Beyond Health Insurance: Public Policy to Improve Health (Advances in Health Economics and Health Services Research)

Citation preview

ADVANCES IN HEALTH ECONOMICS AND HEALTH SERVICES RESEARCH Series Editors: Michael Grossman and Bjo¨rn Lindgren Volume 15:

Health Policy Research in the States – Edited by J. C. Cantor

Volume 16:

Substance Use: Individual Behavior, Social Interaction, Markets and Politics – Edited by M. Grossman and B. Lindgren

Volume 17:

The Economics of Obesity – Edited by J. H. Cawley and K. Bolin

Volume 18:

Evaluating Hospital Policy and Performance – Edited by J. L. T. Blank and V. G. Valdmanis

ADVANCES IN HEALTH ECONOMICS AND HEALTH SERVICES RESEARCH VOLUME 19

BEYOND HEALTH INSURANCE: PUBLIC POLICY TO IMPROVE HEALTH EDITED BY

LORENS HELMCHEN University of Illinois

ROBERT KAESTNER University of Illinois

ANTHONY LO SASSO University of Illinois

United Kingdom – North America – Japan India – Malaysia – China

JAI Press is an imprint of Emerald Group Publishing Limited Howard House, Wagon Lane, Bingley BD16 1WA, UK First edition 2008 Copyright r 2008 Emerald Group Publishing Limited Reprints and permission service Contact: [email protected] No part of this book may be reproduced, stored in a retrieval system, transmitted in any form or by any means electronic, mechanical, photocopying, recording or otherwise without either the prior written permission of the publisher or a licence permitting restricted copying issued in the UK by The Copyright Licensing Agency and in the USA by The Copyright Clearance Center. No responsibility is accepted for the accuracy of information contained in the text, illustrations or advertisements. The opinions expressed in these chapters are not necessarily those of the Editor or the publisher. British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library ISBN: 978-1-84855-180-0 ISSN: 0731-2199 (Series)

Awarded in recognition of Emerald’s production department’s adherence to quality systems and processes when preparing scholarly journals for print

LIST OF CONTRIBUTORS Rosemary J. Avery

Department of Policy Analysis and Management, Cornell University, Ithaca, NY, USA

John Cawley

Department of Policy Analysis and Management, Cornell University, Ithaca, NY, USA

Gautier Duflos

University of Paris, Boulevard de l-Hopital, Paris – Cedex, France

Kevin Fiscella

Family Medicine, Community and Preventive Medicine and Oncology, University of Rochester School of Medicine and Dentistry, Rochester, NY, USA

Paul Hughes-Cromwick

Altarum Institute, Ann Arbor, MI, USA

Amalia M. Issa

Program in Personalized Medicine and Targeted Therapeutics, University of Houston, Houston, TX, USA

Donald Kenkel

Department of Policy Analysis and Management, Cornell University, Ithaca, NY, USA

Craig Lake

Altarum Institute, Ann Arbor, MI, USA

Frank R. Lichtenberg

Courtney C. Brown Professor of Business, Columbia University, New York, NY, USA, and Victoria University, Melbourne, Australia

Dean R. Lillard

Department of Policy Analysis and Management, Cornell University, Ithaca, NY, USA vii

viii

LIST OF CONTRIBUTORS

Alan Mathios

College of Human Ecology, Cornell University, Ithaca, NY, USA

George Miller

Altarum Institute, Ann Arbor, MI, USA

Sendhil Mullainathan

Department of Economics, Harvard University, Cambridge, MA, USA

James B. Rebitzer

Mannix Professor of Economics, Case Western University, Cleveland, OH, USA

Mari Rege

University of Stavanger, Norway

John A. Rizzo

Department of Economics and Department of Preventive Medicine, Stony Brook University, Stony Brook, NY, USA

Charles Roehrig

Altarum Institute, Ann Arbor, MI, USA

Heather Schofield

Harvard Business School, Boston, MA, USA

Jesse D. Schold

Departments of Medicine and Health Services Research, Management and Policy, University of Florida, Gainesville, FL, USA

Christopher Shepard

Veslefrikkveien 16, Stavanger, Norway

Hua Wang

Department of Policy Analysis and Management, Ithaca, NY, USA

Lorens Helmchen Robert Kaestner Anthony Lo Sasso

INTRODUCTION Much of the debate about health policy in the United States has focused on the availability of health insurance coverage and the relatively large number of individuals who are uninsured. While tackling the problem of the uninsured might improve access to and utilization of health care, it would likely have little effect on the health of the population, as there is only a weak connection between health insurance coverage and health. Expanding health insurance coverage alone significantly is unlikely to improve the health of the population or narrow health disparities within the population, given that many of the major causes of poor health such as smoking, obesity, and physical inactivity are largely unaffected by health insurance. The narrow focus on the uninsured in the health policy debate comes at the expense of other policies that could improve health faster and more significantly for every dollar spent. It is well-known that the United States spends approximately twice as much per capita on health care as most other developed nations, but there is little difference in population health between the United States and other developed nations. This suggests that we are on the ‘‘flat part of the curve’’ of health care spending with respect to health and as a result need to pursue other approaches for improving population health. In light of the imbalance in health policy debate in the United States, in November 2007, the Institute of Government and Public Affairs and College of Medicine at the University of Illinois sponsored a conference entitled Beyond Health Insurance: Public Policy to Improve Health. The purpose of the conference was to make available to the public new research on policies that can significantly improve the health of the US population. The conference focused on four areas: reducing racial and ethnic health disparities, preventing disease and promoting health, developing and regulating pharmaceuticals, and improving consumer information.

REDUCING RACIAL AND ETHNIC DISPARITIES IN HEALTH Kevin Fiscella notes that, to date, progress in eliminating racial disparities has been slow. He calls for a comprehensive approach that goes beyond the ix

x

INTRODUCTION

narrow focus of current policy. Given the association between education and health, he advocates greater investments in early childhood education. In light of its broad geographic and demographic reach and role in preventing or delaying the onset of chronic disease, he also proposes to strengthen the delivery of primary care through the network of Federally Qualified Community Health Centers (FQHCs).

PREVENTING DISEASE AND PROMOTING HEALTH George Miller, Charles Roehrig, Paul Hughes-Cromwick, and Craig Lake measure how much the United States currently spends on prevention and wellness. As resources available for health care are necessarily finite, the optimal allocation of health care expenditures will equalize marginal benefit, measured in improved health, of spending an additional dollar on treatment and spending that dollar on prevention. The authors show that the United States spends about 8.4% of its health care dollars on prevention and suggest that by reallocating dollars toward prevention and away from treatment would improve health further than the current allocation. James Rebitzer, Mari Rege, and Christopher Shepard examine whether the adoption of computerized decision-support systems will assist physicians in coping with information overload. They conclude that such systems can enhance the diffusion of new knowledge among physicians but that, due to spillover effects, incentives to invest in this technology are inadequate. The authors argue that government intervention should facilitate these types of investments.

DEVELOPING AND REGULATING PHARMACEUTICALS The growing importance of pharmaceuticals in preventing and treating disease has been widely noted and three papers in this volume focus on developing and regulating pharmaceuticals. One of the most important and controversial developments in this area is direct-to-consumer (DTC) advertising. While DTC advertising might lead to greater education of consumers regarding treatment options available for chronic conditions, it might also distort behavior and interfere with the physician–patient relationship.

xi

Introduction

Rosemary Avery, Donald Kenkel, Dean R. Lillard, Alan Mathios, and Hua Wang ask how DTC advertising affects racial and ethnic disparities in health. They show that older adults, women, and those of low socioeconomic status are disproportionately exposed to DTC advertising. If DTC advertising educates consumers, especially those of low socio-economic status and racial minorities, about the availability and indications of drugs that these consumers would not learn about otherwise, DTC advertising may help to eliminate health disparities by class and race. The authors conclude that limits to DTC advertising may have the unintended consequence of worsening health disparities. Frank Lichtenberg and Gautier Duflos study the health benefits of prescription drugs using data from Australia. For the period 1995–2003, they find that a 5-year increase in mean drug vintage (time since FDA approval) increases mean age at death by almost 11 months. They conclude that the increase in drug vintage accounts for about 65% of the total increase in mean age at death in Australia. John Cawley and John Rizzo study the consequences of withdrawing a drug form the market, which has increased in recent years due to classaction lawsuits and regulatory intervention in the wake of post-marketing studies that uncovered adverse side effects. Although removing a potentially harmful product is likely to be welfare-improving, if an entire class of drugs is tarnished and competition is reduced by the withdrawal, it does not guarantee that health outcomes will improve. The authors ask whether the reduced competition due to the withdrawal of a drug results in higher prices and market share for remaining drugs in that class or lower sales for all drugs in the class because of negative publicity of the withdrawal. They find that ‘‘bad news’’ of a drug withdrawal spills over to all drugs in that class and reduces use of these drugs.

CONSUMER INFORMATION Heather Schofield and Sendhil Mullainathan study the value of nutritional information. They show that consumers over-generalize the information found in health and nutrient claims and fail to differentiate between informative and uninformative health claims. Firms take advantage of the presumably cursory label inspections by consumers and ‘‘hijack’’ the nutritional information to increase sales. This combination of coarse thinking by consumers and message hijacking by firms can negatively impact health by persuading consumers that unhealthy products are healthy.

xii

INTRODUCTION

Jesse Schold reviews the features and likely consequences of population health due to the rapid proliferation of report cards, which are intended to supply consumers with more information about provider quality. He points out that report cards may decrease access to the sickest patients and makes the providers target the wrong outcomes – those measured by the report card instead of better overall health – and concludes with a set of criteria that should be applied in the evaluation of report cards. Amalia Issa describes the benefits and problems associated with personalized medicine and pharmacogenomics. Benefits include better screening of disease, fewer adverse drug reactions, greater patient compliance, and better health at lower cost. However, the regulatory and financial systems that are integral to health care have not kept pace with the science. The author concludes that for this reason there needs to be a concerted effort among insurers and other thirdparty payers, the government (including the Food and Drug Administration), and providers to design an effective, value-based system of personalized medicine in order to make the promise of genomics a reality.

SUMMARY This volume initiates, we hope, a greater focus in the health policy debate on the improvement of health and a departure from the fixation on health insurance. As was observed by Steven Schroeder in the 2007 Shattuck Lecture last year, health care’s proportional contribution to premature death is only 10%, compared with behavioral patterns (40%), genetic predisposition (30%), and social circumstances (15%) (Schroeder, 2007). A lesson from the research presented in this volume is that improving health is difficult, far more difficult than simply spending tax dollars to expand health insurance. It is our hope that follow-up work carries this mantle forward in new and innovative ways.

REFERENCE Schroeder, S. A. (2007). We can do better – improving the health of the American people. NEJM, 357, 1221–1228.

Lorens Helmchen Robert Kaestner Anthony Lo Sasso Editors

QUANTIFYING NATIONAL SPENDING ON WELLNESS AND PREVENTION George Miller, Charles Roehrig, Paul Hughes-Cromwick and Craig Lake ABSTRACT Purpose: We estimate national health expenditures on prevention using precise definitions, a transparent methodology, and a subdivision of the estimates into components to aid researchers in applying their own concepts of prevention activities. Methodology/Approach: We supplemented the National Health Expenditure Accounts (NHEA) with additional data to identify national spending on primary and secondary prevention for each year from 1996 to 2004 across eight spending categories. Findings: We estimate that NHEA expenditures devoted to prevention grew from $83.2 billion in 1996 to $159.8 billion in 2004, in current dollars. As a share of NHEA, this represents an increase from 7.8 percent in 1996 to 8.6 percent in 2004. This share peaked at 9 percent in 2002 and then declined due to reductions in public health spending as a percent of NHEA between 2002 and 2004. Primary prevention represents about half the expenditures, consisting largely of public health expenditures – the largest prevention element. Beyond Health Insurance: Public Policy to Improve Health Advances in Health Economics and Health Services Research, Volume 19, 1–24 Copyright r 2008 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0731-2199/doi:10.1016/S0731-2199(08)19001-X

1

2

GEORGE MILLER ET AL.

Originality/Value of Paper: Our 2004 estimate that 8.6 percent of NHEA goes to prevention is nearly three times as large as the commonly cited figure of 3 percent, but depends on the definitions used: our estimate falls to 8.1 percent when the research component is excluded, 5.1 percent when consideration is limited to primary prevention plus screening, 4.2 percent for primary prevention alone, and 2.8 percent if we count only public health expenditures. These findings should contribute to a more informed discussion of our nation’s allocation of health care resources to prevention.

BACKGROUND Many believe that the US health care system needs to transition from a culture of reactive treatment of disease to one of proactive health promotion and disease prevention (henceforth ‘‘prevention’’). In order to move effectively in this direction, we first need metrics and methods to determine the current distribution of funds between prevention and treatment, and to promote discussion regarding both the amount that should be spent on prevention and how that amount would best be distributed among various prevention activities. These metrics and methods do not currently exist. The NHEA produced by the National Health Statistics Group (NHSG) of the US Department of Health and Human Services (DHHS) is the official source of national healthcare expenditures. Spending is presented according to the type of service (hospitals, physicians, prescription drugs, etc.) and the source of funds (out of pocket, private insurance, public insurance, etc.). However, except for government public health expenditures, spending associated with prevention is not separately identified.1 Furthermore, many expenditures that address prevention (e.g., those associated with transportation safety or environmental monitoring and cleanup) are not considered to be health expenditures and therefore are not included in the NHEA. In previous research, we created a breakout of the NHEA by medical condition (Roehrig, Miller, & Lake, 2007). We used supplementary data sources – such as the Medical Expenditure Panel Survey (MEPS) produced by the Agency for Healthcare Research and Quality of DHHS – to develop this breakout. As part of that work, we generated a preliminary estimate of the portion of Personal Health Care expenditures within the NHEA that corresponds to prevention. This paper summarizes the results of subsequent research to improve this methodology to develop a more complete and accurate estimate of national health expenditures devoted to prevention.

3

Quantifying National Spending on Wellness and Prevention

Hypothesis The hypothesis that motivated this study asserts that current healthcare expenditures inappropriately emphasize treatment activities over prevention efforts. Fig. 1 illustrates this hypothesis conceptually. The horizontal axis represents the division of national expenditures between the treatment of existing conditions and prevention interventions. The vertical axis represents a measure of the effectiveness of these healthcare expenditures (e.g., qualityadjusted life years gained from the expenditures). The (hypothetical) curve labeled ‘‘Effectiveness from Treatment Expenditures’’ is assumed to increase as treatment expenditures increase, but with diminishing returns. (Diminishing returns are consistent with the incremental allocation of additional funds optimally to the next most cost-effective treatment available.) Similarly, the curve labeled ‘‘Effectiveness for Prevention Expenditures’’ increases (with diminishing returns) as a greater portion of the budget is shifted (optimally) from treatment to support prevention interventions. The ‘‘Total Effectiveness’’ curve equals the sum of the two lower curves.2 This total effectiveness reaches a maximum at the point corresponding to Potential Increase in Effectiveness

Effectiveness (e.g.,QALYs Gained)

Total Effectiveness

Effectiveness from Prevention Expenditures

Effectiveness from Treatment Expenditures Expenditures for Treatment Optimal Expenditures for Treatment Current Expenditures for Treatment Budget for Prevention Plus Treatment

Fig. 1.

The Hypothesis.

for Prevention for Prevention

4

GEORGE MILLER ET AL.

‘‘Optimal Expenditures for Treatment.’’ The diagram is constructed so that ‘‘Current Expenditures for Treatment’’ are significantly greater than this optimal level. Thus, total effectiveness would be increased if the share of expenditures going to prevention were increased. The research presented here is limited to estimating the share of national health expenditures going to prevention, and to providing some details regarding the distribution of such spending across different categories of prevention. While this does not answer the larger question of whether the current share is too small, it should help improve the accuracy and precision of public discussions of what we spend on prevention.

Previous Estimates Most published estimates of US prevention spending suggest that we currently allocate 5 percent or less of our health care expenditures to prevention activities. For example, Satcher (2006) notes that ‘‘populationbased prevention . . . is less than 2 percent of our health budget,’’ and OECD (2005) presents data indicating that US expenditures on ‘‘public health and prevention’’ in 2003 were 3.9 percent of our ‘‘current health expenditure.’’ Gerberding (2007) states that ‘‘America spends approximately 2 trillion dollars on health care, and most of this is for treatment of chronic diseases and their myriad complications; less than 2 percent of this amount is spent on preventing these conditions in the first place.’’ The difficulties with these and other estimates is that they use inconsistent language with varying levels of precision to describe what they are measuring, and they are generally based on research conducted by others whose methods and data are not always available. Consider, for example, the following additional quotations:  ‘‘Approximately 95 percent of the trillion dollars we spend as a nation on health goes to direct medical care services, while just 5 percent is allocated to populationwide approaches to health improvement’’ (McGinnis, Williams-Russo, & Knickman, 2002).  ‘‘Only 3% of health care expenditures go toward disease prevention’’ (Woolf, 2006).  ‘‘In 1988, for every 3 cents that US society spent on prevention, it spent 97 cents for curative treatment’’ (Faust, 2005).  ‘‘The US spends 97 cents on curative treatment for every 3 cents spent on prevention’’ (Johns Hopkins/Healthways, 2006).

Quantifying National Spending on Wellness and Prevention

5

 ‘‘The US spends only an estimated 1 to 3 percent of national health expenditures on preventive health care services and health promotion’’ (Lambrew & Podesta, 2006).  ‘‘Efforts to control the underlying causes of preventable death . . . are severely underfunded, representing less than 5% of total health spending’’ (Shodell, 2006).  ‘‘For many years, approximately 96% of expenditures on health in the United States were for medical care rather than public health’’ (Lubetkin et al., 2003). Each of these quotations cites the same 17-year-old study (Brown, Elixhauser, Corea, Luce, & Sheingold, 1991) that employed 1988 data and that is not readily available except in the executive summary (Brown, Corea, Luce, Elixhauser, & Sheingold, 1992).3 The study defined prevention as ‘‘activities that reduce the incidence, prevalence, and burden of disease and injury and enhance health by improving physical, social, and mental wellbeing,’’ and included expenditures in the three categories of health promotion, health protection, and preventive health services. The study reviewed funding sources that included government programs, voluntary health associations, corporations, foundations, worksite programs, and personal prevention services. Various components of prevention expenditures were reported, some of which are not included in the NHEA. Differences in the numbers quoted above can be attributed to differences in the components included in these percentages, as well as to the uncertainties in some of the estimates in the report. If these estimates are correct, it seems almost self-evident that the balance of spending needs to shift toward more spending on prevention, as illustrated hypothetically in Fig. 1. (If prevention is preferable to treatment and yet receives only 3 percent of the funding, we naturally assume it is underfunded.)

Objectives The goal of this study is to improve our understanding of what is spent on prevention through the consistent use of precise definitions, an open description of the methods used to develop our estimates, and a subdivision of the estimates into components so that individual researchers can apply their own concepts of what classes of activities should be designated as prevention. Our estimates are limited to expenditures included in the NHEA and include results for each year from 1996 to 2004.

6

GEORGE MILLER ET AL.

METHODS Definitions Our characterization of prevention activities employs the standard three categories of primary prevention, secondary prevention, and tertiary prevention. While many researchers use these terms, their definitions are often inconsistent or not precisely stated. We have adopted definitions of these three types of prevention based upon our interpretation of McGinnis (2000)4:  Primary prevention consists of interventions to prevent the occurrence of disease or disability. Examples include a broad range of public health activities (e.g., controlling the spread of infectious disease, promoting healthy behaviors, and ensuring preparedness to respond to various types of health crises) as well as immunizations and health behavior counseling delivered in a clinical setting.  Secondary prevention consists of interventions to detect and arrest disease or disability in its early asymptomatic stages. Early detection is accomplished through screening for selected conditions (eye examinations, mammograms, etc.) at recommended intervals rather than having the onset of symptoms generate the encounter.5 When screening detects potential problems that are presymptomatic, subsequent diagnosis and treatment to arrest development are also included as secondary prevention.  Tertiary prevention involves interventions to prevent the progression of disease or disability in persons with symptomatic illness or injury. As is common practice, we exclude tertiary prevention from our prevention expenditure estimates because most interventions within this category are indistinguishable from what would be considered treatment of medical conditions. For purposes of estimation, we divide interventions and activities into public health, medical, dental, and research categories. Then, within each, we identify the prevention components as discussed below.  Public health activities are generally designed to prevent disease and injuries by promoting healthy lifestyles, supporting healthy environments, and preparing for public health crises, such as disease epidemics and other natural and man-made disasters. These activities are generally population-based, and not aimed directly at individuals. Because we limit our estimates to expenditures included in the NHEA, some population health activities that are not conducted by public health entities (such as water fluoridation, pollution control, and some community-wide activities

7

Quantifying National Spending on Wellness and Prevention

Medical Condition

Clinical Activities

Not Present

Counseling & Immunization Screening Diagnosis & Treatment

Fig. 2.

Asymptomatic

Symptomatic

Advanced or with Complications

Primary Prevention Secondary Prevention

Secondary Prevention

Tertiary Prevention

Treatment

Partitioning Clinical Activities into Primary, Secondary, and Tertiary Prevention.

related to tobacco use, obesity, and exercise) are not included in our estimate of public health expenditures. All such expenditures that we do include are considered to be primary prevention.  Medical activities of patient counseling, immunization, screening for disease, and some forms of treatment are partitioned into primary and secondary prevention as a function of the stage of the medical condition that is being prevented (Fig. 2). Counseling and immunization of patients with no known disease constitutes primary prevention; screening for disease in an asymptomatic patient is secondary prevention; and diagnosis and treatment of disease is secondary prevention, tertiary prevention, or non-preventive treatment, depending on the disease stage.  Dental activities include various prevention components. Cleaning and counseling are characterized as primary prevention, while dental examinations and X-rays represent screening that is counted within secondary prevention. We do not include dental treatment such as fillings, although some such treatment might be considered secondary prevention (as when caries are discovered by X-ray or visual examination in otherwise asymptomatic patients), because of the difficulty in isolating this type of treatment from that done in response to a patient complaint.  Research activities are included if they are related to preventive services. Due to data limitations, we do not attempt to distinguish between primary and secondary prevention research activities.

Approach Categories of prevention expenditures that we have quantified within the NHEA framework are depicted in Fig. 3. Of the five principal components

8

GEORGE MILLER ET AL. National Health Expenditures (NHE)

Components of NHEA:

Structures and Equipment

Personal Health Care

Admin. & Net Cost of Priv.Health Insurance

Public Health Activity

Research

Public Health

Prevention Research

Prevention Elements: Medical Preventive Services

Dental Preventive Services

Counseling

Cleaning/ Counseling

Immunizations

Examinations

Screening Diagnose/Treat (Asymptomatic)

Fig. 3.

Approach to Quantifying Expenditures.

of the NHEA, three include prevention elements – Personal Health Care, Public Health Activity, and Research. (For simplicity, we ignore the small portion of the Administration and Net Cost of Private Health Insurance component that is devoted to prevention programs.6) Our estimates of prevention expenditures encompass the eight elements shown at the bottom of the figure, which are lightly shaded if they constitute primary prevention, moderately shaded if secondary prevention, and darkly shaded if they represent a combination of primary and secondary prevention. Medical Expenditures for Counseling, Immunizations, and Screening Within the Personal Health Care component, our previous study to estimate expenditures by medical condition (Roehrig et al., 2007) used supplementary data such as MEPS to separate prevention expenditures from those for diagnosis and treatment. In the current effort, we improved these estimates using additional data sources such as the National Ambulatory Medical Care Survey and the National Hospital Ambulatory Medical Care Survey (NAMCS/NHAMCS). This was a ‘‘bottom-up’’ analysis involving identification of specific clinical preventive services along with expenditure

Quantifying National Spending on Wellness and Prevention

9

estimates from various sources for those services anticipated to have significant expenditures.7 Limitations of available data required us to make a number of assumptions and approximations in generating expenditure estimates for some of these individual services. For example, the frequency with which some services were provided was not always available for each year in our 9year time series, requiring us to use interpolation and extrapolation to fill in estimates for the missing years.8 For expenditures associated with counseling and related patient examinations, we relied heavily on data assembled by Mehrotra, Zaslavsky, and Avanian (2007), who estimated annual costs of preventive health examinations and preventive gynecological examinations (PHE/PGE) using NAMCS/ NHAMCS data and Medicare reimbursement rates. To limit these costs to those involving the primary prevention portions of these examinations, we used information provided by Maciosek (2007) to estimate the fraction of expenditures for such visits that constitute primary prevention, and we excluded all laboratory costs from the estimate. (The costs excluded here were included in screening costs, as described below.) Because Medicare reimbursement rates are generally less than private reimbursement rates, we increased the costs for the estimated fraction of these examinations that were not provided to Medicare beneficiaries, using relative cost factors published by MedPAC (2007). To develop a time series of these estimates for the years 1996–2004, we repeated the Mehrotra analysis for years not included in their analysis. Because Mehrotra’s analysis was limited to visits for adults, we conducted a similar analysis using MEPS data for well-child visits. Our estimate of expenditures associated with immunizations included all immunizations recommended by the Advisory Committee on Immunization Practices for the years 1996–2004, as reported by the US Preventive Services Task Force (2007). These immunizations and the data sources used to estimate their expenditures are listed in Table 1. The table lists immunizations in order of decreasing estimated expenditures for 2004. Immunization coverage rates for childhood immunizations were obtained from the National Immunization Survey, from which we inferred annual numbers of immunizations; we conducted a literature review to identify coverage rates for adult immunizations. Immunization costs include both the cost of the vaccine and an administration fee. Our list of candidate screening services (Table 2) was extracted from the list of services evaluated by the US Preventive Services Task Force (2007), supplemented with additional services that were evaluated by the Partnership for Prevention (2007). To the extent allowed by available data, we

10

GEORGE MILLER ET AL.

Data Sources for Immunization Expenditure Estimates. Immunization Type

Data Sources Cost per series

Influenza (adults)

Maciosek (2007)

Pneumococcal (children) IPV DTaP Hib Hepatitis B TD (adults)

Maciosek (2007)

MMR Varicella Influenza (children) TD (children)

Maciosek (2007) Maciosek (2007) Maciosek (2007) Maciosek (2007) Vaccine cost: 2007 CDC vaccine price list; administration fee: Maciosek (2007) Maciosek (2007) Maciosek (2007) Maciosek (2007)

Vaccine cost: 2007 CDC vaccine price list; administration fee: Maciosek (2007) Hepatitis A Vaccine cost: 2007 CDC vaccine price list; administration fee: Maciosek (2007) Pneumococcal Sisk, Whang, Butler, Sneller, and Whitney (2003) (adults)

Annual frequency National Center for Health Statistics (2006) % Vaccinated: NIS % % % % %

Vaccinated: Vaccinated: Vaccinated: Vaccinated: Vaccinated:

NIS NIS NIS NIS NHIS

% Vaccinated: NIS % Vaccinated: NIS % Vaccinated: NIS % Vaccinated: Jain and Stokley (2007) % Vaccinated: NIS NAMCS/NHAMCS

Note: Populations applied to % vaccinated are from the US Census. Prices were converted to dollars for each year using the Consumer Price Index for medical care services. Abbreviations: NIS, National Immunization Survey; NHIS, National Health Interview Survey; NAMCS/NHAMCS, National Ambulatory Medical Care Survey/National Hospital Ambulatory Medical Care Survey.

excluded from these services any that were associated with diagnosis of symptomatic disease (such as diagnostic mammography). Where information was not readily available to estimate accurately the annual frequency of an individual screening service, but was sufficient for us to infer that national expenditures on the service are small, we omitted that service from our estimates. Our rough estimates suggest that the aggregate impact of these omissions is less than 0.01 percent of the NHEA. Table 3 indicates the data sources used to estimate the expenditures for screening services that are included in our analysis. The table lists services in order of decreasing estimated expenditures for 2004. In addition to the laboratory and procedure costs associated with these services,

Quantifying National Spending on Wellness and Prevention

Candidate Screening Services. Included in expenditure estimates Colorectal cancer screening Breast cancer screening CHD screening Cervical cancer screening Cholesterol screening Prostate cancer screening Gonorrhea screening Vision screening (children and adults) HIV screening Chlamydia screening Osteoporosis screening Glaucoma screening Hepatitis B screening Syphilis screening Diabetes screening Hearing screening Depression screening Genetic risk assessment for breast/ovarian cancer Newborn hearing screening Oral cancer screening Not included in expenditure estimates Aortic aneurysm screening Bacterial vaginosis in pregnancy screening Bacteriura screening Bladder cancer screening Dementia screening Developmental dysplasia of the hip screening Elevated blood lead level screening Family and intimate partner violence screening Genital Herpes screening Gestational diabetes screening Hemochromatosis screening Hepatitis C screening Idiopathic scoliosis screening Iron deficiency anemia screening Lung cancer screening Ovarian cancer screening Pancreatic cancer screening Peripheral arterial disease screening Rh(D) incompatibility screening Skin cancer screening Speech and language screening (preschool) Testicular cancer screening Thyroid disease screening

11

Screening Type

Data Sources Cost per screen

Colorectal cancer screening Breast cancer screening CHD screening Cervical cancer screening Cholesterol screening Prostate cancer screening Gonorrhea screening Vision screening (Adults) HIV screening Chlamydia screening

Maciosek, Solberg, Coffield, Edwards, and Goodman (2006) Medicare reimbursement ratesb Medicare reimbursement ratesb Medicare reimbursement ratesb Medicare reimbursement ratesb Medicare reimbursement ratesb Medicare reimbursement ratesb Medicare reimbursement ratesb Medicare reimbursement ratesb Medicare reimbursement ratesb

Annual frequency NAMCS/NHAMCS

NAMCS/NHAMCSa NAMCS/NHAMCSa NAMCS/NHAMCSa NAMCS/NHAMCSa NAMCS/NHAMCSa % Screened: St. Lawrence et al. (2002) Target population: USPSTF (2007) NAMCS/NHAMCS % Screened: St. Lawrence et al. (2002) Target population: USPSTF (2007) % Screened: St. Lawrence et al. (2002) and Partnership for Prevention (2007); Target population: USPSTF (2007) Medicare reimbursement ratesb NAMCS/NHAMCS Medicare reimbursement ratesb NAMCS/NHAMCS Medicare reimbursement ratesb % of annual birth cohort screened: Burd et al. (1994) Medicare reimbursement ratesb % Screened: St. Lawrence et al. (2002) Target population: USPSTF (2007) Hoerger et al. (2004) NAMCS/NHAMCS Medicare reimbursement ratesb NAMCS/NHAMCS Assumed to be captured as part of PHE/PGE costs Assumed to be captured as part of PHE/PGE costs Assumed to be captured as part of PHE/PGE costs Assumed to be captured as part of PHE/PGE costs Assumed to be captured as part of PHE/PGE costs

Note: Where necessary, prices were converted to dollars for each year using the Consumer Price Index for medical care services. Abbreviations: NAMCS/NHAMCS, National Ambulatory Medical Care Survey/National Hospital Ambulatory Medical Care Survey; USPSTF, US Preventive Services Task Force. a For those not included in PHE/PGE. b Adjusted for ratio of Medicare to private reimbursement rates for estimated fraction of screens not covered by Medicare.

GEORGE MILLER ET AL.

Osteoporosis screening Glaucoma screening Hepatis B screening Syphilis screening Diabetes screening Hearing screening (adults) Depression screening Genetic risk assessment for breast/ovarian cancer Newborn hearing screening Oral cancer screening Vision screening (children)

12

Data Sources for Screening Expenditure Estimates.

Quantifying National Spending on Wellness and Prevention

13

we separately estimated the portions of the PHE/PGE and well-child visit costs associated with ordering the services. This was the complement of the portion of these visit costs captured as counseling and examination, as described above. As noted by Mehrotra et al. (2007), many screening services result from non-preventive physician visits. While the costs of such visits do not constitute primary or secondary prevention expenditures, the costs of the resulting screening activities do represent secondary prevention expenditures, and are included in our estimates. To help ensure that we did not count the costs of diagnostic tests that were conducted in response to symptoms or as part of monitoring existing disease (which constitute diagnosis and treatment rather than screening costs), we included only those screening services that were ordered by a primary care physician. Because Medicare reimbursement rates are generally less than private reimbursement rates, we increased those screening costs that were based on Medicare rates for the estimated fraction of these services that were not provided to Medicare beneficiaries, using relative cost factors published by MedPAC (2007). Other Expenditures Medical diagnosis and treatment services provided to asymptomatic patients to arrest potential problems identified during routine screening are included in our definition of secondary prevention. Although this concept seems straightforward, we found that applying this definition involved substantial subjective judgment. For example, we did not believe that a mastectomy following a routine mammogram that reveals early stage cancer in an asymptomatic patient should be included as secondary prevention. After careful consideration, we arrived at two treatments to be included as secondary prevention in this analysis: hypertension and hyperlipidemia. These two conditions tend to be asymptomatic and are treated in order to avoid the onset of other cardiovascular diseases. We included treatment expenditures on these two conditions in our secondary prevention category except when the patients being treated were symptomatic with other cardiovascular diseases (in such cases, hypertension and hyperlipidemia treatments are more properly viewed as a part of the overall treatment of symptomatic patients). Development of these expenditure estimates began with the estimate of total expenditures for these two conditions from our earlier study (Roehrig et al., 2007). We then used MEPS data to identify the percent of those expenditures devoted to patients who are otherwise free of cardiovascular disease.

14

GEORGE MILLER ET AL.

While there are undoubtedly other treatments that belong in this category, we believe that the two we have included account for most of the expenditures. Future research in this area should include a systematic review of each screening procedure to determine the treatments that normally follow a positive result. Each treatment would then be judged for possible inclusion as secondary prevention. Such a review was beyond the scope of this study. Dental services are a separately reported component of Personal Health Care within the NHEA. We used MEPS data and dental billing data to estimate the portions of dental services expenditures that are devoted to cleaning and counseling patients (primary prevention) and to examinations and X-rays (secondary prevention). In 2004, these portions represented 17 percent and 14 percent, respectively, of NHEA dental expenditures. Public health and prevention research expenditure estimates were based on a top-down approach, in which we identified the fraction of each category that constituted primary or secondary prevention. Based on discussions with public health experts and a review of the documentation of the methods used to estimate public health expenditures within the NHEA (Sensenig, 2007), we concluded that essentially all such expenditures could be considered as primary prevention.9 To estimate the fraction of NHEA research expenditures that are devoted to prevention, we used estimates by the National Institutes of Health (NIH, 2007) of the fraction of the NIH research budget that supports research in prevention. (The NIH budget constitutes roughly 73 percent of total NHEA research expenditures.) In 2004, this amounted to 26 percent of the budget, and we applied this percentage to the annual NHEA research expenditures to provide the research portion of our estimate. While the 26 percent estimate may be somewhat high, it constitutes a relatively small portion of our overall estimated expenditures.10

RESULTS We used the methods summarized above to estimate annual national spending on prevention for each year from 1996 to 2004. Table 4 tabulates these results by year, for each of the eight prevention elements shown in Fig. 3, with expenditures expressed in current year dollars. The table also displays the average annual growth rate for each of the elements. With a few minor exceptions in individual years, each group of expenditures shows steady growth at a rate that varies from 6.2 percent for public health expenditures to 12.3 percent for diagnosis and treatment of hypertension and hyperlipidemia.

15

Quantifying National Spending on Wellness and Prevention

Estimates of Annual Expenditures by Prevention Element (Billions of Dollars). Prevention Activities

Expenditures by Year (Billions of Dollars) 1996 1997 1998 1999

2000

2001

2002

2003

2004

40.7

43.4

46.8

52.4

52.8

52.5

6.2%

3.9 4.1 9.9 23.9

4.4 4.3 10.9 27.9

4.7 4.9 11.9 29.8

5.5 5.7 14.1 35.6

5.6 6.1 14.7 39.5

6.0 6.6 15.8 43.7

7.7% 8.7% 8.2% 12.3%

9.4 7.8 6.0

9.9 8.3 6.6

11.0 9.2 7.4

12.3 10.3 8.4

13.1 11.0 9.2

13.8 11.6 9.9

8.0% 8.0% 10.1%

83.2 91.4 98.0 105.7 115.8 125.6 144.3 152.0 159.8

8.5%

Public health 32.5 34.9 37.6 Medical Preventive Services Counseling 3.3 3.6 4.0 Immunizations 3.4 3.6 3.8 Screening 8.4 9.2 10.0 Diagnosis/treatment 17.3 20.2 21.1 (asymptomatic) Dental Preventive Services Dental cleaning/counseling 7.5 8.2 8.7 Dental examinations 6.2 6.8 7.3 Research 4.6 5.0 5.5 Total

Annual Growth

Discrepancies in totals and annual growth rates are due to rounding.

The overall annual growth rate of 8.5 percent is slightly greater than the average annual growth rate for all national health expenditures of 7.2 percent. Roughly 2 percentage points of these growth rates can be associated with economy-wide price inflation, while population growth averaged 1 percent. The remaining 5.5 percent of growth in prevention expenditures are attributable to changes in patterns of preventive activities, prevalence of asymptomatic disease, and cost per preventive service delivered. Although not shown in the table, we estimated expenditures for specific services within the categories of immunizations and screening. Within the category of immunizations, the highest expenditures in 2004 were for influenza vaccinations for adults, estimated to be $2.0 billion (including vaccine costs and administration fees). Most of the remaining immunization expenditures were for the standard set of childhood immunizations, which accounted for an additional $4.0 billion in 2004. Among screening services, the highest expenditures in 2004 were for three types of colorectal cancer screening (colonoscopies, flexible sigmoidoscopies, and fecal occult blood tests), which we estimated to be $1.4 billion, with breast cancer screening second at $1.3 billion. Table 5 tabulates the same results as Table 4, expressed as a percentage of the NHEA.11 Following gradual but steady growth from 1996 to 2002, there is a slight decline in the overall share in 2003 and 2004. This is primarily due

16

GEORGE MILLER ET AL.

Estimates of Annual Expenditures by Prevention Element (Percent of NHEA). Prevention Activities

Expenditures by Year (Percent of NHEA) 1996 1997 1998 1999 2000 2001 2002 2003 2004

Public health Medical Preventive Services Counseling Immunizations Screening Diagnosis/treatment (asymptomatic) Dental Preventive Services Dental cleaning/counseling Dental examinations Research Total

3.0

3.1

3.2

3.2

3.2

3.2

3.3

3.0

2.8

0.3 0.3 0.8 1.6

0.3 0.3 0.8 1.8

0.3 0.3 0.8 1.8

0.3 0.3 0.8 1.9

0.3 0.3 0.8 2.1

0.3 0.3 0.8 2.0

0.3 0.4 0.9 2.2

0.3 0.4 0.9 2.3

0.3 0.4 0.9 2.3

0.7 0.6 0.4

0.7 0.6 0.4

0.7 0.6 0.5

0.7 0.6 0.5

0.7 0.6 0.5

0.7 0.6 0.5

0.8 0.6 0.5

0.8 0.6 0.5

0.7 0.6 0.5

7.8

8.1

8.2

8.4

8.6

8.5

9.0

8.8

8.6

to the decline in public health expenditures as a percentage of the NHEA during these same 2 years. The share of the NHEA for most other elements remains relatively constant over the 9-year interval, although the percentage associated with diagnosis and treatment of asymptomatic disease has grown significantly. This growth coincides with increased capabilities and more aggressive guidelines for the treatment of hypertension and hyperlipidemia. A more aggregate view of these results is shown graphically in Fig. 4 (in billions of dollars) and Fig. 5 (as a percentage of the NHEA). The charts show annual expenditures by class of preventive activity (primary prevention, secondary preventive activities that include medical screening and dental examinations, secondary preventive activities that involve diagnosis and treatment of asymptomatic disease, and research on either primary or secondary prevention). Primary prevention is the largest contributor to the total, representing slightly more than half of the expenditures in every year except 2004, where it drops to slightly less than half. This is driven by public health expenditures, which represent the largest of our eight individual components of prevention. The diagnostic and treatment component of secondary prevention is the second largest contributor and, as noted earlier, grows rapidly in both dollars and as a percentage of the NHEA over the 9year period. Expenditures for screening activities (including dental examinations) have remained relatively constant as a percentage of the NHEA. Fig. 5 again illustrates the decline in total expenditures on prevention as a percentage of the NHEA in the two most recent years.

17

Quantifying National Spending on Wellness and Prevention $180 $160

Primary and Secondary Prevention (Research) Secondary Prevention (Treatment)

Expenditures (Billions of Dollars)

Secondary Prevention (Screening/Examination) $140

Primary Prevention

$120 $100 $80 $60 $40 $20 $0 1996

Fig. 4.

12% 11% 10%

Percent of NHEA

9%

1997

1998

1999

2000

2001

2002

2003

2004

Estimates of Annual Expenditures by Class of Activity (Billions of Dollars).

Primary and Secondary Prevention (Research) Secondary Prevention (Treatment) Secondary Prevention (Screening/Examination) Primary Prevention

8% 7% 6% 5% 4% 3% 2% 1% 0% 1996

Fig. 5.

1997

1998

1999

2000

2001

2002

2003

2004

Estimates of Annual Expenditures by Class of Activity (Percent of NHEA).

18

GEORGE MILLER ET AL.

DISCUSSION Prevention as a Share of National Health Expenditures It is interesting to compare the estimates developed in this study with the conventional wisdom regarding the share of national health expenditures going to prevention. To assist with this comparison, we have created five ‘‘concepts’’ of prevention, defined as follows: (1) all prevention components identified in this study, (2) all prevention except the research component, (3) primary prevention plus screening, (4) primary prevention, and (5) public health. Table 6 and Fig. 6 present annual estimates of expenditures and shares of NHEA expenditures for each of these concepts. Fig. 7 provides a comparison of the share of NHEA expenditures represented by each concept in 2004. Expenditures for Alternative Definitions of Prevention (Billions of Dollars). Prevention Activities Included in Expenditures

Total Expenditures by Year (Billions of Dollars) 1996 1997 1998 1999 2000 2001 2002 2003 2004

All activities All activities except research Primary prevention and screening Primary prevention Public health

Fig. 6.

83.2 78.6 55.1 46.6 32.5

91.4 86.4 59.4 50.3 34.9

98.0 105.7 115.8 125.6 144.3 152.0 159.8 92.4 99.6 109.2 118.2 135.9 142.8 149.9 64.1 67.9 73.0 79.2 90.0 92.4 94.7 54.1 58.1 62.0 67.3 75.9 77.6 78.9 37.6 40.7 43.4 46.8 52.4 52.8 52.5

Expenditures for Alternative Definitions of Prevention (Percent of NHEA).

Quantifying National Spending on Wellness and Prevention

19

10.0%

Percent of NHEA

8.0%

6.0%

4.0%

2.0%

0.0% All Activities

Fig. 7.

All Activities Except Research

Primary Prevention and Screening

Primary Prevention

Public Health

2004 Expenditures for Alternative Definitions of Prevention (Percent of NHEA).

When all components are included, the prevention share of NHEA expenditures in 2004 is 8.6 percent, which is nearly three times as large as the 3 percent figure most often quoted. This share falls to 8.1 percent when the research component is excluded. When consideration is limited to primary prevention plus screening, the share falls to 5.1 percent. Primary prevention alone claims a 4.2 percent share, while focusing on public health alone gives a 2.8 percent share. From this vantage point, the commonly cited figure of 3 percent as the prevention share of NHEA expenditures applies more accurately to public health activities alone. Public Health Spending Our estimates of public health spending are taken directly from NHEA. Their method of estimation is well documented, and the NHSG is clear about the limitations of currently available data to support such estimates (Sensenig, 2007). While there are some issues with the underlying data, NHEA estimates do provide a time series of expenditures based upon a consistent method over time, and therefore should produce reliable trends. As shown in Fig. 6, public health spending as a share of NHEA grew steadily from 3.0 percent in 1996 to 3.3 percent in 2002, when it then dropped to 3.0 percent in 2003 and again to 2.8 percent in 2004.12 These declines have occurred in years since September

20

GEORGE MILLER ET AL.

11, 2001, during which public health spending for preparedness has grown, which implies that the non-preparedness component has fallen even more rapidly. It seems imperative that we learn more about public health spending, and what is being sacrificed when spending increases are held below those of other components of the NHEA. Limitations Limitations of our methods could cause these estimates to overstate or understate actual spending. Examples of inaccuracies contributing to possible overstatement are our inclusion of all public health expenditures from the NHEA as primary prevention, and our estimate that 26 percent of NHEA expenditures on research are directed at prevention. Inaccuracies that can lead to understatement include our omission of some types of screening and of treatment of some diseases (such as osteoporosis) that might be viewed as secondary prevention. Better data are needed to refine estimates of public health spending and its various components, some of which may not turn out to be suitable for categorization as primary prevention. At the very least, it would be helpful to know the share going to preparedness in recent years. While we believe our estimates of spending on immunizations and screening are reasonable, we are reluctant to include line item detail because of approximations required, particularly for some of the less common services. We find it conceptually difficult to implement our definition of secondary prevention because it includes subsequent treatment for problems identified in asymptomatic patients found through routine screening. A careful examination of the kinds of treatments that follow positive results from each of the screens included in our study, along with an analysis of why they should or should not be counted as secondary prevention, is needed. Our estimates are deliberately restricted to prevention-related expenditures that are included in the NHEA. We therefore exclude expenditures for many activities that address prevention of disease or injury but are not commonly thought of as health expenditures. Concluding Remarks Our objective in this research was to update and clarify estimates of spending on prevention using explicit definitions, transparent methods, and a level of detail that would allow others to include or exclude components

Quantifying National Spending on Wellness and Prevention

21

that are consistent with a particular policy discussion. We hope our findings contribute to a more informed discussion of our nation’s allocation of health care resources to prevention. Our quantification of expenditures does not address whether the money has been well spent. Research has shown that there is wide variation in costeffectiveness across different preventive services. For example, vaccination of individuals over the age of 65 years for pneumonia has been shown (Tufts University, 2007) to be highly cost-effective ($2,800 per qualityadjusted life year saved), while vaccination of 2–4 year olds is much less cost-effective ($210,000 per quality-adjusted life year). One useful extension of our research would be to examine current clinical preventive expenditure patterns in terms of estimated levels of cost-effectiveness. This could be summarized as a weighted average level of cost-effectiveness across all clinical preventive services. We believe that a greater understanding of the incentives that drive current patterns of spending across the varying cost-effectiveness levels is crucial and could lead to innovative approaches to increasing the overall cost-effectiveness of the preventive services mix. A recent study has concluded that the variation in cost-effectiveness across different preventive interventions is similar to that found for different treatment interventions (Cohen, Neumann, & Weinstein, 2008). The authors conclude that across the many interventions studied, ‘‘opportunities for efficient investment in health care programs are roughly equal for prevention and treatment.’’ Though the interventions studied are not necessarily representative of all possible treatments and preventive measures, such findings suggest that opportunities for improving the overall cost-effectiveness of health expenditures through reallocation of resources exist both within and between the broad categories of prevention and treatment. While much research is needed before we have a comprehensive understanding of the optimal allocation of expenditures to prevention, future research should include summarizing what is currently known about the cost-effectiveness of various components of prevention expenditures and examining how reimbursement policies and other incentives impact the quantity and nature of services provided.

NOTES 1. The Organisation for Economic Co-operation and Development (OECD) produces estimates of expenditures on public health and prevention by its member

22

GEORGE MILLER ET AL.

countries, including the US (OECD, 2005). However, these estimates consist only of preventive services provided as parts of organized public and private programs (primarily public health programs), and do not include expenditures for preventive initiatives of individual providers or patients (Orosz, 2005). 2. We recognize that this notional representation glosses over many complexities involving interactions and trade-offs among candidate prevention and treatment interventions. 3. After extensive unsuccessful efforts to obtain a copy of the full study, including conversations with the authors, an anonymous referee located it, in paper copy only, in the National Library of Medicine. 4. These definitions are also consistent with those provided by Turnock (2004). 5. Our use of the term ‘‘screening’’ is consistent with that of the US Preventive Services Task Force (2007), encompassing a wide range of clinical tests conducted on asymptomatic individuals (whether or not in targeted, high-risk populations). Screening is not intended to be diagnostic (though some tests can be used either for screening or diagnosis), and positive or suspicious results are referred to a physician for diagnosis and possible treatment. 6. PricewaterhouseCoopers (2006) estimates that, of the roughly 14 percent of health insurance premiums devoted to the net cost of private insurance, 5 percentage points are allocated to health promotion, wellness, and prevention programs plus all of the following activities: marketing and sales, communications with consumers regarding benefits, disease management programs, care coordination, and investments in health information technology. 7. Details of these computations beyond the summary provided here can be obtained from the authors. 8. The interpolation or extrapolation was adjusted as needed to account for the evolution of the availability and recommendations for use of the services over time. For example, our estimate of expenditures for a recently-introduced service is zero for years, before the service became available. 9. Although some states place Medicaid within their public health departments, NHEA public health spending does not include any Medicaid spending. 10. For example, Toth (2007) indicates that just 14 percent of the Robert Wood Johnson Foundation’s $400 million research budget in 2006 was devoted to prevention. Note also that privately-funded research (such as that conducted by pharmaceutical companies, which is included in the NHEA within prescription drug expenditures) is not included in our estimates. 11. NHEA values used in this study are those published in January 2007. 12. The latest NHEA data show that the decline in the public health share has continued till 2006, although the rate of decline has slowed.

ACKNOWLEDGMENTS We thank Dr. J. Michael McGinnis, Senior Scholar, Institute of Medicine of the National Academies and an anonymous referee for their useful comments on the manuscript.

