tests depending on costs (Mort et al. 1996, Shen et al. 2004, Pham et al. 2007). There is a substant ial literature concerning awareness of diagnostic test costs. A recent s ystematic review by Allen concluded that most doctors have a very limited understanding of diagnostic and non-drug therapeutic costs (Allan and Lexchin 2008). Soods more fo cused review of literature about multiple contextual factors in test ordering tendencies found that cost awareness (among both doctors and patients) was relevant (Sood et al. 2007). In general, when price information is made available to clinicians, they tend to reduce their likelihood to order diagnostic tests (Hoey et al. 1982, Cummings et al. 1982, Long et al. 1983). Bates reported a 5% decrease in clinical laboratory test char ges during inpatient episodes

PAGE 56

56 after price information was routinely displa yed during electronic order entry (Bates et al. 1997). Increasingly, various payers (including M edicare), being aware of the rising costs of outpatient imaging tests, have begun to employ cost-containment measures specifically related to imagi ng (Government Account ability Office (U.S.GAO) 2008). One strategy is to profile indivi dual physicians with respect to imaging (and other resource) utilization and place them into various tiers that give preference to efficient providers in various ways. An emerging trend is for paye rs to contract with one of several imaging benefits management entities (e.g., National Im aging Associates, CareCore National, and others). These companies serve as gat ekeepers for outpatient diagnostic imaging by requiring providers and/or patients to obtain pre-authorization on a case by case basis before tests are scheduled (Otero et al. 2006, Brant-Zawadzki 1994, Bernardy et al. 2009). In addition to simple barrier ef fects mediated by call center and other administrative delays, requests for imaging tests may be denied based on proprietary medical necessity or appropriateness rules. Such arrangements can have considerable impact on the likelihood that a given patient -doctor encounter will result in a scheduled and completed diagnostic imaging test (Blachar et al. 2006, Levy et al. 2006, Smulowitz et al. 2009). Context: Access to Imaging A final category of contextual factor re lates to the facilities and processes that underlie how diagnostic imagi ng examinations are ordered, authorized, scheduled, performed, and interpreted. As described in the Background, the so-called radiology round trip is a complex chain of events that begins with a doctor-patient interaction of some kind that raises a clinical question that might be answered by imaging. In the case

PAGE 57

57 of a completed examination, the round trip generally ends when an interpretative report about the imaging test gets read by the doctor, acted upon, and relayed to the patient. The availability of diagnostic imaging facili ties in terms of proximity and capacity influence the doctors tendency to order and the patients ability to obtain examinations, even after omitting any consideration of te sting facilities owned or operated by the referring physician (self referra l). Nonetheless, ready availa bility and ease of scheduling for various tests will positively influence decis ion making about imaging by both doctor and patient. The means by which tests are ordered and scheduled by the doctor, office staff, and the patient can introduce barrier or enabling effects. For example, changing from written or verbal orders to a system that requires doctors to log on to a computer and order the test personally, may exert barrier effects if doctors believe that more of their time and effort is required to assert the order On the other hand, a robust computerized point of care radiology schedu ling system can allow patients to leave the clinic with their radiology appointment in hand and will increase utiliz ation by virtue of convenience. Patient experience at the diagnostic imaging facility may affect compliance with imaging orders as well as a doctors tendency to order in the first place. Long wait times and other negative experiences at the testi ng facility will become known to the doctor and other patients over time. If doctor and patient are contem plating a diagnostic imaging test, expected difficulty in scheduling and/or long waiting times on the day of examination may be perceived as too much trouble and bother. Radiologist training, skill and style will affect how they interpret any given test and this is manifest in the report that gets sent back to t he referring doctor. If reports tend to be late in arriving,

PAGE 58

58 raise more questions than they answer, and frequently contain recommendations for further testing, doctors may come to rely less on diagnostic imaging. A relatively new development is how the resulting images are handled and distri buted to patients and referring doctors. In modern computerized radi ology practices, patients are given a CD with all the images on them which can be br ought back to the referring doctor for review. Increasingly, images can be view ed on line by the referring doctor along with reports. Imaging providers offer these and other services to increase their market share. Summary The final decision of whet her or not to order and under go an imaging test thus depends on all these factors and the complex interactions among them. As shown in Figure 4-1, the ideal level of utilization is determined by clinical need under conditions of certainty. Adding clinical uncertainty then allows for deviation from the ideal level of utilization, with the variati on potentially being positive or negative. The various contextual factors add additional variation in the observed level of utilization that persists even after case-mix and risk adjustment. For purposes of the empirical analysi s, the conceptual framework can be summarized as: Imaging utilizat ion = f[clinical need, patient factors, physician factors, malpractice environment, practice organi zation, payer, access to imaging] The next chapter will describe the study setting, data sour ces, and variables that will be used in the analysis.

PAGE 59

59 Figure 4-1. Summary diagram of conceptual model for outpati ent imaging utilization in primary care. Clinical Events Co-Morbidities Demographics Evidence Base 1 Necessary Imaging [known, large benefit] Discretionary Imaging [uncertain benefit] Patient Physician Practice Malpractice *Structure *Process *Payer *Cost *Of Imaging 2 + Contingent Imaging Actual Measured Imaging Utilization

PAGE 60

60 CHAPTER 5 SETTING, DATA SOURCES, AND VARIABLES This chapter will describe the institutional setting for this study and then focus on outpatient diagnostic imaging services. The gener al setting includes a large group of primary care physicians practicing in 15 separate locations and a hospital based radiology department with se veral service sites. Patients in the practice who identify wit h a single attending physician as their primary care provider are tracked as a lo yalty cohort. The majo rity of outpatient imaging performed on these pat ients occurs at the asso ciated radiology department. Radiology databases provide co unts of outpatient diagnosti c imaging tests accruing to patients, doctors, and clinics. Other c linical and administrative databases provide information on patient demographics, clinical problems, medical activity (visits, hospitalizations, etc), and physician characteristics. Settings Primary Care Practice This study was conducted at Massachus etts General Hospital (MGH) and the associated Physician Organization (MGPO). In close association with the MGH academic medical center, MGPO is a large multi-specialty fa culty group practice with a full complement of adult prim ary and specialty care. The fu ll time faculty are salaried employees of MGH and the nonradiologists (e.g., primar y care providers) have no direct financial incentives relating to volu me and/or revenues from imaging or laboratory tests that they order. Malpractice insur ance is supplied by the MGPO under a selfinsurance pool arrangement and the providers have limited per sonal liability. However,

PAGE 61

61 malpractice actions are identified at the indi vidual provider level fo r reporting to licensing authorities and government data-banks. This work focuses on the primary care portion of the practice and was aided by an entity within the MGPO called t he Primary Care Operations Improvement (PCOI) group. Physician leaders and staff members in PC OI conduct analyses of various aspects of the practice for internal quality assessment and improvement as well as for presentation and publication in scholarly settings. In suppor t of these efforts, Dr. Steve Atlas (a primary care physician and health services researcher at MGH) and colleagues have devised and validated a method fo r identifying a group of prim ary care patients, doctors, and clinics with stable relationships to each ot her--the result is termed a loyalty cohort. A cohort is identified by year and comprises a list of patients who are considered to be loyal to a single primary care provider by vi rtue of their outpatient visits as documented in the electronic medical record over the three years ending in the cohort year. Loyalty assertions for a given patient-doctor pair are calculated probabilis tically using five variables derived from visit data (Atlas et al. 2006, Lasko et al. 2006, Wasiak et al. 2008, Atlas et al. 2009). These are listed below. Waiting fraction: the total number of days waited for appointments with the given physician, divided by the total waited for all physicians combined. Visit difference: the total number of visi ts that a patient has made to the given physician minus the total to a ll other physicians combined. Days since last visit: the number of da ys since the last visit to the given physician. Future difference: the total number of appointments scheduled for future visits with the given physician, minus the tota l for all other physicians combined. Idle ratio: the number of days since the last visit to the given physician, divided by the number of days since the first visit.

PAGE 62

62 A logistic model was developed and validated on subsets using loyalty assertions as stated by the linked primary care physicians (Lasko et al. 2006). In comparison with this stated (gold-stand ard) loyalty status, model predict ions were over 95% accurate. When the same technique was applied to all r egistered primary care patients, the study (loyalty) cohort described below was the result. Outpatient Radiology The Radiology Department at MGH provides a full range of imaging services for inpatient, emergency room, and outpatient practices. The main department is located at the MGH campus with several ancillary outpatient sites in the Boston area. The whole department is linked via a robust electronic in frastructure and the radiology informatics group is widely recognized as being among the most advanced and sophisticated in the world. The relevance for this study is two-fo ld. First, electronic records of all imaging tests are housed in a data warehouse that has been created and maintained with great care and attention to detail. This means t hat all imaging tests going back to at least 1995 are listed with complete an d accurate information about several items relevant to studying outpatient imaging uti lization, including the identi ty of the doctor ordering the test, dates of ordering/comple tion, modality, body area, and pat ient status at time of examination (e.g., inpatient, outpatient). In total, almost 100 items of information are stored about each test with the o nes just mentioned being most relevant to the present study. Secondly, outpatient radiology ordering by all primar y care physicians at MGPO is performed through the same web-based system This is called Radiology Order Entry (ROE) and it has been in full use for all moda lities since 2004. After selecting from a dynamic menu of all outpatient radiology ex ams, clinicians are required to input

PAGE 63

63 structured information about why the test is being ordered via checkboxes for signs/symptoms/diagnoses supplemented by a free text input field. A popular feature of the ROE system is a patient scheduling module. This allows an appointment for the imaging te st to be made at the point of order (doctors office) without multiple phone calls, faxes, or other efforts. In addition, starting in 2005, a decision support (DS) component was added to the system t hat is triggered with all orders for CT, MRI, and nuclear medicine studies. The DS logic displays a 1-9 utility score based on the test chosen, patient demographics, and the reasons given for the test. The scores are grouped as follows: 1-3=R ed/low, 4-6=Yellow/i ntermediate, and 79=Green/high. A red score does not precl ude going ahead and ordering the test but the clinician must provide a reason for doing so prior to proceeding with scheduling. Several papers have been published about the ROE-DS system with the most relevant describing the effect on total outpati ent CT, MR, and ultrasound volumes at MGH (Rosenthal et al. 2006, Sistrom et al. 2009). We found that, afte r correction for overall practice activity, there were substantial redu ctions in growth rates, especially for CT scans (Sistrom et al. 2009). It should be noted that during the time period (July 2007June 2009) covered by this study of outpati ent imaging utilization in the MGPO primary care practice, the ROE-DS system had been in use for at least two years. Further, no substantial changes were made to the system functionality and only minor alterations were made to the DS scores. Data Sources Loyalty Cohort As described above, the loyalty cohorts are compiled by Dr. Atlas and the PCOI staff every year and the one used in this study is for 2008. This means that MGPO

PAGE 64

64 primary care patients active in 2008 were gathered and their electronic clinic scheduling records going back through 2006 were compiled and analyzed. There were 139,609 unique patients who were candidates for loyalt y status. Based on the algorithm, 87,568 were flagged as being loyal to a single primary care provider in the MGPO. There were 804 patients who were loyal to 26 providers with less than 100 loyal patients in their practices. These were excl uded leaving 86,764 patients. Of these, 1483 were loyal to four providers who had left t he MGPO in late 2008 or the first or second quarters of 2009 and demographic data were not available for 4 of the remaining patients. The analytic sample includes 85,277 patients, loyal to one of 148 primary care physicians who will be characterized below. These phy sicians practice in one of 15 clinics distributed through the greater Boston area. It should be noted that the clinics sometimes do use residents and medical students. However, the ongoing doctor patient relationship is with the identifie d primary care physician. In fa ct, the raison detre of the loyalty cohort methodology is to unambi guously identify this relationship. Patient Details MGH and the MGPO have es tablished a common Research Patient Data Repository (RPDR), which aggre gates numerous sources of information into a single set of databases designed for use in pati ent-centered clinical epidemiology. These include billing and encounter data for inpat ient, emergency department, and outpatient services as well as the contents of inpatient hospital information systems (HIS) and outpatient electronic medical records (EMR). This study was approved by the Institutional Review Board at MGH under an expedited protoc ol for analysis of existing data. Informed consent was not required and was not obtained.

PAGE 65

65 Physician Details The MGPO credentialing database was used to obtain relevant information pertaining to the 148 primary care doctors in the study, including gender, birth year, medical school graduation year and medical school state (or country for foreign medical graduates). By definition, all doctors were licensed in the state of Massachusetts and the publicly available web site for the Massac husetts Board of Medicine was queried to determine whether or not each doctor had been s ued for medical malpractice in the past 10 years. Imaging Utilization The 85,277 patient medical record num bers were queried against the radiology data warehouse to return all diagnosti c imaging tests performed during the study interval (July 1, 2007 through June 30, 2009). The query specifically excluded interventional procedures (e .g., biopsy, drainage, cathet er angiography, embolization, and vascular stenting among others). Also spec ifically excluded we re mammograms, as these are almost all related in some way to breast cancer screening. In addition to patient medical record number modality, body area, place of service (ER, inpatient, outpatient), and the unique prov ider number of the doctor who ordered the exam was obtained. The 221,571 diagnostic imaging procedures performed on cohort patients over the two-year study interval cross tabulated by place of service and modality are shown in Table 5-1. The 157,463 outpatient diagnosti c imaging procedures are cross tabulated by the class of the orderi ng doctor and modality in Table 5-2. Finally, the 60,938 diagnostic imaging procedures performed in the outpatient setting and ordered by the

PAGE 66

66 patients linked (loyal) doctor are cross tabulated by body area imaged and modality in Table 5-3. The unit of observation for this study is the patient. The outcome variable is constructed by aggregating and summing (by patient) the 60,983 out patient imaging procedures ordered by the primary care provi der to whom the patient was loyal in 2008. The remaining diagnostic imaging procedures were performed while patients were in the emergency room (N=34,345), completed while patients were in the hospital (29,763), or ordered as outpati ent by providers other than the patients loyal doctor (other primary care=9,833, specialists=86, 692). These other ca tegories of imaging utilization were also aggregated by patient and summed to produce the other patientlevel imaging utilization (independent ) variables described below. Variables Imaging Utilization (dependent variable) The main patient level outcome variable is called prv_o_cnt and is the count of the number of outpatient diagnostic imaging tests (CT, MR NM, PET, X-Ray, US) ordered by the provider to whom the patient wa s loyal during the study period (July 1, 2007 through June 30, 2009). Univariate statistics for this variable are summarized in Table 5-4. Patient Characteristics There were 35,709 men (41.9%) in the cohort whose ages in 2008 ranged from 17-100 with mean=54.5 years and standard deviati on of 15.1 years. There were 49,568 women (58.1%) whose ages in 2008 ranged from 17-103 with mean=53.2 years and standard deviation of 16.4 years. Patient race was available for all subjects and is shown in Table 5-5.

PAGE 67

67 Each patients payer of record (in 2008) was available in the RPDR as obtained from outpatient billing systems. These were initially categorized into 13 levels (Table 56). For modeling purposes, the 13 payer levels were collapsed into 6 levels as follows: Aetna into Commercial Harvard Pilgrim Healthcare, Neighborhood Health Plan, and Tufts Health Plan into Managed Mass Health Net and Medicaid into State Free Care and Self Pay into Uninsured The resulting 6 level insurance payer ca tegories (Table 5-7) will be used for all subsequent analyses. Clinical Events The RPDR was used to obtain counts of va rious clinical events for each patient. These were summed over the period from July 1, 2007 through June 30, 2009. The events counted all occurred at MG H. Hospital activity variables (summarized in Table 58) included visits to the emergency room inpatient hospital stays, ICU days, and inpatient observation stays. Observation (short) stays are a special category of hospitalization where the patient remains in the hospital for less than 24 hours. Observation stays often occur in concert with an emergency room visit and allow for extended nursing care without a formal admissi on. Observation stays also are used for minor procedures, dialysis, and administration of intravenous medications. Outpatient visit counts for each patient during the st udy period (July 1, 2007 through June 30, 2009) were obtained from t he RPDR and confined to the 15 CPT codes representing outpatient office visits. The professional RVU for each of these CPT codes during 2008

PAGE 68

68 was obtained from the CMS website (see Table 5-9). In addition to counting the number of outpatient visits, the RVU of those visits were also summed to form a separate variable. The visit counts and summed RVU for visits were stratified by type of doctor being visited (prv=provider to whom patient was loyal, pcp=other primary care doctor, spc=specialist). The resulting six vari ables are summarized in Table 5-10. Clinical Problems The EMR systems used by primary care providers to document outpatient care of all patients in the study cohort allow for reco rding of a problem list for each patient. These problems are encoded in one of tw o ways depending on the clinic. The coding systems are internal to MGH and crosswalk t ables are available to parse the problem codes into broad categories, including diabet es, hypertension, heart failure, coronary artery disease, renal failure, cancer, tr auma, obesity, and substance abuse. Active problems not falling into one of these groups were labeled as other problem for purposes of this study. For each of the major categories, a binary variable was constructed with value true/yes/1 when th e patient had at least one active problem listed in their EMR entries falling into that category. That same variable was assigned with false/no/0 when the patient did not have an active problem asserted falling into the category in question. The nine binary clinical problem variables are summarized in Table 5-11. In the cohort of 85,277 patients, 46 ,063 (54.0%) had none of the problem categories listed above, 23,265 (27.3%) had yes for a single category, 10,808 (12.7%) had two of the problem categor ies asserted positively, 3,846 (4.5%) had yes for three categories, 991 (0.30%) had four positive ca tegories, and 304 (0.36%) had yes for five or more.

PAGE 69

69 The EMR systems also included additional problem assertions for many patients that did not fall into one of the categories listed above. Examples include depression, hepatitis, arthritis, and so forth. For each patient, the counts of unmapped problem codes were placed into a variable of other problems (name=oth_prob). Overall summary statistics for the count of ot her problems included median of 6.0, mean of 7.75, and standard deviation of 6.94. This count of ot her problems was zero for 5709 patients and of those, 4941 (5.8% of the whole c ohort) had none of the problem categories enumerated above (e.g., patients clinic al problem list was empty/null). Outpatient Prescriptions The EMR systems also serve prescribing an d drug reconciliation functions for the primary care practice. In the cases in which the patients do not have e-prescribing enabled to their pharmacy and/ or drug benefits program, orders for outpatient medications are still entered into the EMR and printed prescriptions are given to patients. To enumerate the num ber of outpatient medications each patient was taking during 2008, the number of a ctive prescriptions was c ounted starting from the first available entry for any given patient. The quer ies did not count refills of the same drug and dose as new prescriptions. However, switches within a dr ug class and/or dose changes were counted as new prescriptions which could result in over-counting and rendering small differences in the discret e number less meaningful than the general amount each patient was taking. Therefore, we stratified the count of active outpatient medications into four cat egories (variable name=meds_cat) summarized in Table 5-12. Other Imaging Utilization Counts of diagnostic imaging tests perform ed on each patient during the study interval (July 1, 2007 through June 30, 2009) we re stratified by place of service (i.e.,

PAGE 70

70 patient status/location at the time the test was perfo rmed). These include emergency department, inpatient, and outpatient (Table 5-1). Outpati ent exams were further stratified by the category of physician orderin g them; primary care physician (other than the patients own loyal doctor), and speciali st (Table 5-2). Note that the counts of outpatient diagnostic imaging test s ordered by the patients linked (loyal) provider is the outcome variable which has already been described above. Summary statistics for the four strata of (non-outcome) imaging utilization are liste d in Table 5-13. Physician Characteristics The cohort of 85,277 patients were loyal to 148 primary care physicians; 76 (51.3%) women and 72 (48.7% ) men. A variable called prov_age_08 was constructed using each doctors birth year. The ages of the male physicians (in 2008) had minimum=33, maximum=75, mean=49.7, and standard deviation=10.0 years respectively. The female physicians ages (in 2008) had minimum=31, maximum=63, mean=45.4, and standard deviation=8.4 years respectively. T he year of medical school graduation was used to construct a variabl e called prov_exp_08 that quantifies the number of years between medi cal school graduation and 2008. This (prov_exp_08) had minimum=5, maximum=50, mean=19.6, and st andard deviation=9.6 years respectively. Perhaps a better proxy for physician experience would have been years since completion of residency. However, the credentials database had incomplete and inconsistent information in this regard. The number of (loyal and nonloyal) patients in the cohort linked to each doctor was summed into a variable called pr ov_pat_count and it had minimum=172, maximum=2394, mean=801.3, standard deviati on=396.0, and median=751. This variable (prov_pat_count) was categor ized into four levels (variable

PAGE 71

71 name=prov_pat_cat) and will be used as a proxy for how busy the doctor was during the study period (see Table 5-14). The medical school graduation state/count ry was used to construct a variable called prov_fmg that was set to yes (N=8 5.4%) when the doctor had graduated from a medical school outside the US (Argenti na=1, Canada=2, Croatia=1, England=1, Holland=1, Italy=1, Panama=1). The graduate le vel degrees held by each provider were used to construct a variable called prov_md_plus that was set to yes (N=14) when the doctor had obtained a graduate degree in additi on to their M.D. (MPH=6, MSC=1, MSW=1, PHD=6). The assertions about malpract ice cases in the last 10 years found on the Massachusetts Board of Medicine web site were used to construct a variable called mp_flag with yes (N=7) when the doctor had a record of having been sued and no otherwise (N=147). Site Characteristics Each of the 15 sites was labeled with a unique number that will serve as identification for subsequent analysis about inte r-site variability. The only variable that accrues to the clinics themselves is the number of primary care doctors in active practice at each one. This variable along wi th the number of patients and primary care doctors assigned to each site serve as proxie s for the size of the clinics. These are summarized in Table 5-15. Note that the sum of the number of active doctors at all sites is 168 whereas there were only 148 included in this study (St udy Doctors Column). The remaining 20 had less than 100 loyal patients and were not included in the analytic data set.

PAGE 72

72 Patient, Provider, and Clinic Identifiers For multi-level (hierarchical) modeling it is necessary to uniquely identify the individual primary care doctors being studied so as to be able to maintain the linking between them and their loyal pati ents. Therefore, t he MGH provider identifying numbers were sorted and mapped to the corresponding r ank (1, 2, 3, 148). Using this mapping, each of the 85,277 pat ients linked MGH provider identifier was replaced by a unique (though now anonymous) integer. Since each observation in the final analytic data set represents a single pat ient, no identifying information need be retained (e.g., Name, Medical Record Number, and etc) and these were all dropped. As described above, the clinics were already identifi ed by (uninformative ) integers (1-15). Data Integrity: Clinical Activity Variables One way to insure that the queries of outpat ient visits, inpatient stays, emergency room visits, and outpatient imagi ng tests for all patients in the study were complete and consistent is to plot them over time. This was done by counting each event type by month for the whole study cohort and plotting as a time series. The visit counts by month stratified by type of provider render ing the visit are plotted in Figure 5-2. It is reassuring that the counts are relative ly stable and consistent during the study period. This implies that ther e are no large gaps or duplications in the data. As for any secular patterns, this dissertation will not attempt to descri be or explain them. Similarly, the counts of outpatient diagnostic imaging examinations were enumerated by month and stratified by the type of provider order ing the study are plotted in Figure 5-3. As with the counts of outpatient visits, there is appar ent consistency and stability. Further reassurance comes fr om the fact that relative decreases in counts during

PAGE 73

73 December 2007 and August 2008 seem to matc h between outpatient visits and imaging tests. Since the data came from separat e and independent adminis trative sources, there was an actual decrease in outpatient clinical activity during these periods. Visual inspection of the two plots confirms that visits and imaging tests tend to rise and fall together by month. The other main activity variables relate to hospital events and counts of these by month are plotted in Figure 5-4. Again the month to month c onsistency and stability attests to the integrity of the data, which came from two separate databas es (one for emergency room and a second for inpatient/observation stays). Variable Summary A summary of all patient-level independent variables is provided for reference in Table 5-16. A summary of all the clinic and provider level independent variables is provided for reference in Table 5-17. Table 5-1. All diagnostic imaging performed on study cohort during two years of study. Place Of Service CT MR NM PET X-Ray US Total Percent ER 10183 2651 489818608240634345 15.50 Inpatient 4768 1634 944 12819907238229763 13.43 Outpatient 26981 21999 6129 37097665421991157463 71.07 Total 41932 26284 7562384511516926779221571 100 Percent 18.92 11.86 3.41 1.7451.9812.09100 Note: CT=computed tomography, MR=magnet ic resonance imaging, NM=nuclear medicine, PET=positron emission tomogr aphy, X-Ray=radiography, US=ultrasound. The studies performed in the outpatient setti ng are further stratified in Table 5-2.

PAGE 74

74 Table 5-2. Outpatient diagn ostic imaging performed on study cohort during two years of study. Who Ordered CT MR NM PET X-RayUS Total Percent Specialist 15295 13170 3684329442143910686692 55.06 Other primary care doctor 1133 820 15717660511019833 6.24 Patients own loyal doctor 10553 8009 2288398279061178460938 38.70 Total 26981 21999 612937097665421991157463 100 Percent 17.13 13.97 3.89 2.3648.6813.97100 Note: CT=computed tomography, MR=magnet ic resonance imaging, NM=nuclear medicine, PET=positron emission tomogr aphy, X-Ray=radiography, US=ultrasound. The studies ordered by the patients linked (loyal) doctor are fu rther stratified in Table 5-3. Table 5-3. Outpatient diagn ostic imaging ordered by patients linked (loyal) doctor during two years of study. Body Area CT MR NM PET X-Ray US Total Percent Abdomen 4005659 258976240329572 15.71 Cardiac 4210 20030002055 3.37 Chest 4637303 1396110841216145 26.49 Extremity 1181807 001077683413535 22.21 Head/Brain 9762060 042441523436 5.64 Maxillofacial and/or Neck 325249 549557820553356 5.51 Pelvis 256318 00109743105981 9.81 Spine 1932602 00335506150 10.09 Unspecified 11 19311410389708 1.16 Total 105538009 2288398279061178460938 100 Percent 17.3213.14 3.750.6545.7919.34100 Note: CT=computed tomography, MR=magnet ic resonance imaging, NM=nuclear medicine, PET=positron emission tomogr aphy, X-Ray=radiography, US=ultrasound.

PAGE 75

75 Table 5-4. Univariate statistics of the outcome variable (per patient count of outpatient imaging tests ordered by pr imary care provider). N 85277 Minimum 0 (N=53,617) Maximum 15 Mean 0.7146 Standard Deviation 1.2563 Skewness 2.6554 Uncorrected SS 178132 Coeffient of Variation 175.8049 Sum of Observations 60938 Variance 1.5782 Kurtosis 9.8796 Corrected SS 134586 Standard Error of the Mean 0.0043 Note: SS=sum of squares. Table 5-5. Distribution of patient race. Race Frequency Percent White 68432 80.25 Black 4278 5.02 Hispanic 5924 6.95 Other 6643 7.79 Total 85277 100 Table 5-6. Distribution of patients payer categories. Payer Group Frequency Percent Aetna 17172.01 Blue Cross Blue Shield 3076236.07 Commercial 62827.37 Free care 1960.23 Harvard Pilgrim Healthcare 857410.05 Mass Health Net 18902.22 Medicaid 36464.28 Medicare 1955522.93 Neighborhood Health Plan 15731.84 Other 30573.58 Self Pay 10571.24 Tufts Health Plan 69688.17 Total 85277100.00

PAGE 76

76 Table 5-7. Patients payer collapsed into 6 categories. Payer Group Frequency Percent Blue Cross Blue Shield 3076236.07 Commercial 79999.38 Managed 1711520.07 Medicare 1955522.93 Other 30573.58 State 55366.49 Uninsured 12531.47 Total 85277100.00 Table 5-8. Hospital activity variables (per patient). Description Variable Name Minimum Maximum Mean SD Total hours in Emergency room er_hours 0 N=69,4854832.71 10.15 Emergency room visits er_visits 0 N=69,484470.33 1.04 Inpatient admissions inpt_stays 0 N=74,582220.20 0.67 total days in hospital Inpt_los_total 0 N=74,6971790.92 4.70 Days in intensive care units inpt_icu_days 0 N=84,401490.05 0.76 Readmitted within two weeks of inpatient discharge inpt_read_15d 0 N=84,382100.01 0.15 Readmitted within one month of inpatient discharge inpt_read_31d 0 N=84,062170.02 0.21 Observation (short) stays obs_stays 0 N=76,145370.13 0.54 Note: The number of patients with zero c ounts is given under Minimum (where zero).

PAGE 77

77 Table 5-9. CPT codes and relative val ue units for ambulatory office visits. CPT Code Visit Type Complexity RVU 99201 new patient not comprehensive not complex 0.48 99202 new patient not comprehensive mod complex 0.93 99203 new patient not comprehensive high complex 1.42 99204 new patient comprehensive moderate complexity 2.43 99205 new patient comprehensive high complexity 3.17 99211 established not comprehensive not complex 0.18 99212 established not comprehensive mod complex 0.48 99213 established not comprehensive high complex 0.97 99214 established comprehensive moderate complexity 1.50 99215 established comprehensive high complexity 2.11 99241 consultation not comprehensive not complex 0.64 99242 consultation not comprehensive mod complex 1.34 99243 consultation not comprehensive high complex 1.88 99244 consultation comprehensive moderate complexity 3.02 99245 consultation comprehensive high complexity 3.77 Table 5-10. Outpatient visit acti vity variables (per patient). Description Variable Name Minimum Maximum Mean SD Count of outpatient visits to linked (loyal) provider prv_visit_count 0 N=10,39662 3.543.39 Summed RVU of visits to linked (loyal) provider prv_visit_rvu 0 N=10,39676.22 4.954.83 Count of outpatient visits to other primary care doctors pcp_visit_count 0 N=57,11775 0.601.24 Summed RVU of visits to other primary care doctors pcp_visit_rvu 0 N=57,11759.44 0.731.57 Count of outpatient visits to specialists spec_visit_count 0 N=25,771112 3.605.17 Summed RVU of visits to loyal specialists spec_visit_rvu 0 N=25,771165.18 4.967.28 Note: The number of patients with zero count s or RVU is given under Minimum (where zero).

PAGE 78

78 Table 5-11. Binary clinical probl em variables (per patient). Problem category Variable Name Number Yes Percent Yes Diabetes pr_dm 948511.12 Hypertension pr_htn 2521929.57 Heart failure pr_chf 8671.02 Coronary artery disease pr_cad 39404.62 Renal failure pr_crf 10961.29 Cancer pr_can 992511.64 Trauma pr_trm 18042.12 Obesity pr_obs 885510.38 Substance abuse pr_sub 7700.90 Table 5-12. Four level categorization of patient active outpatient medications (per patient). meds_cat Frequency Percent None 4475 5.25 1-5 45536 53.40 6-10 23390 27.43 >10 11876 13.93 Total 85277 100.00 Table 5-13. Summary of other (non-outcome) imaging test utilization variables (per patient). Description Variable Name Minimum Maximum Mean SD Count of imaging tests ordered during emergency room visits all_e_cnt 0 N=73,562 96 0.401.56 Count of imaging tests ordered during inpatient stays all_i_cnt 0 N=79,118 118 0.352.41 Count of outpatient imaging tests ordered by specialists spec_o_cnt 0 N=55,105 61 1.022.28 Count of outpatient imaging tests ordered by other primary care doctors* pcp_o_cnt 0 N=78,289 8 0.120.45 Note: The number of patients with zero c ounts is given under Minimum (where zero). *Not the patients own linked (loyal) doctor. Table 5-14. Four level categorization of t he number of patients cared for by each provider (panel size). prov_pat_cat Frequency Percent <500 37 25.00 500-759 36 24.32 750-999 36 24.32 1K+ 39 26.35

PAGE 79

79 Table 5-15. Site (clinic) characteristics. Site (clinic) ID Active Doctors % Active Doctors Study Patients % Study Patients Study Doctors % Study Doctors 1 74.1734514.0553.38 2 84.7642334.9685.41 3 2011.901192513.981610.81 4 116.5550585.9396.08 5 158.9372208.47138.78 6 52.9832693.8342.70 7 74.1733083.8864.05 8 1710.121038112.171711.49 9 127.1463787.48128.11 10 148.3357286.72149.46 11 95.3650975.9874.73 12 1810.71910910.681610.81 13 52.9811351.3353.38 14 63.5747945.6264.05 15 148.3341914.91106.76 Total 16810085277100148100

PAGE 80

80 Table 5-16. Description and categorization of 33 patient level independent variables. Variable Name Level Class Type Levels Description inpt_stays Patient Activity, Hospital Numeric count of inpatient stays er_hours Patient Activity, Hospital Numeric total hours spend in er er_visits Patient Activity, Hospital Numeric count of er visits obs_stays Patient Activity, Hospital Numeric count of observation stays inpt_read_31d Patient Activity, Hospital Numeric count of readmit within 31 days inpt_read_15d Patient Activity, Hospital Numeric count of readmit within 15 days Inpt_los_total Patient Activity, Hospit al Numeric total days in hospital inpt_icu_days Patient Activity, Hospital Numeric total days in icu all_e_cnt Patient Activity, Other Imaging Numeric count of images done in ER all_i_cnt Patient Activity, Other Imaging Numeric count of images done as inpatient spec_o_cnt Patient Activity, Other Imaging Numeric count of outpatient images ordered by specialists pcp_o_cnt Patient Activity, Other Imaging Numeric count of outpatient images ordered by other primary care doctors prv_visit_rvu Patient Acti vity, Visits Numeric sum of rvu of outpatient visits to linked (loyal) primary care physician prv_visit_count Patient Ac tivity, Visits Numeric count of outpatient visits to linked (loyal) physician spec_visit_rvu Patient Activity, Visits Numeric sum of rvu of outpatient visits to specialists spec_visit_count Patient Activity, Visits Numeri c count of outpatient visits to specialists pcp_visit_rvu Patient Activity, Visits Numeric sum of rvu of outpatient visits to covering pcp pcp_visit_count Patient Activity, Visits Numeric count of outpatient visits to covering pcp age_08 Patient Demographics Numeric Patients age in 2008 Race Patient Demographics Categorical 4 patient identified race Sex Patient Demographics Categorical 2 patient sex payer_group Patient Insurance Categorical 6 category of patient's payer of record in 2008 meds_cat Patient Medications Categorical 4 active outpatient prescriptions in 2008 pr_cad Patient Clinical Problem Categorical 2 coronary artery disease pr_can Patient Clinical Problem Categorical 2 cancer pr_chf Patient Clinical Problem Categorical 2 congestive heart failure pr_crf Patient Clinical Problem Categorical 2 chronic renal failure pr_dm Patient Clinical Problem Categorical 2 diabetes pr_obs Patient Clinical Problem Categorical 2 obesity pr_htn Patient Clinical Problem Categorical 2 hyptertension pr_sub Patient Clinical Problem Categorical 2 substance abuse pr_trm Patient Clinical Problem Categorical 2 trauma oth_prob Patient Clinical Problems Numeric count of active problems other than those separately listed above

PAGE 81

81 Table 5-17. Description and categorization of provider (8) and clinic level (2) independent variables. Variable Name Level Class Type Levels Description site_docs Clinic Characteristic Numeric number of doctors actively practicing at the clinic in 2008 site_id Clinic Identifier Categorical 15 anonymous clinic (site) identifier mp_flag Provider Characteristic Categorical 2 whether provider has been sued in last 10 years prov_md_plus Provider Characteristic Categorical 2 whether provider has a degree beyond MD prov_pat_cat Provider Characteristic Categorical 4 number of patient's in provider practice in 2008 prov_sex Provider Characteristic Categorical 2 provider sex prov_fmg Provider Characteristic Categorical 2 whether provider is foreign medical graduate prov_age_08 Provider Characteristic Numeri c age in years of the provider in 2008 prov_exp_08 Provider Characteristic Numeric number of years after provider MD graduation in 2008 prov_id Provider Identifier Categorica l 148 anonymous provider identifier Figure 5-1. Outpatient imaging tests (per patient) ordered by the linked (loyal) primary care provider. Both the percent (l eft Y axis, bars) and number (right Y logarithmic axis, diamonds and dashes) of observations are shown. 0 10 20 30 40 50 60 012345678910Number Of Imaging TestsPercent Obsrvations1 10 100 1000 10000 100000Number Observation s

PAGE 82

82 Figure 5-2. Number of outpatient visits by all patients in study cohort (by month) over two years of study. Hatched: visits to patients linked (loyal) doctor, Black: visits to another (covering) PCP, White: visits to specialists. Figure 5-3. Number of outpatient imaging test s performed on all patients in study cohort (by month) over two years of study. Outpatient imaging tests ordered by: Hatched: patients linked (loyal) doctor, Black: another (covering) PCP, White: specialists. 0 5000 10000 15000 20000 25000 30000 35000 40000 450002 0 07_ 0 7 20 0 7_0 8 2 0 07_ 0 9 20 0 7_1 0 2 0 07_ 1 1 20 0 7_1 2 2 0 08_ 0 1 20 0 8_0 2 20 08 _0 3 2008_04 20 08 _0 5 2008_06 20 08 _0 7 2008_08 20 08 _0 9 2008_10 20 0 8_1 1 2008_12 20 0 9_0 1 2009_02 20 0 9_0 3 2009_04 20 0 9_0 5 2009_06 spec pcp prv 0 2000 4000 6000 8000 100002 007_ 0 7 200 7_ 08 2 007_09 200 7_ 10 2007 11 20 07_ 12 200 8_ 01 2 008_ 0 2 200 8_ 03 2008_04 200 8_ 05 2008 06 20 08_ 07 200 8_ 08 2 008_ 0 9 200 8_ 10 2008_11 200 8_ 12 2009 01 20 09_ 02 200 9_ 03 2 009_ 0 4 200 9_ 05 2009_06 spec pcp prv

PAGE 83

83 Figure 5-4. Number of hospita l encounters for all patients in study cohort over two years. Hatched: emergency room visits, Black: short (observation) stays, White: inpatient discharges. 0 500 1000 1500 2000 2500 3000 3500 40002 007_ 07 2007_0 8 2 0 07_09 2 007 _10 2007_1 1 2 0 07_12 2 008 _01 2 008_0 2 2 0 08_03 2 00 8_04 2 008_ 05 2 0 08_06 2 00 8_07 2 008_ 08 2008_09 2 00 8_10 2 008_ 11 2008_12 2 00 9_01 2 009_ 02 2009_0 3 2 0 09_04 2 009 _05 2009_0 6 IN OBS ER

PAGE 84

84 CHAPTER 6 METHODS Outcome Variable Distribution Simple univariate statistics of the main outcome variabl e (prv_o_cnt = IMG for this chapter) have been described above in Chap ter 5 (see Table 5-4 and Figure 5-1). Based on the visual inspection, these appear to be count data with a Poisson distribution. This is a discrete distribut ion, with a single parameter, Lambda ( ), that expresses the probability of a number of event s occurring in a fixed period of time if these events occur with a known average rate ( ) and independently of the time since the last event. Each instance of a Poiss on random variable can be expressed as being the result of a single Poi sson experiment and may take any positive integer value (Stahl 1969). An important pr operty of the Poisson distribut ion is that the variance is equal to the mean ( ). The Poisson equation may be written as Equation 6-1. Prob(n) = ( n e) / n! (6-1) In Equation 6-1, e is the base of the natural logarithm, n is the number of occurrences of an event the probability of which is given by the function, and is the positive real number, equal to the expected number of occurrences that occur during the given interval. A more theoretically appealing and possibly informative way to describe the distribution of imaging count s is that of a two stage process with the first stage determining the occurrence of any imaging fr om the whole data set of 85,277 patients and the second stage determining the count of imaging tests for patients that had at least one (IMG > 0, N=31,660) imaging test during the study period. This can be modeled with a logistic regression on all obs ervations followed by Poisson regression