Quantifying National Spending on Wellness and Prevention

23

REFERENCES Brown, R., Corea, J., Luce, B., Elixhauser, A., & Sheingold, S. (1992). Effectiveness in disease and injury prevention. Estimated National Spending on Prevention – United States, 1988. Morbidity and Mortality Weekly Report, 41(29), 529–531. Brown, R. E., Elixhauser, A., Corea, J., Luce, B. R., & Sheingold, S. (1991, June 28). National expenditures for health promotion and disease prevention activities in the United States. Battelle: The Medical Technology Assessment and Policy Research Center, Washington, DC. Burd, L., Chiang, M., Rutherford, G. W., III., Banaie, A., Dutta, S., Faruqi, M., Ho, C., Richman, A., Riester, K., Rohr, C., Yusuf, H., Roome, A., Hadler, J. L., Carlson, R., Craft, W., Keeling, C., Phillips, L., Ryan, R., Satzler, C., Schmid, J., Ummel, M., & Pelletier, A. (1994). Maternal hepatitis B screening practices – California, Connecticut, Kansas, and United States, 1992–1993. Morbidity and Mortality Weekly Report, 43(17), 311, 317–320. Available: http://www.cdc.gov/mmwr/PDF/wk/mm4317.pdf Cohen, J., Neumann, P., & Weinstein, M. (2008). Does preventive care save money? Health economics and the presidential candidates. The New England Journal of Medicine, 358(7), 661–663. Faust, H. S. (2005, May 27). Prevention vs. cure – which takes precedence? Medscape Public Health & Prevention. Available: http://www.medscape.com/viewarticle/504743?rss Gerberding, J. (2007, April 20). CDC Professional Judgment for FY 2008. Available: http:// www.fundcdc.org/documents/CDC_FY2008_PJ.pdf Hoerger, T. J., Harris, R. H., Hicks, K. A., Donahue, K. D., Sorensen, S., & Engelgau, M. (2004). Screening for type 2 diabetes: A cost-effectiveness analysis. Originally in Annals of Internal Medicine, 140, 689–699. Agency for Healthcare Research and Quality, Rockville, MD. Available: http://www.ahrq.gov/clinic/3rduspstf/diabscr/diabcost.htm Jain, N., & Stokley, S. (2007). National vaccination coverage among adolescents aged 13–17 years – United States 2007. Morbidity and Mortality Weekly Report, 56(34), 885–888. Available: www.cdc.gov/mmwr/preview/mmwrhtml/mm5634a3.htm?s_cid=mm5634a3_e Johns Hopkins Medical Institutions/Healthways. (2006). Embracing health: Tools and systems for health promotion and disease prevention. 6th Annual Disease Management Outcomes Summit. November. Lambrew, J. M., & Podesta, J. D. (2006, October 5). Promoting prevention and preempting costs. A new wellness trust for the United States. Center for American Progress. Lubetkin, E. I., Sofaer, S., Gold, M. R., Berger, M. L., Murray, J. F., & Teutsch, S. M. (2003). Aligning quality for populations and patients: Do we know which way to go? American Journal of Public Health, 93(3), 406–411. Maciosek, M. V. (2007, October). Personal communication including data summarizing immunization and other primary prevention costs. Partnership for Prevention. Maciosek, M. V., Solberg, L. I., Coffield, A. B., Edwards, N. M., & Goodman, M. J. (2006). Colorectal cancer screening, health impact and cost effectiveness. American Journal of Preventive Medicine, 31(1), 80–89. McGinnis, J. M. (2000, January 17). Common prevention-related terms & definitions. Institute of Medicine unpublished manuscript. McGinnis, J. M., Williams-Russo, P., & Knickman, J. R. (2002). The case for more active policy attention to health promotion. Health Affairs, 21(2), 78–93. MedPAC. (2007, March). Report to the congress: Medicare payment policy. Medicare Payment Advisory Commission.

24

GEORGE MILLER ET AL.

Mehrotra, A., Zaslavsky, A. M., & Avanian, J. Z. (2007). Preventive health examinations and preventive gynecological examinations in the United States. Archives of Internal Medicine, 167(17), 1876–1883. National Center for Health Statistics. (2006). Health, United States, 2006. DHHS Publication No. 2006-1232. US Department of Health and Human Services, Centers for Disease Control and Prevention. NIH. (2007). Estimates of funding for various diseases, conditions, research areas. Available: http://www.nih.gov/news/fundingresearchareas.htm OECD. (2005). Health at a glance – OECD indicators 2005. Available: http://www.oecd.org/ health/healthataglance Orosz, E. (2005). The OECD system of health accounts and the US National Health Account: Improving connections through shared experiences. Draft paper prepared for the conference on Adapting National Health Expenditure Accounting to a Changing Health Care Environment. Centers for Medicare & Medicaid Services, 21–22 April. Partnership for Prevention. (2007). Preventive care: A national profile on use, disparities, and health benefits. Available: http://www.prevent.org/images/stories/2007/ncpp/ncpp% 20preventive%20care%20report.pdf PricewaterhouseCoopers. (2006, January). The factors fueling rising healthcare costs 2006. Prepared for America’s Health Insurance Plans. Available: http://www.pwc.com/ extweb/pwcpublications.nsf/docid/E4C0FC004429297A852571090065A70B/$File/ahipfactors_fueling_rising_hc-costs.pdf Roehrig, C., Miller, G., & Lake, C. (2007, June 3). NHEA personal health expenditures by medical condition, annual estimates: 1996–2004. Academy Health Annual Research Meeting. Orlando, FL. Satcher, D. (2006). The prevention challenge and opportunity. Health Affairs, 25(4), 1009–1011. Sensenig, A. L. (2007). Refining estimates of public health spending as measured in National Health Expenditures Accounts: The United States experience. Journal of Public Health Management Practice, 13(2), 103–114. Shodell, D. (2006, September 20). Paying for prevention. Medscape Public Health & Prevention. Available: http://www.medscape.com/viewarticle/544651 Sisk, J. E., Whang, W., Butler, J. C., Sneller, V.-P., & Whitney, C. G. (2003). Cost-effectiveness of vaccination against invasive pneumococcal disease among people 50 through 64 years of age: Role of comorbid conditions and race. Annals of Internal Medicine, 138(12), 960–968. St. Lawrence, J. S., Montano, D. E., Kasprzyk, D., Phillips, W. R., Armstrong, K., & Leichliter, J. I. (2002). STD screening, testing, case reporting, and clinical and partner notification practices: A national survey of US physicians. American Journal of Public Health, 92(11), 1784–1788. Toth, R. (2007, October). Personal communication including data describing allocation of research funds. Robert Wood Johnson Foundation. Tufts University. (2007). Cost-effectiveness analysis registry of the center for the evaluation of value and risk in health. Institute for Clinical Research and Health Policy Studies. Available: https://research.tufts-nemc.org/cear/default.aspx Turnock, B. J. (2004). Public health: What it is and how it works. Boston: Jones and Bartlett Publishers. US Preventive Services Task Force. (2007). The guide to clinical preventive services. Rockville, MD: Agency for Healthcare Research and Quality. Woolf, S. H. (2006). The big answer: Rediscovering prevention at a time of crisis in health care. Harvard Health Policy Review, 7(2), 5–20.

ACHIEVING THE HEALTHY PEOPLE 2010 GOAL OF ELIMINATION OF HEALTH DISPARITIES: WHAT WILL IT TAKE? Kevin Fiscella ABSTRACT The second national goal for Healthy People 2010 is the elimination of health disparities related to social disadvantage in the United States. Unfortunately, progress to date has been limited. Our national strategy to achieve this goal has been too narrowly focused on public health. Success will require a broader strategy including alignment of existing national policies in non-health areas that affect the health of the socially disadvantaged such as education, health care, labor, welfare, housing, criminal justice, the environment, and taxation if it is to succeed. Key criteria are needed to begin to prioritize areas for federal investment to achieve this goal. These include the impact of the targeted condition on disparities, evidence base for the intervention, potential impact of the policy on disparities, economic impact, and federal politics. Two ‘‘big ideas’’ offer promise including federal investment in early

Beyond Health Insurance: Public Policy to Improve Health Advances in Health Economics and Health Services Research, Volume 19, 25–41 Copyright r 2008 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0731-2199/doi:10.1016/S0731-2199(08)19002-1

25

26

KEVIN FISCELLA

child education and enhanced primary care within federally qualified community health centers. The proposed criteria are applied to each proposed policy.

INTRODUCTION Socially disadvantaged persons, for example, members of historically disadvantaged racial and ethnic minority groups and persons of low socioeconomic status experience significant worse health status (Department of Health and Human Services, 2000). The Healthy People (HP) 2010 goal, eliminating these health disparities, represents a major national commitment (Department of Health and Human Services, 2000). This paper reviews progress toward this goal, identifies limitations in the current national strategy to achieve it, and proposes key changes in strategy, including implementation of two ‘‘big ideas.’’

PROGRESS IN ELIMINATING HEALTH DISPARITIES HP 2010 includes hundreds of objectives related to achieving the twin goals: (1) Increase quality and years of healthy life for all Americans and (2) Eliminate health disparities (Department of Health and Human Services, 2000). To assess progress toward achieving the second goal, I selected three widely used measures of population health: life expectancy from birth, infant mortality, and health status. Each of these measures is widely used to compare health across different groups including countries (Central Intelligence Agency; Department of Health). In this case, I used these measures to compare progress in closing racial disparities in health. Comparable national data by socioeconomic status for these years are not available. National vital and survey data lag at least 3–4 years behind publication. To assess progress in eliminating racial disparities in health, I compared the most recent data available (2004) with baseline data from 1999, providing a 5-year or midpoint evaluation of HP 2010 progress. I assessed the progress based on the percentage of gap at baseline that has since been closed using the national data (National Center for Health Statistics, 2008). The results (Table 1) show that although some progress has been made toward reducing each of these health gaps by race, progress has been modest and the rate is not sufficient to close the gap by 2010.

27

Achieving the Healthy People 2010 Goal

Progress Toward the HP 2010 Goal of Elimination of Health Disparities: White–Black Health Gaps 1999–2004. Health Measure Life expectancy from birth (years) Infant mortality (per 1,000 births) Fair/poor health status (% adults reporting)

1999

2004

5-Year Change (% Change)

5.9 8.8 6.6

5.2 7.6 6.0

0.7 (12) 1.2 (14) 0.6 (9)

LIMITS OF CURRENT STRATEGY The current national strategy to eliminate health disparities has focused largely on public health promotion and improved access to care for vulnerable populations. Notable examples include the Breast and Cervical Cancer Screening Program funded by the CDC, federal funding for the expansion of federally qualified community health centers (FQHCs), and implementation of the Health Disparities Collaboratives through the Health Resources and Services Administration (HRSA)/Bureau of Primary Health Care (BPHC) (Shi & Stevens, 2005). This strategy has largely ignored what have been referred to as ‘‘the fundamental causes’’ of health disparities (Link & Phelan, 1995). Specifically, the national health strategy is too narrowly focused and fails to address improvements in education, high school graduations rates for poor and minority children, reductions in family poverty, or systematic approaches to improving health care delivery and quality, particularly for members of socially disadvantaged groups. In contrast to European initiatives to address health disparities (Sheena & Joyce, 2006), the U.S. initiative lacks a national plan that aligns national health policies, educational, welfare, housing, labor, criminal justice, environmental, or tax policies with the HP 2010 goal. Because the health of socially disadvantaged populations is typically disproportionately affected by these policies, coordination of these national policies is needed in order to realize this goal. That is, the best strategy for reaching the HP 2010 goal is to examine all national policies that affect health disparities and re-align them toward this national goal.

SELECTING POLICIES A key challenge in identifying key policies for achieving the HP 2010 goal is: Where to start? There are many different polices that could be enacted that

28

KEVIN FISCELLA

could potentially facilitate progress. What criteria might be used to prioritize these policies for funding? Given that the ranks of the 47 million uninsured in the United States are disproportionately socially disadvantaged (U.S. Census Bureau Washington, DC, 2006), implementation of universal health care insurance is an obvious strategy. It would substantially improve health care access to socially disadvantaged populations – a necessary, but not a sufficient condition for reducing health disparities (Adler et al., 1994). Careful analysis suggests that universal health insurance would be cost-effective based on standard metrics (Muennig, Franks, & Gold, 2005). What other national policies besides implementation of universal health insurance might significantly foster progress toward the HP 2010 goal of the elimination of health disparities? In this paper, I discuss two ‘‘big ideas’’ for doing so. These policies are not intended to exclude other critical policies. Ultimately, the elimination of disparities will require multiple strategies implemented at different levels. Nonetheless, we must start somewhere. The following criteria are proposed as a means for examining different policies:  Strength of the evidence linking the problem targeted by the policy to health disparities  Strength of the evidence that the proposed intervention reduces health disparities  The potential impact of the policy on disparities  The economic impact of the policy  Politics surrounding enactment of the policy The two ‘‘big idea’’ policies are: federal funding for early childhood intervention and federal funding to develop a patient-centered medical home (PCMH) within FQHCs. The following sections discuss these policies in the context of these five criteria.

EARLY CHILD INTERVENTION Policy The federal government should provide funding to states to support fullday, high-quality early child education beginning at 6 months of age and continuing until the start of kindergarten and also establish standards for such education. Existing Head Start and state-sponsored programs could

Achieving the Healthy People 2010 Goal

29

be transitioned into this program. The rationale for this policy is that racial/ ethnic and socioeconomic disparities in education are powerful determinants of disparities in health. Educational disparities by social disadvantage, for example, race, ethnicity, and income, emerge prior to kindergarten. Early child education has been shown to ameliorate school readiness disparities and improve outcomes across the life course. Thus, early child intervention represents a necessary condition for eliminating disparities in education. Doing so would have far-reaching effects on health disparities.

Relationship between Education and Health The association between better education and improved health is robust (Cutler & Lleras-Muney, 2006; Gottfredson, 2004). Internationally, rates of maternal literacy are among the most powerful predictors of child and population health (Grosse & Auffrey, 1989). Moreover, there is growing recognition that disparities in educational attainment represent a fundamental determinant of disparities in adult health (Cutler & Lleras-Muney, 2006; Ross & Wu, 1995). Education affects health in multiple ways (Mechanic, 2007). Educational attainment and achievement are key determinants of occupational status and annual income. These in turn are associated with improved health (Mechanic, 2007). Education also affects health indirectly through health-related behavior including smoking, diet, physical activity, sexual behavior, use of health care including adherence, and exposure to injury and accidents. Through effects on everyday decisionmaking, it also affects exposure to stressful circumstances and environments and resilience and coping ability in the face of stress (Gottfredson, 2004). Educational disparities are the primary contributor to racial disparities in mortality (Wong, Shapiro, Boscardin, & Ettner, 2002). A substantial body of evidence, including findings from natural experiments, suggest that the relationship between education and health is at least partly causal (LlerasMuney, 2005).

Educational Disparities Emerge Early Disparities in education begin early in life. Cognitive testing of children less than 1 year old shows minimal disparities by race, ethnicity, or socioeconomic position (Fryer & Levitt, 2007). However, disparities in the acquisition of language (size of vocabulary and complexity of sentence

30

KEVIN FISCELLA

structure) and numeracy soon emerge. By 3–5 years, disparities in cognitive, language, and behavioral development by race, ethnicity, and socioeconomic position range from 0.5 to 1.0 standard deviation (Carneiro & Heckman, 2005). These disparities are affected by child health including birth weight and gestational age at birth, exposure to environmental toxins such as lead and smoking, other health problems that disproportionately affect minority children such as asthma, and most importantly, differences in cognitive stimulation including richness of vocabulary and positive caregiver affect (Magnuson & Waldfogel, 2005; Brooks-Gunn & Markman, 2005; Currie, 2005; Reichman, 2005; Asbury, Dunn, & Plomin, 2006; Fryer & Levitt, 2006; Lanphear, Dietrich, Auinger, & Cox, 2000; Love et al., 2005; Shonk & Cicchetti, 2001; Raikes et al., 2006; Cook & Frank, 2008). There may also be important biological–environmental interactions. That is, high educational status of parents tends to mitigate the effects of low birth weight on cognitive development (Ment et al., 2003).

Early Intervention Works School readiness, defined as a state of development upon entry to school that enables the child ready to learn, differs by social disadvantage (Rock & Stenner, 2005; Brooks-Gunn, Klebanov, & Duncan, 1996). Learning and skill formation build sequentially on previous learning (Cunha, Heckman, Lochner, & Masterov, 2005). Early intervention is far more effective than remediation (Heckman, 2007; O’Connor, Rutter, Beckett, Keaveney, & Kreppner, 2000) because remediation requires that the rate of learning among those needing it exceed those who do not need it. There is robust evidence that high-quality early interventions can reduce gaps in school readiness and improve outcomes across the life course. Longterm followup of randomized controlled trials and quasi-experimental studies demonstrate that high-quality early interventions can reduce gaps in educational achievement and improve adult outcomes, including teenage pregnancy, welfare dependency, arrests, and earnings (Love et al., 2005; Campbell, Ramey, Pungello, Sparling, & Miller-Johnson, 2002; Arnold & Doctoroff, 2003; Gray & McCormick, 2005; Reynolds et al., 2007). The largest benefits have been observed among interventions that begin at 6 months of age, employ college-educated, professionally trained staff, have adequate staff-to-child ratios, and create learning partnerships with parents (Barnett & Belfield, 2006; Karoly, Kilburn, & Cannon, 2005). However,

Achieving the Healthy People 2010 Goal

31

even these interventions have their limits. Little long-term benefits are seen for children born with very low birth weight (McCormick et al., 2006). Findings from these trials can be scaled. A multi-city randomized controlled trial of early head start showed significant improvements in child cognitive and emotional development (Love et al., 2005). State-wide implementation of pre-kindergarten in Oklahoma yielded developmental gains similar to those observed in small, high-quality interventions (Gormley, Gayer, Phillips, & Dawson, 2005).

Impact on Health Disparities Previous estimates suggest that universal enrollment in high-quality child care or preschools for low-income children could reduce the racial and ethnic gaps in reading readiness by 24 and 36%, respectively (Magnuson & Waldfogel, 2005). The impact of early intervention on adult health disparities has yet to be formally quantified, but given the robust relationship between academic achievement and health, the impact is likely to be appreciable. Effects may be even greater than expected if widespread implementation of early childhood intervention alters the learning climate within schools or promotes student body expectations for achievement in inner-city and rural schools.

Economic Impact Careful economic analyzes show that targeted investments in early child intervention would yield large savings over the long term from improved employment and earnings, and reductions in crime and incarceration (Karoly et al., 2005; Heckman, 2006; Lynch, 2007). These estimates are not farfetched. For the first time, California will spend more on criminal justice than on higher education (Legislative Analyst’s Office Site).

Politics Although data on the effectiveness of early child education has been available for a number of years, support for the concept is increasing. The concept has a number of features in its favor. First, it focuses on children. Poor and minority children are often seen by the public as members of the

32

KEVIN FISCELLA

deserving poor. That is, children cannot be held responsible for their poverty. They do not choose their parents or choose to be born poor. Second, the idea of fair competition is deeply ingrained in the American psyche. The finding that many socially disadvantaged children are one to two grade levels behind before they even start school is an affront to American sense of fair play. Third, early intervention offers promise for addressing other societal problems such as teenage pregnancy, crime, and welfare dependency. Fourth, early intervention can be billed as potential investment in human capital rather than a subsidy. Fifth, early intervention builds on the notion of public education. It simply moves the age for starting education backward. Sixth, early intervention would assist states who are challenged to find suitable child care for working mothers transitioning from Temporary Assistance for Needy Families (TANF). Last, it saves money over the long term. Whether sliding fees (with fees waived for families living near or below federal poverty lines) based on household income or full subsidization for all persons (similar to public schools) would be both more feasibly implemented and sustained is not clear.

ENHANCED PRIMARY CARE WITHIN HEALTH CENTERS Policy The federal government should fund the implementation and maintenance of enhanced primary care within FQHCs. Specifically, the HRSA/BPHC would provide sufficient funding for FQHCs to develop and sustain (PCMHs). Funding would support staffing, training, health information technology, and technical support, and provide a mechanism for ongoing funding that would supplement existing sources of revenue. The rationale for this policy is that socially disadvantaged patients face far greater health risks than other groups, but ironically face far less access to high-quality health care. Primary care has been shown to improve population health and reduce disparities. However, primary care delivery and funding have not kept pace with advances in medicine. This gap disproportionately affects socially disadvantaged patients who have greater and more complex health needs. Enhancing primary care within sites serving large numbers of socially disadvantaged patients could substantially improve access to high-quality care for these patients and reduce disparities in outcomes.

Achieving the Healthy People 2010 Goal

33

Lack of Continuous Care Coordination Absence of a regular source of care is strongly associated with disparities (Shi & Stevens, 2005). The U.S. health care system is fragmented and poorly coordinated (McCarthy, 2001). Socially disadvantaged patients face far greater access barriers that extend beyond absent or inadequate health insurance. These barriers include mistrust, low health literacy, limited English proficiency, cultural beliefs, job and child care constraints among others (Shi & Stevens, 2005). Thus, socially disadvantaged patients who because of shared social characteristics confront greater health risks across multiple domains also face greater health care access and coordination barriers (Frohlich & Potvin, 2008). Improving the health of socially disadvantaged populations through evidence-based medical care requires enhancement of primary care.

Primary Care Primary care represents first contact, continuous, comprehensive, and coordinated care through the health care system (Starfield, Shi, & Macinko, 2005). There is robust evidence that primary care improves health outcomes and reduces disparities (Starfield & Shi, 2004). Improving access to primary care for socially disadvantaged patients represents a core strategy for reducing health disparities. This has been the premise behind the expansion of FQHCs and look-alike community health centers (O’Malley, Forrest, Politzer, Wulu, & Shi, 2005). With the exception of funding for the Health Disparities Collaborative, federal funding has primarily supported the numerical expansion of health centers, but provided relatively little funding to improve quality (Fiscella & Geiger, 2006). This is problematic given that primary care even for the average patient is suboptimal (Moore & Showstack, 2003; McGlynn et al., 2003). As medical science has advanced, the demands on primary care have outgrown the traditional delivery structure, the 15-min visit (Grumbach & Bodenheimer, 2002; Goroll, Berenson, Schoenbaum, & Gardner, 2007). Physician payments, health care delivery models, and health information technology has not kept pace with the growing expectations on primary care (Goroll et al., 2007). This has generated a crisis in primary care. The quality of care (McGlynn et al., 2003), physician supply (Phillips, Jr. & Starfield, 1500), and physician satisfaction have suffered (Moore & Showstack, 2003). Socially disadvantaged patients bear the brunt of this crisis (Fiscella,

34

KEVIN FISCELLA

Franks, Gold, & Clancy, 2000). Recruitment of primary care physicians is most difficult in FQHCs, and in inner-city and rural areas (Rosenblatt, Andrilla, Curtin, & Hart, 2006). Health care quality for socially disadvantaged patients is generally worse than it is for other patients (Fiscella et al., 2000; Rosenblatt et al., 2006). Recently, the major primary care organizations have endorsed a set of principles designed to revitalize primary care (Kellerman & Kirk, 2007). These principles represent the PCMH. They include designation of a personal, primary care physician, team-based care directed by a physician, whole person orientation throughout all stages of care, coordination of all facets of care, focus on quality and safety, and enhanced access to care (Kellerman & Kirk, 2007). The seventh principle, payment reform, provides the means for implementing these principles (Kellerman & Kirk, 2007). The central concept is redesign of primary care to meet patients’ needs in the twenty-first century (Grumbach & Bodenheimer, 2002). This means replacing an outmoded delivery model based on a few 15-min, face-to-face visits per year between patients and physicians with one based on multidisciplinary teams lead by physicians that provide care through a variety of modalities in addition to face-to-face visits. These modalities include phone calls, secure email, websites, even text messaging (depending on the level of security required for that particular message). Teams, consisting of physicians, nurse practitioners/physician assistants, nurses, and medical assistants, would allocate tasks based on minimum training and skill needed to effectively accomplish them. For example, nurses and medical assistants would deliver routine preventive care based on standing physician orders (McKibben, Stange, Sneller, Strikas, & Rodewald, 2000). Teams would review patients with quality outcomes and institute action plans to enhance health care processes such as intensifying therapy when needed or linking patients with smoking cessation programs (Neuwirth, Schmittdiel, Tallman, & Bellows, 2008). This could be done through various communication modalities suitable to the patient. This model depends heavily on health information technology including electronic medical records and patient registries (Kellerman & Kirk, 2007; Bodenheimer & Grumbach, 2003). This concept offers tremendous promise for improving care to socially disadvantaged patients by tailoring care to the needs of the patient. Growing evidence, including findings from randomized controlled trials, show that targeted and intensive, team-based care can substantially reduce disparities in health outcomes (Fiscella, 2007). Redesign of primary care creates the potential to implement many of these interventions within the ‘‘real world.’’

Achieving the Healthy People 2010 Goal

35

Implementation of the PCMH within FQHCs will require payment reform (Goroll et al., 2007; Kellerman & Kirk, 2007), and/or external subsidies. Examples of potential reforms include changes in fee-for-service reimbursement, introduction of per member per month payments, sharing of reductions in overall health care costs, and bonuses through improvements in quality (Kellerman & Kirk, 2007).

Inverting the Inverse Care Law Payment reform for primary care in general will not be sufficient to redesign primary care for socially disadvantaged patients. Improving health care outcomes for socially disadvantaged patients requires policies that invert ‘‘the inverse care law.’’ This principle enunciated by Julian Hart suggests that those with greatest health care needs receive the least care (Hart, 1971). Empirical data from the United States support this principle (Blustein, 2008). Improving equity in outcomes requires that health care resources match the needs of the patients. Strategies for doing so have been implemented in several European countries (Bajekal, Alves, Jarman, & Hurwitz, 2001; Verheij, De Bakker, & Reijneveld, 2001; Sundquist, Malmstrom, Johansson, & Sundquist, 2003). For example, in the United Kingdom practices receive additional funding, called deprivation payments, based on the sociodemographic characteristics of the patients served. However, until overall payment reform is implemented, the federal government through HRSA would fund these enhancements in FQHCs.

Impact on Disparities The impact of this policy on disparities is difficult to precisely quantify. However, FQHCs currently serve a substantial proportion of socially disadvantaged patients in the United States (Shi, Stevens, & Politzer, 2007). They serve one quarter of all patients living in poverty, one in seven of the uninsured, one in nine patients receiving Medicaid, and one in ten minority patients (Shields et al., 2007). Even under the currently broken system of primary care, existing data show that FQHCs reduce disparities in care and outperform other sources of care at lower expense (Proser, 2005). However, one does not need to look outside the United States for a successful large-scale, federally financed model of primary care for socially disadvantaged populations. The Veterans Health Administration (VA)

36

KEVIN FISCELLA

health care system represents such a model (Yano, Simon, Lanto, & Rubenstein, 2007). Through a combination of strong leadership, investment in health information technology and implementation of new systems of care, the VA has dramatically improved the quality of care it provides (Jha, Perlin, Kizer, & Dudley, 2003; Kizer, Demakis, & Feussner, 2000). It has become a national leader in primary care quality and has reduced disparities in outcomes among its patients (Jha, Shlipak, Hosmer, Frances, & Browner, 2001). The federal government acting through the Health Services Research Administration and BPHC could effect a similar transformation within FQHCs nationally (Fiscella & Geiger, 2006). Compared to the VA that substantially lagged behind most health systems in quality when their initiative began, FQHCs currently provide comparable, if not better care, allowing for the potential for even greater gains in quality (Proser, 2005).

Economic Impact The economic impact could be potentially far-reaching. Improvements in care coordination, quality, and safety can reduce health care costs through reductions in emergency department visits, specialty visits, duplicated testing, and hospitalizations (Falik et al., 2006). These savings would largely accrue to the federal government through lower Medicaid and Medicare expenses (Fiscella & Geiger, 2006). Improvements in health status can potentially reduce rates of disability and unemployment, thus reducing federal payments for disability and unemployment insurance and improving tax revenues from higher employment. Further research is needed to actually quantify the economic impact. FQHCs also act as key economic engines for economically struggling communities by employing local labor across a range of jobs with varying skill–levels, thus providing a potential economic ladder (Hunt, 2005). Detailed economic analyses are needed to provide actual estimates of both the health care savings and economic impact of this investment in FQHCs.

Politics Historically, funding for FQHCs have enjoyed bipartisan support (O’Malley et al., 2005; Lefkowitz, 2005). Extending increased funding to quality is not a stretch, particularly given the notable success of quality improvement in

Achieving the Healthy People 2010 Goal

37

VA. The geographical distribution of FQHCs further strengthens their political base; there are over 1,000 FQHCs in over 3,600 sites. Furthermore, the HRSA’s Health Disparities Collaborative has invested heavily in FQHCs (The 129th Annual Meeting of APHA, 2001). However, without further investment much of the earlier gains may be lost.

CONCLUSION HP 2010 is unlikely to achieve its goal without a broader approach to addressing health disparities. Success requires a national strategy that aligns federal policies that disproportionately affect socially disadvantaged populations. Key criteria are needed to begin to prioritize areas for federal investment. These include the impact of the targeted condition on disparities, evidence base for the intervention, potential impact of the policy on disparities, economic impact of the policy, and federal politics surrounding implementation. Two ‘‘big ideas’’ that offer promise include federal investment in (1) early child education and (2) enhanced primary care with FQHCs.

REFERENCES Adler, N. E., Boyce, T., Chesney, M. A., Cohen, S., Folkman, S., Kahn, R. L., et al. (1994). Socioeconomic status and health: The challenge of the gradient. American Psychologist, 49, 15–24. Arnold, D. H., & Doctoroff, G. L. (2003). The early education of socioeconomically disadvantaged children. Annual Review of Psychology, 54, 517–545. Asbury, K., Dunn, J. F., & Plomin, R. (2006). Birthweight-discordance and differences in early parenting relate to monozygotic twin differences in behaviour problems and academic achievement at age 7. Developmental Science, 9, F22–F31. Bajekal, M., Alves, B., Jarman, B., & Hurwitz, B. (2001). Rationale for the new GP deprivation payment scheme in England: Effects of moving from electoral ward to enumeration district underprivileged area scores. British Journal of General Practice, 51, 451–455. Barnett, W. S., & Belfield, C. R. (2006). Early childhood development and social mobility. The Future of Children, 16, 73–98. Blustein, J. (2008). Who is accountable for racial equity in health care? JAMA, 299, 814–816. Bodenheimer, T., & Grumbach, K. (2003). Electronic technology: A spark to revitalize primary care? JAMA, 290, 259–264. Brooks-Gunn, J., Klebanov, P. K., & Duncan, G. J. (1996). Ethnic differences in children’s intelligence test scores: Role of economic deprivation, home environment, and maternal characteristics. Child Development, 67, 396–408.

38

KEVIN FISCELLA

Brooks-Gunn, J., & Markman, L. B. (2005). The contribution of parenting to ethnic and racial gaps in school readiness. The Future of Children, 15, 139–168. Campbell, F. A., Ramey, C. T., Pungello, E., Sparling, J., & Miller-Johnson, S. (2002). Early childhood education: Young adult outcomes from the Abecedarian Project. Applied Developmental Science, 6, 42–57. Carneiro, P., & Heckman, J. (2005). Human capital policy. In: B. Friedman (Ed.), Inequality in America (pp. 77–239). Cambridge: MIT Press. Central Intelligence Agency. The World factbook. (Available at http://www.cia.gov/library/ publications/the-world-factbook/index.html, accessed February 16, 2008.) Cook, J. T., & Frank, D. A. (2008). Food security, poverty, and human development in the United States. Annals of the New York Academy of Sciences, x, 1–16. Cunha, T. M., Heckman, J. J., Lochner, L., & Masterov, D. V. (2005). Interpreting the evidence on life cycle skill formation. NBER Working Paper no. 11331. Currie, J. (2005). Health disparities and gaps in school readiness. The Future of Children, 15, 117–138. Cutler, D. M. & Lleras-Muney, A. (2006). Education and health: Evaluating theories and evidence. NBER Working Paper no. 12352. Department of Health and Human Services. (2000). Healthy people 2010: Understanding and improving health (2nd ed.). Washington: U.S. Government Printing Office. Department of Health. Health inequality target monitoring: Update to include data for 2006. (Available at http://www.dh.gov.uk/en/Publicationsandstatistics/Publications/PublicationsStatistics/DH_081280, accessed February 16, 2008). Falik, M., Needleman, J., Herbert, R., Wells, B., Politzer, R., & Benedict, M. B. (2006). Comparative effectiveness of health centers as regular source of care: Application of sentinel ACSC events as performance measures. Journal of Ambulatory Care Management, 29, 24–35. Fiscella, K. (2007). Eliminating disparities in health care through quality improvement. In: R. Williams (Ed.), Eliminating healthcare disparities in America: Beyond the IOM report. Totowa: Humana Press. Fiscella, K., Franks, P., Gold, M. R., & Clancy, C. M. (2000). Inequality in quality: Addressing socioeconomic, racial, and ethnic disparities in health care. JAMA, 283, 2579–2584. Fiscella, K., & Geiger, H. J. (2006). Health information technology and quality improvement for community health centers. Health Affairs, 25, 405–412. Frohlich, K. L., & Potvin, L. (2008). Transcending the known in public health practice: The inequality paradox: The population approach and vulnerable populations. American Journal of Public Health, 98, 216–221. Fryer, R. G., & Levitt, S. D. (2006). The Black–White test score gap through third grade. American Law and Economics Review, 8, 249–281. Fryer, R. G., & Levitt, S. D. (2006). Testing for racial differences in the mental ability of young children, Working Paper, March 2006. (Available at http://www.economics.harvard.edu/ faculty/fryer/files/fryer_levittbabiesrevision.pdf, accessed February 12, 2008.) Gormley, W. T., Jr., Gayer, T., Phillips, D., & Dawson, B. (2005). The effects of universal preK on cognitive development. Developmental Psychology, 41, 872–884. Goroll, A. H., Berenson, R. A., Schoenbaum, S. C., & Gardner, L. B. (2007). Fundamental reform of payment for adult primary care: Comprehensive payment for comprehensive care. Journal of General Internal Medicine, 22, 410–415.

Achieving the Healthy People 2010 Goal

39

Gottfredson, L. S. (2004). Intelligence: Is it the epidemiologists’ elusive ‘‘fundamental cause’’ of social class inequalities in health? Journal of Personality & Social Psychology, 86, 174–199. Gray, R., & McCormick, M. C. (2005). Early childhood intervention programs in the US: Recent advances and future recommendations. Journal of Primary Prevention, 26, 259–275. Grosse, R. N., & Auffrey, C. (1989). Literacy and health status in developing countries. Annual Review of Public Health, 10, 281–297. Grumbach, K., & Bodenheimer, T. (2002). A primary care home for Americans: Putting the house in order. JAMA, 288, 889–893. Hart, J. T. (1971). The inverse care law. Lancet, 1, 405–412. Heckman, J. J. (2006). Skill formation and the economics of investing in disadvantaged children. Science, 312, 1900–1902. Heckman, J. J. (2007). The economics, technology, and neuroscience of human capability formation. Proceedings of the National Academy of Sciences of the United States of America, 104, 13250–13255. Hunt, J. W., Jr. (2005). Community health centers’ impact on the political and economic environment: The Massachusetts example. Journal of Ambulatory Care Management, 28, 340–347. Jha, A. K., Perlin, J. B., Kizer, K. W., & Dudley, R. A. (2003). Effect of the transformation of the veterans affairs health care system on the quality of care. New England Journal of Medicine, 348, 2218–2227. Jha, A. K., Shlipak, M. G., Hosmer, W., Frances, C. D., & Browner, W. S. (2001). Racial differences in mortality among men hospitalized in the veterans affairs health care system. JAMA, 285, 297–303. Karoly, L., Kilburn, M., & Cannon, J. (2005). Early childhood interventions, proven results, future promise. Santa Monica: Rand Corportation. Kellerman, R., & Kirk, L. (2007). Principles of the patient-centered medical home. American Family Physician, 76, 774–775. Kizer, K. W., Demakis, J. G., & Feussner, J. R. (2000). Reinventing VA health care: Systematizing quality improvement and quality innovation. Medical Care, 38, 16–17. Lanphear, B. P., Dietrich, K., Auinger, P., & Cox, C. (2000). Cognitive deficits associated with blood lead concentrations o10 microg/dL in US children and adolescents. Public Health Reports, 115, 521–529. Lefkowitz, B. (2005). The health center story: Forty years of commitment. Journal of Ambulatory Care Management, 28, 295–303. Legislative Analyst’s Office Site. California Spending Plan 2007–2008: The Budget Act and Related Legislation. (available at http://www.lao.ca.gov/2007/spend_plan/spending_ plan_07-08.aspx accessed February 12, 2008). Link, B. G., & Phelan, J. (1995). Social conditions as fundamental causes of disease. Journal of Health Social Behavior (Extra Issue), 80–94. Lleras-Muney, A. (2005). The relationship between education and adult mortality in the United States. Review Economic Studies, 72, 89–221. Love, J. M., Kisker, E. E., Ross, C., Raikes, H., Constantine, J., Boller, K., et al. (2005). The effectiveness of early head start for 3-year-old children and their parents: Lessons for policy and programs. Developmental Psychology, 41, 885–901.

40

KEVIN FISCELLA

Lynch, R. G. (2007). Enriching children, enriching the nation: Public investment in high-quality prekindergarten. Washington: Economic Policy Institute. Magnuson, K. A., & Waldfogel, J. (2005). Early childhood care and education: Effects on ethnic and racial gaps in school readiness. The Future of Children, 15, 169–196. McCarthy, M. (2001). Fragmented US health-care system needs major reform. Lancet, 357, 782. McCormick, M. C., Brooks-Gunn, J., Buka, S. L., Goldman, J., Yu, J., Salganik, M., et al. (2006). Early intervention in low birth weight premature infants: Results at 18 years of age for the infant health and development program. Pediatrics, 117, 771–780. McGlynn, E. A., Asch, S. M., Adams, J., Keesey, J., Hicks, J., DeCristofaro, A., et al. (2003). The quality of health care delivered to adults in the United States. New England Journal of Medicine, 348, 2635–2645. McKibben, L. J., Stange, P. V., Sneller, V. P., Strikas, R. A., & Rodewald, L. E. (2000). Advisory committee on immunization practices. Use of standing orders programs to increase adult vaccination rates. Morbidity & Mortality Weekly Report 2000 [Recommendations & Reports], 49, 15–16. Mechanic, D. (2007). Population health: Challenges for science and society. Milbank Quarterly, 85, 533–559. Ment, L. R., Vohr, B., Allan, W., Katz, K. H., Schneider, K. C., Westerveld, M., et al. (2003). Change in cognitive function over time in very low-birth-weight infants. JAMA, 289, 705–711. Moore, G., & Showstack, J. (2003). Primary care medicine in crisis: Toward reconstruction and renewal. Annals of Internal Medicine, 138, 244–247. Muennig, P., Franks, P., & Gold, M. (2005). The cost effectiveness of health insurance. American Journal of Preventive Medicine, 28, 59–64. National Center for Health Statistics. (2008). Health, United States, 2007. Hyattsville, M.D.: U.S. Department of Health and Human Services, Centers for Disease Control and Prevention, National Center for Health Statistics. Neuwirth, E. B., Schmittdiel, J. A., Tallman, K., & Bellows, J. (2008). Understanding panel management: Comparative case studies of an emerging approach to population care. The Permanente Journal, 11, 11–24. O’Connor, T. G., Rutter, M., Beckett, C., Keaveney, L., & Kreppner, J. M. (2000). The effects of global severe privation on cognitive competence: Extension and longitudinal followup. English and Romanian Adoptees Study Team. Child Development, 71, 376–390. O’Malley, A. S., Forrest, C. B., Politzer, R. M., Wulu, J. T., & Shi, L. (2005). Health center trends, 1994–2001: What do they portend for the federal growth initiative? Health Affairs, 24, 465–472. Phillips, R. L., Jr., & Starfield, B. (2003). Why does a U.S. primary care physician workforce crisis matter? American Family Physician, 68, 1494–1498. Proser, M. (2005). Deserving the spotlight: Health centers provide high-quality and costeffective care. Journal of Ambulatory Care Management, 28, 321–330. Raikes, H., Pan, B. A., Luze, G., Tamis-LeMonda, C. S., Brooks-Gunn, J., Constantine, J., et al. (2006). Mother-child bookreading in low-income families: Correlates and outcomes during the first three years of life. Child Development, 77, 924–953. Reichman, N. E. (2005). Low birth weight and school readiness. The Future of Children, 15, 91–116.

Achieving the Healthy People 2010 Goal

41

Reynolds, A. J., Temple, J. A., Ou, S. R., Robertson, D. L., Mersky, J. P., Topitzes, J. W., et al. (2007). Effects of a school-based, early childhood intervention on adult health and wellbeing: A 19-year follow-up of low-income families. Archives of Pediatrics & Adolescent Medicine, 161, 730–739. Rock, D. A., & Stenner, A. J. (2005). Assessment issues in the testing of children at school entry. The Future of Children, 15, 15–34. Rosenblatt, R. A., Andrilla, C. H., Curtin, T., & Hart, L. G. (2006). Shortages of medical personnel at community health centers: Implications for planned expansion. JAMA, 295, 1042–1049. Ross, C. E., & Wu, C. l. (1995). The links between education and health. American Sociological Review, 60, 719–745. Sheena, A., & Joyce, H. (2006). What works in tackling health inequalities? Pathways, policies and practice through the lifecourse. Bristol: Policy Press. Shi, L., & Stevens, G. D. (2005). Vulnerable populations in the United States. San Francisco: Jossey-Bass. Shi, L., Stevens, G. D., & Politzer, R. M. (2007). Access to care for U.S. health center patients and patients nationally: How do the most vulnerable populations fare? Medical Care, 45, 206–213. Shields, A. E., Shin, P., Leu, M. G., Levy, D. E., Betancourt, R. M., Hawkins, D., et al. (2007). Adoption of health information technology in community health centers: Results of a national survey. Health Affairs, 26, 1373–1383. Shonk, S. M., & Cicchetti, D. (2001). Maltreatment, competency deficits, and risk for academic and behavioral maladjustment. Developmental Psychology, 37, 3–17. Starfield, B., Shi, L., & Macinko, J. (2005). Contribution of primary care to health systems and health. Milbank Quarterly, 83, 457–502. Starfield, B., & Shi, L. (2004). The medical home, access to care, and insurance: A review of evidence. Pediatrics, 113, 1493–1498. Successfully implementing a quality improvement chronic care model for diabetic patients in community health centers, The 129th Annual Meeting of APHA, Atlanta, GA Oct 21–25, 2001. Sundquist, K., Malmstrom, M., Johansson, S. E., & Sundquist, J. (2003). Care need index, a useful tool for the distribution of primary health care resources. Journal of Epidemiology & Community Health, 57, 347–352. U.S. Census Bureau Washington, DC. (2006). Health Insurance Coverage. (Available at http:// www.census.gov/hhes/www/hlthins/hlthin06.html, accessed, February 12, 2008). Verheij, R. A., De Bakker, D. H., & Reijneveld, S. A. (2001). GP income in relation to workload in deprived urban areas in the Netherlands. Before and after the 1996 pay review. European Journal of Public Health, 11, 264–266. Wong, M. D., Shapiro, M. F., Boscardin, W. J., & Ettner, S. L. (2002). Contribution of major diseases to disparities in mortality. New England Journal Medicine, 347, 1585–1592. Yano, E. M., Simon, B. F., Lanto, A. B., & Rubenstein, L. V. (2007). The evolution of changes in primary care delivery underlying the veterans health administration’s quality transformation. American Journal of Public Health, 97, 2151–2159.

INFLUENCE, INFORMATION OVERLOAD, AND INFORMATION TECHNOLOGY IN HEALTH CARE James B. Rebitzer, Mari Rege and Christopher Shepard ABSTRACT We investigate whether information technology (IT) can help physicians more efficiently acquire new knowledge in a clinical environment characterized by information overload. We combine analysis of data from a randomized trial with a theoretical model of the influence that IT has on the acquisition of new medical knowledge. Although the theoretical framework we develop is conventionally microeconomic, the model highlights the non-market and non-pecuniary influence activities that have been emphasized in the sociological literature on technology diffusion. We report three findings. First, empirical evidence and theoretical reasoning suggests that computer-based decision support will speed the diffusion of new medical knowledge when physicians are coping with information overload. Second, spillover effects will likely lead to ‘‘underinvestment’’ in this decision support technology. Third, alternative financing strategies common to new IT, such as the use of marketing dollars to pay for the decision support systems, may lead to undesirable outcomes if physician information overload is sufficiently severe and if Beyond Health Insurance: Public Policy to Improve Health Advances in Health Economics and Health Services Research, Volume 19, 43–69 Copyright r 2008 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0731-2199/doi:10.1016/S0731-2199(08)19003-3

43

44

JAMES B. REBITZER ET AL.

there is significant ambiguity in how best to respond to the clinical issues identified by the computer. This is the first paper to analyze empirically and theoretically how computer-based decision support influences the acquisition of new knowledge by physicians.

1. INTRODUCTION Economists who study incentives in organizations have generally highlighted problems that arise when information is scarce. Indeed the premise of the vast literature on optimal incentives is that principals have incomplete knowledge about agents and their actions. Many important features of organizations, however, emerge from efforts to cope with a superabundance of information, but economic research on these aspects of organizations is less well developed (Brynjolfsson & Hitt, 2000). In what follows we consider economic issues that arise when agents are presented with more information than they can handle. The focus of our study is information overload among physicians. In medicine, the number and variety of diseases and treatments threaten to overwhelm the information processing capacities of individual doctors. This complexity is magnified by the rapid growth of new, medically-relevant knowledge. Failure to cope with this flood of information can cause physicians to overlook important new treatments or new findings concerning the recommended use of existing treatments. The net result is reduced care quality and slow diffusion of new medical innovations (Institute of Medicine Committee on Quality of Health Care in America, 2000, 2001). We investigate how information technology (IT) can help physicians cope with information overload. Our analysis focuses on a particular type of computer system that has drawn considerable attention in the medical community and in public policy debates: a physician decision support tool. The system we study combines information from insurance billing records, pharmacies, and labs to construct an electronic medical record for each patient. Information from this record is then passed through a sophisticated artificial intelligence program that compares treatment received with protocols drawn from the medical literature. If a discrepancy between actual and recommended care is observed, a message is sent to the physician. This message might recommend that a certain patient is a good candidate for a specific lab or test. Alternatively the message might suggest that specific medications should be taken or stopped. The message includes

Information Technology in Health Care

45

citations to the medical literature, allowing the physician to read and assess the relevance of the literature to the patient’s particular circumstances. By sending timely and targeted messages to physicians about evidence-based treatment protocols, the system can help doctors improve care quality and also increase the rate of diffusion of new medical knowledge. Our analysis is both empirical and theoretical. On the empirical side we review evidence about the diffusion of medical knowledge. This review provides necessary context for understanding findings from a randomized prospective trial suggesting that computer-based decision support tools can substantially enhance the diffusion of new medical knowledge among physicians. On the theory side we construct a model of physician learning that is consistent with the stylized facts about knowledge diffusion. Although the framework we develop is conventionally microeconomic, the model highlights the non-market and non-pecuniary influence activities that have been emphasized in the sociological literature on technology diffusion (Skinner & Staiger, 2005). Using this model, we consider both the likely effect of IT systems on physician learning and whether private incentives are sufficient to support optimal investments in IT. This last issue is particularly important for public policy because the Federal government is beginning large scale projects to promote the use of IT-enabled decision support tools in health care (Mullaney, 2005; President’s Information Technology Advisory Committee, 2004).1 The remaining sections are organized as follows. In section two, we briefly review the key empirical findings concerning the diffusion of new medical knowledge. In section three, we present evidence from a randomized prospective trial suggesting that computer-based decision support tools can substantially enhance the diffusion of new medical knowledge among physicians. In section four, we develop a model of physician learning that is consistent with these empirical findings. We then use this model to consider whether private incentives to invest in computer-based decision support technology are sufficient to assure adequate investment. We conclude by assessing the policy implications of our model and suggest future directions for research.