PAGE 85

85 on the non-zero observations. For purposes of estimating patient, provider, and clinic effects; the two step (logistic>Poisson) approach was taken. A final method to model the distribution of outpatient imaging counts combines the logistic portion (any use) with the Poisson assumption for intensity (non zero use) into a single distribution; zero inflated Poisson (ZIP ). The ZIP distribution is almost the same as the standard Poisson when n>0 but has a second portion for n=0 and may be written as Equations 6-2 and 6-3. Prob(n) = + (1) efor n = 0 (6-2) Prob(n) = (1) (n e) / n! for n = 1, 2, (6-3) In Equations 6-2 and 6-3, e is the base of the natural loga rithm, n is the number of occurrences of an event the probability of which is given by the function, is the positive real number, equal to the expected number of occurrences that occur during the given interval, is the real num ber between 0 and 1 ( = 0 is standard Poisson) called the zero in flation parameter. To examine the distribution of this variable, SAS proc GENMOD was used to estimate two null models on all 85,277 observations and a single one for the 31,660 observations with non-zero outcomes (pati ents who had at least one imaging test ordered by their linked/loyal primary care doctor). Specifically the models used were simple Poisson and Zero-Inflat ed Poisson (ZIP) distributions for the whole data set and simple Poisson distribution on the non-zero observations. The intercept coefficients from these models and the 95% confidence intervals were used as estimators of the Poisson Lambda parameters of each proposed distribution. The dispersion for each

PAGE 86

86 modeled distribution was obtained by usi ng the PSCALE and DSCALE option in the model statement. To visually demonstrate the relationshi p between the observed distribution of imaging test counts and the three proposed Poisson distributions, the parameter estimates from the null models described above were submitted to a SAS provided macro called PROBCOUNTS. This software is available for download on the public SAS support web site at the following location (checked May 2, 2010): http://support.sas.com/kb/26/161.html The purpose of this program is succinctly described on the support web site as follows: The PROBCOUNTS macro computes the predicted count and the predicted probabilities of specifi ed counts for Poisson and negative binomial models and for zero-inflated versions of thes e models as fit by PROC COUNTREG in SAS/ETS software and PROC GE NMOD in SAS/STAT software. A plot of the observed counts (from 0 to 15) superimposed on line graphs of the expected counts from each of t he three distributions serves to compare them and will be reproduced in the Chapter 7 as Figure 7-1. The purpose of this was to visually inspect the fit of the proposed distributions wit h the actual imaging counts especially with respect to the upper tail and the t endency of overor under-dispersion. To verify the estimate of Lamda( ) for all three distributions and obtain a direct estimate of Phi( ) for the ZIP distribution requir ed a complementary approach. SAS PROC NLMIXED allows exact specification of the proposed distribution(s) and produces direct estimates for the parameters. Since an iterative approach is used, initial seed values for the parameters of interest (Lam and Phi in the code fragm ents) are supplied. The SAS code for the two simple Poisson models is reproduced below:

PAGE 87

87 proc nlmixed data=input.data_set; parms Lam=1; loglike=IMG*log(Lam)-Lam-lgamma(IMG+1); model IMG~general(loglike); where IMG ne 0; <
PAGE 88

88 Bivariate Relationships Ordinary least squares regression with SAS PROC GLM was used to estimate the relationship between each independent variable and the outcome (count of imaging tests ordered by patients loyal provider) for all observations (N=85,277). Numeric independent variables (N=28) were entered as recorded while categorical variables (N=20) were parameterized using SAS EFFECT method which yields L-1 design variables where L is number of levels in the original variable. This replicates the categorical variable coding meth od (reference cell) that wil l be used in multivariable modeling and yields degrees of freedom that are identical. Basi cally, this translates into an OLS linear regression with numeric variables and a one-way ANOVA with the categorical variables. For each of the 48 variables, the following parameters were obtained to measure the strengt h of bivariate association with the outcome: F Value, RSquared, Correlation (R), and p value from the single variable regression output. Variable Reduction for Modeling Based on the correlation analysis betw een independent variables and evaluating the bivariate relationships with the outcome redundant and/or collinear predictors were omitted from subsequent multiv ariable and multi-level modelin g. The selection heuristic was to choose the one having strongest bivariate relationship with the outcome when independent variables were strongly correlat ed. However, for theoretical reasons, some independent variables were kept despite not having significant bivariate relationship with the outcome (e.g. prov ider malpractice status). Multivariable (logistic) Modeling: Any Imaging Use Multivariate logistic regression was used to analyze a binary outcome derived from the main outcome variable (IMG). This binar y outcome variable (called ANY_IMG) is set

PAGE 89

89 to yes (N=31,660) when that patient had one or more outpatient diagnostic imaging test(s) ordered by the linked (loyal) provi der and no (N=53,617) otherwise. This was done with SAS PROC LOGISTIC with the model ed outcome set to yes. Numeric variables were entered as recorded. The ca tegorical variables were parameterized using the SAS REFERENCE encoding method. This allows explicitly setting the reference level and results in L-1 dummy variabl es where L is the number of levels in the original variable. Provider level variable reference levels were as follows: Sex-Male, FMG-No, MD_Plus-No, Malpractice-No, Provi der Patients-<500. Patient level reference Levels were set as follows: Sex-Male, Race -White, Insurance-Uninsured, MedicationsNone. The remaining patient le vel categorical variables were the binary clinical problem assertions and reference for each of these was set as no. To simplify writing the model, the patient level variables may be represented by a vector P, the provider (docto r) variables by a vector D and the clinic variable(s) by C. The logistic model is ex pressed as Equation 6-4: Logit( [ANY_IMG=yes])i = 0 + pPip + dDid + cCic (6-4) In Equation 6-4, i is the ith patient, p is the pth patient level variable, d is the dth doctor variable, and c is the cth clinic variable. Type III estimates, standard errors, Chi-Squared, p-value, and odds ratios were obtained from the resulting solution output from SAS. These were used to make in ferences and comparisons about joint significance and effect size of all included predictor variables. A Hosmer and Lemeshow (HL) test for goodness of fi t was requested as well.

PAGE 90

90 Multivariable (Poisson) Modeling: Imaging Intensity (non-zero) This portion of the analysis seeks to det ermine the effect of the same predictor variables used in the logistic analysis on the intensity of imaging (given that some has occurred). Only the 31,660 observations (pat ients) where IMG is between 1 and 15 are used in this model. From preliminary analysis of the outcome (IMG) distribution, it was determined to be reasonably represented as a Poisson count variable. To estimate these types of models, SAS provides PROC GENMOD which allows specification of linear models with Poisson error distribut ion and log link function. These may be estimated using maximum likelihood methods. As above, to simplify writing the model, the patient level variables may be represent ed by a vector P, the provider (doctor) variables by a vector D and the clinic variable(s) by C. The Poisson model may be written as Equation 6-5. Log[E(IMGi | Pip, Did, Cic)] = 0 + pPip + dDid + cCic + ei (6-5) In Equation 6-5, i is the ith patient, p is the pth patient level variable, d is the dth doctor variable, and c is the cth clinic variable. The errors (ei) are distributed as Poisson. After estimation, the solution output contains the coefficient estimate, standard error, Chi-Squared, and p-value for each one of the num eric variables and categorical variable levels. The model was re-estimated using DSCALE and PSCALE options to determine if the outcome distribution was overor under -dispersed. As will be described in the Results chapter, the distribution turns out to be underdispers ed, and standard error would tend to be overestimat ed. Therefore, a correction was NOT made for dispersion which results in somewhat conservative inferences about significance of various effect sizes.

PAGE 91

91 Preparation for Multi-Level Mode ling: Imaging Propensity Scores For multi-level modeling, we wish to ans wer questions about variation in imaging utilization between providers (level 2) holding patient factors equal while accounting for clustering of patients (level 1) within provider s. For this dissertation, the possible higher level effect of clinics as aggregations of pr oviders will not be addressed. To simplify the specification, estimation, and interpretation of multi-level modeling results, all patient level factors were collapsed into a single risk adjusted expected imaging (propensity) variable. Previously described evaluation of the overall distribution of the outcome (IMG) determined that a zero-inflated Poisson (ZIP) distribution to be most suitable for a single model. This was done with SAS PROC GE NMOD, all 85,277 observations, and the same patient level independent variables used in multivariable modeling described above. The SAS PROC GENMOD instructions were constructed so as to produce patient level predictions after the initial ma ximum likelihood estimation. This variable will be called IMG_PROP for imaging propensity and represents the num ber of outpatient imaging tests that the average patient with identical values of all independent variables would be expected to have. Another way of describing this te chnique is as regression based risk adjustment for imaging utilization. In typical provider profiling applications, these patient level predictors would be termed the expected imaging utilization and summed across doctors to be compared with the observed count of imaging tests actually performed on those same patients. Multi-Level (Hierarchical) Modeling This part of the analysis will test the relationship between a summary of each patients imaging propens ity (IMG_PROP) and imaging performed (IMG) within each of the 148 primary care doctors practice. This posits that each doctor has his or her own

PAGE 92

92 regression equation with an intercept and slope. The interpretation of the intercept for each doctor is the general tendency to obtai n imaging on the average patient while the slope corresponds to that doctors response to patients with increasing risk adjusted expectation for imaging (IMG_PROP = imagi ng propensity described above). With these interpretations in mind, the analysis will answer the following: What is the average intercept and slope of the 148 provider regression equations? How much do the intercepts vary from doctor to doctor? How much do the slopes vary from doctor to doctor? What is the correlation between intercepts and slopes? Generally, in two-level hierarchical m odeling, where the level 2 intercepts are allowed to vary randomly (in this study for each provider), it is important to consider whether or not to center (offset) the level 1 variable (IMG_PROP) in any way. There are three options; natural metric (no centering), grand mean (e.g ., all patients), group mean (e.g., patients loyal to each doctor). Consequences of these choices affect both the value and interpretation of the inte rcept estimates (Raudenbush and Bryk 2002). Further, the standard errors for estimates of both fixed and rando m effect intercepts differ depending on choice of centering method ( Luke 2004). In health services research applications, centering on the (level 2) group mean (provider in this case) is recommended (Houchens et al. 2007). For this study, t he group means centering approach will be used. One distinct advantage is that it allows interpreting the individual provider intercepts as repr esenting the average tende ncy to order imaging while holding the expected imaging propensity (IMG_P ROP) of their patients equal. On the other hand, provider intercepts estimated with the non-centered appr oach are clinically

PAGE 93

93 less meaningful because they would have to be interpreted as the tendency to order imaging for patients with no clinical need (which numerically could be negative). The model used will have random coeffi cients for intercept and slope and at the level of the patient (level 1) is written as Equation 6-6: IMGij = 0j + 1j (IMG_PROPij IMG_PROP.j) + eij (6-6) In Equation 6-6, IMGij is the Count of outpatient imag es for the ith patient cared for by the jth doctor, 0j is the The intercept for the jth doctor, 1j is the The slope for the jth doctor, IMG_PROPij is the Imaging propensity for the ith patient cared for by the jth doctor, IMG_PROP.j is the Average imaging propensity of the patients cared for by the jth doctor, and eij is the Error (disturbance) for the it h patient cared for by the jth doctor. For simplicity let cIMG_PROPij = IMG_PROPij IMG_PROP.j, where cIMG_PROPij is the centered imaging propensity for the ith patient car ed for by the jth doctor. Therefore, our patient (l evel 1) model may be written as Equation 6-7. IMGij = 0j + 1j (cIMG_PROPij) + eij (6-7) The imaging utilization of each doctors practice is characterized by two parameters: 0j, the intercept for the jth doctor and 1j the slope for the jth doctor. Since the imaging propensity for each patient is centered on the mean fo r their doctor, the intercept is actually that doctors mean imaging. These two parameters vary across doctors in the level-2 model as a func tion of the grand mean and random disturbance such that 0j = 00 + u0j and 1j = 10 + u1j, where 0j is the intercept (mean imaging) for the jth doctor, 00 is the average of the doctor means of imaging use across the population of doctors, u0j is the individual variation from the average interc ept for the jth doctor, 1j is the slope for the jth doctor, 10 is the average imagi ng propensity / imaging

PAGE 94

94 utilization regression slope across the doctors, and u1j is the individual variation from the average slope for the jth doctor Combining the two by substituting the le vel 2 random coefficient models into the level 1 patient level m odel gives Equation 6-8. IMGij = 00 + u0j + 10 (cIMG_PROPij) + u1j (cIMG_PROPij) + ij (6-8) The variance structure can be expre ssed as Equations 6-9 through 6-12. u0j~ N(0, 00) (6-9) u1j~ N(0, 11) (6-10) Cov(u0j,u1j)= 01 (6-11) eij~ N(0, 2) (6-12) In Equations 6-9 through 6-12, 00 is the Var(u0j), 11 is the Var(u1j), and 01 is the the covariance between u0j and u1j. Since the level 2 model has no predictors in either the intercept or slope equation, it is unconditional. Therefor e, we can use multi-level regr ession estimates for variability in intercepts and slopes as show n in Equations 6-13 and 6-14. Var(u0j) = Var( 0j 00) = Var( 0j) (6-13) Var(u1j) = Var( 1j 10) = Var( 1j) (6-14) For estimation with SAS PROC MIXED we divide the combined model into fixed parts (Equation 6-15) and rando m parts (Equation 6-16). 00 + 10 cIMP_PROPij (6-15) u0j + u1j cIMG_PROPij + ij (6-16) Simplified SAS Code is written: proc mixed data=input.data_set;

PAGE 95

95 class prov_id; model IMG = cIMG_PROP; random intercept cIMP_PROP / subject=prov_id; run; To quantify the fraction of variation in imaging utilization attributable to patients and providers it is necessary to obtain patient level residuals (eij ~ 2) from reduced models and the null model. These can be repr esented as follows: Patient level Imaging Propensity only (Equation 6-17), Patient le vel Imaging Propensity and provider slope only (Equation 6-18), Patient level Imaging Propensity and provider intercept only (Equation 6-19), and Null Model (Equation 6-20): IMGij = 00 + 10 (IMG_PROPij) + ij (6-17) IMGij = 00 + 10 (cIMG_PROPij) + u1j (cIMG_PROPij) + ij (6-18) IMGij = 00 + u0j + 10 (cIMG_PROPij) + ij (6-19) IMGij = 00 + eij (6-20)

PAGE 96

96 CHAPTER 7 RESULTS Outcome Variable Distribution The outcome variable (prv_o_cnt=IMG ) was measured for each patient (N=85,277) over two years of study and repr esents the count of outpatient non-invasive diagnostic imaging tests ordered by the patients linked (loy al) primary care provider. Univariate statistics and the distribution were shown (Table 5-4 and Figure 5-1 respectively) in Chapter 5. To estimate the Poisson Lambda for the whole sample (N=85,277 including the 53,617 zero observations), SAS PROC GENMOD was used with a null model (IMG = / di st=poi). Lambda was thus esti mated at 0.71459 with 95% confidence interval of (0. 70718-0.72200). The dispersion (extent to which the observed variance exceeds the expected) was estima ted by Deviance / Degrees of Freedom (145443.64 / 85,276) at 1.7056 whic h means that there are more observations at higher values of n (3-15) than expected by a Po isson distribution with Lambda=0.71459. Such overdispersion can result in underestimat ion standard errors in subsequent modeling. Also using SAS PROC GENMOD and a null model (IMG = / dist=poi), a second estimation, using only the non-zero observations, of Poisson Lambda and dispersion was made in the same manner as described above. In this case, the Poisson Lambda was estimated at 1.92476 with 95% confiden ce interval of (1.90948-1.94005) and the dispersion was 0.7796. Beginning with Version 9.2, SAS PROG GENMOD allowed estimation of ZIP models. In addition to specif ying a model for the outcome counts, a second, so called zeromodel, was specified (IMG = / dist=zip; zeromodel =;). Using the entire sample (N=85,277) with null count and zeromodel st atements, the paramet ers of this ZIP

PAGE 97

97 distribution were estimated. The Po isson Lambda estimate was 1.4917 with 95% confidence interval of (1.4760-1.5074). The dispersion was 1.2428 and the zero inflation intercept was 0.084 with 95% confidence inte rval of (0.0650-0.1028). To obtain a more meaningful value for the ze ro inflation parameter ( ), the output of SAS PROC NLMIXED was used to obtain a direct estima te of 0.521 (CI: 0.516-0.526). At the same time, all three estimates of the Poisson Lam bda from NLMIXED were identical to those obtained from GENMOD which is reassuring. The output from PROBCOUNTS generated using parameter estimates obtained from the null models re present the expected distributions of the simple Poisson, nonzero Poisson and zero-inflated Poisson assumptions. They are plotted in Figure 7-1. Clearly, the best fit appears to be with the Ze ro-inflated Poisson distribution which militates for using it to generate risk adjusted expected imaging (propensity scores) for each patient prior to mu lti-level modeling seeking to char acterize provider variation. However, to estimate the effe ct of individual patient, prov ider, and clinic level factors on imaging utilization the two stage (logistic followed by Poisson) approach were used. One reason is that odds ratios produced by the logistic analysis of any utilization have well understood meanings that are directly inte rpretable. Additionally, since the Poisson distribution of non-zero use has dispersion t hat is less than one, there was no need to correct effect sizes or signific ance levels for intensity of use. If anything, the uncorrected estimates of patient, provider, and clinic factor effects were somewhat conservative. Correlation between Independent Variables Several of the patient level clinical acti vity and other imaging utilization variables are theoretically redundant in that they essentially re flect the same phenomenon. For

PAGE 98

98 example, there are four variables that deriv e from inpatient events; number of inpatient stays (inpt_stays), number of days in t he ICU (inpt_icu_days), and the two readmission measures (inpt_read_15d and inpt _read_31d). The most closely related variable pairs are the counts versus summed RVU of visits to the loyal provider other primary care doctors, and specialists respectively. As t hese are highly correlated, a choice between them (counts or summed RVU) were necessa ry for subsequent mult ivariable modeling. Also, various classes of these clinical activity counts may be correlated even though they are not measuring precis ely the same activity. For ex ample, emergency room visits and imaging tests ordered from the emergency room or inpatient stays and imaging tests ordered in the hospital are likely to be correlated. Table 7-1 lists relevant correlations between clinical activity and other imaging utilization variables. Though there are 18 separate variables (17 in the ro ws plus the first column=all_e_cnt), only 9 columns are shown. The missing columns ha d no correlations > 0.5 and were omitted for brevity. As expected, the correlations between vi sit count and visit RVU for specialists (0.975), linked (loyal) provider (0.994), and covering PCP (0.978) were very high. Therefore, only one set of these would be included in multivariable modeling (visits vs RVU). Emergency room visits and total hours spent in the emergency room were also highly correlated (0.996) as were the number of inpatient stays and total days spent in hospital (0.992). The two readmission measur es (15 and 31 days) were also correlated (0.857). Not included in Table 7-1 is the correlation between the age and experience level of patients linked/loyal primary care doctor. As expected, this was quite high (0.949) and justifies using only one of them for subsequent multivariable modeling.