2. THE DIFFUSION OF INNOVATIONS IN MEDICINE The use of beta-blocker drugs following heart attacks substantially reduces mortality rates. This fact has been established for decades yet in 2000/2001 the level of compliance with appropriate beta blocker use in the median state

46

JAMES B. REBITZER ET AL.

was just 69 percent. In some states the rate of appropriate usage was as low as 50 percent and in others as high as 86 percent (Skinner & Staiger, 2005). This slow and variable take-up of an effective and inexpensive treatment protocol for an important illness is a specific illustration of a more general phenomenon: some innovations appear to diffuse through the health care system much more slowly than one would expect on the basis of their cost and efficacy.2 Perhaps the broadest evidence supporting the slow and uneven diffusion of innovations comes from the vast medical literature on the geographic variation in practice patterns. The central finding in this literature is that the care delivered by doctors varies greatly from one geographic region to another in ways that cannot be accounted for by underlying differences in patient population, medical prices, technology, or other exigencies. In the absence of other explanations, it is hard to escape the implication that these variations are due to the incomplete diffusion of the ‘‘best’’ medical practices to physicians.3 These variations have substantial economic implications as well – some states spend 40–60 percent more per patient on heart attack treatment than others with no improvement in outcomes (Skinner & Staiger, 2005). A Rand study described in the New York Times finds similarly wide geographic divergence in the treatment of various cancers (Altman, 2005). Skinner, Fisher, and Wennberg (2001) find that a large component of Medicare expenditures (nearly 20 percent) is the result of geographic variation in expenditures across 306 hospital referral regions. These expenditures, which are associated with the treatment of the chronically ill, appear to offer no benefits in terms of survival and the effects on quality of life are unclear. If even a portion of this variation is due to cross region productivity differences, then slow diffusion of new bestpractice methods may have very substantial economic consequences. The source of geographic variations in practice styles is poorly understood, but these variations suggest that physicians learn about treatments and protocols from colleagues with whom they interact on a regular basis. These interactions will play an important role in the theoretical analysis we develop in section four, below. Other evidence suggesting an imperfect diffusion of medical knowledge comes from the resources that drug companies spend promoting their products. If the results of clinical studies percolated swiftly and easily to providers, pharmaceutical companies would not need high-powered marketing campaigns and sales forces. Avorn (2004) describes how these campaigns influence the decisions of practicing physicians. He also documents a number of instances where pharmaceutical marketing appears

Information Technology in Health Care

47

to successfully promote the use of favored drugs even when there is good clinical evidence that these drugs were ineffective or less effective than inexpensive alternatives. In Avorn’s cases it was the clinical evidence concerning the ineffectiveness of heavily marketed drugs that appears to spread too slowly.4 A recent econometric study of the diffusion of anti-ulcer drugs finds that equilibrium market shares are strongly determined by the resources drug companies devote to advertising (Berndt, Pindyck, & Azoulay, 2003).5 This study also finds some evidence for consumption externalities that enhance the diffusion of drugs toward their equilibrium market shares. Consumption externalities occur when the use of a brand of drug is enhanced by the fact that many others are using it. The authors argue that these consumption externalities are largely informational in character, i.e., physicians and consumers make inferences about the qualities of a brand by the fact that many others are using it. The existence of consumption externalities is important because they can inhibit the ability of superior new drugs and treatment modalities to enter the market.6 The rate of diffusion of new knowledge is of self-evident importance in an industry as technologically dynamic as health care. Given this, it is surprising that the large and growing literature on the effects of IT systems in health care pays so little attention to the issue. A recent survey of 100 published studies of computer-based decision support programs (Garg et al., 2005) finds that the effects of these programs can be variable. No effort was made in the review, however, to examine whether the variation in efficacy could be linked to how ‘‘new’’ the recommendations were. Another survey of 257 studies of health IT systems (Chaudhry et al., 2006) finds that these systems generally increased adherence to guideline care, but the analysis did not consider potential differences between wellestablished guidelines and guidelines embodying relatively new clinical protocols. One reason for this gap in the literature is that it is often difficult to distinguish protocols embodying relatively new knowledge from those that do not. Consider for example the treatment for diabetes, an important and quite costly chronic disease. Disease management programs for diabetes generally embody well understood treatment protocols (Gertler & Simcoe, 2006; Beaulieu et al., 2006) but as we document in the next section regarding the use of ACE-inhibitors, new discoveries can modify these protocols in subtle but important ways. Tracking the differential effects of similar protocols requires sophisticated electronic medical records systems and these record systems are still quite rare. Survey data collected in 2007–2008

48

JAMES B. REBITZER ET AL.

reveals that only four percent of physicians have a fully functional electronic medical record system in their office and only 13 percent have a basic system (DesRoches et al., 2008). Similarly low rates of IT adoption are reported for electronic medical records in hospital emergency rooms and outpatient departments (Burt & Hing, 2005) and physician order entry systems in hospitals (Cutler, Feldman, & Horwitz, 2005). The problem of studying the effects of the diffusion of new knowledge is further limited by the fact that much of the research on IT comes from a small number of organizations that are likely to be front-runners in adopting new evidence-based protocols. Chaudhry et al. (2006), in their comprehensive review, point out that a large portion of the empirical work on health IT systems comes from the study of only four institutions: the Regenstrief Institute; Brigham and Women’s Hospital/Partners Health Care; the Department of Veterans Affairs; and LDS Hospital/Intermountain Health Care.

3. EMPIRICAL EVIDENCE ON THE INFLUENCE OF COMPUTER-GENERATED MESSAGES Physicians are highly trained and highly motivated professionals. It is reasonable to assume therefore that the slow and uneven distribution of new and effective therapies is not the result of simple inattention. A more likely explanation is that it is difficult for physicians to keep up with the rapidly changing state of medical knowledge and to understand what these changes mean for the treatment of specific patients. If this is the case, then we should expect that an IT-based decision support system could help doctors learn about new treatments. In this section we present empirical evidence from a randomized trial indicating that IT-based decision support tools might enhance the rate of diffusion of new medical treatments. The data we use comes from a randomized controlled trial of a decision support technology.7 In this study patients under age 65 who were members of a single Health Maintenance Organization (HMO) were randomly assigned to a study and a control group. Data from insurance billing records, laboratory feeds, and pharmacies were combined to construct a virtual electronic medical record and the information in these records was passed through a sophisticated program that scanned for clinical mistakes and also deviations from evidence-based, best-practiced protocols.

Information Technology in Health Care

49

For patients in the study group, the information was scanned in real time and if an issue was detected, a message was sent to the physician. The message stated the name of the patient, described the potential issue, and referenced the relevant medical literature. Roughly speaking the messages sent could be divided into three categories: ‘‘start a drug’’, ‘‘stop a drug,’’ or ‘‘do a test.’’ For patients in the control group, the clinical data was saved but not analyzed. After the study was completed, the control group data was analyzed by the software and messages were generated that would have been sent to physicians if the control patient had been in the study group. The trial therefore allowed us to compare the rate of resolution of issues when physicians enjoyed decision support (the study group) with the rate of resolution of the same issues when no such support was available (the control group). The difference in resolution rates between the study and control group is thus a measure of the influence of the IT on physicians.8 For our purposes, we would like to know if the messages from the decision support tool increased the rate of adoption of new medical evidence. We therefore focused our attention on messages concerning the class of medications known as ACE-inhibitors. ACE-inhibitors were first approved by the FDA in 1981 for treatment of severe hypertension. Shortly before the randomized trial began in 2001, several major clinical trials established the beneficial effect of ACE-inhibitors for patients with mild and moderate hypertension, heart failure, past heart attacks, chronic renal disease, certain subgroups of diabetics, and patients at high risk for cardiovascular events. Taken together, these trials greatly expanded the number of patients for whom an ACE-inhibitor was indicated. One particularly important trial was the Heart Outcomes Prevention Evaluation or HOPE trial. The computer system was set up to include recommendations from the HOPE trial among the messages it sent to physicians. By comparing the resolution rates for patients in the study group who triggered the HOPE trial-based recommendations with control group patients who would have triggered those recommendations, we can directly assess the influence of computer-generated messages on newly discovered medical protocols. The sample of participants in column one of Table 1 consists of those who would have qualified for an ACE-inhibitor on the basis of the HOPE trial criteria, but who were not receiving the drug. The dependent variable is a dummy variable equal to one if the issue of ACE-inhibitor use is successfully resolved, i.e., if records indicate that the patient received ACE-inhibitors within 270 days of the message being sent to the physician. The coefficient

50

JAMES B. REBITZER ET AL.

Resolution Rates. Probit

Study Number of patients Number of patients in study group Log pseudo-likelihood

(1)

(2)

Successful resolution ‘‘Add Ace Inhibitor for Hope Qualifier Drug’’ Message [0.141]

Successful resolution for any other ‘‘Add a Drug’’ Message [0.258]

0.130 (2.82) 311 155

0.035 (0.69) 290 166

154.018

158.877

In column 1, the sample consists of participants who would have qualified for an Ace inhibitor on the basis of the HOPE trial criteria, but who were not receiving the drug according to computer records. In column 2 the sample consists of participants who received an ‘‘add a drug’’ message other than that in column 1. The message is successfully resolved if there is evidence in the data base that the patient started the relevant drug within 270 days after the message was sent. Robust z statistics in parentheses. [ ] is mean of dep. var. in the control group in 2001. Coefficients are expressed as ‘‘derivatives.’’ Thus in column 1 members in the study group were 13 percentage points more likely to have resolved the issue successfully than those in the control group. Significant at 5%.

on the variable Study captures the differences in success rates between the study and the control group. Patients in the study group, whose physician received computer-generated messages regarding ACE-inhibitors, were 13 percentage points more likely to take up the ACE-inhibitor than those in the control group. To put this result in context, the resolution rate in the control group whose physicians received no messages was about 0.14. Thus the decision support tool nearly doubled resolution rates. In column two of Table 1, we run analogous probit regressions for participants who received ‘‘add a drug’’ recommendations other than those relating to HOPE trial recommendations. We observe that the rate of resolution of these ‘‘add a drug’’ recommendations was 0.258 in the control group. The effect of being in the study group was to increase the resolution rate by 0.035. This difference is statistically insignificant and behaviorally not very important.

Information Technology in Health Care

51

Comparing columns one and two, it is clear that the decision support system had a more substantial effect on the relatively new ACE-inhibitor recommendation than on all other recommendations that suggested adding a drug. At the time of our study, the new clinical evidence regarding ACEinhibitors was widely promoted in the conventional manner, i.e., via disease management programs and journal articles. We suspect, on the basis of informal conversations with providers, that the computer-generated messages had extra influence because they were reliable, timely, and focused physician attention on a specific issue concerning a specific patient. Put differently what the IT support tool did that other conventional communication channels did not, was to link a useful general recommendation to a specific patient’s situation. The results in Table 1 are consistent with the possibility that IT-based decision support can enhance the diffusion of new medical knowledge, but it is important to observe that they do not constitute a definitive proof of that claim. Non-ACE ‘‘add a drug’’ messages are a mix of many different recommendations. If these messages have heterogeneous effects, pooling the drugs together may create the impression that the entire set of non-ACE ‘‘add a drug’’ messages are ineffective when some of them were in fact quite effective. Pursuing this alternative interpretation of the data would require a larger study that would allow us to better analyze specific subsets of the non-ACE ‘‘add a drug’’ recommendations.

4. INFLUENCE AND PHYSICIAN LEARNING: A SIMPLE ANALYTICAL FRAMEWORK In this section, we develop a model of physician learning in an environment characterized by information overload. The evidence in the preceding discussions suggests that a reasonable model of physician learning must account for three empirical relationships. First, new clinical knowledge does diffuse through the health care system, but the rate of diffusion is often slow. Second, as the geographic variations in practice patterns suggest, physicians learn about new medical treatments from other physicians with whom they interact on a daily basis. Third, IT-based decision support technology can influence physician behavior and this influence is greatest for newly discovered treatment protocols. We introduce slow diffusion of new knowledge into our model by assuming that physicians are hampered by two cognitive limitations: the

52

JAMES B. REBITZER ET AL.

flow of new medical knowledge exceeds the fixed information processing capacities of individual physicians; and it is difficult for physicians to link the new medical knowledge they acquire to the clinical needs of specific patients. To capture the role of informal interactions in the process of knowledge diffusion, we assume that physicians compensate for their cognitive limitations by relying on the recommendations of colleagues with whom they interact on a day-to-day basis. Colleague recommendations are helpful because they can link specific treatments to the clinical needs of particular patients under a physician’s care, but these recommendations are not sufficient to resolve the problem of information overload. After all, physicians making recommendations have limits on their own cognitive abilities and these limits will generally make it hard for them to keep abreast of all the newest procedures. For this reason, physicians will also have to devote time to independent reading in medical journals. Reading journal articles may expose the physician to the newest innovations, but journal articles do not identify for the physician the specific patients for which the innovation applies. The influence of IT-enabled decision support follows naturally in this set-up. In comparison to traditional learning modalities (colleagues’ recommendations and independent reading of medical journals) the computer-based decision support tools are more likely to suggest treatments that are both new and relevant to the care of a specific patient.9 As a result, the new IT will have greater influence on physicians and will under plausible conditions, enhance the rate of diffusion of new knowledge.

4.1. Model Set-up In every period medical research generates s new treatments that are relevant to the care of a physician’s patients. A physicians’ objective when reading about new medical treatments is to maximize the sum of medical benefits from treatment adoption across his patients. We introduce limitations in physician information processing capabilities in the simplest possible way by assuming that in each period physicians can read n articles with nos.10 Physicians are also limited in their ability to apply information about new treatments to specific patients. We capture this aspect of limited cognition by stipulating that when reading about a new treatment in a medical journal, the likelihood that a physician identifies the treatment as relevant to their patients is r, where ro1. The value of r is determined by the relevance of the new treatment to his or her patient population and by the

Information Technology in Health Care

53

probability that the physician will recognize the relevance of the treatment. Because physicians cannot read about all new treatments produced in each period and sometimes fail to see the relevance of a treatment, there will always be a stock of potentially relevant treatments that the physician will have missed in the period in which they were introduced. In fields as technologically dynamic as medicine, the value of innovations tends to depreciate over time. We capture this depreciation through the assumption that patients experience benefits B when the physician adopts a relevant treatment in the same period it was discovered, a benefit Bq, where qo1, when the physician adopts a relevant treatment in the period after discovery, and benefit 0 from treatments that are more than two periods old. Because treatments older than two periods are clinically irrelevant, we can unambiguously refer to established treatments as treatments whose discovery occurred in the prior period. Physicians can learn about the relevance of established treatments from interactions with other doctors. The advantage of learning from other physicians is that colleagues identify treatments that are relevant to specific patients. The disadvantage is that colleagues are also overloaded by new information and therefore cannot keep abreast of the newest procedures. We represent this feature by assuming that colleagues recommend only relevant established treatments – never new ones. Thus, the expected marginal benefit of reading a journal article about a treatment recommended by colleagues is qB. We model the heightened relevance of treatments recommended by colleagues by assuming that qWr. In this way, information flows have a local flavor because the physician always prefers reading about the highly relevant treatments recommended by informal interactions with colleagues than independent reading about new treatments in journals. We focus on an equilibrium, in which the behavior of all doctors in a region can be captured by a single representative physician. Let xt and yt, respectively, denote the number of established and new treatments the physician chooses to read about in period t. Let zt denote the number of recommendations a doctor receives from her colleagues. We endogenize zt by assuming that the recommendations a doctor gives come from the new treatments that he has adopted for his own patients in the previous period. Formally, we capture this by assuming that zt ¼ aryt1, where a is a parameter; and ryt1 simply denotes the average number of new treatments that physicians adopted in the previous period. Since the physicians always prefer reading about treatments recommended by colleagues to reading about new treatments in journals, this implies that physicians learn about

54

JAMES B. REBITZER ET AL.

Period t

Period t + 1

s new treatments

Physicians' independent reading of medical journals

n y= 1 + ar

(1−r) y

arn x= 1 + ar

Physicians' reading after recommendation from colleagues

x

ry

Physicians cannot identify any relevance

Physicians implement

Fig. 1.

The Steady-State Disposition of New Knowledge.

the following number of treatments from colleagues xt ¼ aryt1

(1)

so that information on the remaining treatments comes from reading medical journals. Thus yt ¼ n  xt

(2)

We focus on the steady-state equilibrium, in which a physician does not change her learning choices from one period to the other, i.e., xt ¼ x and yt ¼ y for all t. This, together with Eqs. (1) and (2) implies that in equilibrium physicians read about y ¼ n/1þar new treatments and x ¼ arn/ 1þar established treatments. The flow of new knowledge in our model is depicted graphically in Fig. 1.

4.2. Introducing Computer-Based Decision Support We now introduce computer-based decision support systems. As described above, these systems compare a patient’s treatment with best-practice

Information Technology in Health Care

55

protocols drawn from the medical literature. If the system finds a discrepancy, a message is sent to the physician. New treatments are written into the computer’s software in each period and we assume that the technology only sends messages concerning new treatments that are clinically relevant for a physician’s patients.11,12 The physician can expect to receive ym messages from the decision support technology each period, where m is the number of new treatments entering the data base in each period and y is the fraction of treatments relevant to the physician’s patients. From the physician’s perspective, the benefit of reading about a treatment recommended in a computer-generated message is given by B (it is always new and relevant), whereas the benefit of reading about a treatment recommended by a colleague is qB, where q is the discount factor reflecting the delay in learning about new treatments from colleagues. Thus a doctor will always prefer reading about a computer-generated recommendation over a colleague’s recommendation. It is costly to incorporate new innovations into the decision support software, so it is unlikely that the decision support software will ever be sufficiently comprehensive to serve as the sole source of recommendations for physician learning. Rather than modeling these costs explicitly, we assume that ymon, implying that a doctor who has access to the technology will read about all the treatments recommended in the computer-generated messages and still have enough mental ‘‘shelf space’’ to continue learning about established treatments recommended by other sources. Put differently, the heightened timeliness and specificity of computer-generated recommendations displaces, but does not eliminate the use of other learning modalities. In the US health care system, clinical data about patients rarely sits in a common electronic data base. The IT revolution has been slow coming to hospitals and only in recent years have large expenditures in electronic medical record systems begun. Physicians’ offices are typically paper-based, and where they have moved to electronic records, these systems often do not interface neatly with hospital systems (DesRoches et al., 2008).13 Common standards for electronic medical records systems are only now being put in place. Efficient decision support software requires that patient information reside in a single data base or a set of interconnected data bases, and this sort of integration is precluded by the Balkanized system of medical records maintained by providers. Insurance companies, in contrast, are well positioned to act as aggregators of medical information because most clinical transactions involving their members appear coded into their data bases for billing purposes. This is why the decision support technology we discussed in section three was designed to use insurers’ data and perhaps this

56

JAMES B. REBITZER ET AL.

also explains why the informatics company that developed the decision support technology was eventually purchased by a large health insurer. Insurers’ data systems are also fragmented in that they are confined to individuals who are current policy holders. For this reason, computer-based decision will have both a direct and an indirect effect on physician learning. The direct effect is captured by the ym computer-generated messages a physician receives. The indirect effect results from the way that knowledge acquired from the computer system spills over to other physicians who have no access to the computer system. To capture these spillovers we stipulate that a share b of physicians have access to a particular decision support technology and a share (1b) does not have access to the technology. Both groups of physicians interact with each other so that information that flows directly to physicians with access to the technology can flow via recommendations to other physicians not included in the system.14 Let the subscript iA{1,2} reflect whether the physician has technology access (i ¼ 1) or does not have technology access (i ¼ 2). As in the previous section let z denote the number of recommendations a doctor of type i receives from colleagues, where z ¼ a (the average number of new treatments that physicians adopted in the previous period.). Because a physician will always read about all the care messages generated by the decision support technology and ymon, we have z ¼ aðbðym þ ry1 Þ þ ð1  bÞry2 Þ

(3)

As in the previous section, the expected benefit from reading about a treatment generated by a colleague’s recommendation is greater than the expected benefit from reading independently selected newly published articles (qWr). Thus all physicians will read all colleagues’ recommended treatments before reading independently selected journal articles. We focus our analysis on the equilibrium in which all read about some new treatments, i.e., ynþzon. In this case x1 ¼ x2 ¼ z and the number of independently selected journal articles a doctor chooses to read is given by y1 ¼ n  ym  z

(4)

y2 ¼ n  z

(5)

Eqs. (3)–(5) determine steady-state equilibrium yn1 ; yn2 and xn1 ¼ xn2 ¼ zn . The steady-state disposition of new knowledge with computer support software is depicted in Fig. 2. From this set-up it is easy to prove the following lemma.

y1

m

ry1  (1−)ry2

Physicians' independent reading of medical journals

IT Decision support system

Physicians implement x1

m

m m

Physicians read articals cited in error messages

s new treatments

Physicians with technology access () Physicians with technology access (1−)

y2

Physicians' independent reading of medical journals

m  (1−)ry2 ry1

ry2

Fig. 2.

Physicians' reading after recommendation from colleagues

Information Technology in Health Care

ry1

Physicians reading after recommendation from colleagues

Physicians implement

x2

The Steady-State Disposition of New Knowledge with Computer-Based Decision Support.

57

58

JAMES B. REBITZER ET AL.

Lemma 1. The effects of computer support systems on physician learning: If nWym(1þarþab(1r)), then there exist a steady-state equilibrium in which the doctors with access to the technology will read about all treatments recommended by the technology and by colleagues, and some treatments from independently selected journals. Doctors without the technology will read about all treatments recommended by colleagues and some treatments from independently selected journal articles. Doctors with the technology will read fewer independently selected journal articles than doctors without the technology. Proof. see Appendix We are interested in understanding how the decision support technology affects knowledge diffusion. For this purpose, we define knowledge diffusion among doctors of type i as: Z1 ðmÞ ¼ ryn1 þ ym þ qxn1

(6)

Z2 ðmÞ ¼ ryn2 þ qxn2

(7)

where Zi denotes the total number of relevant treatments that a doctor of type i will learn about and adopt in each period. As before, established treatments are discounted by q. We define the diffusion of knowledge about a treatment included in the technology and a treatment not included in the technology respectively as:       ryn1 ryn2 ryn1 ryn2 þ ð1  bÞ þq b yþ þ ð1  bÞ (8) mt ðmÞ ¼ b y þ s s s s  n   n  ry1 ryn2 ry1 ryn2 mnt ðmÞ ¼ b þ ð1  bÞ þq b þ ð1  bÞ s s s s

(9)

In Eqs. (8) and (9), the first term denotes the likelihood that doctors using the decision support technology adopt the treatment in the first period. The second term denotes the likelihood that doctors without the technology adopt the treatment in the first period. The last term denotes the likelihood that any doctor adopts the treatment in the second period as a result of the recommendations from colleagues. Comparing Eqs. (8) and (9) we can see that adding a new treatment protocol to the decision support data base increases the likelihood of adoption by y in the first period and by qy in the second period. Let o denote total adoption of new knowledge among all

Information Technology in Health Care

59

doctors. Note that Eqs. (6) and (7), and Eqs. (8) and (9) are necessarily related as follows: o ¼ ðs  mÞmnt þ mmt ¼ bZ1 þ ð1  bÞZ2

ð10Þ

By differentiating (6)–(10) with respect to m we establish the following proposition: Proposition 1. Expanding computer-based decision support enhances the rate of knowledge diffusion: Increasing the number of treatments in the technology increases overall knowledge diffusion among all doctors, even among those who do not have access to the technology. This holds, even though the rate of diffusion of treatments in the computer data base will not decline. Proof. see Appendix An important implication of the analysis so far is that the spread of IT-based decision support is likely to reduce geographic variation in practice patterns. To see this assume that: underlying patient clinical needs are the same across regions; and that the medical informatics company offers the same decision support technology in all the regions in which it operates. What differs across regions then are: (i) the degree, to which the technology has penetrated a region (captured by b) and (ii) the degree, to which physicians in a region rely upon their idiosyncratic reading of articles in medical journals. We have shown in Lemma 1 that the presence of IT displaces reading that is not stimulated by recommendations from colleagues or the IT system. Inspection of the Proof of Lemma 1 shows that as m increases, both steady state y1 and y2 decrease. Thus in a steady state as m increases, the idiosyncratic reading of journals (the sort of reading that leads to variation across regions) becomes less important for all physicians. The same proof also shows that as m increases, steady-state recommendations from colleagues (z) increases. By the logic of our model, the only reason that z would increase with m is that as m increases, the IT system identifies more relevant protocols in the first period in which they are released. This source of new information is, however, common across regions. Thus as m increases, cross-regional variation in observed practice patterns should decline.

60

JAMES B. REBITZER ET AL.

4.3. Private Incentives to Invest in Decision Support System Given the central role that health insurers currently play as information aggregators in our health care system, we begin by analyzing their incentives to invest in physician decision support systems. We simplify our analysis by assuming that the insurer is a HMO that must attract physicians to sign on to its network of providers (Cooper & Rebitzer, 2006). Maintaining our assumption that providers derive a direct benefit from offering their patients better care, the HMO will be able to attract physicians to the network at lower cost by investing in a decision support technology that improves care quality.15 The flow of new knowledge with decision support is depicted in Fig. 2. It is clear from this figure that doctors who do not have access to the decision support technology still benefit in period two from the learning induced by the computer technology in period one. This learning spillover leads to Corollary 1. To establish this, let p be the cost of adopting a new treatment. Then, the net private benefits of an insurer with technology access is given by Z1(Bp), whereas the total net benefit to providers is given by (Z1þZ2)(Bp). Thus, the Corollary follows from the fact that both Z1 and Z2 are increasing in m, which was established in Proposition 1. Corollary 1. Insurers will under-invest in computer-based decision support. The underinvestment in the new technology is the result of knowledge spillovers from one set of providers to another. The conventional economic response to positive externalities is to internalize them with subsidies financed by lump sum taxes. This approach is, however, famously difficult to implement. An alternative approach might be feasible in this case because what appears as externalities to health care providers and insurers are actually sources of revenue to other market actors. The pharmaceutical and device manufacturers profit from the informal spread of new information about their products and may therefore have more powerful incentives than providers and insurers to promote the spread of computer-based decision support tools. A pharmaceutical or device manufacturer’s increased profit from including a treatment in the decision support system is given by Dp ¼ ðmt ðm þ 1Þ  mnt ðmÞÞp

Information Technology in Health Care

61

The next proposition shows that for a sufficiently large price of treatment, p, the seller of the treatment is willing to pay more for the inclusion of a new treatment than would a provider or an insurer. Proposition 2. Pharmaceutical and drug makers can have more powerful incentives to invest in computer-based decision support than insurers: There exists a p1oB, s.t. if pWp1 then the seller of a new treatment is always willing to pay more than the HMO for including this treatment into the technology. Proof. see Appendix Proposition 2 has potentially important public policy implications. It suggests that one might create appropriate incentives to invest in computer support technology by allowing pharmaceutical firms and device manufacturers to invest in decision support technology. For concreteness imagine that this investment takes the form of paying the IT company who runs the decision support system to include their new products into the computer system’s data base. This sort of financing might avoid cumbersome government bureaucracy while ensuring that the health care system has a sufficiently comprehensive decision support technology. This strategy raises, however, a concern when one considers the cognitive limitations under which physicians operate. Paying for the inclusion of one or another company’s products in the computer support technology displaces other forms of learning that might lead physicians to adopt different treatments. This crowding out of learning about other products creates negative externalities and therefore creates incentives to over-invest in the decision support technology – a point made in the following proposition: Proposition 3. Pharmaceuticals and device makers may have incentives to over-invest in computer-based decision support technology: There exists a p2oB, s.t. if pWp2 then the seller of a new treatment is willing to pay more than the social marginal benefit for including this treatment in the technology. Other pharmaceuticals will bear the full cost of this overinvestment. Health care providers and insurers will benefit from the overinvestment. Proof. see Appendix If one overlooks the effects of wasted resources resulting from overinvestment by pharmaceutical and device makers, Proposition 3 seems to suggest that private financing of decision support systems may be a sensible

62

JAMES B. REBITZER ET AL.

way to offset the investment shortfall resulting from knowledge spillovers. Just as Google finances enhanced information flow on the internet through the use of marketing dollars, so might the marketing efforts of pharmaceutical and device makers be used to support enhanced information flows about treatment protocols in the health care system. As we demonstrate in the next section, however, this conclusion rests on unrealistically strong assumptions about the unambiguous nature of new knowledge.

4.4. Ambiguity, Influence and Marketing We have so far considered messages whose information content is unambiguously correct. In many settings, however, the right drug or treatment is less clear. Suppose, for example, that a pharmaceutical company has a patent for an anti-ulcer medication, a class of drugs for which there exist a number of competing brands (Berndt et al., 2003). Assume that all of these anti-ulcer drugs are effective, but no single drug dominates the others. Rather each drug works better for a subset of the ulcer population, but it is hard to identify ex-ante which patient would benefit most from which drug. In this setting the pharmaceutical company might profitably subsidize a message that suggests this drug to all physicians with ulcer patients because the message may prove relevant to some subset of physicians. We will refer to this sort of message as a marketing message because its expected relevance to the treatment of a specific ulcer patient is low. As before, assume that the technology sends each physician ym targeted messages. In addition the technology sends marketing messages for k treatments. For each patient, there is some probability, dW0, that the marketing message will prove relevant to some of his patients, but the relevance of the marketing message is less than other messages sent by the system so that doy. Doctors know that the decision support software includes both marketing and non-marketing messages, but cannot identify which individual message is the result of marketing. Thus doctors respond to each message by consulting the cited literature and deciding whether the recommended drug is appropriate for their patient. The likelihood that the message is suggesting a new and relevant treatment is given by l¼

dk þ ym k þ ym

Proposition 4 states that marketing messages can have a negative impact on the rate of diffusion of new knowledge and overall welfare. The intuition

Information Technology in Health Care

63

is very simple: As long as lWr, the expected relevance of messages sent by the decision support tool exceeds that available through other channels. Thus a doctor will prefer reading about a treatment recommended by the decision support tool to any other learning modality, despite his knowledge that the messages include k marketing messages. Getting doctors to read the k marketing messages is profitable for the drug company because some of these drugs will actually prove relevant to a patient and therefore generate sales. Offsetting this private benefit, however, is the opportunity cost of inducing physicians to read about relatively low value suggestions contained in marketing messages. Such reading displaces physicians’ reading about treatments from independently selected journal articles. This displacement has a negative impact on knowledge diffusion if dor, i.e., the expected relevance of a marketing message is lower than the expected relevance of independent reading in journal articles. This reasoning leads directly to the following proposition: Proposition 4. Marketing messages: If dor, then the inclusion of marketing messages has a negative impact on the rate of diffusion of new knowledge and overall welfare, even though the marketing firm will benefit from such a message. Proof. see Appendix This result suggests that although it might be privately profitable to use decision support software as a sophisticated cybernetic drug detailer, doing so could undermine the social benefits produced by the technology.

5. CONCLUSION Physicians are overwhelmed by the task of managing vast amounts of information relating to patient conditions and new treatment protocols. This information overload acts as a drag on the diffusion of new knowledge. ITenabled decision support tools have attracted the attention of health care providers and policy makers because they offer a way to alleviate this drag and hence to improve care quality. Increasing the rate of diffusion of new knowledge might also have the effect of increasing returns to innovative activity and therefore stimulate more rapid innovation. Just as small increases in productivity growth rates accumulate over time to transform living standards; so enhanced innovation rates resulting from the faster diffusion of new knowledge could yield transformations in the quality of health care.

64

JAMES B. REBITZER ET AL.

Because of the fragmented state of the IT infrastructure in the health care system, insurance companies currently play a central role as information aggregators and are therefore in a good position to finance investments in decision support. Our model suggests, however, that insurers’ incentives to invest will generally be inadequate. The problem is the result of knowledge spillovers: doctors who learn about a new treatment protocol from the IT system will likely transmit that information to other physicians through informal interactions. These spillovers improve care quality, but the benefits do not accrue to the insurers financing the decision support software. Our analysis of limited mental ‘‘shelf space’’ and clinical ambiguity suggests that alternative approaches, such as financing investment in improved information flows with marketing dollars, may undermine the social value generated by decision support technology. Given the current surge in investments in health care IT and the positive and negative externalities inherent in these investments, the role of public policy in guiding these investments should be an important area of future research.

NOTES 1. Reference to the use of IT to reduce errors appeared in President Bush’s Economic Report of the President in 2004 (President’s Council of Economic Advisors, 2004) and in his state of the Union Address. ‘‘By computerizing health records, we can avoid dangerous medical mistakes, reduce costs, and improve care.’’(President’s Information Technology Advisory Committee, 2004). 2. Of course not all slow diffusion of innovations is ‘‘too slow.’’ Some new innovations may diffuse gradually because it takes time for providers to fully exploit the potential of the treatment for cost reductions and quality improvement. This appears to be the case, for example, with the use of percutaneous transluminal coronary angioplasty (PTCA), an important alternative to bypass surgery for patients with coronary artery disease (see Cutler & Huckman, 2003). 3. For a discussion of this literature from an economist’s perspective see (Phelps, 1992, 2000). For a comparison of sociological and economic perspectives in the diffusion of innovations see (Skinner & Staiger, 2005). 4. A recent study of the use of anti-psychotics in Medicaid patients (Duggan, 2005) documents the heavy use of expensive new generation anti-psychotics in spite of evidence that they perform no better than older and less-expensive anti-psychotics. 5. The authors report that the elasticity of minutes of drug detailing to equilibrium market share is around one, suggesting that physicians are highly susceptible to this form of influence. 6. In the specific case of anti-ulcer drugs, the authors found that consumption externalities were not strong enough to prevent a superior late entrant, Zantac, from

Information Technology in Health Care

65

overcoming the initial advantage of an entrenched incumbent, Tagamet. Zantac overcame it’s disadvantage due to its superior medical qualities and also due to unusually large expenditures on drug detailing (Berndt et al., 2003) 7. For detailed descriptions of this trial see Javitt et al. (2005); and Javitt, Rebitzer, and Reisman (2008). 8. Although the system, we describe used cutting edge information technology, the source of the information (billing records, and data feeds from pharmacies and labs) ensured that physician’s information about the patient was almost always superior to that available to the computer system. For these reasons, the messages generated by the system were communicated as recommendations that the physician should feel free to ignore. Thus it was not at all certain at the start of the study that the messages would have any influence at all. 9. One might argue that specialists are at least as good as computer data bases in keeping up with the new developments in their discipline. This may be so, but computer-based decision support systems have the advantage of giving primary providers access to the relevant knowledge of specialists in many disciplines. 10. We could easily endogenize n by including doctor’s effort cost of reading about treatments. This would complicate the analysis, but not add any additional insight. 11. Relaxing this assumption by allowing the technology to send messages about both new and established treatments would complicate the analysis considerably, but not affect the results. 12. Assuming instead that the technology also sent out some erroneous messages would not change the results of our analysis as long as the likelihood that a message is useful is larger than q. 13. ‘‘ . . . to be effective, CDSS [clinical decision support systems] diagnostic systems require detailed, patient-specific clinical information (history, physical results, medications, laboratory test results), which in most health care settings resides in a variety of paper and automated datasets that cannot easily be integrated. Past efforts to develop automated medical record systems have not been very successful because of the lack of common standards for coding data, the absence of a data network connecting the many health care organizations and clinicians involved in patient care, and a number of other factors’’ (Institute of Medicine Committee on Quality of Health Care in America, 2001, p. 154). 14. Our analysis would not change if we assume that share (1b) of physicians had access to a different computer-based system. 15. Alternatively we could assume that the purchasers of health insurance would pay more for insurance if the insurer supplied decision support technology to physicians that improved care outcomes. Because physicians in our setup experience the same utility from improved outcomes as patients do, the results under these assumptions would be unchanged.

REFERENCES Altman, L. K. (2005). Studies find disparity in US cancer care. New York Times. Avorn, J. (2004). Powerful medicines: The benefits, risks and costs of prescription drugs. New York: Alfred A. Knopf.

66

JAMES B. REBITZER ET AL.

Beaulieu, N., Cutler, D., Ho, K., Isham, G., Lindquist, T., Nelson, A., & O’connor, P. (2006). The business case for diabetes disease management for managed care organizations. Forum for Health Economics & Policy, 9, 1–36. Berndt, E. R., Pindyck, R. S., & Azoulay, P. (2003). Consumption externalities and diffusion in pharmaceutical markets: Antiulcer drugs. Journal of Industrial Economics, 51, 243–271. Brynjolfsson, E., & Hitt, L. M. (2000). Beyond computation: Information technology, organizational transformation and business performance. Journal of Economic Perspectives, 14, 23–48. Burt, C., & Hing, E. (2005). Use of computerized clinical support systems in medical settings: United States, 2001–2003. Advance Data From Vital and Health Statistics. Hyattsville, MD, National Center for Health Statistics. Chaudhry, B., Jerome, W., Shinyi, W., Maglione, M., Mojica, W., Roth, E., Morton, S. C., & Shekelle, P. G. (2006). Systematic review: Impact of health information technology on quality, efficiency, and costs of medical care. Annals of Internal Medicine, 144, E12–E22. Cooper, D. J., & James, B. R. (2006). Managed care and physician incentives: The effects of competition on the cost and quality of care. B.E. Journals in Economic Analysis and Policy: Contributions to Economic Analysis and Policy, 5(1), 1–30. Cutler, D. M., Feldman, N. E., & Horwitz, J. R. (2005). US adoption of computerized physician order entry systems. Health Affairs, 24, 1654–1663. Cutler, D. M., & Huckman, R. S. (2003). Technological development and medical productivity: The diffusion of angioplasty in New York state. Journal of Health Economics, 22, 187–217. DesRoches, C. M., Campbell, E. G., Rao, S. R., Donelan, K., Ferris, T. G., Jha, A., Kaushal, R., Levy, D. E., Rosenbaum, S., Shields, A. E., & Blumenthal, D. (2008). Electronic health records in ambulatory care – A national survey of physicians. New England Journal of Medicine, 359, 50–60. Duggan, M. (2005). Do new prescription drugs pay for themselves? The case of secondgeneration antipsychotics. Journal of Health Economics, 24, 1–31. Garg, A. X., Adhikari, N. K. J., Mcdonald, H., Rosas-Arellano, M. P., Devereaux, P. J., Beyene, J., Sam, J., & Haynes, R. B. (2005). Effects of computerized clinical decision support systems on practitioner performance and patient outcomes: A systematic review. JAMA: Journal of the American Medical Association, 293, 1223–1238. Gertler, P.J., & Simcoe, T.S. (2006). Disease management. SSRN. Institute of Medicine Committee on Quality of Health Care in America. (2000). In: L. T. Kohn, J. M. Corrigan & M. S. Donaldson (Eds), To err is human: Building a safer health system. Washington: National Academy Press. Institute of Medicine Committee on Quality of Health Care in America. (2001). Crossing the quality chasm: A new health care system for the 20th century. Washington: National Academy Press. Javitt, J. C., Rebitzer, J. B., & Reisman, L. (2008). Information technology and medical missteps: Evidence from a randomized trial. Journal of Health Economics, 27, 585–602. Javitt, J. C., Steinberg, G., Couch, J. B., Locke, T., Jacques, J., Juster, I., & Reisman, L. (2005). Use of a claims data-based sentinal system to improve compliance with clinical guidelines: Results of a randomized prospective study. American Journal of Managed Care, 11(2), 21–31. Mullaney, T. J. (2005). A booster shot for medical data-sharing. Business Week Online. Phelps, C. E. (1992). Diffusion of information in medical care. Journal of Economic Perspectives, 6, 23–42.

Information Technology in Health Care

67

Phelps, C. E. (2000). Information diffusion and best practice adoption. Handbook of health economics (Vol. 1A). Amsterdam: Elsevier Science President’s Council of Economic Advisors. (2004). Economic report of the President. Washington: United States Government Printing Office. President’s Information Technology Advisory Committee. (2004). Revolutionizing health care through information technology. Arlington: Executive Office of the President of the United States. Skinner, J., Fisher, E., & Wennberg, J. E. (2001). The efficiency of medicare. Working Paper Series no. 8395. National Bureau of Economic Research, Inc. Skinner, J. & Staiger, D. (2005). Technology adoption from hybrid corn to beta blockers. Working Paper no. 11251. National Bureau of Economic Research, Inc.

APPENDIX Proof of Lemma 1. Solving Eqs. (3)–(5) for y1, y2, and z implies: n  abymð1  rÞ  ym 1 þ ar n  abymð1  rÞ y2 ¼ 1 þ ar arn þ abymð1  rÞ z¼ 1 þ ar

y1 ¼

This denotes an equilibrium if q1, q2, and zW0. Thus, we must have nWym(1þarþab(1r)). Proof of Proposition 1. Differentiating Eqs. (6)–(10) with respect to m yields: qZ1 1 þ arð1  bÞ þ qab ¼ yð1  rÞ 40 qm 1 þ ar qZ2 qr ¼ abyð1  rÞ 40 qm 1 þ ar qmt 1 þ qa ¼  bryð1 þ aÞ o0 qm sð1 þ arÞ qmnt 1 þ qa ¼  bryð1 þ aÞ o0 qm sð1 þ arÞ qo qa þ 1 ¼ byð1  rÞ 40 qm 1 þ ar

68

JAMES B. REBITZER ET AL.

Proof of Proposition 2. First note that Dp ¼ ðmt ðm þ 1Þ  mnt ðmÞÞp s þ sar  r  ar 40 ¼ byðqa þ 1Þ sð1  arÞ The pharmaceuticals’ willingness to pay for including the treatment is Dp  p, whereas an HMO’s willingness to pay is Z1(Bp). Thus, the pharmaceutical is willing to pay more than the HMO if pW(Z1/DpþZ1)B. Since DpW0 this implies that there exist a p1oB, s.t. if pWp1 then the seller of a new treatment is always willing to pay more than the HMO for including this treatment into the technology. Proof of Proposition 3. The net benefits of the HMOs are given by o(Bp). The net benefits of the pharmaceuticals are op. Thus, the social marginal net benefit is (@o/@m)B. Using the calculations from the Proof of Proposition 1 and 2, the pharmaceutical’s marginal net benefit is larger than the social marginal net benefit if byðqa þ 1Þ

s þ sar  r  ar 1 þ aq p4byð1  rÞ B sð1  arÞ 1 þ ar

This implies pW((1r)s/sþasrrar)B, where ((1r)s/sþasrrar)o1 since sW1. Thus, there exists a p2oB, s.t. if pWp2 then the seller of a new treatment is willing to pay more than the social marginal benefit for including this treatment in the technology. Proof of Proposition 4. Now we need to distinguish between diffusion of three different types of treatments: treatments in the technology that are marketed, treatments in the technology that are not marketed, and treatments those are not in the technology. Diffusion of a treatment in the technology that is marketed is given by:       ryn1 ryn2 ryn1 ryn2 þ ð1  bÞ þq b dþ þ ð1  bÞ m^ t ðmÞ ¼ b d þ s s s s Thus, total adoption is given by o ¼ ðs  m  kÞmnt þ mmt þ km^ t ¼ ðbymð1  rÞ þ nr þ bkðd  rÞÞ

1 þ aq 1 þ ar

Information Technology in Health Care

69

We can see that qo/qko0 if dor. Thus, if dor, then the inclusion of marketing messages has a negative impact on the rate of diffusion of new knowledge. By comparing m^ t ðmÞ to mnt(m) we can see that adding a new treatment protocol to the decision support data base increases the likelihood of adoption by bd in the first period and by qbd in the second period. Thus, the pharmaceuticals have positive willingness to pay for inclusion of marketing messages.

HEALTH DISPARITIES AND DIRECT-TO-CONSUMER ADVERTISING OF PHARMACEUTICAL PRODUCTS$ Rosemary J. Avery, Donald Kenkel, Dean R. Lillard, Alan Mathios and Hua Wang 1. INTRODUCTION Health information drives crucial consumer health decisions and plays a central role in healthcare markets. Consumers who are better-informed about smoking, diet, and physical activity make healthier choices outside the healthcare sector (Kenkel, 1991; Ippolito & Mathios, 1990, 1995; Meara, 2001). Better-informed consumers also interact differently with physicians and other healthcare providers (e.g., Cutler, Landrum, & Stewart, 2006). In addition to the immediate consequences for individual consumers, health economists have long recognized that information also has broader implications for principal–agent relationships and the functioning of healthcare markets.1 More recent lines of research in health economics and medical sociology emphasize the potential role of consumer information in explaining health $

This is a revised version of a paper presented at ‘‘Beyond Health Insurance: Public Policy to Improve Health,’’ November 15–16, 2007, University of Illinois at Chicago.

Beyond Health Insurance: Public Policy to Improve Health Advances in Health Economics and Health Services Research, Volume 19, 71–94 Copyright r 2008 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0731-2199/doi:10.1016/S0731-2199(08)19004-5

71

72

ROSEMARY J. AVERY ET AL.

disparities associated with socioeconomic status (Deaton, 2002; Goldman & Lakdawalla, 2001; Glied & Lleras-Muney, 2003; Link & Phelan, 1995). Both health economists and medical sociologists stress that because of disparities in consumer information, rapid medical progress tends to be accompanied by increased disparities in medical treatment and health outcomes. The link between medical progress and health disparities raises the concern that rapid advances in the pharmaceutical industry lead to increased disparities in pharmaceutical use. As discussed in more detail below in Section 2, research suggests that socio-economically disadvantaged groups and racial/ethnic minority groups are less likely to be treated with newer pharmaceutical products. However, the recent trend toward direct-toconsumer (DTC) advertising provides a source of information for consumers across the socioeconomic spectrum. In general, DTC advertising is most common for new drugs that treat chronic conditions (Donohue et al., 2007). With the advent of DTC advertising on television beginning in the late 1990s, the pharmaceutical industry began to use a powerful and farreaching medium to promote its new products. In this paper, we explore how consumer exposure to DTC advertising on television varies with socioeconomic status and race. As discussed in detail in Section 3, we create and use unique data on individual consumers’ potential exposure to DTC advertisements. Our individual-level data come from 10 waves of the Simmons National Consumer Survey (NCS) administered from 1999 to 2004. The combined waves of the NCS provide rich data for over 80,000 respondents’ consumer and media behavior, including their television-viewing habits. We merge the NCS data with data from TNS Media Intelligence on the airings of DTC advertisements for products that treat conditions ranging from high cholesterol to toenail fungus. We measure an individual’s potential advertising exposure based on the number of DTC advertisements that aired during the specific programs or time slots he or she reports regularly watching. Section 4 present results from descriptive multiple regression models that explore how advertising exposure varies with a consumer’s employment status, race/ethnicity, schooling, income, and health insurance status. Section 5 concludes.

2. BACKGROUND ON HEALTH DISPARITIES, USE OF PHARMACEUTICALS, AND DTC ADVERTISING Health disparities refer to differences in health outcomes and treatments associated with socioeconomic status. Differences in life expectancy

Health Disparities and Direct-to-Consumer Advertising

73

and mortality provide some of the starkest examples. In the National Longitudinal Mortality Study, the life expectancy of people with family incomes below $5,000 in 1980 was 25 percent lower than people with family incomes above $50,000 (Rogot, Sorlie, Johnson, & Schmit, 1992). According to National Vital Statistics data, life expectancy at birth in 2002 for whites was 77.7 years, compared to 72.3 years for Blacks (Arias, 2004). The Healthy People 2010 report cites evidence that heart disease death-rates are more than 40 percent higher for Blacks than whites, and that cancer death-rates are 30 percent higher for Blacks than whites (US Department of Health and Human Services, 2000). Smith (1999) and Deaton (2002) review evidence on the strong gradient between wealth and health outcomes, while Cutler and Lleras-Muney (2006) and Grossman (2006) summarize the large body of research on the gradient between schooling and health. Outside economics, more attention has focused on the link between social class and health. The Whitehall studies document a steep gradient between the employment grade of British civil servants and health outcomes (Marmot et al., 1991). It is notable that this gradient is estimated for a population of mostly office workers with stable employment and who had universal access to Britain’s National Health Service. Reducing health disparities is a major public policy challenge. The Healthy People 2010 public health initiative sets out two major goals: to increase quality and years of healthy life; and to eliminate health disparities. However, Mechanic (2002) and Keppel, Bilhelmer, and Gurley (2007) stress that progress toward these two goals may not necessarily coincide. Based on the mid-course review of the Healthy People 2010 initiative, Keppel et al. report that, for 69 specific objectives, the outcome was progress toward increasing the quality and length of life with little or no change in relative disparity. For ten objectives progress toward the two goals actually moved in opposite directions. The Agency for Healthcare Research and Quality’s National Healthcare Disparities Report finds that ‘‘disparities related to race, ethnicity, and socioeconomic status still pervade the American healthcare system,’’ and documents disparities across all dimensions of quality and access to healthcare (AHRQ, 2006). The causes of health disparities are very difficult to determine, but some researchers suggest that they may partly be the unintended, and perhaps unavoidable, consequence of medical progress. In an influential paper in medical sociology, Link and Phelan (1995, p. 87) stress the importance of ‘‘fundamental causes of disease’’ that ‘‘involve access to resources that can be used to avoid risks or to minimize the consequences of disease . . . .’’ They define resources broadly not only to include money but also knowledge and other intangible resources. Link and Phelan argue that a necessary condition

74

ROSEMARY J. AVERY ET AL.

for fundamental causes to emerge is ‘‘change over time in the diseases afflicting humans, the risks for those diseases, knowledge about risks, or the effectiveness of treatments for diseases.’’ In his article reviewing the gradient between wealth and health Deaton (2002) emphasizes ‘‘the possibility that widening gradients are related to life-saving bursts of technical progress.’’ Theoretical and empirical health economics studies of the relationship between health disparities and medical progress include Goldman and Lakdawalla (2001), Lleras-Muney and Lichtenberg (2002), and Glied and Lleras-Muney (2003). If health disparities are linked to medical progress, the high rate of innovation in the pharmaceutical industry might be expected to result in disparities in pharmaceutical use. The research literature on disparities in healthcare is extremely broad and difficult to summarize. Still, several pieces of evidence are consistent with the prediction that innovations have led to disparities in pharmaceutical use. Lleras-Muney and Lichtenberg (2002) and Wang et al. (2007) use the age of a drug, defined as the number of years since approval by the Food and Drug Administration (FDA), to measure innovation. In data from the Medical Expenditure Panel Survey (MEPS), both studies find evidence of disparities in new drug utilization related to education, race, and insurance status. Nelson, Norris, and Mangione (2002) use data from the Third National Health and Nutrition Examination Survey (NHANES) conducted between 1988 and 1994 to study the utilization of pharmaceutical products to treat high cholesterol. The NHANES data capture drug utilization around the beginning of the market for statins, an important innovation in the treatment of high cholesterol: the first statin was approved in 1987, followed by additional new approvals in 1991 and 1993 (Yang, Mathios, & Avery, 2007). Nelson et al. find that Blacks and Mexican Americans are less likely to be screened for cholesterol. Of those advised to take a prescription drug, Blacks and Mexican Americans were also less likely to be taking a cholesterol drug. The potential of DTC advertising to affect health disparities has emerged because of fairly recent regulatory changes. In 1985 the FDA lifted a moratorium on print DTC advertisements for prescription products. However, until another regulatory change in 1997, it was impractical for most television advertisements to meet the FDA required disclosures of side effects and contraindications. In 1997 the FDA relaxed the disclosure requirements for DTC advertisements on television and radio, marking the beginning of the modern DTC era. Over the decade from 1996 to 2005, the pharmaceutical industry’s expenditures on DTC advertising increased by 330 percent, from $985 million in 1996 to $4.2 billion in 2005.