PAGE 99

99 Bivariate relationships The results of the individual OLS regressi ons are listed in Table 7-2 for patient level demographic, insurance, medications, and problem variables. For patient level clinical activity and other imaging variables, results are summarized in Table 7-3. For provider and clinic variables the results are given in Table 7-4. All of the patient level variables had highly significant linear relationships with the outcome (p < 0.0001). The only exception was substance abuse (binar y problem variable named pr_sub) with p=0.0014. The magnitude of these relationshi ps varied considerably with correlation coefficients ranging from 0.01 up to 0.39, with the strongest being with the summed RVU of visits by the patient to their loyal provider (prv_visit_rvu). Other particularly strong relationships (correlation coefficient s with outcome > 0.2) were exhibited by specialist outpatient visit variables (spec_v is_count, spec_visit_rvu) and the number of active prescriptions for each patient. Finally, the unique identifier for the patients linked (loyal) provider (prov_id on Table 7-4) had a fairly high correlation (0.2268) with the number of outpatient im aging tests ordered by that same doctor. Interestingly, both measures (count and RVU) of visit intensity to the linked (loyal) provider and specialists seemed to have stronger relation ships with the outcome than the actual identity of the patients linked (loyal) provider with all bivariate correlations being > 0.23. Variable Reduction for Modeling As noted above (Table 7-1), the visit counts and RVU variables were highly correlated with each other. The bivariate relati onships with the outcome variable from Table 7-4 were used to guide the choice between them. In each case, the RVU version had slightly higher correlation with the outcome while all of them were significant (p<0.0001). Accordingly, only t he three visit RVU variables (p rv_visit_rvu, pcp_visit_rvu,

PAGE 100

100 and spec_visit_rvu) were carried forward for further analysis. Another pair of variables with high correlation between them was the pr oviders age and experience. From Table 7-4, we see that the provider experience variable was signific antly (p=0.0002) related to the outcome while the provi ders age was not (p=0.2542). This made choosing among them straightforward: select the pr ovider experience variable for subsequent multivariable analysis. Even though the variables representi ng provider foreign medical graduate (FMG) and malpractice status were not significantly related to the outcome in bivariate fashion, they were carried forward due to theoretical considerations. For subsequent modeling, the variables coding ac tual identity of providers (prov_id) and clinics (site_id) were omitted. However, the provider identity variable was used during the final stage of analysis: multi-level hierarchical modeling. In summary, t he following variables were dropped for purposes of multiv ariable logistic and Poisson regression: er_hours (total hours in the ER) inpt_read_15d (count of readmit within 15 days) pcp_visit_count (count of outpati ent visits to covering PCP) prv_visit_count (count of outpatient visits to loyal doc) spec_visit_count (count of outpatient visits to specialists) prov_age_08 (age in years of the provider in 2008) prov_id (provider identifier) site_id (site (clinic) identifier) Multivariable (Logistic) M odeling: Any Imaging Use The logistic model with all 85,277 observations, 28 patient level, 6 provider level and 1 clinic level independent variables was estimated using ANY_IMG as the outcome (yes when the count of im aging tests ordered by patients linked (loyal) doctor was greater than zero and no otherwise). This served to jointly test the effect of each of the 35 independent variables on whether or not the patient had any imaging ordered by their linked (loyal) doctor during the two year s of study. The outcome value of yes was

PAGE 101

101 set to be the event/success level. Subsequent in terpretation of the resulting odds ratios is such that when they are greater than one, that variable/level is associated with a higher probability of imaging. The -2 Log Likelihood was 112,501 for intercept only and 100,678 for the full model. The pseudo R-Squared was 0.13 re scaled to 0.18 and the c Statistic was 0.723. A Hosmer and Lemeshow (HL) test fo r lack of fit was highly significant (p<0.0001) with Chi-Square of 588 on 8 degrees of freedom. However, it should be noted that there is evidence that for large sample sizes (exceeding 50K as in this study) a significant HL test does not entail that a par ticular logistic model is useless or even poorly specified (Kramer and Zimmerman 2007, Bertolini et al. 2000). The only hypothesis affirmed is that t here is a high probability of at least some lack of fit and with the R-Squared of 0.18, this is already established. The individual independent variable resu lts are listed in Tables 7-5 and 7-6 and graphically depicted in Figur e 7-2. In the figure, odds ratios to the right of the reference line (1.0) imply that the variable or leve l was associated with an increased probability that the patient would have any imaging test dur ing the two years of study. Note that the odds ratios for numeric variables (e.g., patient age, clinical activity variables, and provider experience) represent the increase in probability of any imaging with a unit increase in the value of that variable. For example, consider pat ient age. For each additional decade (from 3rd through 9th), the probability of any imaging use increased by 16%. Thus, a 90 year old patient would be about 3 times more likely to have at least one imaging test compared with a 20 year ol d (all else equal). On the other hand, females were only 9% more likely to have imaging than males. Recall, that

PAGE 102

102 mammography has been specifically omitted from this study. Otherwise, that number would likely have been about an order of magnitude higher. Black and Hispanic patients were about 20% more likely to receive imaging as compared to whites (reference level). For the most part, insurance status was either not significant or had a small effect size in the expected direction (greater likelihood of any imaging compared with reference of uninsured). The notable ex ception was Medicare with a significant (<0.0001) and substantial negative effect size (OR = 0.752, ~25% less likely to have any imaging than uninsured--or self-pay--pat ients and even greater compared with other insurance types). The only other signific ant insurance type was Managed and patients were ~18% more likely to have imaging compared with uninsured. The Blue Cross group (BCBS) was marginally significant with a (~14%) positive effect on imaging use. The number of medications patients we re taking had no effect on imaging use. Individual binary clinical problems were signi ficant in six out of nine instances. There was only one clinical problem that seemed to increase lik elihood of imaging and that was Trauma (OR=1.24). The other five problems were associated with decreased imaging when present: Cancer (OR=0.949), Congestive Heart Failure (OR=0.765), Diabetes (OR=0.729), Hypertension (OR= 0.784), and Substance Abuse (OR=0.793). The count variable which subsumed the rema ining clinical problem list entries not categorized above (Other Problems) showed a significant (though small) positive effect on imaging (OR for 1 unit increas e = 1.019). At the median level of 6.0, this would result in ~12% increase in likelihood of any imaging compared with none. Clinical activity variables tested included the summed RVU of outpatient visits. As expected, visits to the pati ents linked (loyal) doctor were strongly related to imaging

PAGE 103

103 (prv_visit_rvu: OR=1.164 for a 1 unit increas e). The maximum value for this variable is 17 RVU. Thus, patients having visits to their linked (loyal) doctor over two years totaling 17 RVU would be at least 13 times more likely to have imaging compared with those having a single visit with fractional RVU. The variable representing visits to other primary care doctors (pcp_visit_rvu) was not significant. Howeve r more visits to specialists (spec_visit_rvu: OR=1.025 for 1 uni t increase) were associated with higher likelihood of imaging (ordered by the patients linked/loyal doctor). Of the six clinical activity variables that measured hospitalizati on, only two were significantly associated with primary care imaging. These were 24 hour observation admissions (obs_stays: OR=1.07 for 1 unit increase) and the total i npatient length of stay (inpt_los_total: OR=0.987 for a 1 unit increase). The seem ing discrepancy makes some sense by speculating that short stays for observati on might indicate and/or engender need for imaging that would be perform ed later (as an outpatient). Two of the additional imaging utilization variables had small effect sizes: outpatient ordered by specialists (spec_o_cnt: OR=1. 025) and inpatient (all_i_cnt: OR=1.020). Emergency room imaging (all_e_cnt) was not significant. The number of outpatient images ordered by other primary care docto rs (pcp_o_cnt: OR=1.157) was positively associated with likelihood of imaging by the patients own (loyal) doctor. For future applications using different data sources (risk adjustment for provider profiling of imaging utilization) these additional imaging variables can probably be omitted with little consequence because the remaining clinical ac tivity variables will capture the same phenomenon. Clearly, hospital ev ents are, by definition, correlated with associated imaging (e.g., ER visits and imaging performed in the ER). Likewise, visits to specialists

PAGE 104

104 and other (covering) primary care doctors could stand in for the outpatient imaging ordered by these same doctors (e.g., omitting pcp_o_cnt might allow pcp_visit_rvu to become significant). Turning to the provider and clinic level variables (Table 7-6) we see that the amount of experience is negatively associ ated with likelihood to obtain imaging (OR=0.997 for each additional year). Howeve r the effect size is rather small. Considering that the range of experience was 5-50 years, this implies that likelihood of any imaging decreases by only about 10% between least and most experienced doctors. The gender of the doctor has a great er effect than expe rience, with women (OR=1.14) being 14% more likely to order imaging on their patients compared with males. Foreign medical graduates (FM G: OR=1.11) and doctors with additional academic credentials beyond M.D. (MD_Plus: OR=1 .37) tend to order tests on more of their patients than American-trained and M.D. only primary care doctors. Malpractice (whether or not the doctor has been sued in past 10 years) has no significant effect. It is certainly possible that the sm all event rate for the malpractice variable (Number Yes=7/148) contributes to the lack of significance. However, the other two (significant) provider variables also had small numbers of yes/true values (FMG=8/148, MD_Plus=14/148). The two variables measuring practice size were both positively related to likelihood of imaging. For the (categoric al) number of patients in each providers panel, the three levels that were greater than reference (<500) had from 10-16% more likelihood to obtain imaging. The number of active provi ders practicing in each of the 15 clinics was

PAGE 105

105 slightly positively associat ed with greater tendency for a ssigned patients to get imaging (OR=1.014) which translates into a 20% increase over the range (5-18). Multivariable (Poisson) Modeling: Imaging Intensity (Non-zero) The Poisson model on the 31,660 observations with non-zero outcomes had Deviance/df of 0.6613 and Chi-Squared/df of 0.7935. When the sa me model was run with the DSCALE and PSCALE options the resulting scale parameters were 0.8132 and 0.8908 respectively. This implies that the Po isson distribution of the outcome (number of outpatient images ordered by the linked/loyal doctor) for these non-zero patients is underdispersed and that the standard errors for the c oefficients might tend to be slightly overestimated. Therefore, inferences about the significance of variables and levels would, if anything, be conservative and c an be discussed with some confidence. These uncorrected coefficients and standard errors ar e presented in Tables 7-7 and 7-8, and Figure 7-3 displays these same results in terms of 95% Wald confidence intervals. When discussing the parameter coefficient estimates in terms of effect size and direction on imaging intensity, it is import ant to recall that t he link function for the Poisson model was log rather than linear. This means that we cant translate the value of the estimate into an additive number of imaging tests per patient for the variable or level in question. However after exponentiati on (far right columns in Tables 7-7 and 78), the values can be interpreted as (multipli cative) incident rate ratios (IRR). To keep these in perspective, recall that the inte rcept for a null Poisson model on the 31,660 non-zero observations is 0.6548 which afte r exponentiation is 1.925. As noted above, this is the Poisson Lambda for the non-zero imaging counts and can be interpreted as the expected number of images that the aver age (non-zero) patient would have in two years.

PAGE 106

106 For the categorical variables, using refer ence cell encoding, the interpretation of the exponentiated coefficients is straightforward. For example, females have about 1.027 times more imaging tests than males (all else equal). Theref ore, for the average female patient that had imaging, the expected count would be 1.925 times 1.027 or 1.977 images over two years. For the numerical variables the exponentiated estimate represents a multiplier applied for each additional unit value. Thus for patient age over the range of 17-103 years there is an 86 y ear difference which translates into Exp(86*0.0053) = 1.58 times more im aging tests between the youngest and oldest patients (all else equal). Moving on to race we note that Bla ck and Hispanic patients tend to have more imaging tests than whit es (the reference level). With patient insurance category, only Medicare is signifi cant compared with uni nsured/self-pay with IRR of ~0.92. Recalling the l ogistic results, we can say t hat patients with Medicare are less likely to have any imaging and those that do tend to have a lower number of imaging tests than patients in other payer categories. As with the logistic analysis for any im aging use, number of active outpatient medications had no effect on number of imaging tests. With three ex ceptions, the binary clinical problem variables were not significa nt. These were chronic renal failure (CRF), diabetes, and hypertension. All else equal, patients with these clinical problems tended to get fewer imaging tests t han those without them. The variable representing the count of other clinical problems was significant ly and positively associated with a larger number of imaging tests as the problem count increased (IRR=1.003 per additional problem).

PAGE 107

107 The clinical activity variable most str ongly associated with the number of imaging tests was the summed RVU of visits to the lin ked (loyal) primary ca re doctor. Over the range of this variable, patient s with 17 RVU worth of visits to their linked primary care doctor had about 1.6 times more imaging tests than those with a single visit (totaling less than one RVU). Visits to specialists were also positively associated with number of imaging tests, though much less strongly wit h IRR for each additional RVU of 1.006 compared with 1.029. On the other hand, when patients saw other (c overing) primary care doctors, the number of imaging tests order ed by their own (linked/loyal) doctor was slightly lower. The only other clinical activi ty variable that was significantly associated with the number of imaging tests was the total inpatient length of stay which had a negative effect. For example, one patient spent a total of 179 days in the hospital over the two years of study. That person w ould be expected to have less than 30% (Exp(179*-0.007)=0.29) of t he number of imaging tests as a patient that had not been in hospital at all during the study. This may seem c ounter intuitive at first, but consider that patients who spend many days in hospital ha ve more imaging tests performed there. Since the results of these tests are available to the patients primar y care doctor, he or she would likely find the answers to their diagnostic questions in tests already performed and not need to order new ones in the outpatient setti ng. Lastly, the variables representing outpatient imaging tests ordered by other docto rs (specialists and covering primary care doctors) and while in the hospital were positively associated with the number of outpatient imaging tests ordered by the pati ents linked (loyal) doctor. Focusing on provider and site level variables (Table 7-8) and recalling the any imaging (logistic) results, a similar pattern emerges for the prov ider experience and

PAGE 108

108 gender variables. Over the observed range of provider experience (5-50 years) the most experienced clinicians ordered about 87% as many tests as the least experienced. Female physicians ordered more tests than males but the difference was only about 7%. Doctors with additional training after their M.D. ordered about 10% more imaging. As with the any imaging (logistic) analysis, th e providers malpractice status had no effect on number of images ordered. The tw o middle practice size categories (500-799, 800-1000) had significant but sm all positive effects on number of images compared with reference (<500). The final provider vari able, foreign medical graduate status (FMG_Yes), had a significant (p=0.023) negativ e effect (4%) on number of images per patient whereas the same variable was a ssociated with a greater likelihood (11%) of ordering any imaging (logistic) on a given patient. Finally, the clinic size (number of doctors) variable had a small but significant positive effect on the number of imaging tests ordered on assigned patients. Over the range of practice sizes of 5-18 doctors, this translates into about 4% more imaging tests. Comparison of Any Imaging and Imaging Intensity Results To compare the effect of the various independent variables and levels on any imaging use (logistic) and imaging intensity (non-ze ro Poisson) it is useful to plot the respective odds ratios and coefficients. Fi gure 7-4 shows odds ratios and coefficients for the 32 variables / levels that were signi ficant for either any imaging use or imaging intensity. Figure 7-5 shows the 9 variables th at had relatively small effect sizes (odds ratio near one, coefficient near zero) wit h the axes scaled down to show the relationships to better advantage. These we re all numeric rather than categorical variables which explains the small effect size s (for a unit change in value) that are still quite significant. Of these, only one (summed RVU of visits to covering primary care

PAGE 109

109 doctors) was significant for imaging intensity (non-zero Poisson) but not for any imaging use (Logistic). In both Figures 7-4 and 7-5, note that the general tendency is for variables and/or levels to be concordant with respect to thei r effect size and di rection (when both are significant) between any imaging use (logistic odds ratio, X Axis) and imaging intensity (Poisson regression coefficient, Y Axis). Ther e is only a single exception and this is with the variable coding for whether or not the linked (loyal) prim ary care doctor is a foreign medical graduate (FMG). It seems that FMG pr imary care doctors we re more likely to order some imaging but ordered fewer imaging tests when they did so. Preparation for Multilevel Modeli ng: Imaging Propensity Scores Patient level predictions from the zero inflated Poisson (ZIP) model using all 28 patient level variables for both the count and zero-model portions were calculated, called IMG_PROP, and stored in a new data set along with the original outcome (IMG) and the (coded/anonymous) provider ID of that patients linked primary care doctor. Table 7-9 compares the raw outcome (IMG) and the predictions from the ZIP model (IMG_PPOP). The mean of IMG_PROP for each provider was subtracted from the original value for each patient to form a centered imagi ng propensity variable (cIMG_PROP), which can be expressed as Equation 7-1. cIMG_PROPij = IMG_PROPij IMG_PROP.j (7-1) In Equation 7-1, i is the ith patient, j is the jth provider, and .j is the average for the jth provider. The result of the above described centering operation is illustrated by plotting the raw imaging propensity scores (IMG_PROP) and centered imaging propensity scores

PAGE 110

110 (cIMG_PROP) against the mean outcome (images per patient) for each provider. These are shown in Figures 7-6 and 7-7 respectively. Multi-Level (Hierarchical) Modeling The SAS PROC MIXED procedure on the full two level model completed in ~20 seconds and converged after 4 iterations. Results from SAS are reproduced in Tables 7-10 through 7-15. These results are summarized in Table 7-16 in terms of fixed (pat ient level) effects and random (patient and prov ider levels) variance. The fixed effect intercept (0.7171) is t he average doctors mean imaging. This is nearly identical to the raw mean value of 0. 7146 obtained by dividing the total number of imaging tests ordered by the pat ients linked provider (N=60,938) by the number of patients in the whole study cohort (N=85,277). A (95%) range of plausible values for doctors mean imaging (intercept) around 0.7171 can be constructed using the variance (0.0835) by 0.7171 1.96(0.0835)1/2 which gives (0.151, 1.283). Note that this is rather wider than t he 95% confidence intervals (0.6695, 0.7648) on the estimate of the fixed intercept prov ided by SAS, which used the standard error. The fixed effect of cIMG_PROP (0.9919) is interpreted as the average doctors response (slope) in number of imaging te sts ordered for a unit change in the imaging propensity score (cIMG_PROP). The fact that this is very cl ose to 1.0 implies that the scale of the imaging propensity score is co rrect (at least around the mean value). In other words, on average, as the expected amount of imaging increases by one unit, actual imaging utiliz ation increased by one extra test pe r patient. As with the intercept,

PAGE 111

111 we can calculate a (95%) range of plausible values for the slopes using the variance (0.1567) by 0.9919 1.96(0.1567)1/2 which gives (0.216, 1.768). As with the intercept, the plausible range of the slope is wider than the SAS calculated 95% confidence interval on the slope parameter estimate (0.925, 1.058). Intraclass correlations (ICC) for intercept and slope were obtained by dividing each component variance (00 for intercept and 11 for slope) by the residual variance (2) plus itself. For intercept, this is 0.0835 di vided by (0.0835+1.1934) which gives 0.065. For slope, the ICC is 0.1567 divided by (0. 1567+1.1934) which gives 0.116. This implies that about 6% of the variance in interc epts is between doctors and about 12% of the variance in slopes is between doctors. It is he lpful to recall that an ICC of 0 would mean that all doctors exhibit the same IMG/c IMG_PROP relationship and clustering of patients by doctor had no effect (i.e., the hier archical modeling not informative). On the other hand, an ICC approaching 1 would mean that any given doctors patients have nearly identical adjusted imaging utilizati on and very small variation between them. Model based estimates (including standard errors and 95% confidence intervals) of individual provider intercept and slope were obtained by requesting the solution for the random portion of the model. Reliability for the individual provider intercept estimates can be calculated as the overall pr ovider intercept variance (0.1567) divided by the providers own variance (standard er ror squared) and the overall reliability for provider intercept estimates is the average of our 148 doctors which is 0.965. Similarly, the individual reliability for each provider s slope estimate is the overall variance (0.0835) divided by the individual variance (standard error squared) with the aggregate reliability being the average of these for the 148 doctors which is 0.939. The high

PAGE 112

112 reliability of both the intercept and slope estimates is reassuring and supports interpreting them as repres enting each providers mean t endency to order imaging tests (intercept) and their response (s lope) to patient level imaging propensity represented by the cIMG_PROP (risk adjusted expected imaging) variable. The correlation between each doctors general tendency to order imaging (intercept) and his or her response to patient imaging propensity (slope) can be expressed as ( 0j 1j) which is estimated by 01/(00 11)1/2. Substituting from Table 7-16 gives 0.0810 / (0.0835 x 0.1567)1/2 which turns out to be 0.7081. This implies a substantial correlation between the aver age tendency to use imagining, and the increase in the number of im ages providers order on their patients with higher imaging propensity (i.e., sicker). A scatter plot of the in tercepts (X axis) and slopes (Y axis) for all 148 providers is shown in Figure 7-8 and se rves to visualize the relationship between them. The quadrants in this slope versus inte rcept plot are labeled A-D and are further detailed in Tables 7-17 and 7-18. Less than 20% of the providers have discordance between their slopes and intercepts (quadrants A and D) with the remaining 80% (quadrants B and C) being concordant with respect to their tendency to image and their response to patient imaging need as represented by the imaging propensity variabl e. Therefore, in general, if a primary care doctor tends to order imaging less than average odds are 4.5:1 that he or she will also increase the amount of images that they obtain on sicker patients less than average For example, provider C.C. obt ained about 29 images per 100 patients over the course of the st udy and increased their image ut ilization by about 15 images per 100 for each unit increase in imaging propensity. It is useful to plot imaging

PAGE 113

113 propensity versus actual imaging utilization for all the patients cared for by Doctor C.C. and this is shown in Figure 7-9. Likewise, if a doctor tends to order imaging more than average odds are 3.7:1 that he or she will increase the am ount of images that they order on sicker patients more than average For example, provider B.B. obt ained about 199 images per 100 patients over the course of the st udy and increased their image utiliz ation by about 264 images per 100 for each unit increase in imaging pr opensity. The observed imaging versus imaging propensity for Dr. B.B. is plotted in Figure 7-10. The discordant doctors (quadrants A and D in Figure 7-8) are not only few in number but tend to cluster near the mean value of the slope (with standard errors overlapping the average slope of zero) such that only three doctors are significantly discordant with one having low in tercept / high slope and two having high intercept / low slope. For example, the provider labeled A.A. in the scatterplot obtained about 50 images per 100 patients over the two years of study and increased imaging by about 119 images per 100 for every unit increase in imaging propensity. Another way of saying this is that Doctor A.A. has a so mewhat higher threshold for obtaining imaging on any given patient but t ends to order more images as patient need for imaging increases. In contrast, provider D.D. obt ained about 85 images per 100 patients and increased their imaging by about 58 images per 100. One might speculate that provider D.D. is generous with imaging in general but less discriminati ng in terms of increasing utilization according to patient need. On the other hand, the two quadrant D providers (D.D. and the one just above on the scatter plot) may be relatively liberal in terms of

PAGE 114

114 both imaging and referral to specialists. This might mean that their sicker patients get less images ordered by the primar y care provider because they tend to refer at a lower threshold and at least some images would be ordered by the specialists rather than themselves. Another interesting visualization is to pl ot the only the intercepts (providers mean imaging) and 95% confidence intervals sorted according to practice site, and this shown in Figure 7-11. Similarly, a pl ot of the slopes (providers change in imaging as imaging propensity increases) with 95% confidence intervals is shown as Figure 7-12. One important observation from the provider slopes plot (Figure 7-12) is that all providers increase their diagnostic image ordering in re sponse to additional patient need (none of the scaled imaging vs. propensity slopes are below zero). The alternative is that some doctors slopes could be negative such that the amount of imaging actually decreased for sicker patients. One expl anation for this (counterfactual ) would be that those with negative slopes, refer sicker patients to s pecialists who themselves obtain the needed imaging tests. The fact that th is did NOT occur at all in t he current study implies that even sick patients who see many specialists cont inue to have at least some of their care rendered by the linked (loyal) primary care pr ovider. This should come as no surprise given that demonstration of this loyalty rela tionship was the main inclusion criteria for both patients and providers. The reduced level 1 model (IMG = IMG_PR OP) yielded a patient level residuals (error variance) of 1.3061 while the null model (IMG=;) gave patient level residuals (error variance) of 1.5782. These are combi ned with the error variance from the full two level model (1.1934) in Table 7-19.