Health Disparities and Direct-to-Consumer Advertising

75

Because the pharmaceutical industry tends to advertise new drugs, DTC advertising has the potential to reduce innovation-related disparities in pharmaceutical utilization. Donohue, Cevasco, and Rosenthal (2007) broadly summarize DTC advertising patterns: ‘‘Drugs that are advertised to consumers are predominantly new drugs used to treat chronic conditions.’’ In particular, they note that of the 20 most-advertised drugs in 2005, 10 were introduced in 2000 or later; and 17 of the 20 advertising campaigns began within a year after FDA approval. The pharmaceutical industry faces strong incentives to advertise new products. Advertisements for new products might be more effective in stimulating demand because they provide consumers with new information. For some products, such as cholesterol medications, DTC advertisements inform consumers that they might have an asymptomatic but treatable condition. For other products, such as anti-depressants, DTC advertisements inform consumers that an effective medical treatment exists for their symptoms. Advertisers generally use ‘‘new and improved’’ claims in advertising; in a content analysis of DTC print advertisements from 1989 to 1998, 40 percent used claims of ‘‘innovativeness’’ (Wilkes, Bell, & Kravitz, 2000). Finally, because pharmaceutical companies earn most of their profits from products under patent protection, they have an especially strong incentive to advertise their newest products.2

3. DATA ON CONSUMER EXPOSURE TO TELEVISION DTC ADVERTISEMENTS 3.1. Measuring Potential Exposure to DTC Advertising To measure individuals’ potential exposure to DTC advertising we link data on individuals’ television-viewing habits with data on television advertisements of pharmaceutical products. The individual-level data come from the Simmons National Consumer Survey (NCS). The NCS is a repeated crosssectional survey, where the sample for each wave is an independently drawn multi-stage stratified probability sample. Because it is a marketing survey, higher-income households were intentionally over-sampled. We use data from 10 NCS waves administered from 1999 to 2004. The data on television advertisements come from TNS Media Intelligence. We have data on televisions advertisements for prescription-only and over-the-counter pharmaceutical products. We use TNS data on advertisements that aired

76

ROSEMARY J. AVERY ET AL.

from 1998 to 2004 on national networks, cable, and spot markets identified by Designated Marketing Areas (DMAs). From 1998 to 2001 the TNS data cover the largest 75 DMAs; from 2002 to 2004 the TNS data cover the largest 100 DMAs.3 We measure consumers’ potential exposure to pharmaceutical advertisements that appeared over the past year during programs or time slots they report regularly watching.4 To match advertisements that appeared in spot markets, we need information on the consumer’s DMA of residence. We therefore limit the NCS sample to the approximately 80,000 consumers with an identified DMA of residence; this is about 70 percent of the full NCS sample for the waves we use.5 Residents of smaller DMAs tend to be excluded from our analysis sample, but otherwise the excluded and included respondents appear to be generally similar. We use responses to several sets of NCS questions about televisionviewing to match aired advertisements to respondents. First, in each survey wave NCS respondents were asked about their viewing habits for a list of about 300–400 broadcast television programs and almost as many cable television programs. Second, NCS respondents were asked about their viewing, by time slot, of broadcast television on a typical weekday and typical weekend. Third, NCS respondents were asked about their viewing, by time slot, of specific cable networks. We measure respondents as potentially exposed to: advertisements that aired during specific programs they watched; advertisements that aired on other broadcast programs during time slots they watched; and advertisements that aired during time slots and on specific cable networks they watched.6 Our measure of advertising exposure captures about 85 percent of aired pharmaceutical advertisements in the TNS data. The degree to which our measure under-states a consumer’s exposure to advertising depends upon his or her television-viewing habits. Our measure under-states advertising exposure more for consumers who watch relatively more television programming not included on the NCS lists of program titles or time slots.7

3.2. Overview of the Data We analyze exposure to advertisements for pharmaceutical products that treat ten categories of health conditions: adults’ allergies; children’s allergies; arthritis; asthma; cholesterol; depression; erectile dysfunction;

Health Disparities and Direct-to-Consumer Advertising

77

insomnia; sexually transmitted diseases; and toenail fungus. Table 1 lists the categories and the most heavily advertised products within each category. We chose the categories to include products that treat a range of health conditions as well as more controversial products. As an example of a more controversial product, Wilkes et al. (2000) use a vignette about a patient asking for a product to treat toenail fungus as an example of one of the drawbacks of DTC advertising. As another controversial example, DTC advertisements for a new class of arthritis medications raised serious concerns after Vioxx was withdrawn from the market due to safety concerns (Vogt, 2005). Our categories include many but not all of the most heavily advertised products. Our list includes 16 of the 20 products with the highest DTC advertising spending in 2000 (Frank, Berndt, Donohue, Epstein, & Rosenthal, 2002). Over the time period and categories we analyze, products that treat adults’ allergies and arthritis are the most heavily advertised. By our measure, over the course of twelve months the average NCS respondent in our sample was exposed to almost 1,500 advertisements for adults’ allergy products and about 750 advertisements for arthritis products. The average NCS respondent was exposed to hundreds of advertisements per year for products in each of most of the remaining categories. Fig. 1 shows the time trend in exposure to television advertisements. We sum advertisements for all products in all 10 categories. Average annual advertising exposure more than doubled, from slightly below 2,000 advertisements in 1999 to nearly 5,000 advertisements by 2003. Although they are not shown in Fig. 1, the trends for some of the individual product categories are even more dramatic. As two examples: from 1999 to 2003 average exposure to advertisements for products that treat cholesterol and for products that treat erectile dysfunction increased roughly fivefold. The different trends across product categories probably reflect different market conditions. In the market for erectile dysfunction products, for example, Viagra’s launch in March 1998 was just prior to our study period, and it was followed by the launches of Cialis and Levitra in 2003. The sharp increase in advertising for these products is consistent with what appears to be a common strategy firms use to heavily advertise new products (Avery, Kenkel, Lillard, & Mathios, 2007a; Donohue et al., 2007). The trend in Fig. 1 for total DTC advertising exposure is an average of more variable trends in advertising in the different product categories. We are not aware of comparable estimates of consumers’ exposure to DTC advertisements for pharmaceutical products. Previous surveys of consumers find that exposure to DTC advertisements is ‘‘nearly universal,’’

78

ROSEMARY J. AVERY ET AL.

Number of TV Ads and Average Exposure. Product Category

Product

Allergy (adult) Allegra Flonase Singulair Zyrtec Other Allergy (children) Children’s claritin al Children’s tylenol che Dimetapp liq Triaminic Other Arthritis Aleve cap Celebrex Tylenol Vioxx Other Asthma Accolate Advair Flovent Serevent Cholesterol Crestor Lipitor Pravachol Zocor Depression Paxil anti Paxil cr Wellbutrin xl Zoloft Other Erectile Dysfunction Cialis ed Levitra Viagra Insomnia Alluna Ambien Simply sleep aid Unisom max Other

Total Ads Number (or % of Total)

Average Exposure (Ads per Year)

1,816,617 17% 15% 13% 18% 37% 70,943 15% 13% 20% 17% 35% 1,205,557 23% 26% 13% 20% 18% 289,652 7% 52% 33% 9% 486,832 3% 56% 7% 34% 531,092 29% 15% 12% 28% 16% 289,849 8% 15% 77% 348,565 7% 45% 7% 29% 12%

1,475

32

752

187

324

279

249

245

79

Health Disparities and Direct-to-Consumer Advertising

(Continued ) Product Category

Product

STDs Valtrex Toenail Fungus Fungi-nail Lamisil

Total Ads Number (or % of Total)

Average Exposure (Ads per Year)

305,943 100% 76,575 1% 99%

162 75

6000 5000 4000 3000 2000 1000 0 Ap rJu 99 nAu 99 gO 99 ct D -99 ec Fe -99 bAp 00 rJu 00 nAu 00 gO 00 ct D -00 ec Fe -00 bAp 01 rJu 01 n Au -01 gO 01 ct D -01 ec Fe -01 bAp 02 rJu 02 nAu 02 gO 02 ct D -02 ec Fe -02 bAp 03 rJu 03 nAu 03 gO 03 ct -0 3

Number of advertisements exposed to in past year

Data Source: National Consumer Survey waves 21–39 and CMR/TNS Media Intelligence’s Pharmaceutical TV Ads Data 1998–2004.

Middle month of survey period

Fig. 1. Overall Exposure to TV Advertisements for Medications Treating 10 Health Conditions. Source: Authors’ calculations from TNS/Media Intelligence advertisements and NCS data.

but do not measure the extent of that exposure (Weissman et al., 2003). Other descriptive and econometric research uses data on DTC advertising expenditures. The general trend in exposure in Fig. 1 is consistent with general trends DTC advertising expenditures (Donohue et al., 2007).

80

ROSEMARY J. AVERY ET AL.

Although not perfect, our measures of individual consumers’ potential exposure to DTC advertisements offer several advantages over data on advertising expenditures. Most importantly for this study, we are able to explore whether consumers of different socioeconomic status are systematically exposed to different levels of DTC advertising. This set of research questions would be very difficult to address using advertising expenditure data, even disaggregated to the market level (DMAs). Average socioeconomic characteristics do not vary that much across DMAs. Moreover, advertising exposure varies much less across market-averages than it varies across individuals within markets. For example, in 2000 the highest marketaverage exposure to arthritis advertising (about 1200 advertisements) was only twice the lowest market-average exposure (about 600 advertisements). In contrast, within some of the largest DMAs, the individual at the 90th percentile of exposure potentially saw over 25 times more arthritis advertisements than the individual at the 10th percentile of exposure (around 1800 advertisements compared to around 70 advertisements).

4. CONSUMER EXPOSURE TO DTC ADVERTISEMENTS 4.1. Main Results In this section, we report the results of descriptive multiple regression models of consumers’ exposure to television advertisements for pharmaceutical products. Each model includes the same set of explanatory variables that measure employment status, age, sex, race/ethnicity, schooling, income, health insurance, region, county size, and survey wave. Table 2 provides the sample means for the explanatory variables. Table 3 presents the results from ten separate regressions, one for each product category. The sample size for the models is 80,615 respondents. In terms of socioeconomic disparities, there are a number of statistically and practically significant results. The results consistently show that consumers not working full-time potentially saw more advertisements in all of the product categories. The disabled are exposed to the most advertisements, followed by the retired, and the unemployed. Another consistent and strong pattern is that Blacks are exposed to more DTC advertisements. In contrast, Hispanics appear to be exposed to fewer DTC advertisements. However, an important caveat is that our measure of

Health Disparities and Direct-to-Consumer Advertising

81

Descriptive Statistics of Socio-Demographic Variables. Variables Age Age Age Age Age Age

Mean

18–24 25–34 35–44 45–54 55–64 65 or older

0.10 0.16 0.22 0.21 0.15 0.16

Female

0.55

White or other races Black Hispanics

0.78 0.07 0.15

Less than high school High school Some college College and higher

0.12 0.27 0.28 0.33

Married Divorced or separated Widow Single

0.64 0.10 0.06 0.19

Number of children in household

0.82

Income Income Income Income Income

0.17 0.20 0.18 0.22 0.23

quintile quintile quintile quintile quintile

1 2 3 4 5

Private health insurance or medicare Medicaid No health insurance

0.68 0.06 0.26

Employed full-time Employed part-time Retired Unemployed Disabled Student Homemaker

0.53 0.13 0.16 0.04 0.03 0.02 0.08

Northeast Midwest South West

0.26 0.24 0.28 0.22

82

ROSEMARY J. AVERY ET AL.

(Continued ) Variables County size A (21 largest metropolitan areas) County size B (W85,000 households) County size CD (other counties) Wave 21 Wave 23 Wave 25 Wave 27 Wave 29 Wave 31 Wave 33 Wave 35 Wave 37 Wave 39 Observations

Mean 0.78 0.16 0.07 0.13 0.13 0.08 0.10 0.08 0.09 0.08 0.07 0.12 0.12 80,615

Source: Simmons NCS waves 21–39.

exposure does not reflect advertisements that appeared on Spanish-language television networks. The college-educated are exposed to fewer advertisements, but the relationship between schooling and exposure is not monotonic for every product category. Exposure across income quintiles does not vary strongly or consistently across all product categories. There is some tendency for low-income consumers to be exposed to more advertisements for some products, such as, for asthma and arthritis products. Finally, consumers who lack health insurance are exposed to fewer advertisements for each of the product categories. Fig. 2A and 2B illustrate the magnitude of the difference in predicted exposure of a reference group and Blacks, the unemployed, and the uninsured. To create the figures, we use the estimated regression coefficients to predict exposure to television advertisements of consumers with different characteristics. The baseline reference group consists of females, aged 35–44, employed full-time, white, college-educated, in the highest income quintile, married, no children, in the South, in the largest metropolitan areas, who were surveyed in NCS wave 29. The other bars in the figures compare the predicted exposure for Blacks, the unemployed, and the uninsured to the reference group, holding all else constant. The illustrated differences are statistically significant; for the relevant standard errors the reader is referred to Table 3. In practical terms, depending on the product category Blacks and unemployed consumers are usually exposed to about one-third more DTC

Health Disparities and Direct-to-Consumer Advertising

83

Regression Results (1). Allergy

Age 18–24 Age 25–34 Age 35–44 Age 45–54 Age 55–64 Female Black Hispanics Less than high school High school Some college Divorced or separated Widow Single Number of children in household Income quintile1 Income quintile2 Income quintile3 Income quintile4 Medicaid No health insurance Employed part time

Allergy (Children)

6.189 256.6 (30.7) (1.20) 6.010 94.84 (25.7) (1.01) 1.568 7.029 (24.5) (0.96) 5.048 39.82 (23.6) (0.92) 3.459 48.55 (22.2) (0.87) 8.734 148.2 (11.2) (0.44) 16.61 416.8 (21.1) (0.83) 398.4 15.88 (17.5) (0.69) 28.08 3.530 (19.8) (0.77) 8.663 178.7 (14.3) (0.56) 7.290 206.3 (13.8) (0.54) 5.747 125.3 (18.6) (0.73) 3.512 0.778 (25.0) (0.98) 5.448 151.0 (17.5) (0.69) 64.85 1.206 (4.98) (0.20) 12.10 0.658 (19.2) (0.75) 13.31 1.332 (17.2) (0.68) 2.300 32.97 (17.0) (0.67) 11.52 0.577 (16.1) (0.63) 0.158 43.63 (24.7) (0.97) 6.728 225.7 (13.3) (0.52) 3.415 156.3 (16.7) (0.66)

Arthritis

Asthma Cholesterol

401.0 26.99 194.6 (15.9) (4.36) (7.52) 299.0 10.70 143.8 (13.3) (3.65) (6.30) 216.5 1.155 103.9 (12.7) (3.48) (6.00) 145.1 2.234 70.89 (12.2) (3.34) (5.77) 65.64 3.773 29.12 (11.5) (3.15) (5.43) 65.82 3.281 1.024 (5.83) (1.60) (2.75) 252.1 99.84 88.12 (10.9) (3.00) (5.17) 151.2 10.28 66.58 (9.09) (2.49) (4.29) 2.376 18.55 29.38 (10.2) (2.80) (4.84) 90.28 29.63 14.60 (7.44) (2.04) (3.51) 89.95 25.03 29.15 (7.17) (1.96) (3.39) 41.64 16.54 17.66 (9.63) (2.64) (4.55) 5.418 19.65 27.24 (13.0) (3.55) (6.12) 48.71 23.08 17.42 (9.09) (2.49) (4.29) 34.96 4.664 16.62 (2.58) (0.71) (1.22) 34.13 10.61 3.851 (9.94) (2.72) (4.69) 26.13 7.901 1.454 (8.94) (2.45) (4.22) 28.22 5.109 3.664 (8.84) (2.42) (4.17) 12.69 1.607 1.149 (8.36) (2.29) (3.94) 10.60 2.626 11.75 (12.8) (3.50) (6.04) 83.9519.63 39.22 (6.90) (1.89) (3.26) 90.96 16.64 35.74 (8.67) (2.37) (4.10)

84

ROSEMARY J. AVERY ET AL.

(Continued ) Allergy

Retired Unemployed Disabled Student Homemaker Midwest South West County size B (W85,000 households) County size CD (other counties) Wave 23 Wave 25 Wave 27 Wave 29 Wave 31 Wave 33 Wave 35 Wave 37 Wave 39 Constant Observations R-squared

343.4 (20.7) 347.2 (27.7) 684.3 (32.4) 46.52 (43.4) 259.5 (21.0) 93.67 (14.9) 36.83 (14.5) 253.7 (15.4) 87.21 (14.6) 73.09 (21.3) 282.9 (20.3) 450.3 (23.5) 451.8 (22.2) 377.9 (24.0) 437.9 (22.7) 594.0 (23.4) 823.5 (24.0) 1130 (21.6) 1009 (21.4) 796.2 (29.5) 80,615 0.10

Allergy (Children)

Arthritis

Asthma Cholesterol

7.234 223.3 39.25 (0.81) (10.7) (2.93) 9.323 192.5 36.17 (1.09) (14.4) (3.93) 20.37 375.3 87.46 (1.27) (16.8) (4.60) 0.873 33.65 4.848 (1.70) (22.5) (6.16) 6.292 151.7 25.63 (0.82) (10.9) (2.97) 0.602 52.45 8.987 (0.58) (7.72) (2.11) 0.167 7.524 5.272 (0.57) (7.51) (2.06) 3.391 116.0 26.76 (0.61) (8.01) (2.19) 2.870 37.58 9.444 (0.57) (7.55) (2.07) 3.955 35.33 10.51 (0.83) (11.0) (3.02) 21.19 137.4 80.55 (0.80) (10.5) (2.88) 48.12 419.9 29.82 (0.92) (12.2) (3.34) 47.67 530.2 36.30 (0.87) (11.5) (3.15) 48.22 503.1 138.3 (0.94) (12.4) (3.41) 24.52 517.8 122.8 (0.89) (11.8) (3.22) 21.77 654.1 58.88 (0.92) (12.1) (3.32) 40.38 571.0 5.453 (0.94) (12.5) (3.41) 4.412 540.0 63.64 (0.85) (11.2) (3.06) 67.02 542.8 81.60 (0.84) (11.1) (3.03) 30.47 403.3 135.4 (1.15) (15.3) (4.18) 80,615 80,615 80,615 0.26 0.16 0.14

Standard errors in parentheses po0.01, po0.05, po0.1.

100.4 (5.06) 81.81 (6.78) 145.6 (7.93) 20.06 (10.6) 58.31 (5.13) 29.96 (3.65) 7.173 (3.55) 45.84 (3.78) 11.97 (3.57) 7.582 (5.20) 49.28 (4.97) 181.5 (5.76) 260.9 (5.44) 280.8 (5.87) 392.6 (5.56) 450.2 (5.73) 295.1 (5.88) 265.3 (5.28) 356.4 (5.23) 169.4 (7.21) 80,615 0.20

85

Number of advertisements exposed to in past year

Health Disparities and Direct-to-Consumer Advertising 2000 1800 1600

Black Unemployed Uninsured Reference group

Number of advertisements exposed to in past year

1708

1361

1400 1200

1135 1027

1000

968 775

800

691

600 400 200 0 Arthritis

A

B

1778

Allergies

450 408

400

425 402

396 368

350

341 320

300

Black Unemployed Uninsured Reference group

315

309

281 268

250 200 150

237 205

249 227

197

194 178

169 149

137

125

100

118 110 81

81

91

56

50 0 Asthma

Cholesterol

Depression

Erectile dysfunction

Insomnia

Sexually Toenail fungus transmitted diseases

Fig. 2. (A) Predicted Exposure to TV Advertisements for Arthritis and Allergy Medications. (B) Predicted Exposure to TV Advertisements for Medications to Treat Other Conditions. Source: Authors’ calculations from TNS/Media Intelligence advertisements and NCS data.

86

ROSEMARY J. AVERY ET AL.

advertisements than consumers in the reference group. Uninsured consumers are usually exposed to about ten percent fewer DTC advertisements. The patterns in Table 3 for the other demographic variables conform with common sense. For most products older consumers are exposed to more advertisements. Not surprisingly, this pattern is especially strong for advertisements of products that treat arthritis, and notable exceptions to the pattern include advertisements for products that treat children’s allergies and sexually transmitted diseases. Women are generally exposed to more advertisements, but men are exposed to more advertisements for products that treat erectile dysfunction and sexually transmitted diseases.

4.2. Discussion and Additional Results The observed patterns in advertising exposure reflect consumers’ televisionwatching and firms’ advertisement-placement decisions. Some of the strongest patterns in Table 3 seem to mainly reflect general differences in television watching. For example, consumers who are not employed have more time available to watch television, so it is not surprising that they are exposed to more DTC advertisements. Similarly, according to Nielsen Media data, in the Fall of 2004 Black households watched 40 percent more television than non-Black households (Steadman, 2005). So again, our finding that Blacks are exposed to at least 30 percent more DTC advertisements in most product categories is not surprising. Although firms’ advertisement-placement decisions are not our focus, the results in Table 3 provide some clues. It might be profitable to target advertisements at poorly informed disadvantaged groups because the marginal returns to providing information are higher.8 An alternative hypothesis is that while DTC advertisements are mainly targeted at advantaged consumers, they hit bystanders who watch a lot of television, including many members of disadvantaged groups. One way to shed light on the ‘‘bystander hypothesis’’ is to explore whether members of a disadvantaged group are exposed to advertisements for relevant products that treat conditions they actually suffer from, or to advertisements for irrelevant products. Because some of the products treat conditions mainly experienced by the elderly, the age of consumers’ provides a rough cut at product relevance. To explore the bystander hypothesis, we re-estimated the model of exposure to advertisements for arthritis products to include interactions between age and race. Fig. 3 compares the predicted exposure of Blacks to the reference group, for consumers ages 18–24 and for consumers over the

87

Health Disparities and Direct-to-Consumer Advertising

Number of advertisements exposed to in past year

1400 1170

1200

996

1000 861 800 600

590

400 200 0 Blacks ages 18-24 Reference group ages 18-24

Blacks ages 65+

Reference group ages 65+

Fig. 3. Predicted Exposure to Television Advertisements for Arthritis Medications. Source: Authors’ calculations from TNS/Media Intelligence advertisements and NCS data.

age of 65. Young Blacks are exposed to almost 50 percent more advertisement than are the young reference group, while the difference for older Blacks is much smaller. Compared to the older reference group, older Blacks are exposed to only 17 percent more advertisements. The fact that young consumers are exposed to a substantial number of advertisements for arthritis medications, and the fact that young Blacks are exposed more than young whites, lend support to the bystander hypothesis. The public health implications of DTC advertising depend upon its ultimate impact on medical treatment, health behaviors, and health outcomes. While an analysis of these effects is beyond the scope of this paper, Table 4 presents some intriguing preliminary patterns. The upper panel of Table 4 shows a strong association between exposure to advertisements for cholesterol medications and several cholesterol-related outcomes. Consumers with the highest quintile of advertising exposure are much more likely to report a diagnosis of high cholesterol and report using a cholesterol prescription medication. In addition, they are more likely to report exercising and being on a diet. These patterns could suggest that DTC advertisements prompt consumers to have their cholesterol measured, and are then given a combination of lifestyle advice and prescription drugs. However, the results in the lower panel of Table 4 suggest other forces may

88

ROSEMARY J. AVERY ET AL.

Exposure to TV Ads and Cholesterol Related Health Outcomes and Behavior among People Age over 45. Exposure to TV Ads for Products High to Treat Cholesterol or Cholesterol Arthritis by Quintiles

Used Prescription Products Treating High Cholesterol

Exercise Regularly

Control Diet

Cholesterol 1st quintile 2nd quintile 3rd quintile 4th quintile 5th quintile Total Sample size

12% 15% 19% 24% 27% 19% 42,720

8% 10% 13% 16% 19% 13% 42,720

43% 46% 48% 50% 52% 48% 41,299

27% 30% 32% 32% 35% 31% 40,697

Arthritis 1st quintile 2nd quintile 3rd quintile 4th quintile 5th quintile Total Sample size

12% 16% 19% 23% 27% 19% 42,720

7% 11% 14% 16% 18% 13% 42,720

45% 46% 48% 50% 51% 48% 41,299

26% 30% 32% 33% 35% 31% 40,697

be at work. There is an equally strong association between exposure to advertisements for arthritis medications and cholesterol-related outcomes. As might be suspected based on the similarity of the patterns in the upper and lower panels of Table 4, exposure to cholesterol advertisements is highly correlated (r ¼ 0.91) with exposure to arthritis advertisements. Instead of advertising exposure causing all of the differences in cholesterol outcomes, the strong associations in Table 4 also appear to reflect some combination of individual television-watching habits and producer targeting of advertisements at certain consumer groups.

5. CONCLUSIONS We find that some disadvantaged groups, most notably Blacks, are exposed to substantially more DTC advertisements. Unemployed consumers and others who do not work full-time are also exposed to substantially more DTC advertisements. Weaker patterns suggest that consumers with less schooling and lower incomes also tend to be exposed to more DTC

Health Disparities and Direct-to-Consumer Advertising

89

advertisements for some products. On the other hand, uninsured consumers are exposed to fewer DTC advertisements, but the magnitude of the difference is relatively small. Just as it is difficult to determine the causes of health disparities in general, the results of our descriptive models do not allow us to determine the causes of the observed differences in advertising exposure. To some extent, socio-economically disadvantaged consumers are probably bystanders who see advertisements targeted at advantaged consumers. Nevertheless, if DTC advertisements provide useful information to disadvantaged consumers, it might not matter too much whether it is by design or by accident. One reason, it might matter, is if the content of advertisements targeted at advantaged consumers is more effective at appealing to and informing the target audience. For example, Kaphingst and DeJong (2004) report that a sample of adults with limited literacy showed poor comprehension of DTC television advertisements for Nasacort (for asthma) and Zocor (for high cholesterol). Based on several studies, Kaphingst and DeJong make a set of recommendations to improve the educational quality of DTC advertisements. The policy implications of the observed differences in advertising exposure across socioeconomic and racial groups are related to more general questions about the desirability of DTC advertising. For some of the product categories we study, DTC advertisements seem to have potential to address important health disparities. For example, as discussed above in Section 2, research suggests that Blacks are less likely to receive prescription drug treatment for high cholesterol. Morris, Gadson, and Burroughs (2007, p. 293) report a survey of members of the National Medical Association, the largest association of Black physicians in the US. They conclude ‘‘our survey reveals that [DTC advertising] has a positive impact on both AfricanAmerican physicians and patients, and, notably, underserved populations . . . . In particular, we were pleased to see that DTC advertising continues to drive patients to visit their doctors. This is very important within the African-American population . . . .’’ The NMA’s recommendations recognize the educational benefit of DTC advertisements while stressing the importance of the patient–physician relationship and the physician’s role as gatekeeper. Pharmaceutical industry studies suggest that many important health conditions are under-treated more generally. For example, Pfizer estimates that only 20 percent of people with dyslipidemia are treated with prescription pharmaceuticals. Moreover, of those patients prescribed a medication for high cholesterol, about half do not take the medication

90

ROSEMARY J. AVERY ET AL.

properly or fail to remain on the therapy after 18 months (Manning, 2006). In this light, DTC advertising for products that treat cholesterol appear to have the potential to improve Americans’ health in general and AfricanAmericans’ health in particular. However, others see DTC advertising in a less favorable light. Critics of advertisements for Lamisil (for toenail fungus) or Vioxx (withdrawn from the market for safety concerns) will not be re-assured by our finding that Blacks saw more advertisements for these products. In December 2004 the FDA requested, and Pfizer agreed to, a voluntary suspension of DTC advertising on Celebrex, an arthritis product in the same class of antiinflammatory drugs as Vioxx (Vogt, 2005). Also influenced by the heavy advertising and subsequent withdrawal of Vioxx, a recent Institute of Medicine report recommends that the FDA should restrict DTC advertising of new products for their first two years (Baciu, Stratton, & Burke, 2006). One notable critic of the pharmaceutical industry goes much further and claims: ‘‘The great majority of DTC ads are for very expensive me-too drugs that require a lot of pushing because there is no good reason to think they are any better than drugs already on the market.’’ Angell (2005, p. 124). Angell (p. 252) recommends a complete ban on DTC advertising. If DTC advertising is more heavily regulated or banned, the pharmaceutical industry might replace it with advertising and promotion that is more targeted at socio-economically advantaged consumers. Despite the rapid growth in DTC advertising, in 2005 expenditures on DTC advertising were only 14 percent of the pharmaceutical industry’s expenditures on advertising and promotion (Donohue et al., 2007). Expenditures on free samples and physician detailing accounted for 62 and 23 percent of total advertising and promotion expenditures, respectively. A recent study suggests that free samples tend to be targeted away from disadvantaged groups. Using data from the 2003 MEPS, Cutrona et al. (2008) found that low-income and uninsured consumers were less likely to have received at least one free sample. Unless these patterns change, moving away from DTC advertising toward free samples could worsen disparities. A growing body of research suggests that DTC advertising increases consumer demand for pharmaceutical products including smoking cessation products, anti-ulcer drugs, and cholesterol-lowering drugs (Avery, Kenkel, Lillard, & Mathios, 2007b; Berndt, Bui, Reiley, Glen, & Urban, 1995; Ling, Berndt, & Kyle, 2002; Wosinska, 2005). Avery et al. (2007b) emphasize the difficult challenge of identifying the causal effect of advertising on consumption when there is also potential reverse causality if firms target advertisements at markets with many consumers. Above we report

Health Disparities and Direct-to-Consumer Advertising

91

intriguing evidence of a strong association between exposure to DTC advertisements for cholesterol products and cholesterol outcomes. However, our preliminary results also reinforce the message that identifying causality will be challenging. Future work needs to isolate suitable sources of variation in advertising exposure that can identify the causal effect of DTC advertising on pharmaceutical use.

NOTES 1. Arrow’s (1963) classic paper contains many of the key insights about the role information plays in medical care markets. McGuire (2001) reviews theoretical and empirical studies of physicians as patients’ agents. 2. While it has also been suggested that advertising can be an entry barrier, ScottMorton (2000) concludes that pharmaceutical advertising is not a barrier to entry by generics. 3. It appears the addition of 25 DMAs was phased in during 2001. In 2001, the TNS data cover some, but not all, of the DMAs covered 2002–2004. While the 2001 data cover advertisements that appeared in up to 95 DMAs, in some of the newly added DMAs the number of advertisements appear unusually low. 4. An example helps clarify the timing of the matching process. Our earliest NCS wave was conducted between October 1, 1999 and April 1, 2000. For respondents to this wave, we match advertisements that aired between October 1, 1998 and 1999. We use a similar timing of the matching process for the other NCS waves. The NCS does not provide the exact interview date. 5. The NCS provides partial information on respondents’ locality of residence by Designated Marketing Areas (DMAs) within state. A DMA is typically identified by its largest city, and includes any surrounding counties where that city’s television broadcasts are most popular. These counties can be within one state or across state boundaries. The NCS has identifiers for each of the 12 most populous DMAs. We use additional information in the NCS to identify respondents who live in 44 smaller DMAs. 6. When we only know that the respondent was watching a broadcast television time slot during which an advertisement appeared, we weight the advertisement by the inverse of an estimate of the number of stations in their DMA. We use the number of stations in each DMA as a proxy for the number of stations a person might watch. 7. Our measure probably tends to under-state advertising exposure more for consumers with unusual viewing habits. The NCS lists of program titles appear to capture the most popular programs. The NCS lists of time slots do not cover the period from 1 a.m. to 5 a.m. for broadcast television or the period from 1 a.m. to 6 a.m. for cable television. 8. Profit-driven targeting could also explain why uninsured consumers are exposed to fewer DTC advertisements, because providing them with information is less likely to result in increased sales of the advertised products.

92

ROSEMARY J. AVERY ET AL.

ACKNOWLEDGMENTS We thank participants at that conference for their comments. We acknowledge financial support from an unrestricted educational grant from the Merck Company Foundation and grant R01 CA113407 from the National Cancer Institute.

REFERENCES Agency for Healthcare Research and Quality. (2006). National healthcare disparities report. Rockville, MD: US Department of Health and Human Services. Angell, M. (2005). The truth about drug companies: How they deceive us and what to do about it. New York: Random House Trade Paperbacks. Arias, E. (2004). United States life tables, 2002. National Vital Statistics Reports, 53(6), 1–6. Arrow, K. J. (1963). Uncertainty and the welfare economics of medical care. American Economic Review, 53, 941–973. Avery, R. J., Kenkel, D. S., Lillard, D., & Mathios, A. D. (2007a). Regulating advertisements: The case of smoking cessation products. Journal of Regulatory Economics, 31(2), 185–208. Avery, R. J., Kenkel, D. S., Lillard, D., & Mathios, A. D. (2007b). Private profits and public health: Does advertising smoking cessation products encourage smokers to quit? Journal of Political Economy, 115(3), 447–481. Baciu, A., Stratton, K., & Burke, S. P. (Eds). (2006). The future of drug safety: Promoting and protecting the health of the public. Committee on the Assessment of the US Drug Safety System, Institute of Medicine. Berndt, E. R., Bui, L., Reiley, D. R., Glen, L., & Urban, G. L. (1995). Information, marketing, and pricing in the US antiulcer drug market. American Economic Review Papers and Proceedings, 85(2), 100–105. Cutler, D. M., Landrum, M. B., & Stewart, K. A. (2006, February). How do the better educated do it? Socioeconomic status and the ability to cope with underlying impairment. NBER Working Paper no. 10240. Cutler, D. M., & Lleras-Muney, A. (2006, June). Education and health: Evaluating theories and evidence. NBER Working Paper no. 12352. Cutrona, S. L., Woolhander, S., Lasser, K. E., Bor, D. H., McCormick, D., & Himmelstein, D. U. (2008). Characteristics of recipients of free prescription drug samples: A nationally representative analysis. American Journal of Public Health, 98(2). Deaton, A. (2002). Policy implications of the gradient of health and wealth. Health Affairs, 21(2), 13–30. Donohue, J. M., Cevasco, M., & Rosenthal, M. B. (2007). A decade of direct-to-consumer advertising of prescription drugs. New England Journal of Medicine, 357, 673–681. Frank, R., Berndt, E. R., Donohue, J., Epstein, A., & Rosenthal, M. (2002, February). Trends in direct-to-consumer advertising of prescription drugs. Menlo Park, California: The Henry J. Kaiser Family Foundation.

Health Disparities and Direct-to-Consumer Advertising

93

Glied, S., & Lleras-Muney, A. (2003, May). Health inequality, education, and medical innovation. NBER Working Paper no. 9738. Goldman, D., & Lakdawalla, D. (2001, June). Understanding health disparities across education groups. NBER Working Paper no. 8328. Grossman, M. (2006). Education and nonmarket outcomes. In: E. Hanushek & F. Welch (Eds), Handbook of the economics of education. Amsterdam: North-Holland, an imprint of Elsevier Science. Ippolito, P. M., & Mathios, A. D. (1990). Information, advertising and health choices: A study of the cereal market. RAND Journal of Economics, 21(3), 459–480. Ippolito, P. M., & Mathios, A. D. (1995). Information and advertising: The case of fat consumption in the United States. American Economic Review Papers & Proceedings, 85(2). Kaphingst, K. A., & DeJong, W. (2004). The educational potential of direct-to-consumer prescription drug advertising. Health Affairs, 23(4), 143–150. Kenkel, D. S. (1991). Health behavior, health knowledge, and schooling. Journal of Political Economy, 99(2), 287–305. Keppel, K., Bilhelmer, L., & Gurley, L. (2007). Improving population health and reducing healthcare disparities. Health Affairs, 26(5), 1281–1292. Ling, D. C., Berndt, E. R., & Kyle, M. K. (2002). Deregulating direct-to-consumer marketing of prescription drugs: Effects on prescription and over-the-counter product sales. Journal of Law and Economics, XL, 691–723. Link, B. G., & Phelan, J. (1995). Social conditions as fundamental causes of disease. Journal of Health and Social Behavior (Extra Issue), 80–94. Lleras-Muney, A., & Lichtenberg, F. R. (2002, September). The effect of education on medical technology adoption: Are the more educated more likely to use new drugs? NBER Working Paper no. 9185. Manning, R. (2006, October 20). Direct-to-consumer advertising. Presentation at Federal Trade Commission Bureau of Economics Roundtable. Pfizer, Inc. Marmot, M. G., Davey Smith, G., Stansfeld, S., Patel, C., North, F., Head, J., White, I., Brunner, E., & Feeny, A. (1991). Health inequalities among British civil servants: The Whitehall II study. Lancet, 1387–1393. McGuire, T. G. (2000). Physician agency. In: A. J. Culyer & J. P. Newhouse (Eds), Handbook of health economics (1A, pp. 461–536). Amserdam: North-Holland. Meara, E. (2001, April). Why is health related to socioeconomic status? The case of pregnancy and low birth weight. NBER Working Paper no. 8231. Mechanic, D. (2002). Disadvantage, inequality, and social policy. Health Affairs, 21(2), 48–59. Morris, A. W., Gadson, S. L., & Burroughs, V. (2007). For the good of the patient, survey of the physicians of the National Medical Association regarding perceptions of DTC advertising, Part, 1.1, 2006. Journal of the National Medical Association, 99(3), 287–293. Nelson, K., Norris, K., & Mangione, C. M. (2002). Disparities in the diagnosis and pharmacologic treatment of high serum cholesterol by race and ethnicity. Archives on Internal Medicine, 162, 929–935. Rogot, E., Sorlie, P. D., Johnson, N. L., & Schmit, C. (1992). A mortality study of 1.3 million persons by demographic, social, and economic factors: 1979–1985 follow-up. Bethesda: National Institutes of Health.

94

ROSEMARY J. AVERY ET AL.

Scott-Morton, F. (2000). Barriers to entry, brand advertising, and generic entry in the US pharmaceutical industry. International Journal of Industrial Organization, 18(7), 1085–1104. Smith, J. P. (1999). Health bodies and thick wallet: The dual relation between health and economic status. Journal of Economic Perspectives, 13(2), 145–166. Steadman, J. (2005, Summer). TV audience special study: African-American audience. Nielsen Media Research. US Department of Health and Human Services. (2000). Healthy People 2010: Understanding and improving Health (2nd). Washington, DC: US Government Printing Office. Vogt, D. U. (2005, March 25). Direct-to-consumer advertising of prescription drugs. Congressional Research Service: The Library of Congress. Wang, J., Zuckerman, I. H., Miller, N. A., Shaya, F. T., Noel, J. M., & Mullins, C. D. (2007). Utilizing new prescription drugs: Disparities among non-Hispanic whites, non-Hispanic Blacks, and Hispanic whites. Health Services Research, 42(4), 1499–1519. Weissman, J. S., Blumenthal, D., Silk, A. J., Zapert, K., Newman, M., & Leitman, R. (2003). Consumers’ reports on the health effects of direct-to-consumer drug advertising. Health Affairs. Web Exclusive W3-82–W3-95. Wilkes, M. S., Bell, R. A., & Kravitz, R. L. (2000). Direct-to-consumer prescription drug advertising: Trends, impact, and implications. Health Affairs, 19(2), 110–128. Wosinska, M. (2005). Direct-to-consumer advertising and drug therapy compliance. Journal of Marketing Research, 42(3), 323–332. Yang, H. K., Mathios, A. D., & Avery, R. J. (2007, June). The impact of direct-to-consumer advertising of cholesterol reducing drugs on diagnosis and treatment of cholesterol. Working Paper, Department of Policy Analysis and Management, Cornell University.

PHARMACEUTICAL INNOVATION AND THE LONGEVITY OF AUSTRALIANS: A FIRST LOOK Frank R. Lichtenberg and Gautier Duflos ABSTRACT The purpose of this paper is to examine the impact of pharmaceutical innovation on the longevity of Australians. The approach utilized involves estimation of difference-in-differences models using longitudinal, diseaselevel data during the period 1995–2003 to determine whether the diseases that had above-average increases in mean vintage (FDA approval year) of drugs had above-average reductions in mortality. Our findings are that the mean age at death increased more for diseases with larger increases in mean drug vintage. A 5-year increase in mean drug vintage is estimated to increase mean age at death by almost 11 months. The number of years of potential life lost before the ages of 65 and 70 (but not before age 75) was reduced by use of newer drugs. During the period 1995–2003, mean age at death increased by about 2.0 years, from 74.4 to 76.4. The estimates imply that, in the absence of any increase in drug vintage, mean age at death would have increased by only 0.7 years. The increase in drug vintage accounts for about 65% of the total increase in mean age at death. Estimated cost per life-year gained from using newer drugs is $10,585. An estimate by previous investigators of the value of a statistical Australian life-year ($70,618) is 6.7 times as large. We acknowledge potential Beyond Health Insurance: Public Policy to Improve Health Advances in Health Economics and Health Services Research, Volume 19, 95–117 Copyright r 2008 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0731-2199/doi:10.1016/S0731-2199(08)19005-7

95

96

FRANK R. LICHTENBERG AND GAUTIER DUFLOS

limitations of this study by discussing several reasons why our estimate of the cost per life-year gained from using newer drugs could be too high or low. The value of this paper’s evidence is primarily due to the government’s Pharmaceutical Benefits Scheme: Australia has much better data on drug utilization than most other countries.

In previous papers, Lichtenberg (2005a, 2005b) has examined the impact of pharmaceutical innovation on longevity in the United States and a group of developed and developing countries. Due to data limitations, the measure of pharmaceutical innovation used in these studies was the cumulative number of drugs launched. These studies provided support for the hypothesis that the introduction of new drugs has played an important role in increasing longevity. In this paper, we will examine the impact of pharmaceutical innovation on the longevity of Australians during the period 1995–2003. In one important respect, the data available for Australia are much better than those used in the previous studies. Rather than merely knowing whether a given drug has been launched in Australia by a certain date, we know how frequently that drug is used in each year. Combining these data with data from other sources enables us to calculate the mean vintage1 of drugs utilized in Australia, by disease and year. Section 1 contains a discussion of the ‘‘embodied technological progress hypothesis.’’ Section 2 describes an econometric model to test this hypothesis. Data sources and descriptive statistics are presented in Section 3. Empirical results are presented in Section 4. Section 5 contains a summary and discussion.

1. EMBODIED TECHNOLOGICAL PROGRESS HYPOTHESIS Economists believe that the development of new products is the main reason why people are better off today than were several generations ago. In their 1993 book, Innovation and Growth in the Global Economy, Grossman and Helpman (1993) argued, ‘‘innovative goods are better than older products simply because they provide more ‘product services’ in relation to their cost of production.’’ In their 1996 book, The Economics of New Goods, Bresnahan and Gordon (1996) stated simply that ‘‘new goods are at the

Pharmaceutical Innovation and the Longevity of Australians

97

heart of economic progress.’’ In a recent paper, Measuring the Growth from Better and Better Goods, Bils (2004) makes the case that ‘‘much of economic growth occurs through growth in quality as new models of consumer goods replace older, sometimes inferior, models.’’ We seek to test the hypothesis that, ceteris paribus, people using newer, or later vintage, drugs will be in better health and therefore live longer (die later). This hypothesis is predicated on the idea that pharmaceuticals, like other R&D-intensive products, are characterized by embodied technological progress. A number of econometric studies (Bahk & Gort, 1993; Hulten, 1992; Sakellaris & Wilson, 2001, 2004) have investigated the hypothesis that capital equipment employed by US manufacturing firms embodies technological change, i.e., each successive vintage of investment is more productive than the last. Equipment is expected to embody significant technical progress due to the relatively high R&D-intensity of equipment manufacturers. The method that has been used to test the equipment-embodied technical change hypothesis is to estimate manufacturing production functions, including (mean) vintage of equipment as well as quantities of capital and labor. These studies have concluded that technical progress embodied in equipment is a major source of manufacturing productivity growth. Although most previous empirical studies of embodied technical progress have focused on equipment used in manufacturing, embodied technical progress may also be an important source of economic growth in health care. One important input in the production of health – pharmaceuticals – is even more R&D-intensive than equipment. According to the National Science Foundation, the R&D-intensity of drugs and medicines manufacturing is 74% higher than that of machinery and equipment manufacturing. Therefore, it is quite plausible that there is also a high rate of pharmaceutical-embodied technical progress. The hypothesis that technical progress is embodied in pharmaceuticals may be tested in two alternative ways. One approach is to investigate whether the health and longevity of people with a given disease is positively related to the number of drugs that have been approved to treat that disease.2 Lichtenberg adopted this approach in several studies (Lichtenberg 2005a, 2005b, 2005c), and in all of them, he found that increases in the cumulative number of drugs improved health. This approach allows one to distinguish between the effects of approval of ‘‘priority-review’’ drugs – drugs that the FDA considers to offer significant improvements over existing therapies – and approval of ‘‘standard-review’’ drugs – drugs that the FDA considers to be similar to the previously approved drugs. In two

98

FRANK R. LICHTENBERG AND GAUTIER DUFLOS

studies, Lichtenberg distinguished between the effects of priority-review and standard-review drug approvals. The results of distinguishing between the two were mixed. Lichtenberg (2005a) found that the approval of standardreview drugs had no effect on longevity, but that of priority-review drugs had a significant positive impact on longevity. But most of the results in Lichtenberg (2005c) indicated that the difference between the effect of priority-and standard-review drugs on ability to work was not statistically significant. The second way to test the hypothesis that technical progress is embodied in pharmaceuticals is to investigate whether the health and longevity of people with a given disease is positively related to the mean vintage (FDA approval year) of drugs used to treat the disease. We believe that the second is superior to first approach. The drugs that have been approved to treat a given disease influence the therapy that a patient could receive, but his health and longevity depends on the therapy he actually does receive. The fact that a drug has been approved does not necessarily mean that it is commonly used. In this paper we will pursue the second approach. Although we believe that mean vintage is a better measure of innovation than the number of previously approved drugs, proper accounting for the distinction between priority- and standard-review drugs when measuring drug vintage, while straightforward in theory, is difficult in practice. Suppose a (standard-review) drug approved in 2008 is ‘‘therapeutically equivalent’’ to a drug approved in 1998. Then the ‘‘effective vintage’’ of the drug is 1998, whereas its actual vintage is 2008. (The effective vintage of a priority-review drug is the same as its actual vintage.) If we could measure the effective vintage of all drugs, we would use mean effective vintage instead of mean actual vintage in our econometric model. However, although the FDA characterizes some drugs as therapeutically equivalent to previously approved drugs, it does not specify the drugs to which they are therapeutically equivalent. Hence measurement of mean effective vintage is not feasible.

2. ECONOMETRIC MODEL To test the hypothesis that pharmaceutical innovation has increased the longevity of Australians, we will estimate the following econometric model:   Sd N_RXdit FDA_YEARd þ ai þ dt þ it Y it ¼ b Sd N_RXdit

Pharmaceutical Innovation and the Longevity of Australians

99

or Y it ¼ bV it þ ai þ dt þ it

(1)

where Yit is a measure based on the age distribution of deaths from disease i in year t; N_RXdit ¼ the number of times drug d was used to treat patients with disease i in year t; FDA_YEARd ¼ the FDA approval year of the active ingredient of drug d; Vit ¼ Sd Ndit FDA_YEARd/Sd Ndit ¼ the mean vintage of drugs used to treat disease i in year t; ai denotes a fixed disease effect; and dt denotes a fixed year effect. There are both practical and theoretical reasons to define the vintage of a drug as the year the drug was approved by the US FDA rather than the year the drug was listed (approved for reimbursement) in Australia’s PBS. Data on the PBS listing dates are quite incomplete. We obtained unpublished data on listing dates of drugs listed by the PBS after 1990.3 Based on a sample of 311 drugs for which both FDA approval dates and PBS listing dates were available, we estimate that the mean lag between FDA approval and PBS listing is 3.6 years. However we believe that the FDA approval date is theoretically superior to the PBS listing date as a measure of vintage (which is intended to indicate the year of (global) market introduction or first use). The vintage of a wine is the year the wine was bottled, not the year it was opened! In principle, health and longevity may be affected by lagged as well as current mean drug vintage. However, including lagged vintage would substantially reduce the size of our sample since we have data on Y and V in only 9 years (1995–2003). Moreover, since vintage tends to be serially correlated, including lagged vintage terms would introduce multicollinearity. We will therefore only include contemporaneous vintage in the model. We will estimate the model using four different dependent variables. The first is the mean age at death of Australians dying from disease i in year t: AGE_DEATHit ¼ Sa ðan N_DEATHait Þ Sa N_DEATHait where N_DEATHait is the number of deaths at age a from disease i in year t. The second is the logarithm4 of potential years of life lost before age 75 from disease i in year t: LPYLL75it ¼ ln½Sa maxð75  a; 0Þn N_DEATHait 

100

FRANK R. LICHTENBERG AND GAUTIER DUFLOS

The Australian Institute of Health and Welfare (AIHW) reports both mean age at death and potential years of life lost before age 75 in its General Records of Incidence of Mortality. It also notes that the limit to life of 75 years is ‘‘arbitrary.’’ We will also estimate models using two lower thresholds, 70 and 65:5 LPYLL70it ¼ ln½Sa maxð70  a; 0Þn N_DEATHait  LPYLL65it ¼ ln½Sa maxð65  a; 0Þn N_DEATHait  All models will be estimated via weighted least squares. For the first model the weight is the number of deaths from disease i in year t: N_DEATHit ¼ Sa N_DEATHait. For the second model, the weight is the mean number of potential years of life lost before age 75 from disease i during the 9 years 1995–2003: (1/9) St exp(LPYLL75it). Analogous weights will be used for the two lower age thresholds. Due to the presence of fixed disease and year effects, Eq. (1) is a difference-in-differences model. If the dependent variable is mean age at death, a positive and significant estimate of b would signify that there were above-average increases in mean age at death for diseases with aboveaverage increases in mean vintage of drugs.

3. DATA SOURCES AND DESCRIPTIVE STATISTICS 3.1. Mortality Data The AIHW has compiled long-term mortality data on selected causes of death by age and sex for each year from the beginning of twentieth century, and published them in its GRIM (General Record of Incidence of Mortality) books. These are interactive excel workbooks updated annually containing comprehensive long-term mortality data on selected causes of death by age and sex for each year. The GRIM books have been grouped together by chapters as adopted by the 10th Revision of the International Statistical Classification of Diseases and Related Health Problems (ICD-10). Each workbook contains mortality data, population data, derived data items (e.g., age-specific and age-standardised rates), summary measures

101

Pharmaceutical Innovation and the Longevity of Australians

(e.g., mean age at death, potential years of life lost, and lifetime risk of dying), birth cohort information, and graphs. The following table shows annual mortality data for all causes of death combined for the period 1995–2003. Year

Number of Deaths

Mean Age at Death

Years of Life Lost before Age 75

Years of Life Lost before Age 75 per 1,000 Population

1995 1996 1997 1998 1999 2000 2001 2002 2003

125,133 128,719 129,350 127,202 128,102 128,291 128,544 133,707 132,292

71.8 72.2 72.4 72.4 72.6 73.0 73.3 73.8 73.9

966,458 963,160 959,548 941,793 938,078 908,058 881,733 876,770 866,298

56.2 55.3 54.6 53.1 52.4 50.2 48.2 47.4 46.4

3.2. Pharmaceutical Utilization Data Data on pharmaceutical utilization were obtained from the National Social Health Statistical Data Library (HealthWIZ)6 a database on CD-ROM that is used to disseminate comprehensive population health-related statistical datasets, across the Australian health services sector, for the purposes of clinical research, policy development, and health services planning, particularly in regional areas. Several datasets contained in HealthWIZ are derived from the Australian Government’s PBS.7 For nearly 60 years, the PBS has provided reliable, timely, and affordable access to a wide range of medicines for all Australians. Many medicines cost the government much more than the price paid by the patient – some cost hundreds, even thousands of dollars, but the government provides a subsidy so that patients pay much less. The patient receives the benefit of this subsidy when she has her prescription for a medicine filled under the PBS. Current provisions governing the operations of the PBS are embodied in Part VII of the National Health Act 1953 together with the National Health

102

FRANK R. LICHTENBERG AND GAUTIER DUFLOS

(Pharmaceutical Benefits) Regulations 1960 made under the Act. The scheme has proven itself to be one of the best drug subsidy systems in the world and around 80% of prescriptions dispensed in Australia are subsidized under the PBS.8 Every time a patient fills a prescription for a PBS medicine, she receives a subsidy. From 1 January 2006, the patient pays up to $29.50 for most PBS medicines or $4.70 if she has a concession card. The Australian Government pays the remaining cost. The PBS covered around 170 million prescriptions in June 2005. This equates to about eight prescriptions per person in Australia for the year. With new and more effective medicines helping us to lead longer and healthier lives, the PBS is growing each year. The cost of the PBS is currently around $6.0 billion per year. HealthWIZ provides data on the number of prescriptions filled under the PBS, by drug and year, 1995–2004. This dataset contained information on approximately 700 drugs. The Anatomical Therapeutic Chemical (ATC) Classification System is used for the classification of drugs. It is controlled by the WHO Collaborating Centre for Drug Statistics Methodology, and was first published in 1976. Drugs are divided into different groups according to the organ or system on which they act and/or their therapeutic and chemical characteristics. In the system drugs are classified into groups of five different levels. There are 14 main groups at the first-level. To illustrate the pharmaceutical utilization data, the following is a list of the top 10 cardiovascular system drugs, ranked by number of prescriptions in 2004:

Drug Atorvastatin Simvastatin Irbesartan Atenolol Irbesartan with hydrochlorothiazide Ramipril Perindopril Amlodipine Pravastatin Perindopril and diuretics

Number of PBS Rx’s in 2004 7,207,717 5,756,278 3,278,440 2,952,209 2,807,419 2,663,857 2,578,733 2,201,328 1,978,913 1,522,659

103

Pharmaceutical Innovation and the Longevity of Australians

The following is a list of the top 10 antineoplastic and immunomodulating agents, ranked by a number of prescriptions in 2004: Drug

Number of PBS Rx’s in 2004

Tamoxifen Methotrexate Leflunomide Azathioprine Goserelin Letrozole Anastrozole Cyclophosphamide Interferon beta-1b Fluorouracil

193,340 149,107 108,144 104,236 53,556 36,837 36,268 35,232 32,282 29,160

3.3. Pharmaceutical Vintage Data We used data from the Drugs@FDA database9 and Mosby’s Drug Consult10 to determine the year in which each active ingredient was first approved by the FDA. 3.4. Descriptive Statistics on the Mean Vintage of PBS Prescriptions As the following table shows, during the period 1995–2004 the mean vintage of PBS prescriptions increased by about 1 year per year, from 1977.8 to 1986.7. The average PBS prescription is for a 17-year-old drug. Year

Number of Rx’s

Mean FDA Approval Year

1995 1996 1997 1998 1999 2000 2001 2002 2003 2004

122,224,901 125,904,079 124,980,656 125,365,284 133,455,864 142,877,869 150,924,801 158,172,125 161,192,358 170,253,375

1977.8 1978.8 1979.8 1980.9 1981.9 1983.1 1984.4 1985.3 1986.1 1986.7

104

FRANK R. LICHTENBERG AND GAUTIER DUFLOS cardiovascular system drugs 1990

Mean vintage

1989 1988 1987 1986 1985 1984 1983 1982 1995

1996

1997

1998

1999

2000

2001

2002

2003

antineoplastic and immunomodulating agents 1980

Mean vintage

1979 1978 1977 1976 1975 1974 1973 1972 1995

Fig. 1.

1996

1997

1998

1999

2000

2001

2002

2003

Mean Vintage of Two Classes of Drugs, 1995–2003.

The level and growth rate of vintage varies considerably across ATC groups. Fig. 1 depicts the mean vintage of two classes of drugs during 1995–2003. Cardiovascular drugs tend to be much newer than antineoplastic and immunomodulating agents; in 2003 the vintage of the latter was almost 10 years lower. The mean vintage of cardiovascular system drugs increased almost twice as much during the first half of this period (1995–1999) as it did during the second half (1999–2003). In contrast, the mean vintage of antineoplastic and immunomodulating agents increased over four times as much in the second half as it did in the first half.

Pharmaceutical Innovation and the Longevity of Australians

105

3.5. Linkage of Drugs to Diseases Estimation of Eq. (1) requires linkage of drugs to the diseases they are used to treat. We used the following linkage of ATC drug groups to ICD-10 causes of death chapters:11 ATC Drug Group(s)

ICD-10 Cause of Death Chapter(s)

Alimentary tract and metabolism (A) þ systemic hormonal preparations, excluding sex hormones and insulins (H) Blood and blood-forming organs (B)

Diseases of the digestive system (XI) þ endocrine, nutritional, and metabolic diseases (IV)

Cardiovascular system (C) Dermatologicals (D) Genitourinary system and sex hormones (G) Anti-infectives for systemic use (J) þ antiparasitic products, insecticides, and repellents (P) Antineoplastic and immunomodulating agents (L) Musculoskeletal system (M) Nervous system (N) Respiratory system (R) Sensory organs (S)

Diseases of the blood and bloodforming organs (III) Diseases of the circulatory system (IX) Diseases of the skin and subcutaneous tissue (XII) Diseases of the genitourinary system (XIV) Certain infectious and parasitic diseases (I)

Neoplasms (II) Diseases of the musculoskeletal system and connective tissue (XIII) Diseases of the nervous system (VI) þ mental and behavioral disorders (V) Diseases of the respiratory system (X) Diseases of the eye and adnexa (VII) þ diseases of the ear and mastoid process (VIII)

4. EMPIRICAL RESULTS Estimates of Eq. (1) with four different dependent variables are shown in Table 1.12 The equations were estimated using annual data for the period 1995–2003 on the 11 groups of diseases shown above: N ¼ 99 (11 diseases  9 years). All equations include disease fixed-effects and year fixed-effects. The dependent variable of the first equation is the mean age at death of the Australians dying from disease i in year t. As shown in line 1, the coefficient on the mean vintage of drugs is positive and statistically significant

106

FRANK R. LICHTENBERG AND GAUTIER DUFLOS

Estimates of Eq. (1). Line

Parameter

Estimate

SE

t Value

p Value

0.091 0.688 0.627 0.569 0.498 0.456 0.374 0.329 0.304 –

1.99 1.09 0.81 0.98 1.16 0.60 0.62 0.46 0.11 –

0.0497 0.2777 0.4219 0.3291 0.2514 0.5487 0.5386 0.6440 0.9112 –

dep. var. ¼ LPYLL75it; weight ¼ (1/9) St exp(LPYLL75it) 11. fda_year 0.015 0.011 12. year 1995 0.008 0.089 13. year 1996 0.001 0.084 14. year 1997 0.013 0.077 15. year 1998 0.000 0.069 16. year 1999 0.032 0.065 17. year 2000 0.023 0.053 18. year 2001 0.025 0.046 19. year 2002 0.008 0.043 20. year 2003 0.000 –

1.36 0.09 0.01 0.16 0.00 0.49 0.44 0.54 0.19 –

0.1787 0.9262 0.9910 0.8708 0.9984 0.6274 0.6640 0.5940 0.8463 –

dep. var. ¼ LPYLL70it; weight ¼ (1/9) St exp(LPYLL70it) 21. fda_year 0.024 0.012 22. year 1995 0.087 0.098 23. year 1996 0.072 0.092 24. year 1997 0.047 0.084 25. year 1998 0.044 0.076 26. year 1999 0.075 0.072 27. year 2000 0.050 0.059 28. year 2001 0.046 0.052 29. year 2002 0.021 0.049 30. year 2003 0.000 –

2.00 0.89 0.78 0.55 0.58 1.05 0.85 0.88 0.43 –

0.0488 0.3763 0.4349 0.5819 0.5609 0.2958 0.3982 0.3816 0.6714 –

dep. var. ¼ LPYLL65it; weight ¼ (1/9) St exp(LPYLL65it) 31. fda_year 0.033 0.013 32. year 1995 0.147 0.107 33. year 1996 0.125 0.101 34. year 1997 0.087 0.093 35. year 1998 0.069 0.083 36. year 1999 0.100 0.079 37. year 2000 0.063 0.066 38. year 2001 0.059 0.058 39. year 2002 0.029 0.055 40. year 2003 0.000 –

2.53 1.37 1.24 0.94 0.83 1.28 0.97 1.01 0.53 –

0.0135 0.1741 0.2175 0.3483 0.4094 0.2058 0.3355 0.3139 0.5961 –

dep. var. ¼ AGE_DEATHit; weight ¼ N_DEATHit 1. fda_year 0.182 2. year 1995 0.752 3. year 1996 0.506 4. year 1997 0.559 5. year 1998 0.576 6. year 1999 0.275 7. year 2000 0.231 8. year 2001 0.153 9. year 2002 0.034 10. year 2003 0.000

107

Pharmaceutical Innovation and the Longevity of Australians

(p value ¼ 0.0497). This indicates that mean age at death increased more for diseases with larger increases in mean drug vintage. The point estimate of b indicates that increasing the mean vintage of drugs by 5 years would increase mean age at death by almost 11 months. Additional implications of the estimates of the first model will be considered below. Before doing that, we will discuss estimates of the other three models. As shown in line 11, when the dependent variable is the logarithm of potential years of life lost before age 75 from disease i in year t, the coefficient on the mean vintage of drugs is negative but not statistically significant (p value ¼ 0.1787). However, as shown in lines 21 and 31, when the age threshold is either 70 or 65, the coefficient on the mean vintage of drugs is negative and statistically significant (p value ¼ 0.0488 and 0.0135, respectively). This implies that using newer drugs has reduced premature mortality – especially mortality before age 65 – in the Australian population.13 The estimates of the three potential years of life lost equations tend to confirm the estimates of the mean age at death equation. We can use our estimates of the first equation to compare the actual increase in mean age at death during the period 1995–2003 to the increase that would have occurred in the absence of any increase in drug vintage.14 As shown in Fig. 2, during this period, mean age at death increased by about 2.0 years, from 74.4 to 76.4. The estimates imply that, in the absence of any increase in drug vintage, mean age at death would have increased by only 0.7 years. The increase in drug vintage accounts for about 65% of the total increase in mean age at death. We can also obtain a rough estimate of the cost per life-year gained from using newer drugs. The calculations are shown in the following table. Year 1995 1. 2. 3. 4.

5.

Rx expenditure Population Rx expenditure per capita ((2)/(1)) Life expectancy (mean age at death) ‘‘Lifetime’’ Rx expenditure per capita ((4)  (3))

Source: OECD Health Database.

Change 2003

$2,672,000,000 18,071,758 $148

$6,268,000,000 19,872,646 $315

75.13

76.36

1.23

$11,109

$24,085

$12,976

108

FRANK R. LICHTENBERG AND GAUTIER DUFLOS 76.5

if no increase in drug vintage actual

76.4

Mean age at death

76.0

75.5 75.1 75.0

74.5 74.4 74.0 1995

1996

1997

1998

1999 Year

2000

2001

2002

2003

Fig. 2. Comparison of Actual Increase in Mean Age at Death to the Increase that would have Occurred in the Absence of any Increase in Drug Vintage.

Line 3 shows that per capita drug expenditure more than doubled in Australia from 1995 to 2003, from $148 to $315. For simplicity, suppose that all of this increase was due to the fact that the drugs used in 2003 were newer than those used in 1995. Line 4 shows the increase in ‘‘life expectancy’’ (mean age at death) that is attributable to increasing drug vintage. Line 5 shows ‘‘lifetime’’ drug expenditure per capita: annual expenditure times life expectancy. Under our assumptions, using newer drugs (increasing drug vintage) increased life expectancy by 1.23 years and increased lifetime drug expenditure by $12,976. The cost per life-year gained from using newer drugs is $10,585 ( ¼ $12,976/1.23). Viscusi (2005), citing Kniesner and Leeth (1991), estimates that the value of a statistical Australian life is 4.2 million USD, which is equal to $A 5.4 million at the current exchange rate (1.2839 $A/USD). This implies that the value of a statistical Australian life-year is $70,618 ( ¼ $A 5.4 million/76.4). This value is 6.7 times as large as our estimate of the cost per life-year gained from using newer drugs.

5. SUMMARY AND DISCUSSION We have examined the impact of pharmaceutical innovation on the longevity of Australians during the period 1995–2003. Due to the government’s PBS,

Pharmaceutical Innovation and the Longevity of Australians

109

Australia has much better data on drug utilization than most other countries. We found that mean age at death increased more for diseases with larger increases in mean drug vintage. The estimates indicated that increasing the mean vintage of drugs by 5 years would increase mean age at death by almost 11 months. The estimates also indicated that using newer drugs reduced the number of years of potential life lost before the ages of 65 and 70 (but not before age 75). During the period 1995–2003, mean age at death increased by about 2.0 years, from 74.4 to 76.4.15 The estimates implied that, in the absence of any increase in drug vintage, mean age at death would have increased by only 0.7 years. The increase in drug vintage accounts for about 65% of the total increase in mean age at death. We obtained a rough estimate of the cost per life-year gained from using newer drugs. Under our assumptions, using newer drugs (increasing drug vintage) increased life expectancy by 1.23 years and increased lifetime drug expenditure by $12,976; the cost per life-year gained from using newer drugs is $10,585.16 An estimate made by other investigators of the value of a statistical Australian life-year ($70,618) is 6.7 times as large as our estimate of the cost per life-year gained from using newer drugs. For several reasons, our estimate of the cost per life-year gained from using newer drugs could be too high or low. Studies based on US data (Lichtenberg, 2001, 2005c, 2006) indicate that the use of newer drugs reduces admissions to hospitals and nursing homes, and increases ability to work. By not accounting for this, we may have overestimated the cost per Australian life-year gained. Use of newer drugs may have cross-disease spillover effects: using newer drugs for one disease may either increase or decrease mortality from other diseases (in part due to ‘‘competing risks’’). Such spillovers could be either negative or positive. For example, using a newer drug to treat cancer might reduce cancer mortality but increase life-years lost due to cardiovascular disease. On the other hand, using a newer drug to treat depression and other mental disorders might lead to better management of cardiovascular disease. Finally, innovation in medical devices and procedures as well as drugs, have undoubtedly contributed to Australian longevity increase.17 The models we have estimated control (via year fixed-effects) for device/ procedure innovation that is common to all diseases, but not for diseasespecific device/procedure innovation: measuring disease-specific device/ procedure innovation is far more challenging than measuring diseasespecific drug innovation. Since device/procedure innovation may either

110

FRANK R. LICHTENBERG AND GAUTIER DUFLOS

substitute for or complement drug innovation, controlling for diseasespecific device/procedure innovation could either decrease or increase our estimate of the cost per life-year gained from using newer drugs. Our findings, which are based on aggregate data, are broadly consistent with previous findings based on individual-level data. Lichtenberg and Virabhak (2007) examined the impact of drug vintage on health and longevity using data on (American) individuals before and after the drugs were consumed. They found that people who used newer drugs had better post-treatment health than people using older drugs for the same condition, controlling for pre-treatment health, age, sex, race, marital status, education, income, and insurance coverage: they were more likely to survive, their perceived health status was higher, and they experienced fewer activity, social, and physical limitations. Most of the health measures indicated that the effect of drug vintage on health is higher for people with low initial health than it is for people with high initial health. This suggests that pharmaceutical-embodied technical progress has a tendency to reduce inequality as well as promote economic growth, broadly defined.

NOTES 1. The dictionary contains different definitions of vintage. The definition we use is: ‘‘a period of origin or manufacture.’’ We define the vintage of a drug as the year in which the US Food and Drug Administration (FDA) first approved the drug’s active ingredient. (The FDA, which has been in existence since 1938, provides the most complete data on drug vintage.) For example, the vintage of PBS items 8213G, 8214H, 8215J, and 8521L is 1997, the year the active ingredient of all these items (atorvastatin calcium) was approved by the FDA. (These items correspond to 10, 20, 40, and 80 mg tablets, respectively.) 2. In his model of endogenous technological change, Romer (1990) hypothesized the production function Y (AL)1a Ka where Y ¼ output, A ¼ the ‘‘stock of ideas,’’ L ¼ labor used to produce output, K ¼ capital, and 0oao1. The cumulative number of drugs approved is analogous to the stock of (FDA-approved) ideas. Health and longevity may be considered outputs of a health production function. 3. We are grateful to Kim Sweeny of Victoria University for sharing these data with us. 4. The logarithmic specification embodies the assumption that equal increases in vintage result in equal percentage reductions in potential years of life lost. 5. The 70-year threshold is the one used in the OECD Health Database for making international comparisons. The 65-year threshold is the ‘‘default choice’’ in the US Center for Disease Control’s Years of Potential Life Lost Reports http:// www.cdc.gov/ncipc/wisqars/fatal/help/definitions.htm 6. http://www.health.gov.au/internet/wcms/publishing.nsf/Content/Healthwiz-1

Pharmaceutical Innovation and the Longevity of Australians

111

7. http://www.health.gov.au/internet/wcms/publishing.nsf/Content/health-pbsgeneral-aboutus.htm 8. Some of the reasons why a medicine may not be available on the PBS are: (1) the manufacturer has not registered its product to treat a particular condition with the Therapeutic Goods Administration; (2) the manufacturer did not apply to the government’s independent expert committee – the Pharmaceutical Benefits Advisory Committee (PBAC) – to list the medicine on the PBS; and (3) the manufacturer has not supplied sufficient evidence, or the evidence supplied does not support a recommendation by the PBAC. http://www.health.gov.au/internet/main/ publishing.nsf/Content/health-pbs-general-faq.htm-copy2 9. http://www.fda.gov/cder/drugsatfda/datafiles/default.htm 10. http://www.mosbysdrugconsult.com/ 11. The following ICD-10 chapters are excluded from our analysis: pregnancy, childbirth, and the puerperium (XV); certain conditions originating in the perinatal period (XVI); congenital malformations, deformations and chromosomal abnormalities (XVII); symptoms, signs and abnormal clinical and laboratory findings (XVIII); injury, poisoning, and certain other consequences of external causes (XIX); external causes of morbidity and mortality (XX); factors influencing health status and contact with health services (XXI); and codes for special purposes (XXII). 12. Data used to estimate Eq. (1) are shown in Table A1. 13. The magnitude of the point estimate of b in line 31 is about 35% larger than the magnitude of the point estimate of b in line 21. But since the number of years of potential life lost before age 70 is about 47% higher than the number of years of potential life lost before age 65, these two models yield similar estimates of the absolute reduction in years of potential life lost from increasing drug vintage. 14. The increase that would have occurred in the absence of any increase in drug vintage is measured by the differences between the year fixed-effects shown in lines 2–10 of Table 1. 15. Lichtenberg (2005a) found that, in the United States, within-disease increases in mean age at death accounted for about 80% of the aggregate long-term increase in mean age at death; the remaining 20% was due to a shift in the distribution of fatal diseases. 16. This is an estimate of the cost per life-year gained from using newer drugs in general. It is likely that the cost per life-year gained from some newer drugs is higher, and that the cost from other newer drugs is lower, than this average. 17. However, the biopharmaceutical industry is much more R&D-intensive than the medical device and equipment industry.

REFERENCES Bahk, B.-H., & Gort, M. (1993). Decomposing learning by doing in new plants. Journal of Political Economy, 101, 561–583. Bils, M. (2004, July). Measuring the growth from better and better goods. NBER Working Paper no. 10606. http://www.nber.org/papers/w10606 Bresnahan, T. F., & Gordon, R. J. (1996). The economics of new goods. Chicago: University of Chicago Press.

112

FRANK R. LICHTENBERG AND GAUTIER DUFLOS

Grossman, G. M., & Helpman, E. (1993). Innovation and growth in the global economy. Cambridge: MIT Press. Hulten, C. R. (1992). Growth accounting when technical change is embodied in capital. The American Economic Review, 82(4), 964–980. Kniesner, T. J., & Leeth, J. D. (1991). Compensating wage differentials for fatal injury risk in Australia, Japan, and the United States. Journal of Risk and Uncertainty, 4, 75–90. Lichtenberg, F. (2001). Are the benefits of newer drugs worth their cost? Evidence from the 1996 MEPS. Health Affairs, 20(5), 241–251. Lichtenberg, F. (2005a). Pharmaceutical knowledge-capital accumulation and longevity. In: C. Corrado, J. Haltiwanger & D. Sichel (Eds), Measuring capital in the new economy (pp. 237–269). University of Chicago Press. Lichtenberg, F. (2005b). The impact of new drug launches on longevity: Evidence from longitudinal disease-level data from 52 countries, 1982–2001. International Journal of Health Care Finance and Economics, 5, 47–73. Lichtenberg, F. (2005c). Availability of new drugs and Americans’ ability to work. Journal of Occupational and Environmental Medicine, 47(4), 373–380. Lichtenberg, F. (2006). The effect of using newer drugs on admissions of elderly Americans to hospitals and nursing homes: State-level evidence from 1997–2003. PharmacoEconomics, 24(Suppl 3), 5–25. Lichtenberg, F., & Virabhak, S. (2007). Pharmaceutical-embodied technical progress, longevity, and quality of life: Drugs as ‘equipment for your health’. Managerial and Decision Economics, 28, 371–392. Romer, P. (1990). Endogenous technical change. Journal of Political Economy, 98, S71–S102. Sakellaris, P., & Wilson, D. (2001). The production-side approach to estimating embodied technological change. Finance and Economics Discussion Series 2001-20, Board of Governors of the Federal Reserve System. Sakellaris, P., & Wilson, D. (2004). Quantifying embodied technological change. Review of Economic Dynamics, 7(1), 1–26. Viscusi, W. K. (2005). The value of life. Discussion paper no. 517, Harvard Law School. http:// www.law.harvard.edu/programs/olin_center/papers/pdf/Viscusi_517.pdf

Data Used to Estimate Eq. (1). Disease

and and and and and and and and and

Endocrine Endocrine Endocrine Endocrine Endocrine Endocrine Endocrine Endocrine Endocrine

Number of Rx’s

Mean Vintage of Rx’s

Number of Deaths

Mean Age at Death

Potential Years of Life Lost before Age 65

1995 1996 1997 1998 1999 2000 2001 2002 2003 1995 1996 1997 1998 1999 2000 2001 2002 2003

16,316,361 17,133,504 17,533,563 17,918,165 19,188,559 20,147,558 20,385,191 22,400,650 23,330,710 1,782,830 1,889,272 1,950,002 2,167,689 2,636,979 3,366,276 3,726,524 4,407,413 5,053,317

1982.8 1983.4 1984.0 1984.7 1985.1 1985.5 1986.1 1987.4 1988.7 1953.2 1953.3 1953.7 1954.2 1954.8 1958.9 1964.2 1967.9 1970.2

7,448 7,783 8,154 7,932 8,321 8,301 8,403 9,125 9,222 871 433 372 436 450 413 408 428 454

73.0 73.1 72.7 72.8 73.4 74.0 74.2 74.2 74.4 54.8 72.4 73.9 72.5 71.1 72.5 72.3 74.1 74.5

23,090 24,213 28,023 27,585 26,993 23,785 24,213 27,298 26,658 13,788 1,995 1,658 2,358 2,655 2,143 1,818 1,870 1,670

113

Digestive Digestive Digestive Digestive Digestive Digestive Digestive Digestive Digestive Blood Blood Blood Blood Blood Blood Blood Blood Blood

Year

Pharmaceutical Innovation and the Longevity of Australians

APPENDIX A

114

(Continued ). Year

Number of Rx’s

Mean Vintage of Rx’s

Number of Deaths

Mean Age at Death

Potential Years of Life Lost before Age 65

Circulatory Circulatory Circulatory Circulatory Circulatory Circulatory Circulatory Circulatory Circulatory Skin Skin Skin Skin Skin Skin Skin Skin Skin Genitourinary Genitourinary

1995 1996 1997 1998 1999 2000 2001 2002 2003 1995 1996 1997 1998 1999 2000 2001 2002 2003 1995 1996

29,274,934 31,445,238 33,112,134 34,601,496 38,246,147 42,380,643 45,401,307 48,340,917 50,585,429 4,158,948 3,935,264 3,189,000 2,748,965 2,919,539 3,003,996 2,969,818 2,870,937 2,757,778 6,272,147 6,239,411

1983.0 1984.1 1984.9 1985.9 1986.9 1987.7 1988.2 1988.6 1989.0 1964.6 1966.3 1964.9 1964.8 1966.3 1967.2 1968.0 1968.1 1968.4 1976.6 1976.6

53,407 53,990 53,636 51,787 51,303 49,687 49,326 50,294 48,835 250 175 240 260 289 252 265 334 305 2,074 2,244

77.6 77.9 78.1 78.2 78.5 78.7 78.8 79.1 79.1 80.5 80.9 78.9 80.5 79.1 80.0 80.6 80.0 80.7 79.2 79.6

65,548 64,778 65,435 62,868 61,273 59,848 61,038 58,803 61,090 260 130 515 178 433 253 203 455 183 1,878 1,890

FRANK R. LICHTENBERG AND GAUTIER DUFLOS

Disease

1997 1998 1999 2000 2001 2002 2003 1995 1996 1997 1998 1999 2000 2001 2002 2003 1995 1996 1997 1998 1999 2000 2001 2002 2003

5,471,427 5,323,027 5,701,087 5,878,884 6,013,677 5,423,044 4,244,974 17,079,435 16,263,453 15,199,509 14,470,051 13,523,718 13,504,891 13,487,633 13,096,864 12,745,328 487,908 534,042 570,489 605,012 697,621 829,760 923,797 1,008,548 1,049,238

1978.3 1979.3 1979.5 1980.0 1980.8 1981.2 1981.4 1972.3 1972.7 1973.1 1973.5 1974.5 1974.7 1974.6 1974.8 1975.1 1972.6 1972.7 1973.2 1973.8 1973.7 1976.1 1977.7 1978.7 1979.3

2,588 2,697 2,768 2,692 2,812 2,983 3,001 1,070 1,638 1,522 1,454 1,603 1,646 1,675 1,790 1,754 34,368 35,252 35,363 35,609 35,856 36,374 37,497 38,426 38,392

80.1 80.3 80.7 80.6 81.0 81.1 80.8 68.7 62.1 67.1 68.7 69.9 70.4 70.6 71.0 72.0 70.1 70.3 70.3 70.5 70.9 71.2 71.3 71.5 71.5

2,095 2,230 2,195 2,043 1,868 1,850 2,350 7,463 18,278 12,308 10,513 10,458 10,290 10,413 10,578 8,958 115,888 118,093 117,395 117,388 114,630 111,575 115,190 115,128 115,990

Pharmaceutical Innovation and the Longevity of Australians

Genitourinary Genitourinary Genitourinary Genitourinary Genitourinary Genitourinary Genitourinary Infectious Infectious Infectious Infectious Infectious Infectious Infectious Infectious Infectious Neoplasms Neoplasms Neoplasms Neoplasms Neoplasms Neoplasms Neoplasms Neoplasms Neoplasms

115

116

(Continued ). Year

Number of Rx’s

Mean Vintage of Rx’s

Number of Deaths

Mean Age at Death

Potential Years of Life Lost before Age 65

Musculoskeletal Musculoskeletal Musculoskeletal Musculoskeletal Musculoskeletal Musculoskeletal Musculoskeletal Musculoskeletal Musculoskeletal Mental and Nervous Mental and Nervous Mental and Nervous Mental and Nervous Mental and Nervous Mental and Nervous Mental and Nervous Mental and Nervous Mental and Nervous Respiratory Respiratory

1995 1996 1997 1998 1999 2000 2001 2002 2003 1995 1996 1997 1998 1999 2000 2001 2002 2003 1995 1996

5,956,861 5,833,568 5,636,378 5,430,948 5,669,394 6,784,075 9,639,186 11,380,343 12,012,146 22,947,117 23,993,807 24,577,524 25,121,805 26,843,646 28,269,319 30,211,670 31,017,021 31,579,863 11,063,249 11,875,706

1978.6 1978.7 1978.9 1978.9 1979.0 1983.7 1990.2 1991.5 1992.0 1973.6 1974.7 1976.6 1978.1 1979.5 1980.9 1982.6 1983.9 1984.8 1981.1 1982.0

734 794 792 751 862 852 896 1,015 999 6,142 6,631 6,591 6,589 6,698 7,113 6,908 7,794 7,565 9,431 10,294

75.5 76.6 75.6 75.1 76.3 76.2 77.4 77.5 77.6 72.5 73.6 71.9 71.8 73.6 73.2 75.9 76.8 76.9 75.6 76.5

1,698 1,438 1,815 1,805 1,735 1,900 1,315 1,715 1,790 36,365 35,608 40,780 41,660 33,388 38,765 25,475 24,618 24,383 17,180 15,228

FRANK R. LICHTENBERG AND GAUTIER DUFLOS

Disease

1997 1998 1999 2000 2001 2002 2003 1995 1996 1997 1998 1999 2000 2001 2002 2003

11,297,442 10,649,965 11,232,232 11,419,935 10,362,935 10,165,471 9,751,515 5,857,233 5,734,316 5,499,593 5,478,050 5,927,943 6,298,017 6,671,412 6,847,396 6,820,565

1982.5 1984.3 1985.2 1986.1 1986.0 1986.9 1988.8 1969.4 1968.4 1968.6 1970.0 1973.2 1975.9 1977.5 1978.7 1979.4

10,349 9,614 9,613 10,907 10,626 11,668 11,892 16 18 9 15 11 10 10 8 15

76.8 76.8 77.2 77.9 77.7 78.3 78.7 47.5 53.1 69.7 71.8 64.3 59.5 74.5 63.1 64.5

15,028 14,543 13,365 13,695 14,515 13,790 14,380 440 390 75 93 143 143 40 113 213

Pharmaceutical Innovation and the Longevity of Australians

Respiratory Respiratory Respiratory Respiratory Respiratory Respiratory Respiratory Eye and Ear Eye and Ear Eye and Ear Eye and Ear Eye and Ear Eye and Ear Eye and Ear Eye and Ear Eye and Ear

117

SPILLOVER EFFECTS OF PRESCRIPTION DRUG WITHDRAWALS John Cawley and John A. Rizzo ABSTRACT Several high-profile prescription drugs have been withdrawn from the U.S. market in the last decade, yet there is no direct evidence of how a prescription drug withdrawal affects consumers’ use of remaining drugs within the same therapeutic class. In theory, remaining drugs in the therapeutic class could enjoy competitive benefits or suffer negative spillovers from the withdrawal of a competing drug. Using the Medical Expenditure Panel Survey, we test for spillovers following prescription drug withdrawals in six therapeutic classes between 1997 and 2001. Results vary, but we find stronger evidence of negative spillovers than competitive benefits. We conclude with a discussion of the characteristics of drugs and classes that may influence how remaining drugs are affected by a withdrawal in the class.

INTRODUCTION The withdrawal of prescription drugs from the U.S. market is a relatively frequent phenomenon; more than 75 drugs have been withdrawn since 1969 Beyond Health Insurance: Public Policy to Improve Health Advances in Health Economics and Health Services Research, Volume 19, 119–143 Copyright r 2008 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0731-2199/doi:10.1016/S0731-2199(08)19006-9

119

120

JOHN CAWLEY AND JOHN A. RIZZO

(Wysowski & Swartz, 2005).1 Some of the more recent withdrawals have been of prominent drugs that represented a large share of the market in their respective therapeutic classes. For example, in 1997 the Food and Drug Administration requested that Wyeth remove from the market its antiobesity drugs Pondimin and Redux that had been used by six million Americans because the drugs caused potentially fatal valvular heart disease (Connolly et al., 1997; Agovino, 2004).2 In 2004, Merck withdrew the pain medication Vioxx, used by an estimated 20 million Americans, for increasing the risk of heart attack and stroke (Agovino, 2004). This paper studies the withdrawals of seven drugs from six therapeutic classes between 1997 and 2001 and answers the following questions: Do remaining drugs in the same therapeutic class enjoy competitive benefits or suffer negative spillovers? Do those taking drugs in the class that were not withdrawn reduce compliance or quit? Do people who previously took the withdrawn drugs quit taking that class of drugs altogether? Are people less likely to initiate use of non-withdrawn drugs following the withdrawal of a therapeutically equivalent drug? On net, how does utilization of the non-withdrawn drugs change? The answers to these questions have important implications for understanding the nature of competition in the pharmaceutical industry and for assessing the economic effects of drug withdrawals. Moreover, the existence of spillover effects can provide important insight into how drugs compete and how product markets should be defined. The present study is timely, given recent withdrawals of ‘‘blockbuster’’ drugs such as Vioxx in 2004 and Bextra in 2005, and the potential for additional withdrawals in the near future. In November 2004, the U.S. FDA’s David Graham questioned whether additional drugs (Meridia, Crestor, Accutane, and Serevent) should be withdrawn from the market (Harris, 2004). These developments highlight a pressing need to understand how consumers respond, and the use of remaining drugs changes, in the wake of drug withdrawals. To our knowledge, this is the first direct study of consumer response to drug withdrawals. Moreover, the related literature implies divergent predictions. Studies of short-run changes in the stock prices of rival firms following drug withdrawals have found evidence of both positive and negative effects (Jarrell & Peltzman, 1985; Dowdell, Govindraj, & Jain, 1992; Ahmed, Gordella, & Nanda, 2002). This paper extends the literature by offering a direct, longer-term test of the impacts of drug withdrawals on spillovers. In addition, the study documents how consumers respond to the withdrawal of a prescription drug

Spillover Effects of Prescription Drug Withdrawals

121

by measuring spillover effects on initiations, quit rates, and the utilization of non-withdrawn drugs. We study a nationally representative patient-level database from the Medical Expenditure Panel Study for 1996 through 2002. Results vary, but we find stronger evidence of negative spillovers than competitive benefits. We then discuss the characteristics of drugs and classes that may influence how remaining drugs are affected by a withdrawal in the class.

CONCEPTUAL FRAMEWORK Our conceptual framework is straightforward. The withdrawal of a drug from the market could result in either competitive benefits or negative spillovers to the remaining drugs in the same therapeutic class. One might expect competitive benefits to the extent that the drugs in a given therapeutic class represent an oligopoly. Entry is regulated by the FDA and is very costly in time and money. The withdrawal of one competitor increases the residual demand and therefore equilibrium quantity supplied by remaining producers.3 If this increase in sales offsets the loss of previous customers who quit because of the withdrawal of the related drug, the remaining drugs enjoy competitive benefits. Several studies in finance have tested whether competitive benefits dominate negative spillovers by examining how pharmaceutical firms’ share prices change in the wake of bad news about a competitor’s product. Ahmed, Gardella, and Nanda (2002) find that competitors’ share prices rose significantly five days after the announcements of drug withdrawals that occurred between 1966 and 1998. Dowdell, Govindaraj, and Jain (1992) find that in the wake of the Tylenol poisonings, the share prices of rival pharmaceutical manufacturers rose relative to the share price of the manufacturer of Tylenol (Johnson & Johnson). On the other hand, one might expect negative spillovers to dominate any competitive benefits. Negative spillovers arise if, for example, consumers become concerned about the safety of the entire class of drugs and decrease their utilization of the non-withdrawn drugs to such an extent that those quits exceed the number of patients switching from the withdrawn to the non-withdrawn drugs. Jarrell and Peltzman (1985) study drug withdrawals during 1974–1982 and find evidence of net negative spillovers; specifically, the share prices of pharmaceutical companies fall an average of 1 percent in the two weeks surrounding the announcement of bad news that led to the withdrawal of a rival drug.4

122

JOHN CAWLEY AND JOHN A. RIZZO

Interestingly, some stock price studies of drug withdrawals find competitive benefits (Ahmed et al., 2002; Dowdell et al., 1992) while another finds negative spillovers (Jarrell & Peltzman, 1985). A further study finds no net effect of drug withdrawals on the share prices of rival pharmaceutical firms (Dranove & Olsen, 1994). These studies of stock price changes following drug withdrawals are informative about investor beliefs (e.g., about changes in the likelihood of regulation or FDA scrutiny of future drugs) but may not reflect changes in actual drug utilization patterns. Moreover, the follow-up period of these studies (ranging from days to weeks) is too brief to assess long-term trends following withdrawals. The fact that the studies come to divergent conclusions about spillovers underscores the need for a direct study of consumer behavior following drug withdrawals. This paper makes a number of contributions to the literature on the competitive effects of prescription drug withdrawals. First, it is the only direct study of how consumers respond to drug withdrawals. Earlier research has focused on investor beliefs, not consumer behavior. Second, this paper uses multiple measures to study competitive effects, including drug utilization, initiations, and quits of non-withdrawn drugs in response to withdrawals in the same class. This approach recognizes the multidimensional nature of competitive responses and provides a sense of the robustness of the results to alternative competitive outcome measures. Third, this paper uses a longer follow-up than previous studies of stock prices; we track consumers from up to three years before to three years after a drug withdrawal. Fourth, it provides a test of whether drugs within the same therapeutic class constitute a product market. While drug product markets are conventionally defined in terms of whether drug utilization patterns respond to significant changes in competitors’ prices or other competitive variables,5 in practice such changes are often difficult to observe. However, studies have found evidence that brand-name drugs within the same therapeutic class do compete. Lichtenberg and Philipson (2002) find that changes in the number of drugs in a therapeutic class affect utilization of drugs within that class. Lu and Comanor (1998) report that a brand’s intertemporal rate of price increase was lower when there were more branded competitors in the market.6 Our study contributes to this literature by examining six natural experiments to investigate whether utilization patterns of drugs within a therapeutic class are linked to the withdrawal of another drug within that class. This paper also relates to a literature in pharmaceutical economics on how consumers respond to information. Studies of the pharmaceutical

Spillover Effects of Prescription Drug Withdrawals

123

industry most commonly focus on the impact of advertising on the sales of the drug that was advertised (Azoulay, 2002; Calfee, Winston, & Stempski, 2002; Rizzo, 1999; Berndt, Bhattacharya, Mishol, Arcelus, & Lasky, 1995, 2002; Hurwitz & Caves, 1988; Leffler, 1981). However, recent research documents that the benefits of advertising may spill over to other drugs in the same therapeutic class. For example, direct-to-consumer advertising (DTCA) for one drug increases the sales of the entire class of drugs (Rosenthal et al., 2003; Iizuka & Jin, 2003). DTCA of one drug also increases compliance among users of other drugs within the same therapeutic class (Wosinska, 2003, 2004). In addition, marketing for prescription drugs has positive spillover effects for same brand over-the-counter (OTC) versions of the drugs, although DTCA for OTC products do not appear to spill over to the same brand in the prescription drug market (Ling, Ernst, & Margaret, 2002). Other research has focused on how physician prescribing behavior responds to various types of information, such as detailing and the results of clinical trials published in professional journals (Azoulay, 2002; Stern & Trajtenberg, 1998). We contribute to this literature by studying how pharmaceutical consumers respond to bad news in general and a drug withdrawal in particular.

METHODS Ideally, we would like to compare the market for a particular therapeutic class of drugs after a withdrawal in the class to its counterfactual: how that same market would look in the same years if the drug had not been withdrawn. Obviously, such information is unavailable. Nor is there any satisfactory ‘‘control’’ group in the form of a therapeutic class with identical trends in unobserved variables but no drug withdrawals (that would permit estimation of a differences-in-differences model). Therefore, we study the impact of prescription drug withdrawal by comparing the consumer use of competing drugs before withdrawal to consumer use of them after withdrawal, controlling for salient observables. One key observable is the number of scrips for all prescription drugs filled, per capita, by year in the respondent’s geographic region; this regressor is particularly important because during the period we study (1996–2002) there was a general upward trend in the use of pharmaceuticals in the U.S. (Banthin & Miller, 2005; Berndt, 2002). Failure to control this trend would bias our results in favor of finding competitive benefits of drug withdrawals.7

124

JOHN CAWLEY AND JOHN A. RIZZO

A limitation of our empirical strategy is that there may be trends in unobserved variables that changed specific drug markets around the time of the drug withdrawal; in other words, there may be omitted variable bias. We believe that such bias is likely to be relatively modest. Drug withdrawals are generally extremely well-publicized and are likely the dominant event in their therapeutic classes. For example, the withdrawal of Pondimin and Redux was accompanied by editorials in the New England Journal of Medicine and JAMA, and prominent coverage in virtually all major U.S. newspapers (the Los Angeles Times won a Pulitzer Prize for its coverage). In addition, we look for common patterns across six drug classes in which a withdrawal occurred between 1997 and 2001. If we see the same pattern across all six drug classes, we can be more confident that it is due to the withdrawals than due to random chance that unobserved changes happened to occur in each class around the time of withdrawal (that ranges across markets from 1997 to 2001). While our model does not control for the prices of drugs or advertising expenditures on such drugs, for our purposes these variables do not cause omitted variable bias. The reason is that we consider how the manufacturers of remaining drugs changed their price and advertising strategies in the wake of drug withdrawal to be a part of the overall impact of drug withdrawal and thus these influences do not represent bias but part of what we wish to measure. We estimate three types of models: (1) utilization, in which the binary dependent variable equals one if the respondent is using a non-withdrawn drug in that class in that year;8 (2) initiation, in which the binary dependent variable equals one if the respondent reports using a non-withdrawn drug in the current interview but did not report using one in the previous interview;9 and (3) quit, in which the binary dependent variable equals one if the respondent reported using a non-withdrawn drug in the previous, but not the current interview. The utilization, initiation, and quit equations are estimated as logit models. The coefficients on year indicator variables provide information about the net effect of the drug withdrawals on remaining drugs. Specifically, we compare the year prior to drug withdrawal to years after drug withdrawal (the year of withdrawal is a mixture of pre- and post-treatment and so is not the primary focus). In the Medical Expenditure Panel Survey (MEPS) data, the number of people with a given condition may fluctuate by chance with the selection of new MEPS sample members. Moreover, there have been trends over time in

Spillover Effects of Prescription Drug Withdrawals

125

some conditions; for example, obesity has doubled in the past 25 years in the U.S. (Hedley et al., 2004). To control for changes in the conditions of sample respondents over time, we control in our models for an indicator variable that equals one if the respondent has the condition treated by the class of drugs under consideration. We control for the following variables in our regressions: the trend in per capita number of scrips in respondent’s geographic area, indicator variables for year, whether the respondent has the condition treated by that class of drugs, gender, African-American, Hispanic, other race/ethnicity, married, whether the respondent has health insurance, whether the respondent’s health insurance includes prescription drug coverage, age categories, urban residence, Census Region categories, income categories, and education categories.

DATA AND EMPIRICAL SPECIFICATION This paper uses 1996–2002 data from the MEPS, which is collected by the Agency for Healthcare Research and Quality (AHRQ). The MEPS database is drawn from the National Health Interview Survey (NHIS) sample, so it is designed to be nationally representative and each year of the MEPS data may be linked to information from the previous year’s NHIS survey. The MEPS database has a complex survey design, which beyond stratifying by sampling units, includes clustering and oversampling of certain subgroups such as minorities. Therefore, our statistical analyses use weights provided in MEPS to correct mean values and appropriate statistical methods in Stata to obtain correct standard errors. The MEPS has an overlapping panel design, in which two calendar years of information are collected from each household through six interviews. We pool both calendar-year observations on each adult (age 18 and older) and pool years 1996–2002. Our final sample is 124,314. We use the first of two calendar years of MEPS data on each person in order to determine whether in the second period the person has initiated or quit using a relevant drug. As a result, we lose half of our data when we estimate models of initiation. The MEPS database consists of a number of files. We linked the FullYear Consolidated File to the Prescribed Medicines File for each year. The Full-Year Consolidated File is at the person-year level and includes information on health care utilization and expenditures, sociodemographic and socioeconomic characteristics, and health insurance status. The

126

JOHN CAWLEY AND JOHN A. RIZZO

Prescribed Medicines File is an event-level file that includes information on specific drug use, the amounts paid for those drugs by patient and insurers, and the length of time that the drug was taken. We convert this event-level data into person-year data and link it to the consolidated MEPS files, which include patient-year level information on the other variables included in this analysis. Given the way the MEPS data are collected, we must study consumer utilization of drugs by calendar year; MEPS asked respondents to list all drugs taken since the last interview up to the end of the year. So, for example, although Redux and Pondimin were withdrawn in September, 1997, any MEPS respondents who had used anti-obesity drugs prior to the withdrawal were still listing them even if they were interviewed in October–December, 1997. Quits can only be ascertained in the next calendar year. We study the seven drugs from six therapeutic classes that were withdrawn between 1997 and 2001 that are listed in Table 1. Pondimin and Redux, both anti-obesity drugs, were withdrawn in 1997 for causing valvular heart disease. Duract was a pain medication withdrawn in 1998 for causing liver failure. Posicor was an anti-hypertensive drug also withdrawn in 1998 for lowering heart rates and causing adverse drug interactions. Propulsid was a heartburn medication withdrawn in 2000 for causing potentially fatal irregular heartbeat. Lotronex treated Irritable Bowel Syndrome (IBS) and was withdrawn in 2000 for causing ischemic colitis (intestinal inflammation due to lack of blood flow). Baycol was a cholesterol-busting drug withdrawn in 2001 for causing fatal rhabdomyolysis (severe adverse muscle reaction that can damage the kidney and other organs). We use the Multum Lexicon File, released in Fall 2004, to identify drugs that remain available in the therapeutic class of each withdrawn drug. These drugs that remained on the market and the name of their therapeutic class are also listed in Table 1. For the sake of clarity, we will refer to these remaining drugs by the indication that the drugs treat. We are forced to exclude from our analysis a few other drugs that were withdrawn between 1997 and 2001. We exclude Seldane and Hismanal, both of which treated seasonal allergies and caused potentially fatal irregular heartbeats, because their withdrawals occurred in consecutive years (1998 and 1999), making it impossible to disentangle the change in utilization due to each withdrawal. Rezulin, a diabetes drug, was withdrawn in 2000 for causing liver failure but we exclude it from our analysis because it was the only drug in its class – there were no competitors to suffer spillovers or reap competitive benefits.

Brand Name of Withdrawn Drug

Indication Treated

Date Withdrawn

Primary Health Risk (Reason Drug Withdrawn)

Name of Class

Pondimin Redux

Obesity

9/15/1997

Valvular heart disease

Anorectics

Duract

Pain

6/22/1998

Liver failure

Non-steroidal antiinflammatory

Posicor

Hypertension

6/8/1998

Calcium channel blockers

Propulsid

Heartburn

7/14/2000

Lotronex Baycol

Irritable Bowel Syndrome High Cholesterol

11/28/2000 8/8/2001

Lowered heart rate, adverse interactions with 26 other drugs Potentially fatal irregular heartbeat Ischemic colitis Fatal rhabdomyolysis

GI stimulants Anti-diarrheals HMG-CoA Reductase Inhibitors

Drugs in Class Remaining on Market

Adipex, Ionamin, Meridia, Phentermine, Diethylpropion, Dexedrine Arthrotec, Naproxon, Ibuprofin, Diclofenac, Daypro, Diclofenac Sodium Adalat, Calan, Cardizem, Covera, Diltiazem Hcl, Norvasc Metoclopramide Hcl, Reglan

Spillover Effects of Prescription Drug Withdrawals

Seven Drug Withdrawals Studied.