PAGE 115

115 The overall amount of explained variation in outpatient imaging utilization after accounting for provider ID and all the patient level variables (risk-adjustment) is 24.4% (0.3848/1.5782). Of that, roughly 70% (0.2721/0. 3848) is attributable to patient level factors as captured in t he imaging propensity variable and the remaining 30% is attributable to provider variation. One imp lication is that about th ree quarters of the variation in the number of out patient imaging tests ordered by a primary care doctor on loyal patients is unexplained. This is despite taking into account a robust and large set of patient factors as well as all between doc tor differences in imaging utilization habits (by directly modeling unique pr ovider identity). Of the ro ughly 25% of variation in primary care outpatient imaging utilization that can be explai ned, the majority (70%) is attributed to factors that mostly relate to each patients clinical need for imaging, regardless of who their primary care doctor is. The remaining 30% arises from differences in the tendency for primary care doctors to order imaging which may be partitioned into intercept (~10%) and slope (~20%) components. The next chapter will summarize and discuss these results in terms of advances in knowledge, applications in imaging utilization management and provider pr ofiling. Directions for future research with these (and similar) data sources as well as some policy implications will be covered as well.

PAGE 116

116 Table 7-1. Spearman correlations between clin ical activity and other imaging variables. Variable Name all_e_cnt all_i_cnt spec_o_cnt spec_visit_count pcp_visit_count prv_visit_count er_visits inpt_stays inpt_read_15d all_i_cnt 0.450 spec_o_cnt 0.226 0.273 pcp_o_cnt 0.095 0.056 0.107 spec_visit_count 0.260 0.278 0.574* pcp_visit_count 0.111 0.048 0.1320.195 prv_visit_count 0.253 0.230 0.2410.3420.125 spec_visit_rvu 0.260 0.279 0.575*0.975*0.1940.348 pcp_visit_rvu 0.115 0.053 0.1350.1990.994*0.128 prv_visit_rvu 0.254 0.239 0.2490.3510.1120.978* er_visits 0.843* 0.423 0.2280.2720.1310.263 er_hours 0.857* 0.429 0. 2290.2730.1300.2670.996* obs_stays 0.295 0.148 0. 2500.2980.0760.1360.290 inpt_stays 0.469 0.726* 0. 3030.3310.0660.2310.455 inpt_los_total 0.474 0.742* 0. 2990.3270.0660.2320.459 0.992* inpt_icu_days 0.203 0.366 0.101 0.1100.0110.0910.184 0.283 inpt_read_15d 0.243 0.343 0. 1060.1210.0280.0990.231 0.302 inpt_read_31d 0.283 0.398 0.126 0.1410.0300.1160.267 0.352 0.857* Key To Variable Names Is Below all_e_cnt count of images done in ER all_i_cnt count of images done as inpatient spec_o_cnt count of outpatient images ordered by specialists pcp_o_cnt count of outpatient images ordered by covering PCP spec_visit_count count of outpatient visits to specialists pcp_visit_count count of outpatient visits to covering PCP prv_visit_count count of outpatient visits to loyal doc spec_visit_rvu sum of RVU of outpatient visits to specialists pcp_visit_rvu sum of RVU of outpatient visits to covering PCP prv_visit_rvu sum of RVU of outpatient visits to loyal doc er_visits count of ER visits er_hours total hours in the ER obs_stays count of observation stays inpt_stays count of inpatient stays inpt_los_total total days in hospital inpt_icu_days total days in ICU inpt_read_15d count of readmit within 15 days inpt_read_31d count of readmit within 31 days NOTE: For brevity, columns where ALL correlations were < 0.5 are omitted. Correlations above 0.5.

PAGE 117

117 Table 7-2. Bivariate relationship between pati ent level variables and outcome (imaging counts). Type Description Variable Name F Value R-Squared Correlation p value Categorical patient identified race Race 29.98 0.0011 0.0324 <0.0001 Categorical patient sex Se x 258.44 0.0030 0.0550 <0.0001 Numeric patient age in 2008 age_ 08 3087.75 0.0349 0.1869 <0.0001 Categorical patient's payer of record in 2008 PayerGroup 343.33 0.0236 0.1536 <0.0001 Categorical Active prescriptions in 2008 meds_cat 1313.04 0. 0442 0.2101 <0.0001 Binary coronary artery disease pr_cad 445.63 0.0052 0.0721 <0.0001 Binary cancer pr_can 476. 77 0.0056 0.0746 <0.0001 Binary congestive heart failure pr_chf 123.35 0.0014 0.0379 <0.0001 Binary chronic renal failure pr _crf 132.79 0.0016 0.0394 <0.0001 Binary diabetes pr_dm 436. 64 0.0051 0.0713 <0.0001 Binary obesity pr_obs 150. 85 0.0018 0.0421 <0.0001 Binary hyptertension pr_htn 628.14 0.0073 0.0855 <0.0001 Binary substance abuse pr_s ub 10.19 0.0001 0.0110 0.0014 Binary trauma pr_trm 48.85 0.0006 0.0239 <0.0001 Numeric count of active problems other than those listed oth_prb 3115.43 0.03 53 0.1877 <0.0001

PAGE 118

118 Table 7-3. Bivariate relationship between clinical activity variables and outcome (imaging counts). Type Description Variable Name F Value RSquared Correlation p value Numeric total hours in the ER er_hours* 1323.130.01530.1236 <0.0001 Numeric count of ER visits er_visits 1412.20.01630.1276 <0.0001 Numeric total days in ICU inpt_icu_days 78.360.00090.0303 <0.0001 Numeric total days in hospital inpt_los_total 753.110.00880.0935 <0.0001 Numeric count of readmit within 15 days inpt_read_15d* 214.520.00250.0501 <0.0001 Numeric count of readmit within 31 days inpt_read_31d 273.090.00320.0565 <0.0001 Numeric count of inpatient stays inpt_stays 1562.360.01800.1341 <0.0001 Numeric count of observation stays obs_stays 647.260.00750.0868 <0.0001 Numeric count of images done in ER all_e_cnt 1281.080.01480.1217 <0.0001 Numeric count of images done as inpatient all_i_cnt 663.80.00770.0879 <0.0001 Numeric count of outpatient images ordered by covering PCP pcp_o_cnt 581.730.00680.0823 <0.0001 Numeric count of outpatient images ordered by specialists spec_o_cnt 2729.80.03100.1761 <0.0001 Numeric count of outpatient visits to covering PCP pcp_visit_count* 436.630.00510.0713 <0.0001 Numeric sum of RVU of outpatient visits to covering PCP pcp_visit_rvu 536.260.00630.0791 <0.0001 Numeric count of outpatient visits to loyal doc prv_visit_count* 14540.960.14570.3817 <0.0001 Numeric sum of RVU of outpatient visits to loyal doc prv_visit_rvu 15228.910.15150.3893 <0.0001 Numeric count of outpatient visits to specialists spec_visit_count* 4786.980.05320.2305 <0.0001 Numeric sum of RVU of outpatient visits to specialists spec_visit_rvu 5683.230.06250.2500 <0.0001 Variables NOT carried forward to multivariable analysis.

PAGE 119

119 Table 7-4. Bivariate relationship between provider and clinic level variables and outcome (imaging counts). Type Description Variable Name F Value RSquared Correlation p value Categorical whether provider has been sued in last 10 years mp_flag 1.050.00000.0032 0.3054 Categorical whether provider is foreign medical graduate prov_fmg 0.170.00000.0000 0.6761 Categorical whether provider has a degree beyond MD prov_md_plus 32.710.00040.0195 <0.0001 Categorical number of patient's in provider practice in 2008 prov_pat_cat 79.40.00280.0528 <0.0001 Categorical provider sex prov_sex 66.90.00080.0279 <0.0001 Numeric age in years of the provider in 2008 prov_age_08* 1.30.00000.0045 0.2542 Numeric number of years after provider MD graduation in 2008 prov_exp_08 14.370.00020.0130 0.0002 Identifier anonymous provider identifier prov_id* 31.430.05140.2268 <0.0001 Numeric number of doctors actively practicing at the clinic in 2008 site_docs 189.380.00220.0471 <0.0001 Identifier site (clinic) identifier site_id* 100.160.01620.1272 <0.0001 Variables NOT carried forward to multivariable analysis.

PAGE 120

120 Table 7-5. Patient level results from multiv ariable logistic model on any imaging use. Type Variable / Level Estimate Standard Error ChiSquare Pr > ChiSq Odds Ratio Demographics Age 0.0160.001483.106<0.0001 1.016 Sex F 0.0830.01822.011<0.0001 1.087 Race Black 0.2120.03536.472<0.0001 1.236 Hispanic 0.2320.03252.404<0.0001 1.261 Other 0.0910.0299.6950.0018 1.095 Insurance BCBS 0.1330.0673.9330.0473 1.143 Commercial 0.0920.0711.7120.1908 1.097 Managed 0.1680.068 6.0650.0138 1.183 Medicare -0.2860.07016.570<0.0001 0.752 State 0.0280.072 0.1500.6982 1.028 Other 0.0160.0780.0410.8390 1.016 Medications 1-5 0.044 0.0371.3810.2400 1.045 6-10 0.0440.0401.2340.2666 1.045 >10 -0.0170.0460.1340.7146 0.983 Problems CAD -0.026 0.0390.4630.4961 0.974 Cancer -0.0520.0254.4910.0341 0.949 CHF -0.2680.07911.4530.0007 0.765 CRF -0.1130.0692.6500.1036 0.893 Diabetes -0.3160.027141.985<0.0001 0.729 Obesity 0.0130. 0260.2670.6056 1.013 Hypertension -0.2430.019160.556<0.0001 0.784 Substance Abuse -0.2320.0857.4400.0064 0.793 Trauma 0.2150.05316.426<0.0001 1.240 Other (count) 0.0180.001193.493<0.0001 1.019 Visits prv_visit_rvu 0.1520.0024253.497<0.0001 1.164 pcp_visit_rvu -0.0060.0061.1150.2910 0.994 spec_visit_rvu 0.0250.002276.583<0.0001 1.025 Other Imaging pcp_o_cnt 0.1460.01962.486<0.0001 1.157 spec_o_cnt 0.0240.00432.949<0.0001 1.025 all_e_cnt 0.005 0.0090.2820.5952 1.005 all_i_cnt 0.020 0.0078.9780.0027 1.020 Hospital er_visits -0.0120.0140.7300.3929 0.989 obs_stays 0.0670.01617.880<0.0001 1.070 inpt_stays -0.0220.0240.8680.3515 0.978 inpt_read_31d -0.077 0.0532.0700.1503 0.926 inpt_los_total -0 .0130.00410.7090.0011 0.987 inpt_icu_days -0. 0130.0150.7360.3910 0.987 Reference Levels: Sex-Male, Race-White, Insurance-Uninsured, Medications-None

PAGE 121

121 Table 7-6. Provider and clinic level result s from multivariable logistic model on any imaging use. Type Variable / Level Estimate Standard Error ChiSquare Pr > ChiSq Odds Ratio Provider Experience -0.0030.00110.9930.00090.997 sex F 0.1330.01855.406<0.00011.142 FMG Yes 0.1080.03410.4120.00131.114 MD Plus Yes 0.3180.029117.859<0.00011.374 Malpractice Yes 0.0200.0400.2620.60851.021 Provider Patients 500-759 0.1230.02818.621<0.00011.130 750-999 0.1500.02731.511<0.00011.162 1K+ 0.0980.02514.9740.00011.103 Clinic size Active Providers 0.0140.00269.051<0.00011.014 Reference levels: Sex-Male, FMG-No, MD_Plu s-No, Malpractice-No, Provider Patients<500

PAGE 122

122 Table 7-7. Patient level results from multiv ariable Poisson model on imaging intensity. Type Variable / Level Estimate Standard Error ChiSquare Pr > ChiSq (RR) Exp Estimate Demographics Age 0.0053 0.0004183.800<0.0001 1.005 Sex F 0.0270.010 8.2200.0042 1.027 Race Black 0.0340.0183.4800.062 1.034 Hispanic 0.0730.01619.500<0.0001 1.075 Other -0.0130.0170.6500.4198 0.987 Insurance BCBS -0.0470.0391.4400.2305 0.954 Commercial -0.0520.0411.5900.2071 0.949 Managed -0.0310. 0400.6200.4305 0.969 Medicare -0.085 0.0404.5200.0335 0.918 State -0.0480. 0411.3600.2435 0.953 Other 0.0000.0450.0000.9932 1.000 Medications 1-5 0.001 0.0240.0000.959 1.001 6-10 0.0100.0250.1700.6787 1.010 >10 0.0280.0271.0600.3028 1.028 Problems CAD -0.028 0.0172.7300.0982 0.972 Cancer 0.0130.0121.2000.2725 1.013 CHF -0.0470.0332.0400.153 0.954 CRF -0.0780.0306.8600.0088 0.925 Diabetes -0.0580.01321.250<0.0001 0.944 Obesity 0.0180. 0132.0600.1508 1.019 Hypertension -0.058 0.01036.160<0.0001 0.944 Substance Abuse 0.0630.0402.5200.1124 1.065 Trauma 0.0130.026 0.2700.6011 1.013 Other (count) 0.0030. 00132.560<0.0001 1.003 Visits prv_visit_rvu 0.0290.0011479.330<0.0001 1.029 pcp_visit_rvu -0.0090.00311.3900.0007 0.991 spec_visit_rvu 0. 0060.00190.200<0.0001 1.006 Other Imaging pcp_o_cnt 0.0260.00810.7400.001 1.026 spec_o_cnt 0.0070.00218.360<0.0001 1.007 all_e_cnt -0.004 0.0031.6400.200 0.996 all_i_cnt 0.0120.00226.340<0.0001 1.012 Hospital er_visits -0.0050.0050.7900.3754 0.996 obs_stays 0.0060.0070.8700.3503 1.006 inpt_stays 0.0100.0101.0300.3101 1.010 inpt_read_31d 0.023 0.0211.2000.274 1.023 inpt_los_total -0. 0070.00217.580<0.0001 0.993 inpt_icu_days -0. 0100.0053.6600.0558 0.990 Reference Levels: Sex-Male, Race-White, Insurance-Uninsured, Medications-None

PAGE 123

123 Table 7-8. Provider and clinic level result s from multivariable Po isson model on imaging intensity. Type Variable / Level Estimate Standard Error ChiSquare Pr > ChiSq (RR) Exp Estimate Provider Experience -0.0030.00128.060<0.0001 0.997 sex F 0.0630.00945.860<0.0001 1.065 FMG Yes -0.0410.0185.1500.0232 0.960 MD Plus Yes 0.0970.01540.980<0.0001 1.102 Malpractice Yes -0.0080.0210.1400.7091 0.992 Provider Patients 500-759 0.0440.0158. 5400.0035 1.044 750-999 0.0840.01435.880<0.0001 1.087 1K+ 0.0260.0143.6300.0567 1.026 Clinic size Active Providers 0.0030.00112.4600.0004 1.003 Reference levels: Sex-Male, FMG-No, MD_Plu s-No, Malpractice-No, Provider Patients<500 Table 7-9. Univariate statistics for raw imaging counts (IMG) and pr edictions from ZIP model (IMG_PROP). Statistic IMG IMG_PROP N 85277 85277 Minimum 0 (N=53,617) 0.0446 Maximum 15*14.8, 15,2, 20.1 Mean 0.7146 0.7164 Standard Deviation 1.256 0.5651 Skewness 2.655 3.844 Coefficient of Variation 176 78 Sum of Observations 60938 61088 Variance 1.578 0.319 Kurtosis 10 46 Standard Error of the Mean 0.004 0.002 NOTE: The highest 3 observations (*) are shown for IMG_PROP Maximum. Table 7-10. Dimensions. Covariance Parameters 4 Columns in X 2 Columns in Z Per Subject 2 Subjects 148 Max Obs Per Subject 2101

PAGE 124

124 Table 7-11. Estimated G correlation matrix. Row Effect prov_id Col1 Col2 1 Intercept 105621.00000.7081 2 cIMG_PROP 105620.70811.0000 Table 7-12. Covariance parameter estimates. Cov Parm Subject Estimate Standard Error Z Value Pr Z UN(1,1) prov_id 0.083450.010038.32<0.0001 UN(2,1) prov_id 0.080980.011956.78<0.0001 UN(2,2) prov_id 0.15670.019538.02<0.0001 Residual 1.19340.005789206.13<0.0001 NOTE: UN(1,1) = 00, UN(2,2) = 11, UN(2,1) = 01 Table 7-13. Fit statistics. -2 Res Log Likelihood 257971.1 AIC (smaller is better) 257979.1 AICC (smaller is better) 257979.1 BIC (smaller is better) 257991.1 Table 7-14. Solution for fixed effects. Effect Estimate Standard Error DF t Value Pr > |t| Intercept 0.7171 0.0241214729.74<0.0001 cIMG_PROP 0.9919 0.0336414729.48<0.0001 Table 7-15. Type 3 tests of fixed effects. Effect Num DF Den DF F Value Pr > F cIMG_PROP 1 147869.19<0.0001 Table 7-16. Results from multilevel random coefficients model. Fixed Effect Symbol Coefficient Standard Error t-value p-value Intercept 00 0.71710.0241229.74 <0.0001 cIMG_PROP 10 0.99190.0336429.48 <0.0001 Random Effect Symbol Variance Component Standard Error z-value p-value Patient Residual eij ( 2) 1.19340.00579206.13 <0.0001 Provider Intercept u0j00) 0.08350.010038.32 <0.0001 Provider Slope u1 j (11) 0.15670.019538.02 <0.0001 Covariance: Intercept, Slope Cov(u0j,u1j) ( 01) 0.08100.011956.78 <0.0001

PAGE 125

125 Table 7-17. Quadrants in intercept versus slope relationship plot. Quadrant Tendency To Order Imaging: Intercept Response To Imaging Propensity: Slope Number Of Providers Percent Of Providers A Low (<0) High (>0) 14 9.46 B High (>0) High (>0) 52 35.14 C Low (<0) Low (<0) 67 45.27 D High (>0) Low (<0) 15 10.14 Table 7-18. Exemplary pr oviders in each quadrant. Provider Tendency To Order Imaging: Intercept Response To Imaging Propensity: Slope A.A. -0.2176 (0.4995) 0.1988 (1.1907) B.B. 1.2708 (1.9879) 1.6451 (2.637) C.C. -0.4268 (0.2903) -0.8387 (0.1532) D.D. 0.1310 (0.8481) -0.4087 (0.5832) NOTE: Numbers in parentheses for intercept are adjusted to the fi xed effect mean by adding 0.7171 and numbers in parenthesis for slope are adjusted to the fixed effect mean by adding 0.9919. Table 7-19. Comparison of null, and r educed model residuals with full model. Patient IMG_PROP Provider Intercept Provider Slope Model Name Residual (error) Absolute Reduction Fraction Reduction Null 1.5782 X Provider Intercept Only 1.49960.0786 0.0498 X Imaging Propensity Only* 1.30610.2721 0.1724 X X Two Level: Provider Slope 1.27270.3055 0.1936 X X Two Level: Provider Intercept 1.23890.3393 0.2150 X X X Two Level: Provider Intercept and Slope 1.19340.3848 0.2438 NOTE: Fraction reduction in the residual is equivalent to R-Squared for that model. *NON Centered. If centered, error=1.3180.