Carafate, Cytotec, Sucralfate Lipitor, Zocor, Pravachol, Lescol, Mevacor

127

128

JOHN CAWLEY AND JOHN A. RIZZO

For each therapeutic class considered, we study the following three outcomes: (1) an indicator that equals one if the respondent in that year had a scrip for a drug within that class; (2) an indicator that equals one if the respondent began taking a drug within that class; and (3) an indicator that equals one if the respondent quit taking a drug within that class. The means of the dependent variables by class are listed in Table 2. Drugs that treat pain are the ones most often used by MEPS respondents; use at some point in a calendar year is reported by 10.8 percent of observations. Hypertension and cholesterol drugs are next most common, used by 6.8 and 6.5 percent of the sample. Drugs that treat obesity, heartburn, and IBS are much less frequently used; only one-half of a percent or less of the sample reports their use in a year. (Note that these do not include use of the withdrawn drug in the years prior to withdrawal.) There exist several measures of or proxies for the out-of-pocket price of drugs, each with its advantages and drawbacks. MEPS respondents list the amount they paid out-of-pocket for each drug, but the prices faced by those who did not buy drugs are not observed. National average wholesale prices are available from Medi-Span, but these are collinear with the year fixed effects. To address patient costs while avoiding problems of multicollinearity, we use two proxies for the out-of-pocket cost of prescription drugs. The first is an indicator variable for whether the respondent lacked Means of Dependent Variables (Refer to Non-withdrawn Drugs in Each Class). Indication Treated by Class that Experienced Drug Withdrawal Obesity Pain Hypertension Heartburn IBS Cholesterol

Utilization

0.005 0.108 0.068 0.004 0.004 0.065

(N ¼ 124,314) (N ¼ 124,314) (N ¼ 124,314) (N ¼ 124,314) (N ¼ 124,314) (N ¼ 124,314)

Initiations

0.003 0.065 0.015 0.002 0.002 0.022

(N ¼ 55,030) (N ¼ 48,852) (N ¼ 51,730) (N ¼ 55,125) (N ¼ 55,133) (N ¼ 51,926)

Quits

0.665 0.592 0.230 0.549 0.632 0.148

(N ¼ 297) (N ¼ 6,475) (N ¼ 3,597) (N ¼ 202) (N ¼ 194) (N ¼ 3,401)

Notes: (1) Data: Medical Expenditure Panel Survey, 1996–2002. (2) Each row refers to non-withdrawn drugs in a therapeutic class that remained after the named drugs were withdrawn. For example, the first row refers to obesity drugs that remained on the market after Redux and Pondimin were withdrawn. (3) Sample size shown in parentheses below mean of the dependent variable.

Spillover Effects of Prescription Drug Withdrawals

129

health insurance; uninsurance would raise the cost of a physician visit to receive a prescription. The second price proxy is an indicator for whether the respondent’s health insurance includes prescription drug coverage, which would lower the cost of filling a prescription. We acknowledge that these indicators for health insurance coverage are endogenous; those who sought to consume large quantities of prescription drugs might be more likely to acquire health insurance and prescription drug coverage. We calculate the number of scrips for all prescription drugs filled, per capita, by year in the respondent’s geographic area, where the area is defined by MEPS sampling areas. There are 258 such areas and we calculate the per capita scrips per year for each. We determine whether the respondent has the condition treated by a given class of drugs using the ICD9 condition codes provided in the MEPS.10 For obesity, we use the FDA and NIH medical criteria for the use of anti-obesity drugs: a body mass index (BMI) of 30 or greater, or BMI of 27 or higher with at least one obesity-related comorbidity (such as hypertension, hyperlipidemia, type II diabetes, coronary heart disease, or sleep apnea). BMI is constructed using self-reported weight and height from the NHIS corrected for reporting error (Cawley, 2004). For those with a BMI of less than 30, we re-classify them as obese if they have ICD9 code 278 (obesity).11

EMPIRICAL RESULTS Tables 3–5 present the results of our regressions for utilization, initiation, and quitting. In the interest of brevity, the tables present only the parameters of interest: the coefficients on the year indicator variables.12 Table 3 includes results from logit regressions of utilization. Each cell of the table includes the odds ratio and below that the t-statistic in parentheses. The odds ratio indicates the odds of utilizing non-withdrawn drugs in a given class in a given year relative to the year prior to the drug withdrawal. Note that the year prior to drug withdrawal differs by class. For obesity drugs it is 1996, for pain and hypertension drugs it is 1997, for heartburn and IBS it is 1999, and for cholesterol it is 2000. Also note that there are empty cells in the table. Because the MEPS began in 1996, there are data for only one year prior to the withdrawal of Pondimin and Redux in 1997. In contrast, for cholesterol drugs, we have many years prior to the withdrawal of Baycol in 2001, but only one year after withdrawal. The first row of Table 3, which corresponds to the anti-obesity drugs, provides evidence of negative spillovers. One year after the withdrawal of

130

Utilization of Non-Withdrawn Drugs in that Therapeutic Class N ¼ 124,314. Odds Ratios and (t-Statistics). Indication Treated

Withdrawal – 3 Years

Year Withdrawn

Withdrawalþ1 Year

Withdrawalþ2 Years

Withdrawalþ3 Years

1.076 (1.79) 1.291 (4.42) 1.094 (0.45) 0.901 (0.49) 0.846 (2.50)

1.104 (0.70) 0.960 (0.96) 0.930 (1.31) 1.249 (1.09) 0.772 (1.16) 1.271 (3.94)

0.673 (2.16) 0.886 (2.94) 0.961 (0.61) 1.537 (2.21) 0.682 (1.86) 1.523 (4.30)

0.664 (1.93) 0.734 (6.63) 0.893 (1.75) 1.408 (1.55) 0.919 (0.40)

0.436 (4.18) 0.684 (7.97) 0.892 (1.81)

Obesity Pain Hypertension Heartburn IBS Cholesterol

0.598 (1.85) 1.229 (0.89) 0.781 (2.49)

Notes: (1) Data: Medical Expenditure Panel Survey, 1996–2002. (2) Cells contain odds ratios and the absolute value of t-statistics in parentheses. (3) Asterisks indicate statistical significance: Significant at 10%; Significant at 5%; Significant at 1%. (4) In addition to time indicators, models control for the following regressors: the trend in per capita number of scrips in respondent’s geographic area, indicator variables for whether the respondent has the condition treated by that class of drugs, gender, African-American, Hispanic, other race/ethnicity, married, whether the respondent has health insurance, whether the respondent’s health insurance includes prescription drug coverage, age categories, urban residence, Census Region categories, income categories, and education categories. (5) Standard errors are cluster-corrected by individual. (6) The STATA command svylogit is used to account for MEPS sample weights and the MEPS survey design (stratum and psu).

JOHN CAWLEY AND JOHN A. RIZZO

Withdrawal – 2 Years

Spillover Effects of Prescription Drug Withdrawals

131

Pondimin and Redux, MEPS respondents were only 67.3 percent as likely to be taking an anti-obesity drug as they were the year before the withdrawals. These negative spillovers are long-lived; the probability of use is only 66.4 percent as likely two years after withdrawal, and only 43.6 percent as likely three years later. Each of these is statistically significant, and the fact that the odds ratios continue to fall suggests that the negative spillovers may not have run their course even after three years. There is a similar pattern of negative spillovers for pain medications. One year after the withdrawal of Duract, utilization is only 88.6 percent as likely as it had been the year before withdrawal. Two years later it is only 73.4 percent as likely, and three years later it is 68.4 percent as likely. The utilization of hypertension drugs does not immediately experience a statistically significant change; one year after withdrawal, utilization is 96.1 percent as likely as it was the year prior to withdrawal (which is not statistically significant). However, the decrease in utilization is statistically significant two years after withdrawal (when it falls to 89.3 percent as likely) and three years after withdrawal (89.2 percent as likely). The pattern of use of IBS drugs is also consistent with negative spillovers, but they quickly dissipate. One year after the withdrawal of Lotronex, use of IBS drugs was only 68.2 percent as likely as it had been the year before withdrawal. The decrease in utilization two years after withdrawal is not statistically significant. The drop in utilization of non-withdrawn obesity, pain, hypertension, and IBS drugs is consistent with the findings in Jarrell and Peltzman (1985) that the stock price of pharmaceutical firms falls after a drug withdrawal by a rival; presumably the lower stock price reflects decreased anticipated market share due to negative spillovers. For two other drug classes there appears to be evidence of competitive benefits, which is consistent with the stock price studies of Ahmed et al. (2002) and Dowdell et al. (1992). The use of non-withdrawn heartburn medications is 53.7 percent more likely one year after the withdrawal of Propulsid than the year before withdrawal. The use of cholesterol drugs is 52.3 percent more likely one year after the withdrawal of Baycol than it was the year before withdrawal. However, the pattern of odds ratios across years for these drugs suggests that these results may reflect long-term trends in unobservables rather than the withdrawals. For example, the point estimates of odds ratios for heartburn medications are rising throughout the period observed: from .598 three years before withdrawal to 1 the year before withdrawal (by construction) to 1.408 two years after withdrawal. This pattern suggests that utilization of heartburn medications was rising

132

JOHN CAWLEY AND JOHN A. RIZZO

throughout this period, and therefore the changes observed after withdrawal should not be attributed solely to the withdrawal. A similar pattern is evident for cholesterol medications. The point estimates of odds ratios rise from .781 three years prior to withdrawal to 1 the year before withdrawal (by construction) to 1.523 the year after withdrawal. For both heartburn and cholesterol drugs, there seems to be a trend toward increasing use over this period; while we control for the trend in scrips per capita by geographic area, the use of cholesterol and heartburn medications may have exceeded that of all drug classes as a whole. As a result of this trend, increases in utilization after withdrawal should not necessarily be interpreted as evidence of competitive benefits. We next study changes in utilization: initiations and quits. A limitation of these regressions is that we are forced to exclude about half of our data; the MEPS includes two observations for each person (each corresponding to a calendar year) and the first of those must be used to determine whether the second period represents an initiation or quit. All observations from 1996 must be dropped for this reason. This is a particular problem for studying the obesity drugs, because 1996 is the only year in the MEPS that is prior to the withdrawal of Pondimin and Redux in 1997. As a result, the best we can do for obesity drugs is compare changes in utilization after withdrawal to those of the year of withdrawal. This is a limitation. For example, there may have been withdrawal-based initiations of nonwithdrawn obesity drugs in 1997 because Pondimin and Redux were withdrawn on September 15, 1997. Thus the treatment effect may already be apparent in our omitted year for obesity drugs. For all other drug classes, dropping 1996 is not a problem because the drug withdrawals in those classes took place in 1998 or later so there remains at least one prewithdrawal year of data in the MEPS. Results for initiations of non-withdrawn drugs are provided in Table 4, which reports odds ratios relative to the year prior to withdrawal. There is evidence of negative spillovers in the form of significantly lower initiations post-withdrawal for obesity, pain, and IBS medications. For obesity drugs, initiation of non-withdrawn drugs was only 44.4 percent as likely one year after withdrawal as the year of withdrawal. Initiations remained significantly lower two and three years after withdrawal, with initiation only 33.2 percent as likely three years after withdrawal. However, these results may overstate the negative spillovers. Because many former users of Pondimin and Redux could have switched to non-withdrawn drugs in the year of the withdrawal (the omitted comparison year in this case), initiations in subsequent years may seem small by comparison. However, it is unknown

Indication Treated

N

Obesity

55,030

Pain

48,852

Hypertension

51,730

Heartburn

55,125

IBS

55,133

Cholesterol

51,926

Withdrawal – 3 Years

1.127 (0.24) 1.175 (0.43) 0.791 (1.55)

Withdrawal – 2 Years

1.468 (0.73) 0.720 (0.97) 0.596 (3.16)

Year Withdrawn

Withdrawalþ1 Year

Withdrawalþ2 Years

Withdrawalþ3 Years

0.915 (1.14) 0.691 (2.30) 1.995 (1.53) 0.607 (1.16) 0.905 (0.61)

0.444 (2.68) 0.877 (1.89) 0.888 (0.66) 2.608 (2.27) 0.458 (2.09) 1.269 (1.41)

0.506 (2.07) 0.703 (4.12) 0.894 (0.74) 1.330 (.059) 0.936 (0.16)

0.332 (3.82) 0.680 (4.46) 0.994 (0.04)

Spillover Effects of Prescription Drug Withdrawals

Initiation of Non-Withdrawn Drugs in that Therapeutic Class Odds Ratios and (t-Statistics).

Notes: (1) Data: Medical Expenditure Panel Survey, 1996–2002. (2) Cells contain odds ratios and the absolute value of t-statistics in parentheses. (3) Asterisks indicate statistical significance: Significant at 10%; Significant at 5%; Significant at 1%. (4) In addition to time indicators, models control for the following regressors: the trend in per capita number of scrips in respondent’s geographic area, indicator variables for whether the respondent has the condition treated by that class of drugs, gender, African-American, Hispanic, other race/ethnicity, married, whether the respondent has health insurance, whether the respondent’s health insurance includes prescription drug coverage, age categories, urban residence, Census Region categories, income categories, and education categories. (5) Standard errors are cluster-corrected by individual. (6) The STATA command svylogit is used to account for MEPS sample weights and the MEPS survey design (stratum and psu).

133

134

JOHN CAWLEY AND JOHN A. RIZZO

the degree to which the negative spillovers are overstated. The results for utilization (for which the omitted year was that prior to withdrawal since it did not need to be used to assess the change in utilization) also indicated substantial negative spillovers lasting at least three years however. Initiations of pain medications also reflect negative spillovers. Relative to the year before withdrawal, initiations were only 87.7 percent as likely one year after withdrawal, 70.3 percent as likely two years after, and 68 percent as likely three years after. The negative spillovers for IBS drugs are shorterlived. Only one year after withdrawal is initiation significantly lower (45.8 percent as likely as the year before withdrawal). Initiations of non-withdrawn hypertension drugs fell to only 69.1 percent of their previous level in the year of withdrawal. In subsequent years the odds ratios remain below 1 but the change relative to the year before withdrawal is not statistically significant. Only one drug class, for heartburn, has results consistent with competitive benefits. One year after withdrawal, the initiation of non-withdrawn heartburn drugs was 160.8 percent more likely than it was the year before withdrawal. There are no statistically significant benefits in the following year (two years after withdrawal) however. Quits of non-withdrawn drugs are detailed in Table 5. Quits of nonwithdrawn obesity drugs rise to 113.3 percent the year after withdrawal. One year after that the change in quits is not statistically significant. No other drug class exhibits negative spillovers in quitting behavior. In contrast, hypertension and cholesterol drugs appear to enjoy competitive benefits as a result of the withdrawal of a rival drug. Quitting of non-withdrawn hypertension drugs is only 53.8–61.3 percent as likely in any of the three years following withdrawal as it was the year before withdrawal. Quits of non-withdrawn cholesterol drugs are only 70.2 percent as likely one year after as one year before withdrawal. However, the cholesterol results are curious because quitting was also significantly less likely three years before and two years before withdrawal; it may be that by chance (or because news of Baycol’s adverse effects was disseminating) the year before withdrawal had an unusually high quit rate and as a result in every other year quits were significantly less likely. We also examined a special kind of quitting: quitting of the entire class of drugs by those who were taking the withdrawn drugs in the year they were withdrawn. This is a very small subsample of our overall sample, because we can only study those who in their first MEPS observation are taking the withdrawn drug in the year it was withdrawn (and therefore at risk of quitting). Across all drug classes, 70.8 percent of those taking

Indication Treated

N

Obesity

297

Pain

6,475

Hypertension

3,597

Heartburn

202

IBS

194

Cholesterol

3,400

Withdrawal – 3 Years

1.948 (0.97) 1.508 (0.55) 0.665 (1.99)

Withdrawal – 2 Years

0.988 (0.02) 0.533 (0.80) 0.696 (1.77)

Year Withdrawn

Withdrawalþ1 Year

Withdrawalþ2 Years

Withdrawalþ3 Years

0.871 (1.55) 0.603 (3.38) 2.322 (1.31) 1.305 (0.34) 0.717 (1.84)

2.133 (1.72) 1.066 (0.66) 0.554 (3.53) 0.634 (0.74) 1.763 (0.68) 0.702 (1.72)

1.188 (0.33) 1.134 (1.35) 0.613 (3.30) 0.569 (0.83) 3.233 (1.29)

0.743 (0.66) 1.242 (2.01) 0.538 (3.66)

Spillover Effects of Prescription Drug Withdrawals

Quits of Non-Withdrawn Drugs in that Therapeutic Class Odds Ratios and (t-Statistics).

Notes: (1) Data: Medical Expenditure Panel Survey, 1996–2002. (2) Cells contain odds ratios and the absolute value of t-statistics in parentheses. (3) Asterisks indicate statistical significance: Significant at 10%; Significant at 5%; Significant at 1%. (4) In addition to time indicators, models control for the following regressors: the trend in per capita number of scrips in respondent’s geographic area, indicator variables for whether the respondent has the condition treated by that class of drugs, gender, African-American, Hispanic, other race/ethnicity, married, whether the respondent has health insurance, whether the respondent’s health insurance includes prescription drug coverage, age categories, urban residence, Census Region categories, income categories, and education categories. (5) Standard errors are cluster-corrected by individual.

135

136

JOHN CAWLEY AND JOHN A. RIZZO

the withdrawn drug in the year it was withdrawn (N ¼ 79) did not switch to a drug that remained available in the same class; that is, they quit the entire class. For the two classes with the largest number of observations in this analysis, 48.7 percent of the 37 MEPS respondents who took Baycol in 2001 quit taking any cholesterol drug, and 94.7 of the 19 MEPS respondents who took Pondimin or Redux in 1997 quit taking any antiobesity drug.

PREDICTING NEGATIVE SPILLOVERS OR COMPETITIVE BENEFITS Our results indicate that there is no definitive pattern following drug withdrawal, though most results point to negative spillovers. In this section, we briefly explore what factors might determine whether a drug class enjoys competitive benefits or suffers negative spillovers. Because the MEPS data cover only seven withdrawals in six classes, we lack power to formally test our class-level hypotheses, but we can check to see if the results in this paper are consistent with our predictions. The withdrawal of a drug is expected to impose both gross negative spillovers and gross competitive benefits. The gross negative spillover is that non-withdrawn drugs will lose customers they would have otherwise had as a result of the withdrawal, both in the form of increased quits and decreased initiation; perhaps because the patients fear that the adverse effects of the withdrawn drug are shared by the remaining drugs in the class. The gross competitive benefit is that the non-withdrawn drugs gain some business as a result of the withdrawal; some of those who were previously taking the withdrawn drug will switch to other drugs that remain available in the same therapeutic class after their drug is withdrawn from the market. In order for there to be competitive benefits on net, the number of former users of the withdrawn drugs who switch to remaining drugs must exceed the number of previous users of non-withdrawn drugs who quit the class plus the number of people who would have initiated use in the absence of the withdrawal but will not initiate because of the withdrawal, or: gross increase4gross decreases switchers4additional quits þ lost initiators N W ð1  xÞ4N NW  y þ N NONE  z

Spillover Effects of Prescription Drug Withdrawals

137

Where NW is the number of people who were taking the withdrawn drug, NNW is the number of people who were taking the non-withdrawn drugs, and N NONE is the number of people not using any drug in the class (i.e., the number at risk of initiating use). Letting x denote the percentage of people previously taking the withdrawn drug who quit the class entirely after the drug withdrawal, (1x) is the percentage of those who previously used the withdrawn drug who switch to a drug that remains on the market in the therapeutic class. The increase in rate of quits of non-withdrawn drugs as a result of the withdrawal is denoted as y, and z is the decrease in the initiation rate of non-withdrawn drugs due to the withdrawal. According to these equations, competitive benefits are more likely if ceteris paribus, the following conditions hold. First, the number of people taking the withdrawn drug is large relative to the number of people already taking the non-withdrawn drugs in the same class and the number of lost initiators. Competitive benefits are more likely when the withdrawn drug had a high market share because it ensures that a large number of people are at risk of switching to non-withdrawn drugs in the same class. The market share of the withdrawn drugs we study varies dramatically. In the MEPS Prescribed Medicines File, Propulsid represented 57.7 percent of the heartburn medication market at the time it was withdrawn, and Pondimin and Redux jointly represented 46.9 percent of the anti-obesity drug market. In contrast, Lotronex accounted for 8.1 percent of the IBS market, and Baycol for 4.7 percent of the cholesterol drug market. The other two withdrawn drugs had trivial shares of the market: Duract was 1.2 percent of the pain medication market when it was withdrawn and Posicor was only 0.3 percent of the hypertensive market. The large market shares of withdrawn drugs in the obesity and heartburn markets suggests that we should be more likely to find net competitive benefits of the withdrawals in the obesity and heartburn classes, ceteris paribus. In the heartburn class, we do in fact find evidence of net competitive benefits in the overall utilization (Table 3) and initiation (Table 4). However, for obesity drugs we find strong evidence of net negative spillovers in those same two outcomes. We also find evidence of negative spillovers in overall utilization and initiation for two classes in which the withdrawn drug had small market share: pain and hypertension medications. Duract and Posicor were used by so few that the number switching from them to non-withdrawn drugs in the same class was insufficient to offset the quits of non-withdrawn

138

JOHN CAWLEY AND JOHN A. RIZZO

drugs. In both cases, overall utilization of non-withdrawn drugs fell after withdrawal. However, market share of the withdrawn drug is not the only variable that matters. Competitive benefits are more likely the smaller are x (the percentage of previous users of the withdrawn drug that quit the entire class after withdrawal), y (the increase in rate of quits by users of non-withdrawn drugs after withdrawal), and z (the decrease in initiations of non-withdrawn drugs after withdrawal). Each of these variables will be smaller when the remaining drugs are perceived by consumers and physicians to be unlikely to share the adverse health events that led to the withdrawal of the other drug in the class, especially when the remaining drugs are efficacious or when there exists few over-the-counter or non-pharmacologic treatments for the condition. While we do not have quantitative measures of these factors, we consider this to be an important direction for future work, so the FDA can better predict how consumers will respond if they withdraw a given drug from the market.

GENERALIZABILITY Several factors should be considered when generalizing the results associated with these drug withdrawals to other classes. First, the withdrawal of Redux and Pondimin was extremely well-publicized and this may have led to greater response by consumers than is typical. Second, one non-withdrawn obesity drug, Phentermine, was both a substitute to and a complement for the withdrawn drugs. It was a substitute because it could be prescribed in the place of Redux or Pondimin, but was also a complement in that it was the other ingredient in the drug cocktail fen-phen. The complementary nature of Phentermine suggests a smaller increase in utilization after the withdrawal of Redux and Pondimin than one would predict if Phentermine were exclusively a substitute for the withdrawn drugs. Markets in which remaining drugs are exclusively substitutes may exhibit greater competitive benefits and weaker negative spillovers. Third, the withdrawn IBS drug Lotronex is a special case because it was returned to the market in November of 2002, and approved for only a small segment of the patient population. There is no respondent in the 2002 MEPS data who reports taking Lotronex, but the fact that it was returned to market at all makes it a special case and may limit its generalization to permanent drug withdrawals in other classes.

Spillover Effects of Prescription Drug Withdrawals

139

CONCLUSIONS Our findings complement a literature in finance that documents changes in the prices of shares in pharmaceutical companies after a rival firm’s drug is withdrawn from the market. Some papers in that literature document negative spillovers (Jarrell & Peltzman, 1985) while others document competitive benefits (e.g., Ahmed et al., 2002; Dowdell et al., 1992). Our results vary by drug class. For three classes (obesity, pain, and IBS) we find evidence of negative spillovers, while in two others (heartburn and cholesterol) we find evidence of competitive benefits. In another class (hypertension) we find evidence of both negative spillovers (in the form of decreased utilization and initiations) and competitive benefits (in the form of lower quits). Across all drug classes and outcomes, the evidence for competitive benefits is weaker because it is also consistent with trends in unobserved factors or idiosyncratic comparison years. Overall, we find stronger evidence of negative spillovers than competitive benefits. These results also have implications for drug product market definition. Both the positive and negative spillovers in the wake of drug withdrawal that we document suggest that drugs within the same therapeutic class are to some extent substitutable and hence compete within the same product market. Finally, our paper also relates to a recent literature in pharmaceutical economics that documents consumer responses to positive information such as direct-to-consumer advertising (Rosenthal et al., 2003; Iizuka & Jin, 2003). Our findings establish that the effects of bad news also can spill over throughout a drug class.

NOTES 1. The roughly 75 drugs withdrawn between 1969 and 2002 represent about 1 percent of all marketed drugs (Wysowski & Swartz, 2005). 2. Although the Food and Drug Administration (FDA) does request that a manufacturer withdraw a drug from a market, the FDA can mandate the withdrawal if necessary. When asked by the FDA to withdraw a drug because of safety concerns, manufacturers have agreed in all cases except one: Ceiba-Geigy refused to voluntarily withdraw the anti-diabetic drug Phenformin in 1976. If a company refuses the FDA’s request, the FDA can begin procedures to compel withdrawal and it was through this process that Phenformin was taken off the market (Meadows, 2002). 3. Overall sales of the class could fall even if the sales of the remaining drugs are higher after the withdrawal.

140

JOHN CAWLEY AND JOHN A. RIZZO

4. Examples from outside the pharmaceutical industry, in which the negative spillovers of information dominated competitive benefits include: decreases in the prices of shares for airlines after a crash by a competing airline (Bosch, Eckard, & Singal, 1998) and decreases in the price of shares for nuclear energy firms after the 1979 core meltdown at the Three Mile Island nuclear power plant (Hill & Schneeweis, 1983). 5. More specifically, in defining product markets, the Department of Justice guidelines recommend that all relevant information be considered, including ‘‘evidence that sellers base business decisions on the prospect of buyer substitution between products in response to relative changes in price or other competitive variables . . . .’’ See: www.usdoj.gov/atr/public/guidelines/horiz_book/01.html at section 1.11. 6. For a review of the evidence on competition between brand-name drugs see: Congressional Budget Office (1998): How Increased Competition from Generic Drugs has Affected Prices and Returns in the Pharmaceutical Industry, ch. 3 (available at www.cbo.gov/showdoc.cfm?index ¼ 655&sequence ¼ 0). 7. Another potentially important unobservable is the change in the FDA’s policy allowing direct-to-consumer advertising, which occurred during the period of our study. This provides a further motivation for controlling for scrip use over time to the extent that this policy change affected drug use. We thank an anonymous referee for calling this point to our attention. 8. If utilization were the only outcome in which we were interested this could be measured roughly by using aggregate sales data. However, we also wish to test how drug withdrawals affect initiations and quits, and this requires longitudinal micro data. Even with respect to utilization, we wish to control for correlates of demand such as income and insurance status, which also requires micro data. 9. There are several steps involved for a patient to receive a prescription drug. First, the patient must decide whether to visit a physician. Second, the physician must determine whether to prescribe any drug, and then which drug to prescribe. Third, the patient must decide whether to fill the prescription. We set aside explicit consideration of the agency relationship between consumer and physician and study consumer use of drugs as an outcome; thus our results reflect consumer behavior under the average agency relationship. 10. Specifically, respondents are considered to have pain if they have ICD9 code 204 ( joint pain), 205 (back problems), or 84 (headache); hypertension if they have code 98 (hypertension); heartburn if they have code 787 (heartburn and others); IBS if they have code 565 (IBS and others); and to have high cholesterol if they have code 53 (lipid disorders including high cholesterol). There are limitations to the use of ICD9 codes to classify respondents as having the condition treated by a particular drug class. For some ICD9 codes, false positives are a concern: ICD9 53 includes other lipid disorders than high cholesterol. Likewise, the ICD9 codes for heartburn and IBS are broader than we would like. On the other hand, there are likely false negatives for pain; the ICD9 codes permit us only to control for particular sources of pain (back, joint, head). While there are limitations to the use of ICD9 codes to identify respondents with a given condition, they remain the best data for that purpose available in the MEPS.

Spillover Effects of Prescription Drug Withdrawals

141

11. We determine obesity-related comorbidities using ICD9 codes, but sleep apnea is not one recorded by MEPS. 12. The full set of regression results is available from the authors on request.

ACKNOWLEDGMENTS This research was supported with an unrestricted educational grant from the Merck Company Foundation, the philanthropic arm of Merck & Co., and a grant from the Cornell Institute for the Social Sciences. We thank Rena Conti, Dhaval Dave, Julie Donohue, Richard Frank, David Grabowski, Henry Grabowski, Caroline Hoxby, Robert Kaestner, Margaret Kyle, Sara Markowitz, Sean Nicholson, David Ridley, Judy Shinogle, and seminar participants at Harvard University, Princeton University, Duke University, the University of Illinois at Chicago, the American Economic Association meetings, the International Health Economics Association World Congress, the NBER Summer Institute in Health Care, the Association for Public Policy Analysis and Management meetings, the Southeastern Health Economics Study Group Conference, and the Eastern Economic Association meetings for helpful comments. We also thank Rebecca Friedkin who provided excellent programming assistance.

REFERENCES Agovino, T. (2004). Merck faces huge fallout over Vioxx suits. Washington Post, November 4. Ahmed, P., Gardella, J., & Nanda, S. (2002). Wealth effect of drug withdrawals on firms and their competitors. Financial Management, 31(3), 21–41. Azoulay, P. (2002). Do pharmaceutical sales respond to scientific evidence? Journal of Economics and Management Strategy, 11, 551–594. Banthin, J., & Miller, G. (2005). Trends in prescription drug expenditures by Medicaid enrollees. Working paper: Agency for healthcare research and quality. Paper presented at 2005 Association for Public Policy Analysis and Management Meetings, Washington, DC. Berndt, E. (2002). Pharmaceuticals in US health care: Determinants of quantity and price. Journal of Economic Perspectives, 16, 45–66. Berndt, E., Bhattacharya, A., Mishol, D. N., Arcelus, A., & Lasky, T. (2002). An analysis of the diffusion of new antidepressants: Variety, quality, and marketing efforts. The Journal of Mental Health Policy and Economics, 5, 3–19. Berndt, E. R., Linda, B., David, R. R., & Glen, L. (1995). Information, marketing, and pricing in the US antiulcer market. American Economic Review, 85(2), 100–105.

142

JOHN CAWLEY AND JOHN A. RIZZO

Bosch, J. C., Eckard, E. W., & Singal, V. (1998). The competitive impact of air crashes: Stock market evidence. Journal of Law and Economics, 41, 503–519. Calfee, J. E., Winston, C., & Stempski, R. (2002). Direct-to-consumer advertising and the demand for cholesterol-reducing drugs. Journal of Law and Economics, 45, 673–690. Cawley, J. (2004). The impact of obesity on wages. Journal of Human Resources, 39, 451–474. Congressional Budget Office. (1998). How increased competition from generic drugs has affected prices and returns in the pharmaceutical industry. Washington: Congressional Budget Office. Connolly, H. M., Crary, J. L., McGoon, M., et al. (1997). Valvular heart disease associated with fenfluramine-phentermine. New England Journal of Medicine, 337(9), 581–588. Dowdell, T. D., Govindaraj, S., & Jain, P. C. (1992). The Tylenol incident, ensuing regulation and stock prices. Journal of Financial and Quantitative Analysis, 27(2), 283–301. Dranove, D., & Olsen, C. (1994). The economic side effect of dangerous drug announcements. Journal of Law and Economics, 37, 323–348. Harris, G. (2004). FDA failing in drug safety, official asserts 11/19/2004. New York Times. Hedley, A. A., Ogden, C. L., Johnson, C. L., Carroll, M. D., Curtin, L. R., & Flegal, K. M. (2004). Prevalence of overweight and obesity among US children, adolescents, and adults, 1999–2002. Journal of the American Medical Association, 291(23), 2847–2850. Hill, J., & Schneeweis, T. (1983). The effect of Three Mile Island on electricity utility stock prices: A note. Journal of Finance, 38, 1285–1292. Hurwitz, M., & Caves, R. (1988). Persuasion or information? Promotion and the shares of brand name and generic pharmaceuticals. Journal of Law and Economics, 31, 299–320. Iizuka, T., & Jin, G. (2003). The effects of direct-to-consumer advertising in the prescription drug markets. Unpublished manuscript, University of Maryland. Jarrell, G., & Peltzman, S. (1985). The impact of product recalls on the wealth of sellers. Journal of Political Economy, 93, 512–536. Leffler, K. (1981). Persuasion or information? The economics of prescription drug advertising. Journal of Law and Economics, 24, 45–74. Lichtenberg, F., & Philipson, T. (2002). The dual effects of intellectual property regulations: Within- and between-patent competition in the US pharmaceuticals industry. Journal of Law and Economics, 45, 643–672. Ling, D. C., Ernst, R. B., & Margaret, K. K. (2002). Deregulating direct-to-consumer marketing of prescription drugs: Effects on prescription and over-the-counter product sales. Journal of Law and Economics, 45, 691–723. Lu, J., & Comanor, W. (1998). Strategic pricing of new pharmaceuticals. Review of Economics and Statistics, 80, 108–118. Meadows, M. (2002, Jan./Feb.). Why drugs get pulled off the market. Food and Drug Administration Consumer, http://www.fda.gov/FDAC/features/2002/102_drug.html Rizzo, J. (1999). Advertising and competition in the ethical pharmaceutical industry: The case of antihypertensive drugs. Journal of Law and Economics, 42, 89–116. Rosenthal, M., et al. (2003). Demand effects of recent changes in prescription drug promotion. Henry, J. Kaiser Family Foundation Report, http://www.kff.org/rxdrugs/upload/ Demand-Effects-of-Recent-Changes-in-Prescription-Drug-Promotion-Report.pdf

Spillover Effects of Prescription Drug Withdrawals

143

Stern, S., & Trajtenberg, M. (1998). Empirical implications of physician authority in pharmaceutical decisionmaking. Working Paper no. 6851. National Bureau of Economic Research. Wosinska, M. (2003). Advertising and optimal consumption path: The case of prescription drugs. Unpublished manuscript, Harvard University Business School. Wosinska, M. (2004). Direct-to-consumer advertising and patient therapy compliance. Unpublished manuscript, Harvard University Business School. Wysowski, D., & Swartz, L. (2005). Adverse drug event surveillance and drug withdrawals in the United States 1969–2002. Archives of Internal Medicine, 165, 1363–1369.

THE PSYCHOLOGY OF NUTRITION MESSAGES Heather Schofield and Sendhil Mullainathan ABSTRACT The purpose of this paper is to explore consumer thinking about nutrition decisions and how firms can use consumers’ awareness of the links between nutrients and health generated by public health messages to market products, including ones, which have little nutritional value. We approach this issue by tracking the development of public health messages based on scientific research, dissemination of those messages in the popular press, and use of nutrition claims in food advertisements to assess whether firms are timing the use of nutrition claims to take advantage of heuristic-based decision-making. Our findings suggest that the timing of the development of nutrition information, its dissemination in the press, and use in advertising accords well with a heuristic processing model in which firms take advantage of associations between nutrient information and health in their advertisements. However, the demonstrated relationships may not be causal. Further research will be needed to provide stronger and more comprehensive evidence regarding the proposed message hijacking process. If the message hijacking framework is borne out: (1) simple overall health rating scales could significantly improve consumer decision-making, (2) the impact of misleading advertisements could be mitigated by encouraging a multidimensional view of nutrition,

Beyond Health Insurance: Public Policy to Improve Health Advances in Health Economics and Health Services Research, Volume 19, 145–172 Copyright r 2008 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0731-2199/doi:10.1016/S0731-2199(08)19007-0

145

146

HEATHER SCHOFIELD AND SENDHIL MULLAINATHAN

and (3) more intensive regulation of product labeling could limit the impact of hijacked messages. Overall, this paper considers a novel hypothesis about the impact of public health messages on nutrition and health.

INTRODUCTION The increasing prevalence and high cost of treatment of chronic disease have motivated governments and advocacy groups to produce public health messages intended to encourage consumers to adopt healthy lifestyles that can prevent or delay the onset of chronic disease. One area in which public health messages have increased dramatically in the past 30 years is nutrition (Weimer, 1999). Nutrition and diet have received significant attention because of the dramatic rise in obesity rates, which is a risk factor for many chronic diseases, and in diseases such as diabetes, which are heavily influenced by dietary factors. A variety of theories have been put forth to explain the rise in obesity and rapid change in food consumption patterns that began in the late 1970s (Cutler, Glaeser, & Shapiro, 2003; Nestle, 2002; Olshansky et al., 2005). Ironically, public health messages about nutrition are largely overlooked, but potentially important factor that may contribute to the observed changes in nutrition. It is commonly assumed that public health messages will improve health. These messages contain information that should allow consumers to make more informed decisions. Public health campaigns often promote simple catchphrases and cultivate broad associations to ensure that messages are remembered and readily applied. For example, the ‘‘5 a day for better health’’ campaign broadly links high fruit and vegetable consumption with improved health in an easy to remember phrase (Produce for Better Health Foundation, 2008). The media often further publicizes these health messages with headlines ranging from the prosaic, ‘‘New study demonstrates value of low-fat diet,’’ to the colorful, ‘‘Trans fat: Taste buds cry ‘yes!’ but arteries demur’’ (Waldholz, 1990; Balu, 1998). The prediction that public health messages will have a positive effect on behavior and health relies on strong assumptions about consumer behavior. Specifically, this prediction rests on the assumption that consumers are rational actors with detailed knowledge about nutrition and the motivation to apply their knowledge. If these assumptions are not correct and consumers instead make decisions using heuristic-based thinking, promoted and reinforced by public health messages broadly linking specific nutrients and health, it becomes

Psychology of Nutrition Messages

147

possible for firms to ‘‘hijack’’ public health messages. When consumers rely on heuristic-based thinking, firms may be able to appropriate nutritional messages for profit in situations where the net informational value of the message is scientifically questionable. For example, Twizzlers licorice, which is labeled as ‘‘low-fat,’’ can hardly be considered healthy. This message by itself would not have unintended consequences in a world where consumers had full information, complete rationality, and sufficient time to consider decisions carefully. Individuals would simply base their purchase decision on factors such as the cost of the good and utility they derive from its consumption while ignoring the ‘‘low-fat’’ message as relatively uninformative with regard to the overall health benefits of the product. However, broad and simple public health messages encourage consumers to engage in heuristic processing rather than considered rational analysis. For example, public health messages about the health benefits of a low-fat diet build strong associations in consumers’ minds between the message ‘‘low-fat’’ and idea of health. Hence, a heuristic-based thinking consumer viewing the Twizzlers may associate ‘‘low-fat’’ with ‘‘healthy’’ without seriously evaluating the food’s overall contribution to one’s health, thereby increasing the likelihood of purchase.1 In other words, when using heuristic processing, consumers may easily be persuaded by ingrained associations about specific aspects of the product that were generated by public health activities and media attention, rather than the overall objective value of the product. Hence, if firms do indeed ‘‘hijack’’ these public health messages in inappropriate situations and consumers fall into the trap that these messages set through heuristic processing, public health messages regarding healthful nutrition could potentially have negative impacts on population health through diets with lower nutritional quality and an increase in obesity. This paper begins a preliminary examination of this possibility.

CONSUMER COGNITION Broadly speaking, psychologists have broken human thought and decisionmaking into two categories or systems. The first type of thought in these dual-process models is referred to as ‘‘System 1’’ and it is quick, associative, and intuitive. This type of thought is automatic and effortless and a decision is made without one being aware of the process. Heuristics, which can be thought of as rules of thumb, are one such type of thought. They are used to make judgments or decisions and are simple, practical, and easily applied. For example, when individuals are faced with a

148

HEATHER SCHOFIELD AND SENDHIL MULLAINATHAN

complex problem or difficult question such as ‘‘is this product good for my health?’’ they may substitute an easier question such as ‘‘is this product low in fat?’’ in place of the original question without consciously recognizing the substitution (Kahneman & Frederick, 2002). In this example the individual tries to assess the ‘‘target attribute,’’ health, but instead assesses a quality that is more readily evaluated, the ‘‘heuristic attribute,’’ fat content. While the use of heuristics has the advantage of saving time and cognitive effort, it can also lead to a number of systematic biases. The most widely known and documented biases include the availability heuristic (Tversky & Kahneman, 1973, 1974), representativeness heuristic (Kahneman & Tversky, 1973; Tversky & Kahneman, 1974), and anchoring heuristic (Tversky & Kahneman, 1974). The second type of thought, ‘‘System 2,’’ is slow, controlled, and deliberate. With this type of thought the individual rationally weighs the options and applies logic and reasoning to decisions (Kahneman, 2003; Kahneman & Frederick, 2002). In the previous example, an individual assessing the health qualities of a product would consider not only the fat content to be one factor contributing to the overall health consequences of the product, but would also weigh other aspects such as sodium levels, caloric content, and types of fat in the product. It is generally thought that the quick associative thinking of System 1 is the default decision-making process, but the decisions made in this way are monitored by and can be modified or overridden by System 2 (Kahneman & Frederick, 2002). The relative contribution of the two systems is determined by a number of factors. Factors that encourage automatic heuristic-based System 1 processing with little interference from System 2 include time pressure (Finucane, Alhakami, Slovic, & Johnson, 2000), other cognitive burdens (Gilbert, 1989), strong visceral emotion (Loewenstein, 2000), and positive affect or happy mood (Bless et al., 1996; Isen, 1987; Clore, Schwarz, & Conway, 1994). On the other hand, logical thought is encouraged by general statistical training as well as task-specific training among those with little general mathematical or statistical background (Nisbett, Krantz, Jepson, & Kunda, 1983; Agnoli & Krantz, 1989; Agnoli, 1991). In general, however, the oversight by System 2 is quite often relatively weak (Kahneman, 2003). In addition, these two types of thought are not isolated. As an individual becomes more skilled at a given cognitive operation, either through innate ability or training, the thought process shifts from laborious reasoned thought (System 2) to automatic and effortless thought (System 1) (Kahneman & Frederick, 2002). In fact, as Kahneman (2003, p. 1451) notes ‘‘the difference in effort provides the most useful indication of whether a given mental process should be assigned to System 1 or System 2.’’

Psychology of Nutrition Messages

149

Because System 1 is the default system and oversight by System 2 is typically weak, individuals are likely to make many decisions that rely more on ‘‘instinct’’ than logical reasoning. If an individual has invested significant time and effort to train themselves to make complex decisions in a given domain, such as nutrition, these System 1 choices may be as good as System 2 choices and they will save both time and cognitive effort. On the other hand, for complex choices in which an individual is not especially experienced, the System 1 choices will still save on time and effort, but may not reflect optimal decision-making. When this is the case, if an individual has sufficient monitoring to realize that their decision-making produces sub-optimal results they can evaluate whether it is worth investing additional time to improve their decision-making either through practice to improve their System 1 decision-making, or by allocating time and cognitive effort to System 2 thought. However, without adequate monitoring by System 2, the individual may not realize the negative impact of those decisions. Although it is possible for individuals to thoroughly and rationally evaluate claims made on food packaging before making a purchase decision, a number of factors suggest that more associative or heuristic-based decision-making is likely to occur. First, these decisions are typically considered mundane, and hence are unlikely to spur additional monitoring by System 2. In addition, consumers are often in a hurry or cognitively overloaded with many choices and competing demands for their attention in a grocery store (Underhill, 1999). Further, other studies have demonstrated that seemingly inconsequential manipulations to the shopping environment, including location on the shelf and signs regarding purchase limitations have meaningful impacts on consumer behavior in grocery stores (Underhill, 1999; Wansink, Kent, & Hoch, 1998). These shifts in behavior are consistent with automatic heuristic processing rather than careful planned and rational thought. Although providing time savings, this heuristic processing in turn opens the door to the possibility that firms can exploit one aspect of their product to encourage consumers to purchase a product that is, on the whole, unhealthy. In addition, there is evidence that consumers do change their likelihood of purchase in response to health claims on food products. The first piece of evidence is simply that many firms use such claims. Assuming that firms are profit maximizing and recognizing that there is some cost to the development and printing of a health message, the firms must believe that the use of health messages increase revenues. Further, Ippolito and Mathios (1995) found that the consumers shifted their consumption patterns to

150

HEATHER SCHOFIELD AND SENDHIL MULLAINATHAN

decrease overall fat consumption following awareness of public health messages linking lowered fat consumption and health and an increase in advertising using these claims. Researchers have also found that Kelloggs was able to gain significant market share in the cereal market by promoting the health benefits of high-fiber cereals (Nestle, 2002; Ippolito & Mathios, 1990). These shifts in consumption habits can potentially improve nutrition. However, if heuristic processing is used, as it is likely to be given the factors described previously, consumers may also shift their consumption in reaction to empty or even misleading messages.

THE HISTORY OF NUTRITION MESSAGES IN THE UNITED STATES Knowledge regarding links between nutrition and health are an essential component of this application of heuristic processing. It is only recently that many of these links have been discovered and widely publicized. While the first dietary guidelines generated by the United States Government were published in 1894, these guidelines were quite general and recommended a balanced and moderate diet (Davis & Saltos, 1999). At this time, vitamins and minerals had not yet been discovered. Fifty years later, much progress had been made in terms of knowledge and Recommended Dietary Allowances (RDAs) had been developed for nine essential nutrients. However, nutritional emphasis was still on ensuring that individuals consumed a balanced diet that was centered on seven food groups (Davis & Saltos, 1999). Importantly, at this time, guidance about specific nutrients such as sugars and fats was very limited. Although the exact groupings changed, emphasis on sufficient2 and varied nutrition from all of the major food groups continued until the late 1960s and early 1970s. In the 1970s, health messages began to shift toward specific nutrient recommendations and explicitly linked aspects of one’s diet and health outcomes3 (Davis & Saltos, 1999). While categorical recommendations are still used (e.g., the food guide pyramid), many of the recommendations in recent government guidelines center on specific nutrients rather than food categories (Nestle, 2002). Consumer knowledge linking nutrients and health outcomes has grown correspondingly. For example, in 1978 approximately 12% of the population was aware of the relationship between sodium and hypertension. However, following a public awareness campaign, knowledge about the relationship increased threefold4 (Guthrie, Derby, & Levy, 1999). Similar increases in awareness have also occurred for other nutrients such as

Psychology of Nutrition Messages

151

fat. Indeed, ‘‘concern about the fat content’’ of one’s diet skyrocketed from less than 10% in the early 1980s to nearly two-thirds of the population by 1995 (Guthrie et al., 1999).

FIRM INCENTIVES TO HIJACK Firms have used health claims to promote products for years. However, the shift from general nutritional information centering on a ‘‘balanced diet’’ including multiple food groups to a focus on consuming or avoiding specific nutrients is important to a firm’s ability to appropriate messages and was followed quite closely by the beginning of the shift in nutritional habits and rise in obesity rates. Guidelines suggesting a ‘‘balanced diet’’ allow firms to advertise their foods as a part of a balanced diet, but only a part. By themselves, these products could not be touted as leading to health. Thus, no food in and of itself could be advertised as a key to health. However, links between specific nutrients and either risk reduction for a specific disease or general health allow firms to associate their products with health directly. For example, many unhealthy products such as sugary candies were not included in food categories or considered to be a part of a balanced diet. However, nearly any product can cite some nutrient it contains (or does not contain) as a source of health, allowing firms to much more selectively advertise healthy aspects of products which are, on the whole, unhealthy. For examples, products such as ‘‘low-fat’’ licorice, brownies with ‘‘real milk chocolate,’’ and artificial juice (i.e., flavored sugar water) that is a ‘‘good source of vitamins A and C,’’ are quite common in grocery stores today.5 In the presence of heuristic processing, these selective advertisements can potentially lead to poor food purchase decisions, especially because profit maximizing firms are likely to select advertisements that are effective, regardless of whether or not the claim is actually reflective of the underlying health qualities of a product. Further, if a product does not contain (or contains too much of) a nutrient, the food can be modified in order to be able to make a claim. In fact, approximately one-fourth of all products coming to market in recent years are ‘‘functional foods’’ that have been nutritionally modified in such a way that they are able to make a nutritional claim even though many of these foods would often be classified as ‘‘junk foods’’ (Nestle, 2002). Further, many of the newly introduced foods are branded and packaged to facilitate nutritional marketing through the name of the product (e.g., Lean Cuisine) or labeling on the package (Cutler et al., 2003; Nestle, 2002).

152

HEATHER SCHOFIELD AND SENDHIL MULLAINATHAN

A firm’s incentive to shoulder the cost of producing a message (say through advertising or claims on the packaging) is influenced by consumers’ beliefs about that message. If consumers find the message to be highly informative or persuasive about the advantages of the product, then the benefit of using that message increases because it is likely to stimulate purchases (Ackerberg, 2001; Deighton, Henderson, & Neslin, 1994; Chang & Kinnucan, 1991). Under a full information model with System 2 processing, an advertiser’s presentation of true information about the health content of a product should be sufficient to increase the appeal of the product.6 Hence, in this model, firms would be able to utilize scientific information regarding nutrition as soon as it becomes available and one would expect to see concurrent rises in scientific information about links between nutrients and health and advertisements relying on those links. However, under a heuristic-based processing (System 1) model, the information will only increase the likelihood of a sale if the information has a positive association with health. In other words, associative thinking relies on ingrained associations and their related fixed-action patterns rather than the objective level of information contained in the message. So, firms are most likely to profit from heuristic thinking when consumers have had significant exposure to a message that builds an ingrained association. Hence, under this model, one would expect to see a lag between dissemination of scientific information and the use of that information in advertisements while popular press and public health institutions build an association between that nutrient and health. Of course, it is possible for firms to undertake an extended advertising campaign to make consumers aware of public health messages that can then be used in advertising campaigns.7 However, this approach is costly and unless the expected increase in revenue due to the campaign is quite large, this strategy is unlikely to be profit maximizing. A less costly approach that shifts the costs of forming the association between a nutrient and health to other parts of society still reaps the benefits of health associations that highlight product attributes that have had an association with health formed through public health messages and popular media. Hence, one would expect that firms that wish to use heuristic processing to their advantage would undertake advertising campaigns following closely on the heals of broad dissemination of a given health message. For example, the use of ‘‘low-fat’’ claims followed the rise in public health messages and media coverage relating to dietary fat intake in the late 1970s and early 1980s with a slight lag. In short, under the rational model (System 2 thinking), advertising campaigns based on the health appeal of a product should be related to the dissemination of new scientific information about the health qualities of

Psychology of Nutrition Messages

153

nutrients. However, under the heuristic thinking model one would expect to see a flow of information from scientific publications, to the popular press and finally, with a slight lag, to advertising campaigns. Predictions for the extended use of these health messages in advertising would also vary under the two models. In the rational model, the information should remain consistently informative and one would not predict a decline in advertisements using a given health message unless the evidence for the link is reversed.8 Predictions under the heuristic processing model would depend upon one’s assumptions about the strength and duration of the associations formed. If one believes that the associations remain strong even after discussion of the health message is no longer at the forefront of popular media, then one would expect advertising based on established messages to remain strong, as under the rational model. However, if one assumes that associations diminish over time, then one would expect to see declines in the number of advertisements using messages that are no longer ‘‘in the public eye.’’ In summary, we make two predictions about the differences in patterns of information generation and dissemination under System 1 and System 2 processing models. First, under System 1 processing models, advertising will follow the generation of information with a lag during which the activities of the popular press and public health institutions form associations between that nutrient and health outcomes. Under System 2 models, on the other hand, advertisements will instead follow closely on the heals of the generation of new scientific knowledge. Second, with System 2 processing, one should not observe declines in advertising about a nutrient unless new scientific information reverses previous knowledge about the health impacts of that nutrient because the message remains consistently informative. However, under a System 1 model, if associations between a nutrient and health decline because of declines in public health messages or popular press coverage, one would expect similar declines in advertisements using claims about that nutrient.

RESEARCH DESIGN AND METHODS In order to assess whether advertising patterns are consistent with a model of heuristic thinking and message hijacking, we tracked the development of public health messages resulting from scientific research, dissemination of those messages and findings into the popular press, and food advertisements. This analysis consists of gathering information about the use of 10 health messages from 1975 to 2004 from three sources: a large academic

154

HEATHER SCHOFIELD AND SENDHIL MULLAINATHAN

public health and medical literature database, PubMed; a large popular press database, LexisNexis; and food advertisements in magazines.

Pubmed The first part of this analysis was conducted by searching PubMed for citations including specific nutrition-related phrases. PubMed was used for this analysis because its database of 16 million citations from a variety of health and medical journals dating from the 1950s to the present is one of the most comprehensive sources of health information available (National Library of Medicine and National Institutes of Health, 2007). Searches were conducted for specific nutrition message phrases in 5-year periods between 1975 and 2004. Table 1 provides a list of phrases that were tracked. For each message and each 5-year period the total number of articles in the database containing the phrase in the full text was recorded.9

Phrases Used in Literature Searches and Advertisement Coding. Nutrient

PubMed and LexisNexis

Calcium

Calcium and osteoporosis

Cholesterol

‘‘low cholesterol diet’’ or ‘‘low-cholesterol diet’’ ‘‘low fat diet’’ or ‘‘low-fat diet’’ ‘‘high fiber diet’’ or ‘‘high-fiber diet’’ ‘‘low carb diet’’ or ‘‘low-carb diet’’, ‘‘low carbohydrate diet’’ or ‘‘low-carbohydrate diet’’ ‘‘Saturated fat’’ and ‘‘heart disease’’ ‘‘low sodium diet’’ or ‘‘low-sodium diet’’ Sugar and (‘‘diabetes prevention’’ or ‘‘diabetes control’’) ‘‘Trans fats’’ and ‘‘heart disease’’ (‘‘Whole grain’’ or ‘‘whole-grain’’) and ‘‘diet’’

Fat Fiber Low carb

Saturated fat Sodium Sugar Trans Fat Whole grain

Advertisements Calcium and osteoporosis or bone-health claim Cholesterol claim Fat claim Fiber claim Carbohydrate claim

Saturated fat claim Sodium claim Sugar claim Trans fat claim Whole grains claim

Note: Similar wording combinations (e.g., ‘‘saturated fat’’ and ‘‘cardiovascular disease’’) were also examined to ensure that the findings were not due to unusual trends related to a specific choice of words. Patterns were relatively robust regardless of the exact phrasing.

155

Psychology of Nutrition Messages

LexisNexis Although academic studies may suggest relationships between certain nutrients, or types of diets, and health outcomes, the readership of these journals is limited. Hence, in order for the messages to reach average consumers, they must be taken up and disseminated by the popular press and other media. These other media would include other forms of news such as television and radio, and explicit public health education campaigns by governments and other public service organizations. Studying the impact of public health education efforts by themselves would provide valuable information about how to influence consumption because these messages are more easily influenced than media coverage in its many forms and often influence the media coverage received. However, in practice, it would very difficult to parse the effects of public health messages by governments and other authoritative bodies because distinguishing the impact of the ‘‘official’’ public health messages from the impact of other news media would present a significant challenge. Further, many public health campaigns tend to be local and difficult to track, suggesting that exposure to broad national media coverage about health messages may have a much more significant impact on the messages to which large numbers of individuals are exposed. Hence, exposure to information about public health messages through popular media was used as an indication of exposure to public health messages more broadly. Specifically, popular press in the form of newspapers is used as a proxy. Using the volume of printed media as a proxy for exposure to a public health message may not accurately represent the exposure of all demographic groups. However, all forms of media that could potentially be studied are subject to some similar form of bias. This medium was used because the historical records of printed media are much more complete and readily assessed than records of other forms of media. In this analysis, the LexisNexis database was used to track publications (LexisNexis, 2007). LexisNexis was chosen due to its broad coverage of popular press. With print news from approximately 6,000 sources including leading newspapers such as the Washington Post, New York Times, and other high circulation publications. LexisNexis is a good source of information regarding the news to which the majority of the population is exposed (LexisNexis, 2007). A method similar to the method used to track the dissemination of information through PubMed was used to track the dissemination of information through the popular press. A guided news search restricting the

156

HEATHER SCHOFIELD AND SENDHIL MULLAINATHAN

search to the ‘‘headline, lead paragraph, and terms’’ of newspaper articles in the general news category was used.10 The articles were grouped into the same 5-year categorizations as were used to examine PubMed citations. Food Advertisements To track the use of health messages in advertisements, an analysis of the content of food advertisements in leading women’s magazines was undertaken. The magazines included are Good Housekeeping, Woman’s Day, and McCall’s. These magazines were chosen because all three are popular magazines in publication during the period of interest and the density of food advertisements allowed for sufficient data collection to track trends over time.11 A total of 732 advertisements were coded.12 Although these magazines may be more likely to cite health benefits of the products that are advertised, or may market certain characteristics of the products more frequently (e.g., calcium content or diet products) because they market to women, sampling from this group of magazines may actually reflect the messages reaching the individuals purchasing food in grocery stores more precisely than magazines targeted at a broader audience. In fact, women were and are significantly more likely go and spend more time grocery shopping than men (Blaylock & Smallwood, 1987; Hamrick & Shelley, 2005). Hence, studying advertisements that target women captures information that is likely to be used in many household food purchase decisions. Coding System The magazine coding system was designed to capture a number of aspects of the advertisements and their claims, with a focus on health and nutrient claims (a complete codebook is available from the authors upon request). The first information collected on each advertisement was simple tracking information such as the year of publication, issue number, product brand, and product category (such as meat, vegetable, starch, dessert, etc.). These data were collected for reliability checks (see below) and analysis purposes. The content of the health message was also captured in the coding system. The system captured six main categories of claims: general nutrition, sugars and fats, vitamins and minerals, endorsements, diet claims, and miscellaneous claims. Each of these categories of claims contained a number of variables. For example, the fats category included ‘‘saturated,’’ ‘‘unsaturated’’ or ‘‘omega 3,’’ ‘‘trans,’’ ‘‘olestra,’’ ‘‘cholesterol,’’ ‘‘general’’ (just ‘‘fat’’ used), and ‘‘other.’’ In

Psychology of Nutrition Messages

157

addition, a binary variable indicating whether any type of fat was mentioned was included. The coding system also allowed for the use of multiple claims of each type. For example, a claim such as ‘‘low in saturated fats and high in unsaturated fats’’ would be coded as containing two claims about fats. Both the type of fat mentioned and claim (i.e., reduced, zero, low in, etc.) were recorded. An additional notes section was also available to allow coders to note features of the advertisements that they considered relevant, but that were not captured in the coding system. For example, common notes included information about implied health claims, such as ‘‘golden goodness’’ that were not to be explicitly coded because they do not make any formal health claim.13 Each magazine was sampled once per year. Random draws were used to determine which issue of the magazine was used. Sampling of specific advertisements in each magazine was also done in a random fashion. Specifically, 8 random numbers between the number 1 and 200 (the average length of each of the magazines) were drawn. The coder would then examine each page that was selected. If the page did not contain an advertisement, the coder moved forward in the magazine until the next food advertisement was located and recorded the appropriate information for that advertisement. If the end of the magazine was reached before an advertisement was found, the coder cycled back to the beginning of the magazine.14 In cases where two randomized pages were in close succession and an advertisement was not located between the two randomized numbers, the next two advertisements following the second randomized page number were sampled. When more than one advertisement was located on a page, both advertisements were coded. Similarly, when two products were advertised together, the products would be coded as two advertisements on the same page. This process resulted in a total number of 732 advertisements. Because the advertisements were coded by more than one individual, steps were taken to ensure sufficient inter-coder reliability. Specifically, both individuals coding advertisements began by jointly coding a series of advertisements. After this was done, each individual coded a new set of 20 advertisements separately and the results were compared. In 97% of the cases, the information regarding the content of the messages was consistent. Further, reliability was insured during the coding process by having one coder independently recode work done by the other coder. During this second check of inter-coder reliability, results were similar with approximately 2% of the data discrepant between coders. In cases when the coding was discrepant, the original advertisement was examined again by both coders and agreement was reached about the appropriate coding.

158

HEATHER SCHOFIELD AND SENDHIL MULLAINATHAN

RESULTS In order to analyze the data, time-series graphs of the utilization of the messages and advertising claims were constructed. Because the scales for the three types of media varied significantly, plots were generated using the natural logarithm of the number of citations or uses in advertisements during the period. The absolute magnitudes of the counts are of less relevance than the relative differences in magnitude and slopes of each of the plots given the difference in the scales between the three data sources. As discussed previously, a distinguishing feature for determining whether these data support heuristic thinking that is conducive to message hijacking is the point at which advertisers begin to use a health message relative to its dissemination in the medical literature and popular press. Specifically, under a full information, rational model one would expect advertisers to use a health message at a relatively constant-level once knowledge about the relationship between a nutrient and disease risk has been established in the medical literature. However, under the heuristic-based decision-making model, one would expect that advertisers would not use a message until after the message has been established in the popular press as well as medical literature.15 Hence, under this model, one would expect the advertising curves to remain fairly flat (or be non-existent) until after the PubMed curve has increased (or remained high and flat or increasing) for some time and/or the LexisNexis curve has increased (or again remained high for some time). This rough pattern is observed for a majority of 10 plots (Figs. 1–6, 10). Consider the graphs of the use of messages regarding low-fat diets (Fig. 1: Fat) and calcium in one’s diet (Fig. 2: Calcium). In these graphs, the number of scientific publications regarding these messages is fairly high and upwardsloping in the late 1970s. This trend is closely followed by a distinct rise in the number of popular press publications about low-fat diets and calcium in one’s diet during roughly the same time frame, but perhaps with a slight lag. However, the use of advertising messages about fat and calcium remains fairly flat (fat) or non-existent (calcium) for the first 15 years of observation and then increases dramatically.16 Although these patterns support the proposed heuristic-based decisionmaking model, some of the 10 graphs do not show a similar pattern. The most striking example that goes against the predicted pattern is found in the plot regarding sugar messages, which is the only plot that appears to follow the expected pattern under a full information rational actor (System 2 processing) model. In this plot (see Fig. 7: Sugar), although the medical

159

Psychology of Nutrition Messages

Natural Log of Count

7 6 5 4 3 2 1 0 75-79

80-84

85-89 90-94 Year

PubMed

95-99

00-04

LexisNexis Advertisements

Fig. 1. Fat.

Natural Log of Count

10 8 6 4 2 0 75-79

80-84

85-89

90-94

95-99

00-04

Year PubMed

LexisNexis Advertisements

Fig. 2.

Calcium.

curve begins to trend upward first, the popular press curve lags significantly behind the advertising curve.17 Further, in the graph related to fiber, the rise in messages is concurrent across all three-publication sources (see Fig. 8: Fiber). The pattern observed for fiber is likely due to Kellogg’s large advertising campaign regarding a new line of high-fiber cereals in the 1980s.

160

HEATHER SCHOFIELD AND SENDHIL MULLAINATHAN

Natural Log of Count

8 6 4 2 0 75-79

80-84

85-89

90-94 Year

PubMed

95-99

00-04

LexisNexis Advertisements

Low Carb.

Fig. 3.

Natural Log of Count

6 5 4 3 2 1 0 75-79

80-84

85-89 90-94 Year

PubMed

95-99

00-04

LexisNexis Advertisements

Fig. 4.

Sodium.

This campaign is notable in that it preceded general awareness of the health benefits of fiber, and in conjunction with public institutions such as the National Cancer Institute, generated an awareness of that relationship (Nestle, 2002; Ippolito & Pappalardo, 2002; Guthrie et al., 1999).

161

Psychology of Nutrition Messages

Natural Log of Count

5 4 3 2 1 0 75-79

80-84

85-89 90-94 Year

PubMed

95-99

00-04

LexisNexis Advertisements

Fig. 5.

Trans Fat.

Natural Log of Count

6 5 4 3 2 1 0 75-79

80-84

85-89 90-94 Year

PubMed

95-99

00-04

LexisNexis Advertisements

Fig. 6.

Whole Grain.

The saturated fat plot is also of interest due to its unusual pattern in which PubMed and LexisNexis citations increased, but were not followed by an increase in claims in advertisements (see Fig. 9: Saturated Fat). This lack of reaction by firms may have been caused by limited change in public

162

HEATHER SCHOFIELD AND SENDHIL MULLAINATHAN

Natural Log of Count

6 5 4 3 2 1 0 75-79

80-84

85-89 90-94 Year

95-99

00-04

LexisNexis

PubMed Advertisements

Fig. 7.

Sugar.

Natural Log of Count

5 4 3 2 1 0 75-79

80-84

85-89

90-94 Year

PubMed

95-99

00-04

LexisNexis Advertisements

Fig. 8.

Fiber.

awareness, despite the increase in publication about the relationship. Guthrie et al. (1999) suggested that the lack of awareness about this relationship may have been driven by the perceived complexity of the message. Consumers may see little benefit on differentiating between types of fat if they simply aim for a diet that is low in all types of fat.

163

Psychology of Nutrition Messages

Relevant to our second hypothesis, a number of the messages show declines in advertising use after a period of 10–15 years during which they are increasing. The most prominent example of this shift is found for cholesterol (see Fig. 10: Cholesterol). Here, as awareness of the message

Natural Log of Count

6 5 4 3 2 1 0 75-79

80-84

85-89

90-94 Year

PubMed

95-99

00-04

LexisNexis Advertisements

Saturated Fat.

Fig. 9.

Natural Log of Count

5 4 3 2 1 0 75-79

80-84

85-89 90-94 Year

00-04

LexisNexis

PubMed Advertisements

Fig. 10.

95-99

Cholesterol.

164

HEATHER SCHOFIELD AND SENDHIL MULLAINATHAN

increases, advertisements follow closely. However, after publication regarding the relationship between cholesterol and cardiovascular health declines, the use of health claims regarding cholesterol declines as well. Similarly, in the ‘‘fat’’ and ‘‘fiber’’ plots previously presented, advertisements featuring these claims appear to decline as publication regarding the claims levels off or declines. These declines would suggest that firms are most likely to use health messages that are ‘‘fresh’’ in consumer’s minds or the latest ‘‘fad’’ in health to sell their product. This observation corresponds with the predictions of the heuristic-based decision-making model because rational consumers would find the information equally informative across time, while consumers using heuristic thinking would be most swayed by the message when the associations between the message and health is strongest (i.e., when it is in being publicized heavily in the media).18

FURTHER RESEARCH The evidence described previously demonstrates that food advertisements often attempt to leverage associations built by public health knowledge and media about the health impact of specific nutrients, opening the door for message hijacking. However, this evidence is not causal and does not prove that these messages are used inappropriately or that consumers react to inappropriately placed messages. Hence, further research will be needed to provide stronger and more comprehensive evidence regarding the proposed message hijacking process. The first aspect of this research will need to examine the relationship between the use of health and nutrition messages and claims and actual nutrient content of the food products. If foods with limited nutritional value use claims more frequently than healthier products or products make very selective claims that highlight the positive aspects of products that also have many negative nutritional qualities, it will provide additional evidence of the ‘‘hijacking’’ of claims. In addition, while the use of heuristic processing increases the likelihood that consumers react to hijacked messages, more concrete evidence is needed to determine whether consumers react to these misleading claims. Specifically, experimental evidence examining the use of dubious health claims on grocery store products will provide concrete evidence regarding consumer reactions to health claims.

Psychology of Nutrition Messages

165

POLICY IMPLICATIONS Although further research is warranted before making policy decisions, there are a number of important policy implications if further evidence supports the message hijacking hypothesis. First, public health institutions should consider the development of simple overall health rating scales because an overall health-scale that simplified the decision-making process and aggregated health information would significantly limit firms’ ability to hijack messages in a way that confers ‘‘overall health’’ benefits to products.19 It would undoubtedly be challenging to create agreement about the essential components of this scale. However, should the impact of message hijacking be significant, it is likely that individuals at public health institutions will realize that an imperfect scale may be better than no scale at all if it prevents message hijacking and improves consumers’ nutritional choices. Of note, there has already been some movement in this direction by private firms. For example, the super-market chain, Hannaford Brothers, has recently developed a health rating system for the products it stocks (Martin, 2006). However, these scales are likely to have increased credibility if they are developed by public health institutions. Second, public health institutions could also change their messaging to mitigate the impacts of heuristic thinking by creating a mental model of ‘‘health’’ that contains subcategories rather than simply associating a nutrient (or the lack of a nutrient) with health broadly. For example, health messages could foster a model of health with categories that focus on factors related to diseases with significant morbidity and mortality burdens such as ‘‘low calories,’’ ‘‘nutrient-rich,’’ and ‘‘heart healthy.’’ With this mental model of ‘‘health,’’ advertisements about the benefits of a product would be categorized first into one of these sub categories and then further aggregated into an overall concept of health, mitigating the impact of the claim on the perception of ‘‘overall health benefits.’’ For example, with this slightly more complex model, a claim about vitamin and mineral content of a food would influence an individual’s perception of the ‘‘nutrient-rich’’ category, and in turn slightly improve the ‘‘overall health’’ aggregate assessment, rather than directly feeding into an overall assessment that may lead to an ‘‘overreaction’’ of the claim.20 This strategy would allow firms to continue to advertise the benefits of the products that they offer, while mitigating the impact of claims that are misused. Finally, in order to mitigate the effects of heuristic thinking, closer regulation of food labeling may be appropriate. Currently, nutrient claims are much more lightly regulated than disease claims. Policies such as

166

HEATHER SCHOFIELD AND SENDHIL MULLAINATHAN

requiring functional foods that highlight a positive change to their product also highlight (at least at visibly) negative changes that have been made would encourage firms to only produce and label products that have net health benefits. Further, requiring firms to inform consumers about common misconceptions caused by labeling could limit the impact of heuristic thinking. For example, requiring products with reduced fat labels to also clarify whether they are reduced in calories would allow consumers to more accurately judge the net benefits of the product. Although this study suggests more powerful regulation of the use of health claims, it also acknowledges that there can be many benefits to health claims. While some functional foods are only marginally healthier than the original versions, certain functional foods are indeed an improvement over the original versions. For example, Ippolito and Mathios (1990) noted that significant improvements in a number of nutritional dimensions for cereals following the increase in nutrient advertising (especially fiber) in the late 1980s. Allowing firms to make these claims provides them with an incentive to improve their product and compete on nutritional factors. However, regulations will need to strike a balance between providing an incentive to improve a product and the provision of sufficient information to prevent the abuse of the marketing claims that create confusion.21

CONCLUSION Changing nutrition habits including increased caloric intake and increased consumption of nutrients that have significant potential to harm health, such as refined sugars, have the potential to substantially increase morbidity and mortality in the coming years. There are clearly a number of factors that influence these trends. One factor that may play an important role, yet has been largely overlooked, is health messages and the knowledge they generate. While knowledge about health usually improves decision-making, under the right circumstances it could potentially harm health. In this context, the broad associations formed by public health messages may enable and encourage consumers to think in heuristic-based terms, allowing firms to exploit the associations between specific nutrients and health by hijacking the messages for use in uninformative or inappropriate situations. This paper examines the timing of the use of a variety of nutrient-based health claims to assess whether firms are timing the use of health messages to take advantage of heuristic-based decision-making strategies. Specifically, the timing of the use of health messages and claims is suggestive of a possible role

167

Psychology of Nutrition Messages

in shifting nutritional habits; firms typically use health claims only after the public health community and popular press have publicized the link between a nutrient and health outcomes. After awareness of the health link begins to fade, as publications about the relationship drop off, advertising resting on the health message decreases as well. Under a rational model there would be little reason for a firm to behave in this manner. However, these observations accord well with a heuristic-based decision-making model where the salience of associations between a health claim and health is critical. Although the evidence we presented is preliminary and further research needs to be conducted before policy changes are considered, if the message hijacking framework is borne out a number of policy implications would follow. First, if message hijacking does occur there would be significant benefit in the development of one simple overall health rating scale that is sanctioned by reputable public health authorities. Second, public health institutions could mitigate the impact of misleading messages by providing health education that encourages a slightly more complex (multidimensional) view of nutrition, isolating the impact of hijacked messages within a subcategory of health/ nutrition and limiting the net effect of any one message on an individual’s overall perception of the health benefits of a product. Finally, more intensive regulation of the labeling of functional foods that requires disclosure of negative changes as well as positive changes would encourage the development of healthier products and limit the impact of hijacked messages.

NOTES 1. Other firms and consumers may react to these messages and adjust their advertising strategy regarding the ‘‘healthfulness’’ of foods labeled with a health claim; however the long-run equilibrium is beyond the scope of this paper. 2. It should be emphasized that individuals were generally provided with dietary minimums during this time, with the intention of ensuring adequate intake. 3. Because firms began to use fat and cholesterol claims as soon as research indicated the nutrients might impact health, a ban on the use of nutrient claims on food labeling was put into effect briefly from the late 1960s till 1973 when regulations were established (Ippolito & Mathios, 1995). 4. While there is some debate regarding the impact of sodium intake on health outcomes (Freedman & Petitti, 2001), public health institutions still advocate reductions in sodium in the average American’s diet (NHLBI, 2008). 5. The FDA does limit claims relating nutrient content and specific diseases on foods containing very high-levels of unhealthy nutrients. However, these foods are not banned from making nutrient content claims such as ‘‘reduced fat’’ or ‘‘a good source of vitamins and minerals.’’

168

HEATHER SCHOFIELD AND SENDHIL MULLAINATHAN

6. Of course, this information would need to be viewed as credible to be influential. However, because the government regulates the messages, and consumers would be aware of this regulation in a full information setting, the messages would be viewed as credible as long as the regulation is seen as sufficient. 7. In fact, Nestle (2002) documents a case in which Kelloggs did popularize the association between fiber and health and gained significant market share in the cereal market by doing so. 8. One could imagine that new evidence regarding the benefits of other nutrients that are more important to one’s health could arise and companies would shift to this new message. However, there would be nothing to prohibit the firm from using both messages instead of substituting between the messages because under the rational model more positive information would improve the likelihood of sales and the marginal cost of continuing to use the message is virtually zero once the message has been developed. 9. It would be ideal to format the data as percentage deviations from the general trend in the total number of articles in the database during that 5-year period to ensure that trends were not simply due to an expansion of the database. However, while it is possible to ascertain the number of articles available on PubMed during a 5-year period, this information was not available for LexisNexis. Hence, in order to be consistent, neither database was scaled by total publications. Despite this shortcoming, a number of the health messages tracked do experience declines in the number of articles using the relevant phrase. Further, the increases, and timing in the increases in the number of messages are not consistent across messages suggesting that the observed trends are not simply due to trends in the total number of publications. 10. This option was chosen because full text searches resulted in query responses that were larger than the 1,000 article limit, limiting the ability to distinguish trends in the data. Search queries of magazine headlines were also examined to test the robustness of the results. Trends between newspapers and magazines were quite similar. 11. Publication of McCall’s was halted in 2001. Hence, Good Housekeeping and Woman’s Day were over-sampled in the last 5-year period to obtain the necessary number of advertisements. Use of any type of health message was not significantly different across the three magazines ( p ¼ 0.89). However, two of the 10 specific messages examined (calcium claims and fiber claims) were significantly different between the magazines. 12. Similar studies of food advertising content have been conducted before, the largest being conducted by the Federal Trade Commission (Ippolito & Pappalardo, 2002). However, for all datasets located, the data was unavailable for public use and/ or did not cover the necessary time span. 13. Implied health claims were fairly common. However, the claims could frequently be interpreted in other ways. For example, ‘‘golden goodness’’ could also imply good flavor. Because of this ambiguity it was difficult to code these claims reliably and the value of the recording the claims was questionable. Hence, the implied claims were not included in the analysis, but they were noted in order to examine the data for other unanticipated trends. 14. In three cases, a sampled magazine did not contain eight health advertisements. When this occurred, all advertisements in the magazine were coded. These

Psychology of Nutrition Messages

169

instances were largely balanced by cases in which multiple advertisements occurred on one page so that the total number of advertisements in each 5-year period were was between 118 and 128 with a median value of 122. 15. This particular result could also be obtained in a bounded rationality model in which consumers do not have full information. However, this type of model would not explain the findings presented below regarding corresponding declines in the use of media coverage and advertisements. 16. Adding validity to the coding system used for advertisements in this study, a much larger study conducted by researchers at the FTC notes similar trends for nutrients such as fats (Ippolito & Pappalardo, 2002). Data from that study could not be used because the raw data are not publicly available, not all of the nutrients studied in this paper were tracked, and the study did not cover the necessary 30-year period from 1975 to 2004. 17. Curves for advertising claims tend to be less smooth than the other curves due to the lower number of total claims. 18. It would be possible to observe a similar decline in advertisements under a rational model if new information contradicting previous research were brought to light. However, one would expect that this new evidence would likely be to spur the increased publication in medical journals and newspapers as well, a trend that is not observed. Further, reductions in cholesterol consumption are still recommended by federal guidelines and trusted institutions such as the American Heart Association (AHA), suggesting that contradictory evidence is limited (USDHHS & USDA, 2005; AHA, 2007). Hence, it seems unlikely that the pattern of reduced publication and advertising would be observed under the rational full information model. 19. While no holistic health metrics sponsored by reputable public health organizations exist, some reputable agencies do sponsor symbols relating to specific aspects of health. For example, the AHA has developed a ‘‘Heart Check’’ symbol to designate heart healthy products (American Heart Association, 2008). However, these symbols may still be ‘‘hijacked’’ because they focus on a limited number of factors rather than a holistic view of health. In addition, a variety of logos sponsored by self-interested firms have become more prominent in recent years. The Pepsi brand ‘‘Smart Spot’’ logo (a green check mark surrounded by the words ‘‘Smart Choices Made Easy’’), is a common example of logos sponsored by firms that include products with questionable health value, such as Gatorade, and seem likely to be designed to maximize profits rather than consumer health (Pepsi Co., 2008). 20. An analogy might be to think about buying a car. When assessing the purchase of a car a person may think about the durability, safety, and gas mileage. An advertisement regarding side air-bags is likely to impact the individual’s perception of the safety of the car, but not its gas mileage or durability. Hence, the net impact of the claim regarding air-bags is limited. 21. Other equilibrium affects may also occur over time. These effects could occur on both the consumer and producer side and take a number of forms. For example, if a firm producing vegetable oil begins to label its product as ‘‘naturally cholesterol free’’ other firms are likely to do the same in order to remain competitive. This type of competition is likely to be a driving force leading to the exceptionally common use of health claims (nearly three quarters of products carry a claim). While the ubiquitous labeling may simply encourage consumers to consume more, it may also

170

HEATHER SCHOFIELD AND SENDHIL MULLAINATHAN

increase skepticism or cause consumers to down-weight the informativeness of overused messages. The current increases in consumption suggest that the first reaction is more common. However, as firms stretch the use of health claims further, the second may become more likely. Any formal analysis of equilibrium effects is beyond the scope of this paper.

REFERENCES Ackerberg, D. (2001). Empirically distinguishing informative and prestige effects of advertising. The RAND Journal of Economics, 32(2), 316–333. Agnoli, F. (1991). Development of judgmental heuristics and logical reasoning: Training counteracts the representativeness heuristic. Cognitive Development, 6, 195–217. Agnoli, F., & Krantz, D. (1989). Suppressing natural heuristics by formal instruction: The case of the conjunction fallacy. Cognitive Psychology, 21, 515–550. American Heart Association. (2007, April). Our 2006 diet and lifestyle recommendations. Cited April, 2007. /http://www.americanheart.org/presenter.jhtml?identifier ¼ 851S. American Heart Association. (2008, June). Food certification program. Cited June, 2008. /http://www.americanheart.org/presenter.jhtml?identifier ¼ 2115S. Balu, R. (1998, June 8). Trans fat: Taste buds cry ‘yes!’ but arteries demur. The Wall Street Journal. Blaylock, J., & Smallwood, D. (1987). Intrahousehold time allocation: The case of grocery shopping. The Journal of Consumer Affairs, 21(2), 183–201. Bless, H., Clore, G., Schwarz, N., Golisano, V., Rabe, C., & Wolk, M. (1996). Mood and the use of scripts: Does a happy mood really lead to mindlessness? Journal of Personality and Social Psychology, 71(4), 665–679. Chang, H., & Kinnucan, H. (1991). Advertising, information, and product quality: The case of butter. American Journal of Agricultural Economics, 73(4), 1195–1203. Clore, G., Schwarz, N., & Conway, M. (1994). Affective causes and consequences of social information processing. In: R. Wyer & T. Srull (Eds), Handbook of social cognition (pp. 323–417). Hillsdale: Lawrence Erlbaum Associates. Cutler, D., Glaeser, E., & Shapiro, J. (2003). Why have Americans become more obese? National Bureau of Economic Research Working Paper no. 9446. Davis, C., & Saltos, E. (1999). Dietary recommendations and how they have changed over time. In: E. Frazao (Ed.), America’s eating habits: Changes and consequences (pp. 33–50). Agriculture Information Bulletin 750. United States Department of Agriculture. /http:// www.ers.usda.gov/Publications/AIB750/S. Deighton, J., Henderson, C., & Neslin, S. (1994). The effects of advertising on brand switching and repeat purchasing. Journal of Marketing Research, 31, 28–43. Finucane, M., Alhakami, A., Slovic, P., & Johnson, S. (2000). The affect heuristic in judgments of risks and benefits. Journal of Behavioral Decision Making, 13, 1–17. Freedman, D., & Petitti, D. (2001). Salt and blood pressure: Conventional wisdom reconsidered. Evaluation Review, 25(3), 267–287. Gilbert, D. (1989). Thinking lightly about others: Automatic components of the social interference process. In: J. Uleman & J. Bargh (Eds), Unintended thought (pp. 189–211). Englewood Cliffs: Prentice-Hall. Guthrie, J., Derby, B., & Levy, A. (1999). What people do and do not know about nutrition. In: E. Frazao (Ed.), America’s eating habits: Changes and consequences (pp. 243–280).

Psychology of Nutrition Messages

171

Agriculture Information Bulletin 750. United States Department of Agriculture. /http:// www.ers.usda.gov/Publications/AIB750/S. Hamrick, K., & Shelley, K. (2005). How much time do Americans spend preparing and eating food? Amber Waves, 10–11. Ippolito, P., & Mathios, A. (1990). Information, advertising, and health choices: A study of the cereal market. RAND Journal of Economics, 21, 459–480. Ippolito, P., & Mathios, A. (1995). Information and advertising: The case of fat consumption in the United States. American Economic Review, 85(2), 91–95. Ippolito, P., & Pappalardo, J. (2002). Advertising nutrition and health: Evidence from food advertising 1977–1997. Bureau of Economics Staff Report, Federal Trade Commission. Isen, A. (1987). Positive affect, cognitive processes, and social behavior. In: L. Berkowitz (Ed.), Advances in experimental social psychology (Vol. 20, pp. 203–253). New York: Academic Press. Kahneman, D. (2003). Maps of bounded rationality: Psychology for behavioral economics. American Economic Review, 93(5), 1449–1475. Kahneman, D., & Frederick, S. (2002). Representativeness revisited: Attribute substitution in intuitive judgment. In: T. Gilovich, D. Griffin & D. Kahneman (Eds), Heuristics and biases: The psychology of intuitive thought (pp. 49–81). New York: Cambridge University Press. Kahneman, D., & Tversky, A. (1973). On the psychology of prediction. Psychological Review, 80(4), 237–251. LexisNexis. (2007). Guided news search. Cited Feb., 2007. /http://web.lexis-nexis.com.ezp2. harvard.edu/universe/form/academic/s_guidednews.htmlS. Loewenstein, G. (2000). Emotions and economic theory and economic behavior. American Economic Review, 90(2), 426–432. Martin, A. (2006, Nov. 6). The package may say healthy, but this grocer begs to differ. New York Times. National Heart Lung and Blood Institute (NHLBI). (2008). Your guide to lowering high blood pressure. Cited June, 2008. /http://www.nhlbi.nih.gov/hbp/prevent/sodium/ sodium.htmS. National Library of Medicine and National Institutes of Health. (2007). ‘‘Pubmed’’, Cited Feb., 2007. /http://www.ncbi.nlm.nih.gov/entrez/query.fcgiS. Nestle, M. (2002). Food politics: How the food industry influences nutrition and health. Berkeley: University of California Press. Nisbett, R., Krantz, D., Jepson, C., & Kunda, Z. (1983). The use of statistical heuristics in everyday inductive reasoning. Psychological Review, 90(4), 339–363. Olshansky, S., Passaro, D., Hershow, R., Layden, J., Carnes, B., Brody, J., Hayflick, L., Butler, R., Allison, D., & Ludwig, D. (2005). A potential decline in life expectancy in the United States in the 21st century. New England Journal of Medicine, 352(11), 1138–1145. Pepsi Co. (2008). Smart spot. Cited June, 2008. (http://www.smartspot.com). Produce for Better Health Foundation. (2008). Five a day the color way. Cited Jan., 2008. (http://www.5aday.org/index.htm). Tversky, A., & Kahneman, D. (1973). Availability: A heuristic for judging frequency and probability. Cognitive Psychology, 5, 207–232. Tversky, A., & Kahneman, D. (1974). Judgment under uncertainty: Heuristics and biases. Science, New Series, 185(4157), 1124–1131. Underhill, P. (1999). Why we buy. New York: Simon & Schuster.

172

HEATHER SCHOFIELD AND SENDHIL MULLAINATHAN

United States Department of Health and Human Services and United States Department of Agriculture. (2005). Dietary guidelines for Americans. Cited May, 2007. /www. healthierus.gov/dietaryguidelinesS. Waldholz, M. (1990, Mar. 1). New study demonstrates value of low-fat diet. The Wall Street Journal. Wansink, B., Kent, R., & Hoch, S. (1998). An anchoring and adjustment model of purchase quantity decisions. Journal of Marketing Research, 35(1), 71–81. Weimer, J. (1999). Accelerating the trend toward healthy eating: Public and private efforts. In: E. Frazao (Ed.), America’s eating habits: Changes and consequences (pp. 385–401). Agriculture Information Bulletin 750. United States Department of Agriculture. /http:// www.ers.usda.gov/Publications/AIB750/S.

EVALUATION CRITERIA FOR REPORT CARDS OF HEALTHCARE PROVIDERS Jesse D. Schold ABSTRACT Report cards, performance evaluations, and quality assessments continue to penetrate the lexicon of the healthcare sector. The value of report cards is typically couched as enhancing consumerism among patients, increasing accountability among healthcare providers, and more broadly increasing the transparency of healthcare information. This paper discusses the potential benefits and pitfalls of these performance assessments. This paper briefly reviews empirical evidence regarding the impact of report cards for healthcare providers and synthesizes the role and limitations of these performance measures into distinct evaluation criteria. The rapid proliferation of report cards for healthcare providers suggests a growing need to develop mechanisms and tools to evaluate their impact. The risks associated with utilizing report cards for provider oversight include the deleterious impact on vulnerable populations and a failure to accurately measure quality of care. The capacity to create report cards should not be the sole criterion to develop and utilize report cards to evaluate healthcare providers. Rather, careful consideration of the benefits and risks should accompany the implementation and

Beyond Health Insurance: Public Policy to Improve Health Advances in Health Economics and Health Services Research, Volume 19, 173–189 Copyright r 2008 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0731-2199/doi:10.1016/S0731-2199(08)19008-2

173

174

JESSE D. SCHOLD

utilization of report cards into regulatory processes. This report proposes an evaluation checklist by which to assess the role of report cards in a given healthcare context.

INTRODUCTION The increased volume and availability of data, heightened emphases on publicly reporting healthcare information, and expanded technological capacity have all contributed to the proliferation of report cards for healthcare providers. The notions of accountability, pay for performance, and evidence-based medicine have also gained significant popularity over the recent era. The related goals of inducing quality of care with various incentives and basing care practices on performance benchmarks are now mainstream concepts. At the same time, reports that the American healthcare system is besieged with medical errors resulting in increased mortality, morbidity, and inefficiencies have motivated efforts for greater introspection of processes and delivery of healthcare (Institute of Medicine, 2000). Efforts to collect and amalgamate data from a variety of healthcare contexts have led to unprecedented opportunities in which to assess outcomes at the level of individual provider, but, precisely how or when to appropriately integrate this new information into the public domain or regulatory processes is not as well understood. The notion of performance evaluations of healthcare providers with report cards has gained significant momentum in the United States. Independent evaluations of providers that may serve as snapshot indications of quality can potentially inform consumer decision-making without requiring sophisticated understanding of research methods or medical terminology. However, this synthesis of information may come at a cost that is not yet fully understood that may also be context dependent. A review of report cards conducted by the RAND organization suggested that while report cards have clearly proliferated, the ‘‘evaluation of report cards has not kept pace with their development’’ (RAND Health, 2007). As suggested by this report, further emphases are needed to examine the appropriate role and impact of report cards in healthcare. The concept that patient outcomes may be attributable to provider characteristics can be traced back more than 100 years to the work of Florence Nightingale. However, the proliferation and subsequent debate concerning the utility of report cards for healthcare providers has largely

Evaluation of Report Cards

175

arisen in the last two decades and the issues and controversies are now as salient as ever (Lee, Torchiana, & Lock, 2007). The public debate in the medical literature reached prominence in the context of cardiac bypass surgery (CABG) (Green & Wintfeld, 1995; Mukamel, Weimer, Zwanziger, Gorthy, & Mushlin, 2004; Schneider & Epstein, 1999). At the center of these controversies was whether risk adjustment strategies sufficiently accounted for patient characteristics and whether report cards provide disincentives to treat higher acuity patients who may deleteriously influence outcomes. Additional research suggests that few patients either know about or utilize these report cards to select providers, which are presumably one of the primary goals of their implementation (Schneider & Epstein, 1998). Another prominent source of controversy has been the application of the Health Plan Employer and Data Information Set (HEDIS) outcome measures. An important observation with use of the HEDIS measures was that health maintenance organizations were significantly less likely to disclose outcome information in circumstances in which they were likely to be assessed with poor quality of care (McCormick, Himmelstein, Woolhandler, Wolfe, & Bor, 2002). Other sources of debate have been discordance of information based on the method of data acquisition and the lack of correlation between different measures of quality for the same providers (Parkerton, Smith, Belin, & Feldbau, 2003; Thompson et al., 2001). However, an important contribution of this standardized metric for quality has been facilitating the identification of disparities in care to certain patient populations (Schneider, Zaslavsky, & Epstein, 2002; Trivedi, Zaslavsky, Schneider, & Ayanian, 2006). Further evidence suggests that quality measures such as HEDIS may instigate significant improvements in objective and meaningful care practices (Epstein, Lee, & Hamel, 2004). Similar debates on observations have extended to many other healthcare contexts including evaluations of overall hospital performance, managed care plans, and providers treating patients with a variety of chronic and acute health conditions (Baker et al., 2002; Epstein, 1998; Gandhi et al., 2002; Hochhauser, 1998; Zaslavsky et al., 2000). The extensive research interest focused on healthcare report cards underlies their significance and to some extent the contentious nature of implementation into practice. Overall, there is little doubt that variability in the quality of providers of care exists, but at the same time there is sufficient evidence that risks can be associated with implementing report cards into practice. The open fundamental question that is yet unresolved is what role should report cards have in regulating health providers and informing consumer choice? At one extreme, the argument is that given the obvious

176

JESSE D. SCHOLD

importance of quality healthcare, providers require rigorous oversight, and specific objective mechanisms need to be in place to identify and exclude poor providers of care. Furthermore, despite any difficulties in measurement of quality, even imperfect report cards provide an important utility given the alternative of an absence of information, and furthermore these reports can be adjusted on an ongoing basis. At the other extreme, a view exists that report cards cannot adequately evaluate quality of care, consumers are unwilling or unable to utilize this information appropriately, providers are able to manipulate data to their benefit and selected, typically vulnerable patients, are less likely to be treated as an ancillary product of report card dissemination. This perspective suggests that the risks associated with report cards may in fact outweigh the potential benefits. However, it is also conceivable that the truth lies not within these polarized views, but that there is some reasonable middle ground. That is, report cards may have an important role in certain healthcare contexts but not others. Moreover, it is possible that utilizing report cards as an indicator for certain quality measures is appropriate, while measuring other quality indicators is beyond reasonable capacity or too subjective to rely on. Similarly, providing report cards on a public forum may be applicable in certain healthcare contexts to inform consumers while in other circumstances this data may be best limited to use by insurers or agencies regulating patient safety. These concepts raise the challenging dilemma of how to distinguish circumstances in which report cards may have wide appeal and utility or conversely when report cards are endemic with perils to the populations served. The remaining difficult questions therefore, are how to evaluate the potential utility of report cards and to determine if there are means by which to understand the level of risk associated with report cards in a given context. This report will review some of the potential risks and benefits of report cards that have been discussed to date and then will attempt to outline certain fundamental criteria by which report cards may be evaluated.

POTENTIAL BENEFITS OF REPORT CARDS There are numerous cited utilities of report cards for healthcare providers as summarized in Table 1. One potential advantage of report cards is the increased transparency of provider outcomes that may allow patients and caregivers to select providers based on past performance. Allowing patients to select healthcare providers based on quality indicators may increase autonomy in a process which remains complex and frustrating to many

177

Evaluation of Report Cards

Risks and Benefits of Report Cards for Healthcare Providers. Benefits Identification of good and poor processes of care Enhance transparency of information and ‘‘consumerism’’ among patients and caregivers Create accountability among providers

Increase competition between providers based on quality measures Improved regulatory processes to identify safety concerns Providing objective feedback to providers for internal quality control

Encourages mandatory data collection

Risks Decreased access to care among vulnerable populations ‘‘Cream-skimming’’ – providers actively selecting ideal patients to enhance performance standards Providers target ‘‘artificial outcomes’’ that are measured in report cards that may not be concordant with other important outcomes Self-reported information or subjective coding which may be misleading Costs associated with data collection and reporting Misleading identification of either high or low quality of care due to exogenous factors (e.g., unmeasured patient characteristics or environmental factors) Inappropriate dissemination of information Failure to reach consumers and patients or to have information reach only selected individuals

individuals and ultimately result in better individualized care (Guadagnoli et al., 2000). These reports are already prolific in certain healthcare contexts, conducted by independent and regulatory agencies evaluating individual physicians, hospitals, and healthcare plans using various measurement criteria (Consumers’ Checkbook, 2008; Healthgrades, 2008; THE LEAPFROG GROUP, 2008). Certainly the notion of providing patients with objective information in a consolidated form to integrate with other personal preferences in order to guide provider selection is an attractive utility of report cards. Regulators may also utilize information gathered for reports to monitor safety concerns among providers and potentially leverage funding streams to mandate improved performance among providers. Agencies such as the Joint Commission may use outcomes data consolidated into report cards to justify the investigation of providers that demonstrate lower than expected results. In this regard, report cards may serve as objective means to concentrate regulatory efforts and potentially protect patients from care settings with poor practices. These efforts also may serve an important utility by providing objective feedback to providers allowing them to use the

178

JESSE D. SCHOLD

information for internal quality control initiatives which otherwise may not have been identified. In this sense, mandating data collection for the purpose of report cards has the ancillary effect of emphasizing certain outcome measures to providers and enhanced internal efforts to achieve quality of care goals. Accountability for patient outcomes, as measured by report cards, may also induce competition (on quality) among providers and ultimately result in improved care (Epstein, 1996). There is significant evidence that healthcare providers may be influenced by competitive practices and additional evidence to suggest a positive association of competition and quality of care (Gozvrisankaran & Town, 2003; Pawlson & O’Kane, 2002). Thus, report cards may facilitate the role of the marketplace to achieve enhanced quality measures. These arguments are particularly salient given a mechanism to disseminate report cards openly to patients and in circumstances with multiple service providers available to particular patient population. Certainly, providers utilizing report cards as marketing instruments has already become common practice. Data identifying provider quality may also lead to greater understanding of processes of care and subsequently the ability to reduce the incidence of medical errors. Relative to evaluation of typical ‘‘hard endpoints,’’ understanding the association of processes of care as an indicator quality has proved daunting. However, using report cards as an indicator of quality, may enhance the ability to examine care practices that are consistent with performance levels. In many ways, processes of care are more difficult to quantify as compared to objective measures such as mortality rates, but may in fact be a critical factor toward understanding the quality of care delivered (Ferguson et al., 2003; Peterson et al., 2006). As such, our understanding of better processes of care in healthcare can be identified and subsequently more readily evaluated as a result of efforts to generate report cards. Similarly, one of the ancillary benefits of the proliferation of report cards is the establishment and extension of data collection systems that are utilized to capture patient characteristics and outcome measures. These data may serve as valuable for research purposes and quality control initiatives independent of the report cards for which they are intended.