PAGE 126

126 Figure 7-1. Comparison of imaging c ounts with three Poisson distributions. 0 10000 20000 30000 40000 50000 60000 012345678910Number Of Imaging TestsCount Observed Counts Simple Poisson Non-zero Poisson Zero-inflated Poisson

PAGE 127

127 Figure 7-2. Logistic regression results for any imaging utilizatio n. Horizontal bars represent 95% confidence intervals on odds ratio for each variable/level. Patient variable reference levels: Sex-Male, Race-White, InsuranceUninsured, Medications-None, Problem s-No. Provider variable reference levels: Sex-Male, FMG-No, MD_Plus-No, Malpractice-No, Provider Patients<500. 1.014 1.103 1.162 1.130 1.021 1.374 1.114 1.142 0.997 0.987 0.987 0.926 0.978 1.070 0.989 1.020 1.005 1.025 1.157 1.025 0.994 1.164 1.019 1.240 0.793 0.784 1.013 0.729 0.893 0.765 0.949 0.974 0.983 1.045 1.045 1.016 1.028 0.752 1.183 1.097 1.143 1.095 1.261 1.236 1.087 1.0160.50.60.70.80.91.01.11.21.31.41.5CLINIC SIZE [Active Providers ] [1K+] [750-999] PROVIDER PATIENTS [500-759] [Malpractice Yes] [MD Plus Yes] [FMG Yes] [sex F] PROVIDER [Experience ] [inpt_icu_days ] [inpt_los_total ] [inpt_read_31d ] [inpt_stays ] [obs_stays ] HOSPITAL [er_visits ] [all_i_cnt ] [all_e_cnt ] [spc_o_cnt ] OTHER IMAGING [pcp_o_cnt ] [spc_visit_rvu ] [pcp_visit_rvu ] VISITS [prv_visit_rvu ] PROBLEMS [Other (count)] [Trauma] [Substance Abuse] [Hypertension] [Obesity] [Diabetes] [CRF] [CHF] [Cancer] PROBLEMS [CAD] [>10] [6-10] MEDICATIONS [1-5] [Other] [State] [Medicare] [Managed] [Commercial] INSURANCE [BCBS] [Other] [Hispanic] RACE [Black] [Sex F] DEMOGRAPHICS [Age]Odds Ratio

PAGE 128

128 Figure 7-3. Poisson regression results for ( non-zero) imaging intensity. Horizontal bars represent 95% confidence intervals on estimated coefficient for each variable/level. Patient variable refe rence levels: Sex-Male, Race-White, Insurance-Uninsured, Medications-None, Problems-No. Provider variable reference levels: Sex-Male, FMG-No, MD _Plus-No, Malpractice-No, Provider Patients-<500. -0.2-0.1-0.10.00.10.10.2CLINIC SIZE [Active Providers ] [1K+] [750-999] PROVIDER PATIENTS [500-759] [Malpractice Yes] [MD Plus Yes] [FMG Yes] [sex F] PROVIDER [Experience ] [inpt_icu_days ] [inpt_los_total ] [inpt_read_31d ] [inpt_stays ] [obs_stays ] HOSPITAL [er_visits ] [all_i_cnt ] [all_e_cnt ] [spc_o_cnt ] OTHER IMAGING [pcp_o_cnt ] [spc_visit_rvu ] [pcp_visit_rvu ] VISITS [prv_visit_rvu ] PROBLEMS [Other (count)] [Trauma] [Substance Abuse] [Hypertension] [Obesity] [Diabetes] [CRF] [CHF] [Cancer] PROBLEMS [CAD] [>10] [6-10] MEDICATIONS [1-5] [Other] [State] [Medicare] [Managed] [Commercial] INSURANCE [BCBS] [Other] [Hispanic] RACE [Black] [Sex F] DEMOGRAPHICS [Age]Beta Estimate

PAGE 129

129 Figure 7-4. Comparison of significant variables for any imaging use and imaging intensity. Unless underlined, all variable/ levels were significant for both any imaging use and imaging intensity. T he underlined variable/levels were not significant for imaging intensity, except for chronic renal failure (CRF), which not significant for any imaging use. The variables indicated by open circles (near the origin) are show n again in Figure 13 with appropriate axis scaling. observation days imaging by covering pcp linked (loyal) doc visits race: black race: hispanic race: other pat female BCBS managed medicare cancer CHF CRF diabetes hypertension substance abuse trauma provider is FMG provider has extra degree pract size=500-799 pract size=750-000 pract size=1K+ provider is female -0.15 -0.10 -0.05 0.00 0.05 0.10 0.15 0.50.60.70.80.91.01.11.21.31.41.5Odds Ratios Any Imaging Estimates Imaging Intensity observation days imaging by covering pcp linked (loyal) doc visits race: black race: hispanic race: other pat female BCBS managed medicare cancer CHF CRF diabetes hypertension substance abuse trauma provider is FMG provider has extra degree pract size=500-799 pract size=750-000 pract size=1K+ provider is female -0.15 -0.10 -0.05 0.00 0.05 0.10 0.15 0.50.60.70.80.91.01.11.21.31.41.5Odds Ratios Any Imaging Estimates Imaging Intensity

PAGE 130

130 Figure 7-5. Comparison of significant variables for any imaging use and imaging intensity (small effect sizes). Unle ss underlined, all variable/levels were significant for both any imaging use and imaging intensity. The summed RVU of visits to covering primary care doctors (underlined) was not significant for any imaging use.

PAGE 131

131 Figure 7-6. Imaging propensity score distributions by provider. Each providers (N=148) mean imaging propensity (diamonds) along with 10th and 90th percentiles (error bars) are plotted on the Y axis against the observed mean number of images per patient for that provider (X axis). The grand mean of imaging propensity (0.73) is indicat ed by the horizontal line. Figure 7-7. Centered imaging propensity score distributions by provider. Each providers (N=148) centered imaging propens ity (diamonds) along with 10th and 90th percentiles (error bars) are plotted on t he Y axis against the observed mean number of images per patient for that pr ovider (X asis). The overall centered mean of imaging propensity (zero) is indicated by the horizontal line. 0.0 0.5 1.0 1.5 2.0 2.5 0.00.51.01.52.0Observed Images / patientIMG_PROP -1.0 -0.5 0.0 0.5 1.0 1.5 0.00.51.01.52.0Observed Images / PatientCentered IMG_PROP

PAGE 132

132 Figure 7-8. Plot of intercept and slopes for all 148 providers obtained from multi-level model of imaging utilizati on. Error bars represent standard error for each providers intercept (horizontal) and slope (vertical). -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 -1.0-0.50.00.51.01.5InterceptSlope A C B D D.D. A .A. B.B. C.C.

PAGE 133

133 Figure 7-9. Imaging utilization versus c entered imaging propensity for a low utilizing doctor (C.C. in Figure 7-8). Each diam ond represents a sing le patient and the dashed line is the linear equation defined by the multi-level model intercept and slope for provider C.C. y = 0.1532x + 0.2903-2 0 2 4 6 8 10 12 14 -2-1012345Centered Imaging PropensityImaging Tests Performed

PAGE 134

134 Figure 7-10. Imaging utilization versus centered imaging propensity a high utilizing doctor (B.B. in Figure 7-8). Each diamond represents a single patient and the dashed line is the linear equation defined by the multi-level model intercept and slope for provider B.B. y = 2.637x + 1.9879-2 0 2 4 6 8 10 12 14 -2-1012345Centered Imaging PropensityImaging Tests Performed

PAGE 135

135 Figure 7-11. Provider means sorted by ascending order within each site. Site (clinic) numbers shown along the bottom. Each providers intercept is scaled by adding the model coefficient for each pr ovider to the fixed intercept (0.7171) and multiplied by 100 to represent adjus ted images per 100 patients over the two year study interval. The solid horizontal line is at 71.7 images per 100 patients which is the gr and mean number of images per 100 patients. Error bars are 95% confidence intervals scal ed up the same way. The individual providers (A.A. D.D.) as discussed in the text are labeled). 0.0 50.0 100.0 150.0 200.0Intercept (images / 100) .. .. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 A.A. B.B. C.C. D.D. 0.0 50.0 100.0 150.0 200.0Intercept (images / 100) .. .. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 A.A. B.B. C.C. D.D.

PAGE 136

136 Figure 7-12. Provider slopes sorted by ascending order within each site. Site (clinic) numbers shown along the bottom. Each pr oviders sloped is scaled by adding the model coefficient for each provider to the fixed imaging propensity effect (0.9919) and multiplied by 100 to represent adjusted images per 100 patients over the two year study interval. Error bars are 95% confidence intervals scaled up the same way. The individual providers (A.A. D.D.) as discussed in the text are labeled). 0.0 50.0 100.0 150.0 200.0 250.0 300.0Slope (images / 100) .. .. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 A.A. B.B. C.C. D.D. 0.0 50.0 100.0 150.0 200.0 250.0 300.0Slope (images / 100) .. .. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 A.A. B.B. C.C. D.D.

PAGE 137

137 CHAPTER 8 DISCUSSION Summary of Key Results Utilization of outpatient diagnostic imaging in a cohor t of 85,277 patients was evaluated over a two year per iod extending through June of 2009. The adult patients in this study were cared for in a stable medi cal home as defined by regular visits to a primary care doctor practicing in one of 15 clinics. The institutional setting is an academic health center located in Boston. In general, the study revealed that older female patients had more imaging as did t hose who had many medical problems listed in the clinical record system. Also, patient s who visited doctors, were admitted to the hospital, or seen in the emergency room more frequently had more imaging. Doctor factors associated with a great er tendency to order imaging tests were less experience, female gender, and having a medium size practice (500-1000 patients). A special statistical technique (hierarchical modeling) that accounts for all patient factors allowed creation of profiles scoring each of the 148 doctors on t heir general tendency to order imaging tests on the average patient and how many more tests were ordered on patients with greater comparative need for diagnostic imaging. On a per patient basis, the amount of im aging ordered by their own (linked / loyal) primary care doctor ranged fr om 0-15 examinati ons, with average of 0.7146 (60,938 / 85,277). This translates into 35.7 images per 100 patient years. For comparison, during the same period, the cohort had 50.8 outpati ent images per 100 pati ent years ordered by specialists caring for them and 37.6 im ages per 100 patient years performed while they were in the emergency department or hospitalized. This study concentrated exclusively on variability in and factors contri buting to the outpatient imaging ordered by

PAGE 138

138 the linked (loyal) primary care doctor, wh ich accounts for about 27.5% of the total imaging received by the entire patient cohort. With the patient experience over 2 years as the unit of analysis and the count of primary care ordered outpatient images as the outcome, fitting a simple Poisson distribution yields mean ( ) of 0.715 (CI: 0.707-0.722). The outcome distribution also fits quite well to a Zero-Inflated Poisson (ZIP) distribution with mean ( ) of 1.492 (CI: 0.3890.411) and zero-inflation parameter ( ) of 0.521 (CI: 0.516-0.526). This (ZIP) distribution was used for model based creation of patient level imaging propens ity expected number of images using available risk adjustment variables. An alternate way to model the same phenomenon is in two stages; any imaging versus none (logistic process) followed by imaging intensity (Poisson process) for patients with at least one imagi ng test. For the logistic pr ocess, the overall success rate (patient had some imaging) was 37.13% (31,660 / 85,277). The Poisson mean ( ) for the non-zero imaging (N=31,660 patients) was 1.925 (CI: 1.910-1.940). This two stage approach was used to test joint effect of 28 patient, 6 provider, and 1 clinic variable(s) on any imaging utilization (logistic regression) followed by imaging intensity (Poisson regression). Patient demographic and clinical factors significantly associated with a greater likelihood of any imaging and greater imaging intensity incl uded: increasing patient age, female sex, Hispanic race (compared with wh ite), and more clinical problems. Patient level clinical activity variables associated with greater likelihood of any imaging and greater imaging intensity included: office visits to the linked (loyal) primary care doctor, office visits to specialists, inpatient im aging tests, outpatient imaging ordered by other

PAGE 139

139 (covering) primary care docto rs, and outpatient imaging ordered by specialists. Provider and clinic level factors associated with a gr eater likelihood of any imaging and greater imaging intensity included: an extra degree (a fter M.D.) held by the patients linked (loyal) primary care doctor, patients linked (loyal) doctor was female, mid-level practice size (500-799 patients), and increa sing size of the clinic (number of practicing doctors). Patient level factors associated with decreased likelihood of any imaging and decreased imaging intensity incl uded: insurance with Medicare (compared to self-pay), diabetes, and hypertension. Patients with longer total length of stay in the hospital actually had lower likelihood of any imaging and lower imaging intensity. As the primary care providers experience increased, both likelihood and intensity of imaging for their linked (loyal) patients decreased. The providers place of M.D. traini ng (foreign medical graduate=FMG) had a discordant effect on imaging utilization. When the doctor had FMG status of yes, they tended to be more likely to order at least some imaging on their patients but the am ount of imaging tests was less than American trained (FMG=no) counterparts. One factor that did not have any significant effect during multivariable modeling of imaging utilization (any imaging or imaging in tensity) was the amount of outpatient prescription medications (in 4 ordinal cat egories) each patient was taking. This is notable because there was a strong bivariat e relationship between this medication variable and the outcome (outpatient imaging ordered by primary care doctor). This will be discussed below. When all patient level factors were combin ed in a single ZIP model, the predictions for each patient were used as an (expected) imaging propensity score for subsequent

PAGE 140

140 multi-level hierarchical m odeling. This imaging risk adjustment model had R-Squared of 0.17 at the patient level. Adding the providers unique (anonymous) identity as a predictor in a two level hierarchical model brought the full model R-Squared up to 0.24. Since the patient level imaging propensity sco res were centered on each providers mean, the provider intercept can be interpreted as the av erage imaging utilization for that doctor (all else equal). The plausible range of these intercepts was 0.151 to 1.283 images per patient over two years with mean of 0.7171. This translates into 35.9 images per 100 patient years with plausible range of 7.6 to 64.2. At the same time, the individual provider slopes from the hierarchical model can be interpreted as the extent to which each doctor responds to a unit increase in patient imaging propensity (which also ranged from 0 to 15 with a single hi gher score of ~20). The mean was 0.9919 and the plausible range was 0.216 to 1.768. Scaling these up to images per 100 patient years gives mean of ~50 and plausible range of 10.8 to 88.4 for every unit increase in patient imaging need. These estimates of provider imaging ut ilization parameters (average/intercept and slope) ar e quite precise with calculat ed reliability of 0.965 and 0.939 respectively. Discussion of Key Results Perhaps the most vexing and interesting quest ion arising from this study has to do with the fact that no more t han 25% of the variation in the number of primary care imaging tests per patient is explainable using a quite robust and complete set of patient, doctor, and clinic variables. This holds true for even the most comp lete models tried. For example, the ZIP model wit h all 28 patient level variabl es that produced the imaging propensity scores for multi-level m odeling can be modified to include all variation due to providers and clinics by placing the unique identity of each patients doctor and clinic

PAGE 141

141 into the model (as class variables). The R-Squared of this fullest model is still just over 0.24 and about 30% of the expl anatory power comes from knowing who the doctors and clinics are. Further, t he empirically determined vari ation in the 148 providers tendency to order imaging on the average patient (intercepts and slopes from multi-level modeling) is quite substantial. This conundr um goes to the very heart of philosophical considerations concerning causes and consequenc es of variations in medical resource utilization. These questions have been posed occasionally in the health services research literature but remain una nswered (Cain and Diehr 1992, Diehr et al. 1990). With the preceding in mind, consider that the practice setting (MGH) is among the most sophisticated and consistent with respect to the processes of outpatient imaging ordering, scheduling, and provider feedback. As mentioned in the Chapter 5, virtually all outpatient imaging was order ed and scheduled via a web-based radiology order entry (ROE) system. During the entire study period, the ROE system had fully functional and complete real-time appropriateness decision support (DS) feedback for all CT, MRI, and nuclear medicine tests. Additionally, the prim ary care doctors included in this study were all given periodic (bi-yearly) feedback about their utilization of outpatient imaging compared with peers. It can be persuasively argued that the practice examined herein is as good as it gets with respect to outpat ient imaging decision support and utilization management. This implies that the amount of variability in primary care outpatient imaging utilization accruing to doctors (~30%) is a lower bound. Therefore, if similar studies were to be conducted elsewhere, the absolute amount of variability between doctors would be greater and the fracti on of total variation also larger.

PAGE 142

142 Limitations of Study In terms of generalizability, the primary care practice being studied is not widely representative of other priv ate, or even other academic settings. As mentioned above, near complete use of electronic order ent ry with imaging-specif ic decision support coupled with active utilization management means that most other practice settings and locales will differ on several axes. The absol ute number (and distribution over modality) of images obtained per patient in other pr imary care settings may be substantially higher or lower. Other pr imary care doctors/groups may actually obtain advanced imaging less often without easy access to elec tronic ordering, scheduling, and clinical decision support. For, example they may more often refer co mplex patients to specialists and defer imaging to them resu lting in lower apparent utilization by the primary care provider(s). On the other hand, without barrier effects of a formal order entry system and decision support, which sometimes recommends against imaging, overall utilization could be much higher. In ei ther event, variability of utilization between patients (after risk adjustment ) and providers will almost ce rtainly be greater in other settings. This would manifest in a lower frac tion of overall explained variation (25% in this setting), less effective risk adjustment at the patient level, and a greater fraction of variation attributable to providers. Thes e observations should not discourage others from using risk adjusted benchmar king of imaging utilization. Qu ite to the contrary, such provider profiles under conditions of greater variation in imaging utilization will have potentially greater impact. The main study outcome, outpatient imaging utilization, was quantified by counting imaging tests. Clearly, not all imaging tests are equal and so me cost more others. For utilization management efforts that seek to understand an d control expenditures,

PAGE 143

143 resource use (cost) of each imaging procedu re is important. The relative value unit (RVU) of each imaging procedure is the obvious c hoice as a proxy for cost and is in fact often used to calculate reimbursement. T hus, the summed RVU of outpatient imaging tests performed on each patient has a potentia l advantage over simple counts if cost were to be the main focus of analysis. For example, an o ld school doctor who ordered chest X-Ray on most patients having respir atory symptoms might seem to have the same level of utilization (by simple counti ng of procedures) as a doctor that ordered chest CT scans much more often. In compari ng relative contribution of primary care and specialist ordered tests to outpatient imagi ng expenditures, summed RVU (as opposed to simple counts) would account for differences in the type of tests (modalities) that get ordered. Also, the summed RVU as outco me approach would allow model based prediction/speculation about pot ential cost savings that could be realized by reducing provider variation and/or curtailing utilization among the high outliers. This study only examined outpatient imaging ordered by each patients linked primary care doctor which accounts for less than half (~40%) of all outpatient imaging. The vast ma jority of the rest is ordered by the (sometimes many) specialists caring for the same patient. However, the relationship between pr imary care and specialist or dered imaging was partly addressed by including the amount specialist ordered imaging as a patient level predictor. The assumption about distribution of errors in modeling counts of imaging was that they were Poisson or ZIP. In nature, true Poisson data generating processes have identical probability of events during time t+1 independent of the cumu lative number of events through time t. It can be argued that in actual patient care, this independence

PAGE 144

144 assumption may not hold because sicker pa tients tend to have more imaging tests in the future. At the same time, patients who have been ill may already had had more imaging tests in the past than otherwise hea lthy peers such that prior imaging is a marker for having been sick. However, the distribution of imaging test counts over patients has an empirical shape that is quite we ll fitted by Poisson or ZIP distributions. At the same time, one of the main reasons for selecting non normal error distribution for count modeling is to avoid deflation of standard errors of es timates and resulting mistakes in hypothesis tests about t hem. Thus, even if the data generating phenomenon is not a perfect nat ural Poisson or ZIP proce ss, these distributions may still be most suitable for error fitting. The multivariable modeling of any imaging use (logistic) and imaging intensity (Poisson on non-zero observations) did not a ccount for the nested st ructure of the data so that the doctor and clinic factors were r epeated over all patients in each respective unit. This may result in biased estimation of the effect size and somewhat lower standard errors for these coefficients. The subsequent hierarchical modeling helps to address this shortcoming. Despite selecti ng one from each group of highly correlated variables, there may be additional problems with multi-colinearity as evidenced by the behavior of the variable measuring patients outpatient medication s (significant at bivariate analysis but not significant when analyzed jointly with other variables). Also, additional work will better characterize the re siduals from the logistic, Poisson, ZIP, and multi-level models used at various stages of the analysis. Alternat e error distribution assumptions for the count of imaging tests per patient might also be worthy of exploration, including Negative Binomial.

PAGE 145

145 There were 10,396 (12.2%) patients that had no recorded visits to the linked (loyal) provider during the two year study period. These were retained in the analytic data set. This raises the concern that the linked (loyal) provider visit RVU variable (which was strongly associated with the number of imaging tests ordered by that doctor) was confounded in some way or that the loyalty attribution methodology was flawed. By definition of the (2008) loyalty cohort, all patients had at least one visit to their linked provider from 2006-2008 and this was confirm ed. All 10,396 patients had at least one visit to the linked provider occurring between January 1, 2006 (start of the loyalty cohort definition period) and July 1, 2007 (start of the study period). The patients with no visits during the study period were distributed acro ss 143 of the providers. That is, only 5 providers saw all their loyal patients at least once during the study. A by provider distribution of the percent patients with no visit during t he study period had mean=11.8, median=11.3, and standard deviation=8.25 in a nearly normal distribution (skew=1.11) of loyal patients who did not visit them during the study period. This is reassuring in that it would seem to reflect act ual practice variation rather than a substantial data integrity problem (e.g., provider identifier mismatch). There were 804 patients excluded from t he analytic data set because they were cared for by one of 26 physicians that had le ss than 100 loyal patients. In general, these were doctors who had mixed practices, work ed part time, or left practice during the study period. For example, an endocrinologist (cardiologist, gastr oenterologist) that worked out of a primary care clinic might have a few patients identified as loyal to them by the Atlas methodology. Alternatively, a doctor with administrative, teaching, or research commitments taking up most of their time might attend in one of the primary

PAGE 146

146 care clinics a few days per month. Of the 804 excluded patients 593 (74%) had no images ordered by the linked (loyal) provider, 764 (95%) had two or less, and 40 (5%) had from 3-7 images. Exclusion of patients (and doctors) in these small and/or mixed practices may reduce generalizabilit y. However, the specific aim of this research was to evaluate imaging utilization by actively practi cing primary care doctors and including this small number of patients and their doctors mi ght have biased the results without adding any additional useful information. The hierarchical modeling was only carried out with two levels (patients and doctors) which discounts the effect of havin g patients nested within doctors that are in turn nested in clinics. Perhaps a single gr and three level (patient, doctor, and clinic) model that incorporates all relevant predictors individually and handles error distribution robustly (e.g. ZIP) might provide great er insight into the phenomenon and produce superior estimates of parameters. This w ould be a very complex undertaking and might well involve a full ZIP specification (prima ry and zero-model) at each of 3 hierarchical levels (patient, doctor, and clinic). Policy Implications Two overarching policy concerns attach to outpatient diagnostic imaging; substantial and rapidly rising costs as we ll as increasing populat ion radiation burden from medical imaging (especially due to CT scans) with attendant risk of cancer induction. In addition, the situation of pr imary care doctors deciding between test (imaging), treat, observe, or refer when faced with their pat ients varying clinical presentations is a classic par adigm of medical decision maki ng. Any attempts to curtail imaging utilization growth in general or to tar get high users for remediation or sanction must be informed by proper modeling of driv ers of variation at patient, provider, and

PAGE 147

147 perhaps higher levels. The empiric findings of this study will help to understand the manner and extent to which risk adjustment and variation analysis methods can help with outpatient imaging utilization management. For example, there is a growing body of literature reflecting a vigorous debate about the re liability, utility, and fairness of provider efficiency profiles promulgated by payers (Adams et al. 2010). Unlike typical observed / expected metrics, the multi-level modeling descr ibed herein, directly produces highly reliable (>95%) and much more meaningful provider level measures about average utilization (intercept) and response to patient clinical need variables (slope). Further, each provider intercept and slope has its own standard error which allows much more meaningful comparison with individual peer s and the overall average utilization. As stated above, the clinical leadership fo r the large group practice serving as the setting for this study has engaged in qui te robust and longstanding utilization management efforts specific to outpatient im aging utilization and appropriateness. Also, the doctors studied were all salaried employees of the group practice and none had any financial incentives (or disincentives) asso ciated with diagnostic imaging. For example, there was almost no growth in the use of CT scans by this practice for 4 years which include the period of study (Sistrom et al. 2009). At the same time double digit rates of growth in CT volumes have occurred in m any other settings in the U.S. Thus, the average yearly number of imagi ng tests ordered by these pr imary care doctors (~36 per 100 patients) would seem to be a lower bound es timate of what occurs nationally. At the same time, there was substantial variati on between doctors in their average use of imaging (intercept from mu lti-level model) and their re sponse to patient propensity (clinical need) for imaging (slopes from multi-level analysis).

PAGE 148

148 There are policy implications from this study that relate to high technology in medical care more generally. Advanced imagi ng is a poster child for a broad array of diagnostic and therapeutic interventions made possible by advances in basic sciences, engineering, and informatics. These devic es and techniques are attractive and compelling to health care providers, patient s, and lay public such that hospitals now actively compete to obtain and advertise extensively about the latest and greatest advances. Further, the regulation of m edical devices by the Food and Drug Administration and other agencies is much le ss stringent than for dr ugs. Specifically, there is little or no requirement that developers and vendors demonstrate clinical effectiveness; only safety and functionality. Comparative effectiveness evaluations to determine appropriateness of various devices and technologies for different clinical purposes will be needed to guide reimburse ment determinations. These have two separate stages: first, whether or not to allow claims for a new device/technology at all, and second, what clinical situations warrant reimbursement on a case by case basis. As with imaging, there are many contextual factors surrounding ut ilization of high technology medical interventions that operat e in concert (or opposition) with clinical need (appropriateness). Contribution to Literature The use of the MGH/Atlas loyalty cohort methodology provides a unique population of patients and docto rs participating in a stable medical home type of primary care practice. As described above, the robust medical and imaging informatics infrastructure at MGH provides an optimal situation for standardizing the appropriateness and intensity of outpatient imaging utilizat ion. Thus the practice under

PAGE 149

149 observation is quite likely at or near optim um with respect to variation between primary care providers with respect to diagnosti c imaging. This same rich informatics environment also means that the available empirical data about patients, doctors, and the clinical activity they engage in (includi ng but not limited to imaging tests) is unparalleled in fidelity and completeness. The es timates of the effect size and direction of many patient and provider level factors on imaging utilization should be useful in themselves. Despite increasing popularity and application of hierarchical techniques to health outcomes, cost analyses, and riskadjustment, this is the first study in which hierarchical modeling has been used to study outpatient imaging utilization in primary care. Preparatory risk-adjustment for imaging propensity at t he patient level using ZIP modeling is also unique. T he combination of these meth ods yields important and interesting insights into how doctors differ in both their general tendency to use imaging on the average patient but how they respond to changing need for imaging in their own panel of patients. Future Research The current data source incl udes individual records for every outpatient visit with CPT code indicating type and intensity of vi sit, ICD-9 codes for t he visit reason(s), patient ID, rendering provider ID, and date of service. T he CPT codes were already used to create summed RVU of visits by provider type for the completed study. Combining these, it should be possible to use the visit data in much more robust ways. For example, the date, prov ider, and patient information common to both visit records and outpatient imaging events can be submitt ed to attribution logic which matches imaging tests to visits. A simple set of rules serves to do th is. Also, it is important to

PAGE 150

150 have at least two contiguous years worth of visits and three years of imaging data on the same population (which is available for this cohort). This visit-based method has been va lidated on a separate set of neurology outpatient visits and associated imaging tests with successful attribution of more than 90% of imaging tests to a visit. The anal ysis becomes a visit-based rather than panelbased and the measure becomes images per visit rather than images per patient year. By modeling visits and grouping by provider, comparison between providers concerning their relative tendency to order imaging c an be performed. Using the same set of patients and primary care doctors over the same time frame would allow comparing visit-based and panel-based imaging utilization profiles to see if the visit-based method gives similar results (e.g., ranks providers in the same order in terms of imaging utilization intensity). If a visit-based method is acceptable, it can be generalized to data sets that are less granular and robust (e.g., Medicare claims). The other advantage of visit-based provider profiles is that they will work with specialists who are much less likely--than primary care doctors--to have a stable panel of patients. Another interesting set of questions arising from the current data source has to do with the relationship between out patient imaging ordered by t he patients linked (loyal) primary care doctor and other doctors, mostly s pecialists. Specifically, in this cohort of patients the majority (about 60%) of all outpat ient imaging was ordered by specialists. As described above, primary care doctors fac ed with clinical uncertainty have a limited set of options: observe, treat, (imaging) test, or refer to specialist. The current study lumps three of the choices; observe, treat or refer into a no imaging category and evaluates factors relating to the single alte rnate choice: order an imaging test. Parallel

PAGE 151

151 and/or simultaneous modeling ( perhaps with multivariate techniques) of the imaging performed by specialists could add an additiona l dimension to our characterization of primary care doctor behavior. It may turn out that some provider s who seem to be conservative with respect to imaging are act ually liberal in terms of referring to specialists and this behavior may costlier ov erall compared with pr oviders who order more of their own imaging tests.

PAGE 152

152 LIST OF REFERENCES Adams, J. L., A. Mehrotra J. W. Thomas, et al. 2010. Physician Cost Profiling Reliability and Risk of Misclassification. New England Journal of Medicine 362:1014. Allan, G. M., and J. Lexch in. 2008. Physician Awareness of Diagnostic and Nondrug Therapeutic Costs: A Systematic Review. International Journal of Technology Assessment in Health Care 24:158. Andersen, R. M. 1995. Revisiting the Behav ioral Model and Access to Medical Care: Does it Matter? Journal of Health and Social Behavior 36:1. Anthony, D. L., M. B. Herndon, P. M. Gallager, et al. 2009. How Much Do Patients Preferences Contribute to Resource Use? Health Affairs (Millwood) 28:864. Atlas S. J., Y. Chang, T. A. Lasko, et al. 2006. Is this My Patient? Development and Validation of a Predictive M odel to Link Patients to Primary Care Providers. Journal of General Internal Medicine 21:973. Atlas, S. J., R. W. Grant, T. G. Ferris, et al. 2009. Patient-Physi cian Connectedness and Quality of Primary Care. Annals of Internal Medicine 150: 325. Axt-Adam, P., J. C. van der Wouden, and E. van der Does. 1993. Influencing Behavior of Physicians Ordering Laboratory Tests: A Literature Study. Medical Care 31:784. Baicker, K., E. S. Fisher, and A. Chandra. 2007. Malpractice Li ability Costs and the Practice of Medicine in the Medicare Program. Health Affairs (Millwood) 26:841 852. Bates, D. W., G. J. Kuperm an, A. Jha, et al. 1997. Does the Computerized Display of Charges affect Inpatient Ancillary Test Utilization? Archives of Internal Medicine 157:2501. Bernardy, M., C. G. Ullrich, J. V. Rawson, et al. 2009. Stra tegies for Managing Imaging Utilization. Journal of the American College of Radiology 6:844. Berrington de Gonzalez, A., M. Mahesh, K. P. Kim, et al. 2009. Projected Cancer Risks from Computed Tomographic Scans Performed in the United States in 2007. Archives of Internal Medicine 169:2071. Bertolini, G., R. DAmico, D. Nardi, et al. 2000. One Model, Several Results: The Paradox of the HosmerLemeshow Goodness-of-Fit Te st for the Logistic Regression Model. Journal of Epidemiology and Biostatistics 5:251.

PAGE 153

153 Bhargavan, M., and J. H. Sunshine. 2005. Ut ilization and Radiology Services in the United States: Levels and Trends in M odalities, Regions, and Populations. Radiology 234:824. Blachar, A., S. Tal, A. Mandel, et al. 2006. Preauthorization of CT and MRI Examinations: Assessment of a Managed Care Preaut horization Program based on the ACR Appropriateness Criteria and the Royal College of Radiology Guideliness. Journal of the American College of Radiology 3:851. Blumberg, M. S. 1987. Inter-area Variat ions in Age-adjusted Health Status. Medical Care 25:340. Bodenheimer, T., and A. Fernandez. 2005. High and Rising Health Care Costs. Part 4: Can Costs be Controlled While Preserving Quality? Annals of Internal Medicine 143:26. Boutwell, R. C. and J. B. Mi tchell. 1993. Diffusion of New Te chnologies in the Treatment of the Medicare Population. Implic ations for Patient Access and Program Expenditures. International Journal of Technology Assessment in Health Care 9:62. Brant-Zawadzki, M. 1994. Cost Cont ainment Drives Utilization Management. Diagnostic Imaging (San Francisco) 16:37. Brenner, D. J., and E. J. Ha ll. 2007. Computerized Tomogr aphyAn Increasing Source of Radiation Exposure. The New England Journal of Medicine 357:2277. Britt, H., A Bhasale, D. A. Miles, et al. 1996. The Sex of the General Practitioner: A Comparison of Characteristics, Pa tients, and Medical Conditions Managed. Managed Care 34:403. Brook, R. H., and K. N. Lohr. 1985. Efficacy, Effectiveness, Variations, and Quality. Boundary-crossing Research. Medical Care 23:710. Bugter-Maessen, A. M., R.A. Winkens, R. P. Grol, et al. 1996. Factors Predicting Differences among General Practitioner in Test Ordering Behaviour and in the Response to Feedback on Test Requests. Family Practice 13:254. Burkhardt, J. H., and J. H. Suns hine. 1996. Utilization of Radiol ogic Services in Different Payment Systems and Patient Populations. Radiology 200:201. Cain, K. C., and P. Diehr. 1992. Testing the Null Hypothesis in Small Area Analysis. Health Services Research 27:267. Casparie, A. F. 1996. The Ambiguous Rela tionship between Practice Variation and Appropriateness of Care: An Agenda for Further Research. Health Policy 35:247 265.

PAGE 154

154 Chang, W. and S. B. McCra cken. 1996. Applying Case Mix Adjustment in Profiling Primary Care Physician Performance. Journal of Health Care Finance 22:1. Chassin, M. R. 1993. Expl aining Geographic Variation. The Enthusiasm Hypothesis. Medical Care 31:YS37YS44. Chassin, M. R., J. Kosecoff, R. E. Park, et al. 1987. Does Inappropriate Use Explain Geographic Variations in the Use of Health Care Services? A Study of Three Procedures. Journal of the Americ an Medical Association 258:2533. Childs, A. W., and E. D. Hunter. 1972. Non-medical Fa ctors Influencing Use of Diagnostic X-ray by Physicians. Medical Care 10:323. Cohen, B. J. 1996. Is Expect ed Utility Theory Normative for Medical Decision Making? Medical Decision Making 16:1. Couchman, G. R., S. N. Forjuoh, M. H. Rajab, et al. 2004. Nonclinical Factors Associated with Primary Care Physi cians Ordering Patterns of Magnetic Resonance Imaging/Computed Tomography for Headache. Academic Radiology 11:735. Couchman, G. R., S. N. Forjuoh, M. D. Re is, et al. 2005. Variation in MRI/CT Utilization among Family Physicians and General Inte rnists in a Multi-specialty Group Practice. Medical Science Monitor 11:MT19MT25. Cummings, K. M., K. B. Frisof M. J. Long, et al. 1982. The Effects of Price Information on Physicians Test-ordering Behavior. Ordering of Diagnostic Tests. Medical Care 20:293. Davis, P., B. Gribben, R. Lay-Y ee, et al. 2002. How Much Variation in Clinical Activity is There Between General Practitioners? A Mu lti-Level Analysis of Decision-Making in Primary Care. Journal of Health Services Research & Policy 7:202. Davis, P., B. Gribben, A. Scot t, et al. 2000a. Do Physician Pr actice Styles Persist Over Time? Continuities in Patterns of Clinical Decision-Making Among General Practitioners. Journal of Health Services Research & Policy 5:200. Davis, P., B. Gribben, A. Scott, et al. 2000b. The "S upply Hypothesis" and Medical Practice Variation in Primary Care: Test ing Economic and Clinical Models of InterPractitioner Variation. Social Science & Medicine 50:407. DeKay, M. L. and D. A. Asch 1998. Is the Defensive Use of Diagnostic Tests Good for Patients, or Bad? Medical Decision Making 18:19. Deyo, R. A. 2002. Cascade Effects of Medical Technology. Annual Review of Public Health 23:23.

PAGE 155

155 Diehr, P., K. Cain, F. Connell, et al. 1990. What is too Much Variation? The Null Hypothesis in Small-Area Analysis. Health Services Research 24:741. Eddy, D. 2007. Reflections on Science, Judgment, and Value in Evidence-Based Decision Making: A Conversation with David Eddy by Sean R. Tunis. Health Affairs (Millwood) 26:w500w515. Eddy, D. M. 1984. Variations in Physician Practice: The Role of Uncertainty. Health Affairs (Millwood) 3:74. Eddy, D. M. 1990. Clinical Decision Making: From Theory to Practice. Guidelines for Policy Statements: The Explicit Approach. Journal of the American Medical Association 263:2239, 2243. Eeckhoudt, L. 1996. Expected Utility Theory--I s it Normative or Simply "Practical"? Medical Decision Making 16:12. Eisenberg, J. M. 1985. Physi cian Utilization. The State of Research about Physicians' Practice Patterns. Medical Care 23:461. Eisenberg, J. M. 1986a. The Physician as Self-Fulfilling Practitioner. in Doctor's Decisions and the Cost of Medical Care Health Administration Press Perspectives, Ann Arbor, MI, pp. 29. Eisenberg, J. M. 1986b. The Physi cian as the Patient's Agent. in Doctor's Decisions and the Cost of Medical Care Health Administration Pre ss Perspectives, Ann Arbor, MI, pp. 57. Eisenberg, J. M., and D. Ni cklin. 1981. Use of Diagnostic Services by Physicians in Community Practice. Medical Care 19:297. Epstein, A. M., and B. J. McNeil. 1985a. Physician Charac teristics and Organizational Factors Influencing Use of Ambulatory Tests. Medical Decision Making 5:401. Epstein, A. M., and B. J. McNeil. 1985b. The Effects of Patient Characteristics on Ambulatory Test Ordering. Social Science & Medicine 21:1071. Epstein, A. M., and B. J. Mc Neil. 1986. Relationship of Beliefs and Behavior in Test Ordering. The American Journal of Medicine 80:865. Epstein, A. M., and B. J. Mc Neil. 1987. Variations in Ambul atory Test Use: What Do They Mean? Medical Clinics of North America 71:705. Fazel, R., H. M. Krumholz, Y. Wang, et al. 2009. Exposure to Low-Dose Ionizing Radiation from Medical Imaging Procedures. The New England Journal of Medicine 361:849.

PAGE 156

156 Fenn, P., A. Gray, and N. Rickman. 2007. Li ability, Insurance and Medical Practice. Journal of Health Economics 26:1057. Fertig, A., M. Roland, H. King, et al. 1993. Understanding Vari ation in Rates of Referral Among General Practitioners: Are Inappropriate Referrals Important and Would Guidelines Help to Reduce Rates? British Medical Journal 307:1467. Fisher, E. S., D. E. Wennberg, T. A. Stukel, et al. 2003a. The Implications of Regional Variations in Medicare Spending. Part 1: The Content, Quality, and Accessibility of Care. Annals of Internal Medicine 138:273. Fisher, E. S., D. E. Wennberg, T. A. Stukel, et al. 2003b. The Implications of Regional Variations in Medicare Spending. Part 2: Health Outcomes and Satisfaction with Care. Annals of Internal Medicine 138:288. Folland, S., and M. Stano. 1989. Sources of Sma ll Area Variations in the Use of Medical Care. Journal of Health Economics 8:85. Folland, S., and M. Stano. 1990. Small Area Variations: A Critical Review of Propositions, Methods, and Evidence. Medical Care Review 47:419. Franks, P., and K. Fiscella. 2002. Effect of Patient Socioeconomic St atus on Physician Profiles for Prevention, Disease Managem ent, and Diagnostic Testing Costs. Medical Care 40:717. Franks, P., C. Mooney, and M. Sorbero. 2000a. Physician Referral Rates: Style Without Much Substance? Medical Care 38:836. Franks, P., G. C. Williams, J. Zwanziger, et al. 2000b. Why do Physicians Vary so Widely in Their Referral Rates? Journal of General Internal Medicine 15:163. Franks, P., J. Zwanzig er, C. Mooney, et al.1999. Variati ons in Primary Care Physician Referral Rates. Health Services Research 34:323. Gazelle, G. S., E. F. Halpern, H. S. Ryan, et al. 2007. Utilization of Diagnostic Medical Imaging: Comparison of Radiologist Refe rral Versus Same-Specialty Referral. Radiology 245:517. Goel V., K. Iron, and J. I. Williams. 1997. Enthusiasm or Unce rtainty: Small Area Variations in the use of Mammography Services in Ontario, Canada. Journal of Epidemiology & Community Health 51:378. Goldstein J. N., C. A. Cama rgo, Jr., A. J. Pelletier, et al. 2006. Headache in United States Emergency Departm ents: Demographics, Work-up and Frequency of Pathological Diagnoses. Cephalalgia 26:684.

PAGE 157

157 Government Accountability Offi ce (U.S.GAO). 2008. Medicare Part B Imaging Services: Rapid S pending Growth and S hift to Phy s ician Office s Indicate Need for CM S to Consider Additional Mana gement Practice s GAO-08-462. Greene, B. R., J. Barlow, and C. Newm an. 1996. Ambulatory Care Groups and the Profiling of Primary Care Ph ysician Resource Use: Examining the Application of Case Mix Adjustments. Journal of Ambulatory Care Management 19:86. Grytten, J., and R. Sorensen. 2003. Practice Variation and Physician-Specific Effects. Journal of Health Economics 22:403. Hartley, R. M., A. M. Epstein, C. M. Harris, et al. 1987. Differences in Ambulatory Test Ordering in England and America. Role Of Doctors' Beliefs and Attitudes. The American Journal of Medicine 82:513. Health Services Research Group. 1992. Sm all-Area Variations: What are They and What do They Mean? Health Services Research Group. Canadian Medical Association Journal 146:467. Hermer, L. D., and H. Brody 2010. Defensive Medicine, Cost Containment, and Reform. Journal of General Internal Medicine Hillman, B. J. 2004. Self-Referral: The Time for Whining is Over. Journal of the American College of Radiology 1:795. Hillman, B. J., C. A. Joseph, M. R. Mabry, et al. 1990. Frequency and Costs of Diagnostic Imaging in Office Practice --A Comparison of Self-Referring and Radiologist-Referring Physicians. The New England Journal of Medicine 323:1604. Hillman, B. J., G. T. Olson, P. E. Griffith, et al. 1992. Ph ysicians' Utilization and Charges for Outpatient Diagnostic Imagi ng in a Medicare Population. Journal of the American Medical Association 268:2050. Hoey, J., J. M. Eisenberg, W. O. Spitzer, et al. 1982. Physician Sensitivity to the Price of Diagnostic Tests: A U.S.-Canadian analysis. Medical Care 20:302. Holtgrave, D. R., F. Lawler, and S. J. Spann. 1991. Physicians' Risk Attitudes, Laboratory Usage, and Referral Decisions: The Case of an Academic Family Practice Center. Medical Decision Making 11:125. Houchens, R., B. Chu, and C. Steiner. 2007. Hierarchical Modeling Using HCUP Data., 2007 edn. Isaacs, D. M., J. Marinac, and C. Sun. 2004. Radiograph Use in Low Back Pain: A United States Emergency Depar tment Database Analysis. The Journal of Emergency Medicine 26:37.

PAGE 158

158 Katz, S. J., L. F. McMahon, and W. G. Manning. 1996. Com paring the Use of Diagnostic Tests in Canadian and US Hospitals. Medical Care 34:117. Kessler, D. P., and M. B. McClellan. 2002. How Liability Law Affects Medical Productivity. Journal of Health Economics 21:931. Kessler, D. P., N. Summerton, and J. R. Graham. 2006. Effects of the Medical Liability System in Australia, the UK, and the USA. Lancet 368:240. Knottnerus, J. A., J. Joosten, and J. Daam s. 1990. Comparing the Quality of Referrals of General Practitioners with High and Average Referral Rates: An Independent Panel Review. British Journal of General Practice 40:178. Kramer, A. A., and J. E. Zimmerman. 2007. Assessing the Calibration of Mortality Benchmarks in Critical Care: The Hosmer-Lemeshow Test Revisited. Crit Care Med 35:2052. Kravitz, R. L., and S. Gr eenfield. 1995. Variations in Resource Utilization Among Medical Specialties an d Systems of Care. Annual Review of Public Health 16:431 445. Landon, B. E., J. Reschovsky, M. Reed, et al. 2001. Personal, Organizational, and Market Level Influences on Physicians' Prac tice Patterns: Results of a National Survey of Primary Care Physicians. Medical Care 39:889. Lasko, T. A., S. J. Atlas, M. J. Barry, et al. 2006. Automated Identification of a Physician's Primary Patients. Journal of the American Medical Informatics Association 13:74. Lawthers, A. G., A. R. Localio, N. M. Laird et al. 1992. Physicians Perceptions of the Risk of Being Sued. Journal of Health Politics, Policy and Law 17:463. Leape, L. L., R. E. Park, D. H. Solomon, et al. 1990. Does Inappropriate Use Explain Small-Area Variations in the Use of Health Care Services? Journal of the American Medical Association 263:669. Levy, G., A. Blachar, L. Goldstein, et al 2006. Nonradiologist Utilization of American College of Radiology Appropria teness Criteria in a Preauthorization Center for MRI Requests: Applicability and Effects. American Journal of Roentgenology 187:855 858. Long, M. J., K. M. Cummings, and K. B. Fris of. 1983. The Role of Perceived Price in Physicians' Demand for Diagnostic Tests. Medical Care 21:243. Luke, D. A. 2004. Multilevel Modeling 07-143 edn. Sage Publications, Thousand Oaks, Calif.

PAGE 159

159 Lysdahl, K. B., and I. Borretzen. 2007. Geographica l Variation in Radiological Services: A Nationwide Survey. BMC.Health Services Research 7:21. Majeed, A., A. B. Bindman, and J. P. Weiner. 2001a. Use of Risk Adjustment in Setting Budgets and Measuring Performance in Primary Care I: How it Works. British Medical Journal 323:604. Majeed, A., A. B. Bindman, and J. P. Weiner. 2001b. Use of Risk Adjustment in Setting Budgets and Measuring Performance in Primary Care II: Advantages, Disadvantages, and Practicalities. British Medical Journal 323:607. Medicare Payment Advisory Commission. 2003. MedPAC: Variation and Inovation in Medicine. Report to congress, June. Medicare Payment Advisory Commissi on. 2005. MedPAC Recommendations on Imaging Services. Testimony Before Congress, March 17 Medicare Payment Advisory Commissi on. 2006. MedPAC Recommendations on Imaging Services. Testimony Before Congress, July 18 Miller, W. L., R. R. McDani el, Jr., B. F. Crabtree, et al. 2001. Practice Jazz: Understanding Variation in Family Pr actices Using Complexity Science. The Journal of Family Practice 50:872. Mitchell, J. M. 2007. The Prevalence of Physician Self-Referral Arrangements After Stark II: Evidence from Advanced Diagnostic Imaging. Health Affairs (Millwood) 26:w415w424. Mold, J. W., and H. F. Stein. 1986. The Cascade Effect in the Clinical Care of Patients. The New England Journal of Medicine 314:512. Mort, E. A., J. N. Edwards, D. W. Emmons, et al. 1996. Physician Response to Patient Insurance Status in Ambulatory Care Clinical Decision-Making. Implications for Quality of Care. Medical Care 34:783. National Council on Radiation Protection and Measurements. 2009. NCRP Report No. 160: Ionizing Radiation Exposure of the Population of the United States. Journal of Radiological Protection 29:465. Neilson, E. G., K. B. Johnson, S. T. Ros enbloom, et al. 2004. The Impact of Peer Management on Test-Ordering Behavior. Annals of Internal Medicine 141:196 204. Nyweide, D. J., W. B. Weeks, D. J. Gottlieb, et al. 2009. Relationship of Primary Care Physicians' Patient Caseload with Measurem ent of Quality and Co st Performance. The Journal of the Amer ican Medical Association 302:2444.

PAGE 160

160 O'Neill, L., and J. Kuder. 2005. Explaining Va riation in Physician Practice Patterns and Their Propensities to Recommend Services. Medical Care Research and Review 62:339. Otero, H.J., S. Ondategui-Parra, E. M. Na thanson, et al. 2006. Utilization Management in Radiology: Basic Concepts and Applications. Journal of the Amer ican College of Radiology 3:351. Parchman, M. L. 1995. Small Area Variati on Analysis: A Tool for Primary Care Research. Family Medicine Journal 27:272. Pekoz, E. A., M. Shwartz, L. I. Lezzoni, et al. 2003. Comparing the Importance of Disease Rate Versus Practice Style Variations in Explaining Differences in Small Area Hospitalization Rates for Two Respiratory Conditions. Statistics in Medicine 22:1775. Pezzin, L. E., P. M. Ke yl, and G. B. Green. 2007. Disparities in the Emergency Department Evaluation of Chest Pain Patients. Academic Emergency Medicine 14:149. Pham, H. H., G. C. Alexander and A. S. O'Malley. 2007. Physician Consideration of Patients' Out-of-Pocket Costs in Making Common Clinical Decisions. Archives of Internal Medicine 167:663. Phelps, C. E., C. Mooney, A .I. Mushlin, et al. 1994. Do ctors Have Styles-and They Matter! Working paper. Phillips, K. A., K. R. Morri son, R. Andersen, et al. 1998. Understanding the Context of Healthcare Utilization: Assessing Environm ental and Provider-Related Variables in the Behavioral Model of Utilization. Health Services Research 33:5716. Quintana, J. M., D. Goldmann, and C. Homer. 1997. Social Disparities in the use of Diagnostic Tests for Childr en with Gastroenteritis. International Journal for Quality in Health Care 9:419. Raudenbush, S. W., and A. S. Bryk. 2002. Hierarchical Linear Models Applications And Data Analysis Methods 2nd ed edn. Sage Publications, Thousand Oaks. Reinhardt, U. E. 1999. The Economist 's Model of Physician Behavior. Journal of the American Medical Association 281:462. Restuccia, J., M. Shwartz, A. Ash, et al. 1996. High Hospital Admission Rates and Inappropriate Care. Health Affairs (Millwood) 15:156. Rosen, M. P., R. B. Davis, and L. G. Le sky 1997. Utilization of Outpatient Diagnostic Imaging. Does the Physician's Gender Play A Role? Journal of General Internal Medicine 12:407.

PAGE 161

161 Rosenthal, D. I., J. B. We ilburg, T. Schultz, et al. 2006. Radiology Order Entry with Decision Support: Initial Clinical Experience. Journal of the American College of Radiology 3:799. Salem-Schatz, S., G. Moore, M. Rucker, et al. 1994. The Ca se for Case-Mix Adjustment in Practice Profiling. W hen Good Apples Look Bad. Journal of the American Medical Association 272:871. Schulman, K. A., J. A. Berlin W. Harless, et al. 1999. T he Effect of Race and Sex on Physicians' Recommendations for Cardiac Catheterization. The New England Journal of Medicine 340:618. Shen, J., R. Andersen, R. Brook, et al. 2004. The Effects of Payment Method on Clinical Decision-Making: Physician Responses To Clinical Scenarios. Medical Care 42:297. Sirovich, B., P. M. Gallagher, D. E. W ennberg, et al. 2008. Discretionary Decision Making by Primary Care Physicians and the Cost of U.S. Health Care. Health Affairs (Millwood) 27:813. Sirovich, B. E., D. J. Gottli eb, H. G. Welch, et al. 2006. R egional Variations in Health Care Intensity and Physician Perceptions of Quality of Care. Annals of Internal Medicine 144:641. Sistrom, C.L. 2009. The Appropriateness of Imaging: A Compr ehensive Conceptual Framework. Radiology 251:637. Sistrom, C.L., P. A. Dang, J. B. Weilburg, et al. 2009. Effect of Computerized Order Entry with Integrated Decision Support on the Growth of Ou tpatient Procedure Volumes: Seven-Year Time Series Analysis. Radiology 251:147. Sloan, F.A., and J. J. S hadle. 2009. Is there Empiri cal Evidence for "Defensive Medicine"? A reassessment. Journal of Health Economics 28:481. Smith-Bindman, R. J. Lipson, R. Marcus, et al. 2009. Radiation Dose Associated with Common Computed Tomography Examinations and the Associated Lifetime Attributable Risk of Cancer. Archives of Internal Medicine 169:2078. Smulowitz, P. B., L. Ngo, and S. K. Epstein. 2009. The Effect of a CT and MRI Preauthorization Program on ED Utilization. The American Journal of Emergency Medicine 27:328. Solomon, D. H., H. Hashimoto, L. Dalt roy, et al. 1998. Tec hniques to Improve Physicians' Use of Diagnostic Tests: A New Conceptual Framework. Journal of the American Medical Association 280:2020.

PAGE 162

162 Sood, R., A. Sood, and A. K. Ghosh. 2007. Non-evidenc e-based Variables Affecting Physicians' Test-ordering Tenden cies: A Systematic Review. The Netherlands Journal of Medicine 65:167. Stahl, F. W. 1969. The Mechanics of Inheritance 2d edn. Prentice-Hall, Englewood Cliffs, N.J. Stano, M. 1991. Further Issues in Small Area Variations Analysis. Journal of Health Politics, Policy and Law 16:573. Stano, M. 1993. Evaluating the Policy Role of the Small Area Variations and Physician Practice Style Hypotheses. Health Policy 24:9. Thrall, J. H. 2005a. Reinventing Radiology in the Digital Age. Pa rt II. New Directions and New Stakeholder Value. Radiology 237:15. Thrall, J. H. 2005b. Reinventing Radiology in the Digital Age. Part III. Facilities, Work Processes, and Job Responsibilities. Radiology 237:790. Thrall, J. H. 2005. Reinventi ng Radiology in the Digital A ge. Part I. The All-digital Department. Radiology 236:382. Tucker, A. M., J. P. Weiner, S. Honigfeld, et al. 1996. Pr ofiling Primary Care Physician Resource Use: Examining the Application of Case Mix Adjustment. Journal of Ambulatory Care Management 19:60. Verrilli, D., and H. G. Welch. 1996. The Impact of Diagnos tic Testing on Therapeutic Interventions. Journal of the Americ an Medical Association 275:1189. Wasiak, R., G. S. Pransky, and S. J. Atla s. 2008. Who's in Charge? Challenges in Evaluating Quality of Primary Care Treatment for Low Back Pain. Journal of Evaluation in Clinical Practice 14:961. Welch, H. G., M.E. Miller, and W. P. Welch. 1994. Physician Profiling. An Analysis of Inpatient Practice Patterns in Florida and Oregon. The New England Journal of Medicine 330:607. Welch, W. P., M.E. Mille r, H.G. Welch, et al. 1993. Geographic Variation in Expenditures for Physicians' Services in the United States. The New England Journal of Medicine 328:621. Wennberg, D. E., J. D. Dickens, Jr., L. Biener et al. 1997. Do Physicians Do What They Say? The Inclination to Test and its A ssociation with Coronar y Angiography Rates. Journal of General Internal Medicine 12:172. Wennberg, J., and A. Gittelsohn. 1973. Small Area Variations in Health Care Delivery. Science 182:1102.

PAGE 163

163 Wennberg, J. E. 1987a. Population Illne ss Rates Do Not Explain Population Hospitalization Rates. A Comment on Ma rk Blumberg's Thesis that Morbidity Adjusters are Needed to Interp ret Small Area Variations. Medical Care 25:354 359. Wennberg, J. E. 1987b. The Par adox of Appropriate Care. Journal of the American Medical Association 258:2568. Wennberg, J. E. 1993. Future Directi ons for Small Area Variations. Medical Care 31:YS75YS80. Wennberg, J. E. 2002. Unwarranted Variations in Healthcare Delivery: Implications for Academic Medical Centres. British Medical Journal 325:961. Wennberg, J. E. 2004a. Practice Variation: Implications fo r Our Health Care System. Managed Care 13:3. Wennberg, J. E. 2004b. Practi ce Variations and Health Care Reform: Connecting the Dots. Health Affairs (Millwood) Suppl Web Exclusives, VAR140VAR144. Wennberg, J. E., B. A. Barnes, and M. Zubk off. 1982. Professional Uncertainty and the Problem of Supplier-Induced Demand. Social Science & Medicine 16:811. Westert, G. P., and P. P. Groenewegen. 1999. Medical Prac tice Variations: Changing the Theoretical Approach. Scandinavian Journal of Public Health 27:173. Whiting, P., M. Toerien, I. de Salis, et al. 2007. A Review Identifies and Classifies Reasons for Ordering Diagnostic Tests. Journal of Clinical Epidemiology 60:981 989. Williams, S. V., J. M. Eisenberg, L. A. Pascale, et al. 1982. Physicians' Perceptions about Unnecessary Diagnostic Testing. Inquiry 19:363. Wu, G. 1996. The Strengths and Limi tations of Expected Utility Theory. Medical Decision Making 16:9.

PAGE 164

164 BIOGRAPHICAL SKETCH After serving in the U.S. Army Signal Corps as a cryptographer from 1973-1977, Chris Sistrom, MD, MPH obtained his undergraduate degree in Computer Science (1980) from the Universi ty of Oregon in Eugene. He attended at Oregon Health Sciences University (MD in 1984) and complet ed radiology residency at the University of Virginia in 1988. He is now Associat e Chairman of Radiology, Chief Information Officer for Radiology, and Associat e Professor at the Universi ty of Florida, College of Medicine. Dr. Sistrom obtained an MPH degr ee in epidemiology and health policy in 2003 from the University of Florida, and is in final stages of a PhD in Health Services Research there. The topic of his dissertatio n is Imaging Utilization in Primary Care. The research goal is to quantify and model various factors that affe ct the intensity and mixture of outpatient imaging performed on primary care pati ents. The resulting models will be useful in practitioner profiling at institutional and r egional levels. The eventual goal is to produce a Map of Imaging al ong the lines of the Dartmouth Atlas of Healthcare and to create risk adjusted popul ation based estimates of optimal and appropriate imaging utilization.