POTENTIAL RISKS OF REPORT CARDS Ostensibly, the benefits of report cards are significant, have substantial potential to increase quality of care and have gained such momentum that

Evaluation of Report Cards

179

cautionary efforts may be ineffectual. In many ways, the argument against developing a more transparent healthcare system seems counterintuitive and may replace circumstances in which there is a complete absence of data to guide consumers and regulators. Thus, to understand the risks of report cards may require more foresight into the practical limitations associated with their development, implementation, and long-term ramifications. Table 1 provides a summary of potential risks which may accompany the use of report cards which are described in detail in the following discussion. At base, report cards are intended to measure an aspect of providers’ quality of care. In this regard, it is critical to the development of report cards that there is some consensus on the definition of quality of care. One of the challenges for report card development is to identify an outcome measure that is an accurate reflection of quality of care specifically attributable to the role of a provider. However, to some degree the specific definition of quality of care is as ambiguous as our notion of health. Whether quality of care should comprise physical and mental health status, patient satisfaction or utilization of certain treatments with known efficacy remains an open question and is often context dependent. Report cards may utilize one or multiple measures of quality; however whether these criteria are reflective of the most important patient outcomes in some settings may not be clear. Similarly, specific quality indicators may create mixed incentives for treatment strategies (e.g., maximizing early outcomes at the expense of future prognoses) depending on the healthcare context. Outcome measures such as mortality (if collected rigorously) lack subjectivity, however other measures, such as certain medical diagnoses, patient satisfaction levels or quality of life are often more difficult to assess rigorously. In particular, concerns that exist regarding subjective endpoints include the role of measurement error, the opportunity for changes in coding practice to elevate ratings and generally the degree to which measures are truly indicative of provider quality of care (Iezzoni, Shwartz, Ash, Mackiernan, & Hotchkin, 1994). There are also potential dangers associated with healthcare contexts in which self-reported data or non-uniform data is utilized for outcome measures or risk adjustment purposes (Gandhi et al., 2002). In these situations, providers may have unintended incentives to portray their cohort with a higher case-mix in order to enhance risk adjusted outcomes. In addition, contexts in which data reporting is not mandatory may evoke selective samples of data that do not accurately portray the full spectrum of available providers and quality of care. As a result, report cards may inadvertently create disincentives to disseminate provider-level information

180

JESSE D. SCHOLD

given the potential deleterious ramifications of reporting lower than expected outcomes. Cumulatively, certain risks associated with report cards include an inability to accurately portray provider quality of care, the ability to manipulate outcome data (particularly those measured on subjective scales), outcomes measures may not be strongly correlated with long-term results or patient preferences, and the additional oversight associated with report cards may induce disincentives to report outcomes. One of the well-documented potential perils of report cards is that vulnerable populations may be discriminated against based on the increased likelihood that they will deleteriously impact provider ratings (D’Oronzio, 1995; Davies, Washington, & Bindman, 2002). Discrimination in this context refers to selectively refusing care to patients that are perceived to have poor prognoses, especially given that these characteristics may not be fully accounted for by traditional risk adjustment methodology. This risk pertains to healthcare contexts in which patient selection is logistically feasible, and can be exhibited in various manners, such as providers’ unwillingness to treat patients lacking comprehensive insurance coverage. Given the realities that poor report cards can affect providers’ competitive advantage or funding streams these risks are not to be taken lightly. At base, the degree to which this risk is germane is centered upon how much information a provider can ascertain from initial patient screening experiences as compared to the statistical power of models to control for ‘‘important’’ indicators. Physicians may be able to evaluate patients’ likelihood for compliance of treatment regimens, for instance, based on brief interactions. Moreover, risk adjustment is only applicable for variables that are routinely collected, and in certain contexts, there may be many additional factors that are known to impact outcomes measured by report cards but are not captured. For example, providers may be well-aware that socioeconomic status or a history of comorbidities is a significant determinant of patient outcomes. However, in many cases these types of variables are not incorporated into risk adjustment models. In these circumstances, providers are in a compromising situation in which certain patients are more likely to detract from performance ratings given inadequate risk adjustment, yet the given treatment may be in the best interest of the patient. In a similar fashion, use of report cards may motivate providers to seek out particularly ‘‘good’’ patients, such as those that are healthier or otherwise have good prognoses based on numerous factors that cannot be fully accounted for, and prioritize treatment for these patients. Perhaps the most contentious aspect concerning use of report cards to evaluate healthcare providers is whether risk adjustment is adequate to fully

Evaluation of Report Cards

181

account for differences in patient characteristics (Iezzoni, 1997). There are few circumstances in which a comparison of provider outcomes would be valid without accounting for characteristics of the population. Therefore, the method and predictive power of risk adjustment models are critical to developing fair assessments of provider outcome performance. Empirically, the wide variability in the predictive ability of models to assess provider performance has been demonstrated in various healthcare contexts ranging from very strong in some circumstances to relatively weak in others (Iezzoni et al., 1994). The specific risk of application of report cards in contexts with low predictive power is that differences in outcomes may be due to exogenous factors, those not captured in statistical models, and potentially unrelated to quality of care (Schold, Srinivas, Howard, Jamieson, & MeierKriesche, 2008). In these cases, providers may be inappropriately given poor ratings that are not associated with quality of care and others given artificially high ratings. Donabedian is credited with coining the term attributional validity with respect to quality of care between providers (Donabedian, 1988). This concept raises the question as to whether observed differences in provider outcomes are likely to be attributable to quality of care or conversely, whether provider quality of care is unlikely to independently explain observed differences. In general, the appropriateness of report cards in a given healthcare context may depend on the availability of appropriate risk adjustment variables and the ability to accurately attribute provider outcomes to the level and quality of care received based on analytical and contextual understanding. By design, report cards are typically available online or through various media outlets. While there are potential merits for widely publicizing this information, there may be little to discourage providers, insurers or other industries to misuse the data for advertisement or competitive purposes for which they were not originally intended. Oversight of the dissemination of reports is a necessary component for the agencies conducting provider evaluations. There are additional concerns that report cards rarely reach the intended populations or that the information only reaches selected portions of the population (Nair & Valuck, 2004; Schauffler & Mordavsky, 2001). Given the substantial resources that may be associated with data collection, analysis, and reporting processes, empirical evidence is required to demonstrate an effect on consumerism if that is a pre-conceived goal (Baker et al., 2002). Moreover, concerns exist that data from report cards are not fully comprehended by the intended audiences and that the proper interpretation of the information comprising report cards, including the limitations, are not (or cannot) be explained sufficiently to a broad

182

JESSE D. SCHOLD

population (Hochhauser, 1998). In general, even in circumstances in which report cards represent accurate quality measures, the results may have little impact on the intended populations accompanied by significant healthcare resources required to produce them. There is emerging evidence that the methodology utilized to generate report cards is associated with content. That is, based on different statistical approaches, research has demonstrated that report cards may give substantially variable assessments of providers’ quality of care (Gandhi et al., 2002; Glance, Dick, Osler, Li, & Mukamel, 2006). The implication of these findings is largely to illustrate the difficulty and variability in assessing quality of care, even with sound, but different, methodological approaches. These inconsistencies raise doubts as to whether the measurement tools are sufficient to provide fair assessments. Additional controversy relates to the relationship between statistically significant versus clinically relevant results (Iezzoni, 1997). Objective quantitative measures of provider outcomes may not always be indicative of differences that are meaningful in a clinical setting and alternatively, clinically relevant differences may not always be captured by performance evaluations alone. One common method of risk adjustment is indirect standardization, basing assessments of quality relative to outcomes of a larger population. One of the pitfalls associated with this methodology is that there is substantial likelihood that a relatively fixed proportion of providers will have lower than expected results regardless of the clinical relevance or overall level of quality in the population. Thus, using this methodology, rather than some known benchmark that is indicative of quality, assessments are based on a relative basis and may denote providers with clinically acceptable outcomes but relatively poor performance. In general, there exist non-trivial methodological challenges in the development of report cards that may significantly impact the value of these assessments.

EVALUATION CRITERIA FOR REPORT CARDS Based on the empirical evidence, a broad generalization about the role of report cards in healthcare is likely not warranted. Rather, our understanding of the utility of report cards may depend on the context in which they are used and how they are implemented into practice. For instance, report cards that are used as determinants of public funding may have severe implications and require greater scrutiny as opposed to assessments by private agencies. In general, given that there are a host of potential

Evaluation of Report Cards

183

concerns about the functionality, appropriateness and misuse of report cards in healthcare, careful ongoing examination of the use of report cards is clearly warranted. Despite the limitations, the positive momentum of report cards in healthcare is undeniable and further, among the report cards in existence, there is evidence that some may serve an important utility. Therefore dissuading the proliferation or utilization of report cards as a general principle is not likely practical or a rational approach. Instead, what is likely needed is a greater understanding of the role of report cards are (and should be), what defines a ‘‘quality’’ report card, and how to gauge whether report cards are appropriate in a given context. Further research concerning the impact of report cards and mechanisms to disseminate results appropriately also require more intensive examination. To further elucidate the types of information that should be gathered and the questions that should be asked for both existing report cards and proposals for new report cards in healthcare, objective criteria for report card evaluations would therefore be helpful. Table 2 lists a number of questions that can be useful for evaluating the quality or potential caveats of report cards in a given healthcare context, which will be discussed in the subsequent material. These criteria must be evaluated with appropriate clinical and research

Evaluation Criteria for Report Cards for Healthcare Providers. (1) Is there opportunity for ‘‘patient selection’’ among providers? If so, are there objective measures for these behaviors? (2) Is there opportunity for patients to select providers based on report cards? If so, are there objective measures for evaluating this behavior? (3) What is the predictive value of models utilized to generate report cards? (4) What is the reproducibility of factors utilized in report cards? Do different statistical approaches significantly alter the conclusions of report cards? (5) Are data collected on a mandatory basis and are data self-reported? (6) What is the variability in morbidity and mortality in the underlying populations of interest between providers and are there systematic differences in factors that are associated with outcomes between providers? (7) What is the level of variation in patient outcomes between providers and is this level likely to be indicative of the quality of care? (8) Are differences in performance evaluations applicable for only certain subsets of the population (e.g., high-risk patients)? (9) Is the information synthesized by report cards disseminated in a responsible manner relative to the potential limitations? (10) Are there ways by which to measure the impact of report cards on prospective quality of care?

184

JESSE D. SCHOLD

considerations and as well as with respect to the logistical aspects of patient care. One of the clear limitations of report cards is the role of selection bias toward influencing performance evaluations. The mechanisms of selection bias may differ depending on the healthcare context, but as with any observational investigation, selection bias cannot be discounted. This may be manifested from the perspective of the provider and the ability to select ‘‘better’’ patients (Item #1). The definition of ‘‘better’’ patients is context dependent but an important question to understand is whether mechanisms of care lend themselves to allow providers to avoid higher risk patients or in contrast if there is little control of providers to define which patients are treated. In this sense, an initial evaluative step is to determine if this type of selection is possible. If it is a reasonable concern, then there should also be mechanisms for measuring the degree of this behavior to infer the impact on report card evaluations. Failure to measure this or have data that defines this behavior may encourage biased report cards based on patient selection or even more critical, the denial of care to patients that are deemed as highrisk. Alternatively, selection bias may also influence performance evaluations from the perspective of patients selecting providers (Item #2). That is, whether ‘‘better’’ patients systematically select certain providers based on non-codified or unavailable information is critical to evaluating the quality of care between providers. Again, if this behavior is likely, there must be data and mechanisms for evaluating its influence on report card outcomes. Failure to do so may present significant bias in performance evaluation such that providers of care receive better ratings simply by treating better patients. An important utility of report cards is the ability for patients to utilize the information to guide decisions to select providers. However, the ability to utilize report cards to inform provider selection may not always be feasible. For example, in the context of cardiac bypass surgery, research suggests that few patients had sufficient time to research surgeons report card evaluations based on the time between recommendation for treatment and the procedure itself (Schneider & Epstein, 1998). As one of the primary roles of report cards is to inform patients, these types of logistical considerations must also be taken into account to maximize or even justify the use of report cards. Rarely in any medical context are the patient populations being served at equal risk between providers of care. Therefore, in order to generate and evaluate differences between providers of care, the utilization of risk adjustment is clearly indicated. However, the degree of risk adjustment and

Evaluation of Report Cards

185

the ability to predict outcomes based on available information also varies substantially between medical contexts. This variation affects the ability to make inferences about the cause of outcomes and the potential impact of random or systematic variation in the applicable populations. A measure of the predictive value of a particular risk adjustment model is one manner by which to gauge utility (Item #3). Models with low predictive power (e.g., as measured by r2 statistics or concordance indices) may imply a failure to capture certain influential factors that are associated with health outcomes. In contrast, models that have high predictive power give some assurance that models explaining outcomes capture important confounding variables. An additional indicator of the robust level of risk adjustment is to examine the stability of adjustment factors over time (Item #4). In the case that certain influential risk adjustment factors have highly variable impact over time, this may imply that either exogenous factors are influencing outcomes or that currently utilized factors are insufficient to account for differences in the baseline population. Moreover, evaluating the role of various methodological approaches toward measuring quality of care is an important tool. Inconsistencies of report card ratings associated with different evaluative methods may suggest significant caution in their interpretation. In general, risk adjustment per se is not a panacea to eliminate the effect of confounders on models assessing provider performance levels. The techniques and utility of risk adjustment varies and should be carefully examined toward providing fair assessments of provider quality of care. A significant contributing factor to potential bias in provider assessments is the degree of mandatory data collection (Item #5). Failure to disclose full information or ‘‘informed’’ loss of information between providers promotes significant potential biases, specifically, that poor outcomes are less likely to be reported. In a similar fashion, all providers of care within a given health context must report data on regular intervals, or misaligned incentives may exist that providers select intervals to report data specifically corresponding to performance. Significant problems may be associated with self-reported data in general, in which mixed incentives exist concerning data that may place the provider at risk for diminished evaluations and vice versa. An additional characteristic which may indicate the appropriateness of report cards in a given health context is the morbidity and mortality rate of the ‘‘baseline population’’ (Item #6). In conjunction with previously listed items, for populations with extremely diverse risk profiles, the ability to equate report cards with quality of care becomes more tenuous. In these cases, the role of environmental factors or underlying morbidity prior to

186

JESSE D. SCHOLD

treatment may have a more predominant impact on outcomes than the quality of care that patients receive. This is epitomized by efforts to compare quality of care between nations with highly diverse wealth, mortality, and healthcare structure (among other factors). In these cases attributing patient outcomes to strictly quality of care provided by specific providers, despite attempts to adjust for patient characteristics, is a daunting undertaking. Ultimately, the primary justification of report cards is to improve quality of care. Therefore, the existence of objective measures of the impact of report cards on patient care is another important criterion for evaluation (Item #7). Inability to assess the impact of report cards on patient outcomes diminishes, if not eliminates, the rationale for their development. Particularly given evidence that some report cards may be associated with deleterious ramifications, there must exist an analytical plan to evaluate the impact on patient outcomes and there must be information available to conduct such analyses. A portion of this analytical plan should also specifically address the impact of report cards on various portions of the population including an evaluation health outcomes and delivery of care to patients that might be characterized as vulnerable (Item #8). An additional criterion for the appropriate implementation of report cards is that the dissemination of the results reach the intended audiences and are coupled with clear language describing the interpretation of the results as well as the limitations (Item #9). As one of the greatest purported rationales for report cards is to inform patients, the language of report cards must be clearly understood by patients and the limitations listed in a coherent manner. This information must also be accessible to patients. Ongoing investigation assuring that report cards are being utilized by patients should be an important component of a plan for implementing them into practice (Item #10). Report cards that are not widely accessible to patients cannot hope to achieve the intended goal of providing transparent information to consumers of care. These evaluation criteria are certainly not intended to be exclusive but should be considered prior to and ongoing for report card development. Failure to meet criteria may indicate that report cards are an inefficient utilization of resources or even potentially deleteriously impact public health. The ability to meet these criteria likely depends on the healthcare context and while it may be difficult to compare use in highly diverse settings, research, and policy makers should also incorporate the important lessons that prior experience with report cards have provided. Careful consideration of the parties who have incentive to develop and use report cards and whether the incentives for various groups (i.e., patients, providers,

187

Evaluation of Report Cards

physicians, and insurance and government agencies) are concordant with patient care must also be understood. Ultimately, the value of report cards should be the broad impact on public health (not only restricted to the patients that actually receive care) and failure to consider the intermittent steps in evaluating this impact may render report cards ineffective or even harmful.

THE BOTTOM LINE: REPORT CARDS – ARE THE BENEFITS WORTH THE RISKS? This is the question that many clinicians, policy makers, and researchers want to have resolved. Despite the call for a solution to this question, the evidence to date is mixed and without the benefit of more experience and more sophisticated evaluation tools, the debates will likely continue. Based on the empirical evidence, there are tangible risks associated with report cards and such, the availability of data and the ability to create report cards is not sufficient justification to implement them into practice. Rather a careful analytical plan evaluating the utility and potential ramifications of the use of report cards is necessary. Clearly report cards may have differential impact between healthcare settings and while lessons can be learned from prior use of report cards, considerations of the particular nuances in a novel setting are important. Therefore, rather than focus on the broader question, many incremental gains can be made by evaluating the impact of report cards and building consensus for defining the evaluative tools by which to gauge the appropriateness of report cards more broadly. Given the many significant potentials benefits and perils of report cards in healthcare, continued attention to evaluative techniques and crosscontextual research will be critical in the years to come.

REFERENCES Baker, D. W., Einstadter, D., Thomas, C. L., Husak, S. S., Gordon, N. H., & Cebul, R. D. (2002). Mortality trends during a program that publicly reported hospital performance. Medical Care, 40(10), 879–890. Consumers’ Checkbook. (2008, February 5). http://www.checkbook.org/. Ref Type: Electronic Citation. Davies, H. T., Washington, A. E., & Bindman, A. B. (2002). Healthcare report cards: Implications for vulnerable patient groups and the organizations providing them care. Journal of Health Politics, Policy and Law, 27(3), 379–399.

188

JESSE D. SCHOLD

Donabedian, A. (1988). The quality of care: How can it be assessed? JAMA, 260(12), 1743–1748. D’Oronzio, J. C. (1995). ‘‘Unexpected’’ death and other report cards on access and ethics. Cambridge Quarterly of Healthcare Ethics, 4(4), 549–552. Epstein, A. M. (1996). The role of quality measurement in a competitive marketplace. Baxter Health Policy Review, 2, 207–234. Epstein, A. M. (1998). Rolling down the runway: The challenges ahead for quality report cards. JAMA, 279(21), 1691–1696. Epstein, A. M., Lee, T. H., & Hamel, M. B. (2004). Paying physicians for high-quality care. The New England Journal of Medicine, 350(4), 406–410. Ferguson, T. B., Jr., Peterson, E. D., Coombs, L. P., Eiken, M. C., Carey, M. L., Grover, F. L., & DeLong, E. R. (2003). Use of continuous quality improvement to increase use of process measures in patients undergoing coronary artery bypass graft surgery: A randomized controlled trial. JAMA, 290(1), 49–56. Gandhi, T. K., Francis, E. C., Puopolo, A. L., Burstin, H. R., Haas, J. S., & Brennan, T. A. (2002). Inconsistent report cards: Assessing the comparability of various measures of the quality of ambulatory care. Medical Care, 40(2), 155–165. Glance, L. G., Dick, A., Osler, T. M., Li, Y., & Mukamel, D. B. (2006). Impact of changing the statistical methodology on hospital and surgeon ranking: The case of the New York state cardiac surgery report card. Medical Care, 44(4), 311–319. Gozvrisankaran, G., & Town, R. J. (2003). Competition, payers, and hospital quality. Health Services Resolution, 38(6 Pt 1), 1403–1421. Green, J., & Wintfeld, N. (1995). Report cards on cardiac surgeons: Assessing New York State’s approach. The New England Journal of Medicine, 332(18), 1229–1232. Guadagnoli, E., Epstein, A. M., Zaslavsky, A., Shaul, J. A., Veroff, D., Fowler, F. J., Jr., & Cleary, P. D. (2000). Providing consumers with information about the quality of health plans: The consumer assessment of health plans demonstration in Washington State. Joint Commission Journal on Quality Improvement, 26(7), 410–420. Healthgrades. (2008, January 16). http://www.healthgrades.com/. Ref Type: Electronic Citation. Hochhauser, M. (1998). Can consumers understand managed care report cards? Managed Care Interface, 11(11), 91–95. Iezzoni, L. I. (1997). The risks of risk adjustment. JAMA, 278(19), 1600–1607. Iezzoni, L. I., Shwartz, M., Ash, A. S., Mackiernan, Y., & Hotchkin, E. K. (1994). Risk adjustment methods can affect perceptions of outcomes. American Journal of Medical Quality, 9(2), 43–48. Institute of Medicine. (2000). To err is human: Building a safer health system. Washington, DC: National Academy Press. Lee, T. H., Torchiana, D. F., & Lock, J. E. (2007). Is zero the ideal death rate? The New England Journal of Medicine, 357(2), 111–113. McCormick, D., Himmelstein, D. U., Woolhandler, S., Wolfe, S. M., & Bor, D. H. (2002). Relationship between low quality-of-care scores and HMOs’ subsequent public disclosure of quality-of-care scores. JAMA, 288(12), 1484–1490. Mukamel, D. B., Weimer, D. L., Zwanziger, J., Gorthy, S. F. H., & Mushlin, A. I. (2004). Quality report cards, selection of cardiac surgeons, and racial disparities: A study of the publication of the New York State cardiac surgery reports. Inquiry – the Journal of Health Care Organization Provision and Financing, 41(4), 435–446.

Evaluation of Report Cards

189

Nair, K. V., & Valuck, R. J. (2004). Consumer responses to a pharmacy benefit report card. Journal of Health Care Finance, 31(1), 55–72. Parkerton, P. H., Smith, D. G., Belin, T. R., & Feldbau, G. A. (2003). Physician performance assessment: Non-equivalence of primary care measures. Medical Care, 41(9), 1034–1047. Pawlson, L. G., & O’Kane, M. E. (2002). Professionalism, regulation, and the market: Impact on accountability for quality of care. Health Affairs (Millwood), 21(3), 200–207. Peterson, E. D., Roe, M. T., Mulgund, J., DeLong, E. R., Lytle, B. L., Brindis, R. G., Smith, S. C., Jr., Pollack, C. V., Jr., Newby, L. K., Harrington, R. A., Gibler, W. B., & Ohman, E. M. (2006). Association between hospital process performance and outcomes among patients with acute coronary syndromes. JAMA, 295(16), 1912–1920. RAND Health. Report Cards for Health Care. (2007, October 7). http://www.rand.org/pubs/ research_briefs/RB4544/index1.html. Ref Type: Electronic Citation. Schauffler, H. H., & Mordavsky, J. K. (2001). Consumer reports in health care: Do they make a difference? Annual Review of Public Health, 22, 69–89. Schneider, E., & Epstein, A. (1999). Public performance reports for cardiac surgery: Reply. JAMA, 281(2), 135. Schneider, E. C., & Epstein, A. M. (1998). Use of public performance reports-a survey of patients undergoing cardiac surgery. JAMA, 279(20), 1638–1642. Schneider, E. C., Zaslavsky, A. M., & Epstein, A. M. (2002). Racial disparities in the quality of care for enrollees in medicare managed care. JAMA, 287(10), 1288–1294. Schold, J. D., Srinivas, T. R., Howard, R. J., Jamieson, I. R., & Meier-Kriesche, H. U. (2008). The association of candidate mortality rates with kidney transplant outcomes and center performance evaluations. Transplantation, 85(1), 1–6. THE LEAPFROG GROUP. (2008, February 12). http://www.leapfroggroup.org/. Ref Type: Electronic Citation. Thompson, B. L., O’Connor, P., Boyle, R., Hindmarsh, M., Salem, N., Simmons, K. W., Wagner, E., Oswald, J., & Smith, S. M. (2001). Measuring clinical performance: Comparison and validity of telephone survey and administrative data. Health Services Resolution, 36(4), 813–825. Trivedi, A. N., Zaslavsky, A. M., Schneider, E. C., & Ayanian, J. Z. (2006). Relationship between quality of care and racial disparities in medicare health plans. JAMA, 296(16), 1998–2004. Zaslavsky, A. M., Hochheimer, J. N., Schneider, E. C., Cleary, P. D., Seidman, J. J., McGlynn, E. A., Thompson, J. W., Sennett, C., & Epstein, A. M. (2000). Impact of sociodemographic case mix on the HEDIS measures of health plan quality. Medical Care, 38(10), 981–992.

EVALUATING THE VALUE OF GENOMIC DIAGNOSTICS: IMPLICATIONS FOR CLINICAL PRACTICE AND PUBLIC POLICY Amalia M. Issa ABSTRACT An important current trend in health care is the move toward personalized medicine. Personalized medicine includes diagnostic and therapeutic interventions, with risk defined through genetics. The key paradigm shift brought about by the advent of personalized medicine is the increased use of in vitro genomic diagnostics. These tests offer the potential of being able to predict which patients are likely to respond to a particular drug, or which patients are likely to develop adverse reactions to a drug. The focus of this paper is the use of genomic diagnostics, and how the increasing development and translation into clinical practice of diagnostic – drug combination products will be adopted into health care delivery. The meaning of value and how to measure it is considered from different perspectives. A novel framework for evaluating the value of genomic diagnostics is proposed. Finally, the implications for regulatory approval and policy are discussed using an illustrative case study.

Beyond Health Insurance: Public Policy to Improve Health Advances in Health Economics and Health Services Research, Volume 19, 191–206 Copyright r 2008 by Emerald Group Publishing Limited All rights of reproduction in any form reserved ISSN: 0731-2199/doi:10.1016/S0731-2199(08)19009-4

191

192

AMALIA M. ISSA

The delivery of personalized medicine involves translating the science of pharmacogenomics into applications that can improve healthcare outcomes for patients. One key paradigm of personalized medicine applications is the combined use of diagnostics and therapeutics, designed to allow for more precise prescribing. This basic difference from the more traditional ‘‘blockbuster’’ (generally defined as drugs that meet a threshold of over one billion dollars in worldwide sales) model of drug development and distribution can lead to important implications for the economics of drug discovery and development, manufacturing, regulatory approval and distribution of combined genomic diagnostics, and targeted therapeutics as well as the clinical applications of the combined products. The wellknown cholesterol reducing drug atorvastatin (Lipitors) sold by Pfizer, with some $12 billion in sales, is an example of blockbuster. The use of so-called blockbusters, however, does not take into consideration inter-individual patient variability in diseases or drug responsiveness. Personalized medicine has the potential to impact patient health outcomes by reducing adverse drug reactions and increasing the effectiveness of therapies (Issa, 2008). In addition, personalized medicine may present opportunities for more efficient allocation of scarce health care resources if genomic technologies are found to be more cost-effective than existing alternatives (Phillips, Veenstra, Van Bebber, & Sakowski, 2003; Stallings et al., 2006). Clinical applications of personalized genomic medicine are occurring at a relatively slow pace. Nevertheless, a number of clinical applications have already been identified (Vogel et al., 2002; Kirchheiner et al., 2004; Wajapeyee & Somasundaram, 2004), and pharmacogenomic diagnostic tests and drugs are poised to enter more widespread clinical use (Shah et al., 2004; Wajapeyee & Somasundaram, 2004; Burczynski et al., 2005; De Leon, Susce, & Murray-Carmichael, 2006). Most recently, in August 2007, the United States Food and Drug Administration (FDA) approved the relabeling of the anticoagulant drug warfarin, to include the use of genomic diagnostic tests to determine a patient’s sensitivity to warfarin, and thus to improve the initial estimates that are being made concerning dosage based upon testing for the vitamin K epoxide reductase complex 1 (VKORC1) and cytochrome-P450 2C9 subtype (CYP2C9) genotypes (US FDA, 2007). This action by the FDA represents a significant step in the application of genomic diagnostics in health care. The focus of this paper is the use of genomic diagnostics and their use for clinical decisions of targeted therapeutics as an application of personalized medicine. A novel framework for the evaluation of genomic diagnostics is proposed. The particular attributes of defining and measuring the value of

Evaluating the Value of Genomic Diagnostics

193

genomic diagnostics for adoption into clinical practice and health care delivery are considered. Finally, the policy implications are considered using a case study as an illustrative example.

THE ADVENT OF GENOMIC DIAGNOSTICS In vitro diagnostics can be broadly defined as ‘‘those reagents, instruments, and systems intended for use in diagnosis of disease or other conditions, including a determination of the state of health, in order to cure, mitigate, treat or prevent disease or its sequelae. Such products are intended for use in the collection, preparation, and examination of specimens taken from the human body’’ (Food and Drug Administration, 21 CFR 809.3). According to one estimate some $56 billion were spent on diagnostic services in 2005 (Lewin Group, 2005). Molecular diagnostics, which can be considered a subset of in vitro diagnostics, can be categorized into several types depending on the purpose and function (Ross & Ginsburg, 2003):  Risk assessment markers, which are designed to measure disease susceptibility;  Screening markers, which can be used to screen large populations to enable the discrimination between healthy and asymptomatic disease states;  Prognostic markers, which are used to facilitate prediction of the course of a given disease, once a particular disease has been established, to aid in the determination of treatment modalities. A substantial portion of the inter-individual variability in drug response is believed to be genetic in nature (Issa, 2002, 2008; Freudenberg-Hua et al., 2005; Evans & McLeod, 2003; Evans & Relling, 2004; Kirchheiner et al., 2004). For the purposes of this paper, the term ‘‘genomic diagnostics’’ is proposed and defined as the use of genomic approaches to detect genetic sequences that are predictive of drug response and that can be associated with a targeted therapeutic in an underlying cause–effect relationship. The anticipated benefits from using genomic diagnostics in combination with therapeutics along with associated potential challenges are shown in Table 1. Numerous authors have suggested that pharmacogenomics could have an impact on healthcare expenditures and outcomes (Veenstra, Higashi, & Phillips, 2000; Robertson, Brody, Buchanan, Kahn, & McPherson, 2002; Phillips & Van Bebber, 2004; Phillips, Veenstra, Ramsey, Van Bebber, & Sakowski, 2004; Stallings et al., 2006).

194

AMALIA M. ISSA

Selected Benefits and Challenges of Genomic Diagnostic and Targeted Therapeutic Combination Products. Benefits Improved screening for drug response Less number of medications patients must take to find an effective therapy Less adverse drug reactions

More accurate methods of determining appropriate drug dosages Improved efficacy of the drug development process Reducing length of clinical trials

Faster drug approval times by FDA Reduced overall health care costs

Challenges Variable degrees of penetrance for drug response and adverse events Consequences of not performing a test if available Consequences of information provided by pharmacogenomic testing if no drug is available Changes needed to health care system infrastructure Changes needed to health insurance industry Uncertainties regarding interpretation of information provided by pharmacogenomic testing Regulatory oversight Fairness in access to genomic technologies

MEASURING VALUE: THE ECONOMIC APPROACH It is useful to consider what is meant by value. In economics, value refers to exchange value (Varian, 1993). Since money is the medium of exchange, the value of the benefit is generally determined by price – that is, the quantity of money that will be exchanged. However, the value of a benefit is not simply the price of that product in the open market. It is rather the worth of that benefit to a potential buyer. This is often measured in economic terms as willingness to pay (WTP). There are a number of approaches that economists use to measure value (Table 2). One widely used approach to measure the value of healthcare interventions is cost-effectiveness (Task Force, 1995). In the context of pharmacogenomics analyzing cost-effectiveness involves comparing disease, the specific genotype of interest, the genomic diagnostic available for testing, and the therapeutic modality to standard of care. Economic studies of the pharmacogenomic approach have focused mainly on cost-benefit and costeffectiveness analyses. The cost-effectiveness of pharmacogenomics has been discussed in the literature (Veenstra, Higashi, & Phillips, 2000; Phillips et al., 2003; Phillips & Van Bebber, 2004; Van den Akker-van Marle, Gurwitz, & Detmar, 2006). A major limitation for cost-effectiveness and

Evaluating the Value of Genomic Diagnostics

195

Economic Measures of Value. Cost of illness (COI) Measures economic burden of disorders Can be used to calculate potential benefits of prevention Cost benefit analysis (CBA) Health outcomes converted to currency (e.g., dollars) Value of a ‘‘statistical life’’ Productivity gains Willingness-to-pay Net benefit or benefit-cost ratio Cost effectiveness analysis (CEA) Costs and effectiveness of alternatives compared using ratio of incremental costs to incremental effect. Cost minimization analysis Outcomes of two or more comparators assumed equal Assessment based solely on comparative costs Cost-utility (CU) Health outcomes expressed in life years or quality adjusted life years (QALYs), disability adjusted life years (DALYs), or health year equivalents (HYEs) Cost per unit of outcome ratios can be derived that depict costs required to obtain one QALY

cost-benefit analyses is that one generally requires population-based data, such as can be derived from randomized controlled trials (RCTS) or cohort studies (Sullivan & Weiss, 2001), which in the still relatively nascent field of pharmacogenomics, is not yet widely available. Cost-utility analysis measures quality-adjusted life years (QALYs), and therefore is a better measure of value of diagnostic-drug combinations in economic terms.

MEASURING VALUE: THE EPIDEMIOLOGIC APPROACH A number of criteria have been developed and used by epidemiologists to measure the value of diagnostics, most commonly, sensitivity and specificity as well as predictive value (positive and negative). Several epidemiological approaches to measure value are delineated in Table 3. Sensitivity refers to the proportion of persons with the disease who are correctly identified by a screening test. Specificity refers to the proportion of persons without a disease who are correctly identified by a test. The

196

AMALIA M. ISSA

Epidemiologic Measures of Value. Measure Sensitivity

Specificity

Positive predictive value Negative predictive value Positive diagnostic likelihood ratio

Negative diagnostic likelihood ratio

Description The probability that a diagnostic or screening test will generate a true positive result when used on a population with the condition The probability that a diagnostic or screening test will generate a true negative result when used in the absence of a condition The probability that the genotype or condition of interest is present when the diagnostic is positive The probability that the genotype or condition of interest is absent when the diagnostic is negative The odds ratio that a positive diagnostic test result will be observed in a population with the genotype of interest as compared with the odds of a positive result in a population without the genotype or condition The odds ratio that a negative test result will be observed in population with the condition of interest as compared to the odds that a negative result will be observed in a population without the condition of interest

Formula TP/(TPþFN)

TN/(TNþFP)

TP/(TPþFP) TN/(TNþFN) TP=ðTP þ FNÞ FP=ðFP þ TNÞ

FN=ðTP þ FNÞ TN=ðFP þ TNÞ

Abbreviations: TP, number of true positives; FN, number of false negatives; TN, number of true negatives; FP, number of false positives.

specificity is the number of true negative results divided by the sum of the numbers of true negative plus false positive results. Negative predictive value (NPV) refers to the probability of not having the disease given a negative diagnostic test. It requires an estimate of prevalence. Positive predictive value (PPV) specifies the probability of having the disease given a positive diagnostic test. It requires an estimate of prevalence. An additional useful measure that is not frequently reported in the literature, but which may be valuable for genomic diagnostics is the use of diagnostic likelihood ratios (DLR; Halkin, Reichman, Schwaber, Paltiel, & Brezis, 1998). In the context of genomic diagnostics, a positive DLR represents the odds ratio that a positive test result will be observed in a population with the genotype of interest as compared with the odds of a positive result in a population without the specific genotype of interest. The DLR can thus be applied to determine the usefulness of a particular diagnostic since by definition, a diagnostic test exhibiting a larger positive DLR will be more useful than tests exhibiting smaller positive DLRs.

Evaluating the Value of Genomic Diagnostics

197

On the other hand, negative DLRs represent the odds ratio that a negative test result will be observed in population with the specific genotype of interest as compared to the odds that a negative result will be observed in a population without the genotype of interest. In this case, a negative DLR approximating zero represents a useful diagnostic, whereas less useful diagnostics will exhibit higher negative DLRs.

MEASURING VALUE AND THE IMPORTANCE OF CLINICAL UTILITY Although these epidemiologic measures are routinely used for diagnostic tests and provide some measure of clinical utility, by themselves they are not enough to provide a true measure of the value of genomic diagnostics. For example, the predictive value of a diagnostic test, such as measuring PPV or NPV, is not a real measure of its clinical value. Given that the value of diagnostics in general depends on whether the information they provide leads to actionable outcomes, particularly improvements in patient outcomes, it is not surprising that concern about clinical utility of diagnostics and how best to measure it have been expressed in the literature and debated for some time (Fryback & Thornbury, 1991; Lord, Irwig, & Simes, 2006; Guyatt, Tugwell, Feeny, Haynes, & Drummond, 1986; Barratt et al., 1999; Busse et al., 2002; Dauzat et al., 1997; Tatsioni, Zarin, Aronson, & Samson, 2005). Indeed, Fryback and Thornbury (1991) suggested a six-tiered hierarchical model of evaluating the efficacy of diagnostic imaging techniques with the first tier focused on technical efficacy or proficiency of a test followed by diagnostic accuracy, the ability of the test to alter the physician’s thinking about the diagnosis in question, the ability to elicit a change in the treatment plan of the patient, to change patient outcomes, and finally the economic effect from a societal viewpoint. Although useful for certain types of screening methods and diagnostic tests such as radiographic imaging, for example, this representation is less useful for assessing genomic diagnostics. The main difficulty with this model is the requirement of RCTs for data related to patient outcomes, which is not always feasible or practical in the venue of genomic diagnostics and drug combinations. In the context of genetics, the term ‘‘clinical utility’’ has been defined as ‘‘the ability of a screening or diagnostic test to prevent or ameliorate adverse health outcomes such as mortality, morbidity, or disability through the adoption of efficacious treatments conditioned on test results’’ (Grosse &

198

AMALIA M. ISSA

Khoury, 2006). Under the auspices of the Centers for Disease Control (CDC), the ACCE (Analytic validity, Clinical validity, Clinical utility, and Ethical, legal, and social implications) project was launched and suggested as a framework to specifically evaluate genetic screening and diagnostic tests broadly (i.e., not specifically limited to genomic diagnostics for the purpose of prescribing targeted therapeutics) further elaborating the concept of clinical utility (Grosse & Khoury, 2006; Khoury, 2003). Various groups, including the United Kingdom’s Public Health Genetics Unit have adopted the ACCE (Sanderson, Zimmern, Kroese, & Higgins, et al., 2005); however specific endpoints and outcomes are not specified for clinical utility.

A COMPREHENSIVE FRAMEWORK OF EVALUATING THE VALUE OF GENOMIC DIAGNOSTICS A more comprehensive approach to measuring value of novel genomic diagnostics would have to encompass epidemiological, socioeconomic, and clinical data to provide an actionable framework for health policy decisionmaking at both the macro and micro levels. My proposed framework is illustrated in Table 4. The framework proposed here is designed to be a simple yet powerful tool for making policy decisions at the macro or micro level about the value and utilization of novel genomic diagnostics. It is important to point out that despite its hierarchy of levels; it is not intended simply to be a decision analytic algorithm for go/no go type decisions. Rather, it is intended to be used at multiple stages of the diagnostic and drug development process and translation into clinical practice in an iterative manner. The model framework guides clinicians and policymakers to assess 12 attributes of a genomic diagnostic and integrates the population perspective, characteristics to measure test accuracy, including the strength of the genotypephenotype association, a particularly critical factor in evaluating genomic diagnostics, as well as regulatory and socioeconomic attributes. Particularly noteworthy are the levels pertaining to pertinent subgroups and the presentation of indefinite results, both attributes that are uniquely applicable to novel genomic diagnostics. Each of the 12 levels of attributes is probed further using a set of supporting questions for effectiveness. The next step is to test this novel comprehensive framework using a Delphi panel of experts to validate its usefulness in real world current examples of genomic diagnostics.

199

Evaluating the Value of Genomic Diagnostics

Hierarchy of Attributes for Evaluating Measures of Genomic Diagnostics. Level

Attribute

Measure(s)

1

Priority for maximizing the health of individuals or populations

Prevalence incidence; morbidity/mortality disability

2

Economic significance of relevant disease or condition

For example, COI

3

Genomic diagnostic and drug combination characteristics

Existence of relevant biomarker; availability of diagnostic; availability of drug

4

Accuracy of genomic diagnostic

5

Precision of diagnostic test results

6

Pertinent sub-groups

Sensitivity; specificity of genomic diagnostic; PPV; NPV; strength of genotype–phenotype association; positive and negative DLRs Reliability of sensitivity and sensitivity depends on number of patients evaluated Indices of accuracy for different sub-groups

Description How frequent is the disease in the population? What are the morbidity and mortality outcomes associated with the disease? How does it impact quality of life? Are there high expenditures on inpatient or outpatient care for the condition in question? Does it affect a significant proportion of the population? A condition may have economic significance if it affects a large number of people or has high per-person costs Does a biological marker that has been validated exists that can be specifically used to genotype drug response? If a genomic diagnostic is available, how well does it meet the criteria of analytical and clinical validity? Is there a drug that is available to justify using the diagnostic? To determine effectiveness of a genomic diagnostic, it is important to determine the level of penetrance and how well the genotype predicts a phenotypic manifestation Stated point estimate should have confidence intervals (or standard errors) regardless of magnitude Sensitivity and specificity are criteria that might be representative of average values for a given population. Unless the genotype for which a diagnostic is to be used is narrowly defined, then the indices may vary in different drug response sub-groups

200

AMALIA M. ISSA

(Continued ) Level

Attribute

Measure(s)

Description Test–retest reliability can be evaluated by repeated measurement within specified timeframe; inter-rater agreement is measured and reported using the kappa statistic; validation of instrument reliability can be analyzed to ensure internal consistency of testing Diagnostic influences treatment; clinician able to take action on data provided by use of diagnostic; physicians are able to use the information produced to improve patient care Diagnostic tests, particularly genomic diagnostics will not always provide a go/no go decision or yes versus no answer. The frequency of indefinite results from testing will limit the clinical utility of a test or increase costs due to the need for further confirmatory tests. How indefinite test results are used in determining a diagnostics efficacy is critical to the value added that a diagnostic represents Has the diagnostic received regulatory approval? Is regulatory approval contingent upon combination or integration with a targeted therapeutic? Have appropriate comparative analyses been done?

7

Reproducibility of diagnostic test

Instrument variability and calibration; test–retest reliability; inter-observer reliability

8

Clinical effectiveness and utility

Therapeutic efficacy; potentially actionable to user; meaningful and interpretable results

9

Presentation of indefinite test results

10

Regulatory status

Regulatory approval

11

Economic analyses

12

Patient outcomes and practicability

Societal costeffectiveness and cost-benefit ratios Specific patient outcome measures, such as QALYs; confidentiality of data is appropriately guarded; equitable distribution of benefits among population

Evaluating the Value of Genomic Diagnostics

201

SOME POLICY IMPLICATIONS: AN ILLUSTRATIVE CASE STUDY To better understand the policy implications for patients and society, and demonstrate both the potential opportunities and the current challenges of the utilization of genomic diagnostics in the development and utilization of combined diagnostics and drugs, it is useful to consider the illustrative example of trastuzumab (Herceptins). Trastuzumab was approved by the FDA in 1998 (Food and Drug Administration, 1998). The discovery that the human epidermal growth factor receptor-2 (HER-2) protein on the surface of breast cancer cells is over-expressed in approximately 25–30% of breast cancer patients led to the development of a therapeutic antibody that could target HER-2 (Albanell & Baselga, 1999; Vogel et al., 2002). Trastuzumab was initially approved in combination with paclitaxel for patients with metastatic breast cancer whose tumors over-expressed HER-2 and who had not received any treatment for their disease. Trastuzumab, as a single agent, is indicated for patients with metastatic breast cancer whose tumors over-expressed HER-2 and who had not responded to other chemotherapeutics. The case of trastuzumab offers several lessons for understanding how the value of genomic diagnostics combined with targeted therapeutics will be appraised and applied in society. One critical lesson is the importance of identifying appropriate biomarkers in the quest for improving the development and shepherding of novel therapeutics through the pipeline. The defining characteristic of trastuzumab is essentially a biomarker: the over-expression of HER-2. A second lesson is the value of codeveloping accurate diagnostic tests. The case of trastuzumab’s pathway through the pipeline from development to regulatory approval and more widespread clinical use illustrates that the value of an integrated drug and diagnostic combination cannot be over-estimated. However, despite the success of trastuzumab through the pipeline and onto market, difficulties in data interpretation due to a lack of standardization in HER-2 screening have been acknowledged. Several tests have been approved for HER-2 screening. The HercepTests, a semiquantitative immunohistochemical assay, developed by DAKO, utilizes only the membrane staining score (in increasing order of positivity 0, 1þ, 2þ, 3þ) (DAKO, 1998). Moreover, according to the Herceptins product insert, the HercepTests ‘‘has not been directly studied for its ability to predict HERCEPTIN treatment effect’’ (Genentech, Inc., 2008). Rather, the HercepTests was evaluated retrospectively for concordance with the investigation-only immunohistochemistry clinical trial assay (CTA) that

202

AMALIA M. ISSA

had been used in the two pivotal trials that led to approval of trastuzumab. It is important to note that 42% of the specimens that tested 2þ positive by the HercepTests were considered negative by the original CTA and thus would not have met clinical trial eligibility criteria (DAKO, 1998). DNA probe kits using fluorescent in situ hybridization (FISH) technology for the detection and quantification of HER-2/neu gene amplification, in which paraffin-embedded formalin-fixed tissue can be used, have also been approved by the FDA (PathVysion Her-2 Package Insert, 1998; Oncor, 1998). These different diagnostic tests have spawned debate in the literature regarding specificity, sensitivity, and the provision of prognostic data of HER-2 over-expression and clinical utility (Wang et al., 2000; Hammock, Lewis, Phillips, & Cohen, 2003; Luftner et al., 2004). Thus, an important lesson to keep in mind is that the co-development of a validated, sensitive, and specific diagnostic, a likely key requirement for approval of a novel pharmacogenomic-based therapeutic, would have to be simultaneous with the development of the drug. A considerable degree of validation data would have to be completed during phase II clinical trials, in order for the sponsor to be ready to provide diagnostic data such that it will impact the filing for the first indication. A third lesson comes from economics. Despite trastuzumab ostensibly being a small-market drug rather than the traditional ‘‘blockbuster,’’ annual sales have steadily grown from an initial $188 million in 1998 to over $1.2 billion in 2006 (the last year for which sales data are publicly available) (Genentech, Inc., 2006). With expanded use as a post-surgical supplementary treatment for breast cancer and in cases of ovarian and lung cancer with HER-2 over-expression, it appears that the prospects for trastuzumab may continue to grow. Other drugs that appear to have followed a similar path through the pipeline such as imatinib meslate (Gleevecs) and rituximab (Rituxans) offer similar lessons. In patients with Philadelphia chromosome positive chronic myeloid leukemia (CML), Gleevecs is a protein tyrosine kinase inhibitor that inhibits the bcr-abl tyrosine kinase (Cohen, Moses, & Pazdur, 2002; Deininger, 2004). Rituxans is indicated for CD20-positive, B-cell non-Hodgkin’s lymphoma (Grillo-Lopez, Hedrick, Rashford, & Benyunes, 2002). The lessons that can be drawn from these examples, however, also yield clues regarding the scientific, clinical, and economic uncertainties inherent in the application of pharmacogenomic approaches to address the policy issues inherent in combined genomic diagnostics and therapeutic products. A key issue that is critical for consideration relates to the vast amounts of genomic data that has been generated since the completion of the Human

Evaluating the Value of Genomic Diagnostics

203

Genome Project. The challenge, of course, is to identify the targets that will produce effective and safe drugs in an efficient and cost-effective manner. Although the technological advances in combinatorial chemistry and high throughput screening are making it possible to screen many more polymorphisms, target selection remains an inherent challenge to informing drug development programs. The lessons drawn from the examples briefly discussed above suggest that decisions regarding the implementation of pharmacogenomic approaches will be driven by economic factors as well as the scientific and clinical aspects, and the need for a comprehensive evaluative framework for clinical and policy decision-making.

CONCLUSIONS In this paper, a number of methods that can be applied to measure the value of genomic diagnostics have been reviewed. Each of the approaches has its relative advantages and disadvantages. Importantly, a novel framework that integrates different aspects of the various approaches has been proposed. In this framework, a genomic diagnostic can be evaluated at several levels using 12 principal attributes, ranging from the population perspective and diagnostic accuracy to cost-effectiveness and cost-benefit analysis, which can be further, elaborated using the associated measures and probing questions. This model framework remains to be tested empirically before it can be proposed for wider adoption and diffusion.

REFERENCES Albanell, J., & Baselga, J. (1999). Trastuzumab, a humanized anti-HER 2 monoclonal antibody, for the treatment for breast cancer. Drugs Today, 35(12), 931–946. Barratt, A., Irwing, L., Glasziou, P., Cumming, R. G., Raffle, A., Hicks, N., et al. (1999). Users’ guides to the medical literature: XVII. How to use guidelines and recommendations about screening. Evidence-based medicine working group. Journal of the American Medical Association, 281(21), 2029–2034. Burczynski, M. E., Oestreicher, J. L., Cahilly, M. J., Mounts, D. P., Whitley, M. Z., Speicher, L. A., & Trepicchio, W. L. (2005). Clinical pharmacogenomics and transcriptional profiling in early phase oncology clinical trials. Current Molecular Medicine, 5(1), 83–102. Busse, R., Orvain, J., Velasco, M., Perleth, M., Drummond, M., Gurtner, F., et al. (2002). Best practice in undertaking and reporting health technology assessments. Working group 4 report. International Journal of Technology Assessment in Health Care, 18(2), 361–422.

204

AMALIA M. ISSA

Cohen, M. H., Moses, M. L., & Pazdur, R. (2002). Gleevec for the treatment of chronic myelogenous leukemia: US Food and Drug Administration regulatory mechanisms, accelerated approval, and orphan drug status. The Oncologist, 7(5), 390–392. DAKO. (1998). HercepTest Package Insert. DAKO Corp., Carpineteria, California. Available from: http://www.dakocytomation.com/ Dauzat, M., Laroche, J. P., Deklunder, G., Ayoub, J., Quere, I., Lopez, F. M., et al. (1997). Diagnosis of acute lower limb deep venous thrombosis with ultrasound: Trends and controversies. Journal of Clinical Ultrasound, 25(7), 343–358. Deininger, M. W. (2004). Basic science going clinical: Molecularly targeted therapy of chronic myelogenous leukemia. Journal of Cancer Research and Clinical Oncology, 130(2), 59–72. De Leon, J., Susce, M. T., & Murray-Carmichael, E. (2006). The AmpliChip CYP450 genotyping test: Integrating a new clinical tool. Molecular Diagnosis & Therapy, 10(3), 135–151. Evans, W. E., & McLeod, H. L. (2003). Pharmacogenomics-drug disposition, drug targets, and side effects. The New England Journal of Medicine, 348(6), 538–549. Evans, W. E., & Relling, M. K. (2004). Moving towards individualized medicine with pharmacogenomics. Nature, 429(6990), 464–468. Food and Drug Administration. (1998). Letter of Approval. Available from: www.fda.gov/ cder/approval/index Food and Drug Administration Department of Health and Human Services. Part 809 – In Vitro Diagnostic Products for Human Use. 21CFR809.3, http://www.accessdata.fda.gov/ scripts/cdrh/cfdocs/cfcfr/CFRSearch.cfm?fr ¼ 809.3 Freudenberg-Hua, Y., Freudenberg, J., Winantea, J., Kluck, N., Cichon, S., Bruss, M., Propping, P., & Nothen, M. M. (2005). Systematic investigation of genetic variability in 111 human genes-implications for studying variable drug response. Pharmacogenomics Journal, 5(3), 183–192. Fryback, D. G., & Thornbury, J. R. (1991). The efficacy of diagnostic imaging. Medical Decision Making, 11(2), 88–94. Genentech, Inc. (2006). Annual report. Available from: http://www.gene.com/gene/ir/financials/ annual-reports/2006/2006annualreport.pdf Genentech, Inc. (2008). Drug prescribing information: Herceptin (Trastuzumab). Available from: http://www.gene.com/gene/products/information/pdf/herceptin-prescribing.pdf Grillo-Lopez, A. J., Hedrick, E., Rashford, M., & Benyunes, M. (2002). Rituximab: Ongoing and future clinical development. Seminars in Oncology, 29(1 Suppl. 2), 105–112. Grosse, S. D., & Khoury, M. J. (2006). What is the clinical utility of genetic testing? Genetics in Medicine, 8(7), 448–450. Guyatt, G. H., Tugwell, P. X., Feeny, D. H., Haynes, R. B., & Drummond, M. (1986). A framework for clinical evaluation of diagnostic technologies. Canadian Medical Association Journal, 134(6), 587–594. Halkin, A., Reichman, J., Schwaber, M., Paltiel, O., & Brezis, M. (1998). Likelihood ratios: Getting diagnostic testing into perspective. QJM: An International Journal of Medicine, 91(4), 247–258. Hammock, L., Lewis, M., Phillips, C., & Cohen, C. (2003). Strong HER-2/neu protein overexpression by immunohistochemistry often does not predict oncogene amplification by fluorescence in situ hybridization. Human Pathology, 34(10), 1043–1047. Issa, A. M. (2002). Perspectives in pharmacogenomic profiling in the drug development process. Nature Review Drug Discovery, 1(4), 300–308.

Evaluating the Value of Genomic Diagnostics

205

Issa, A. M. (2008). Clinical applications of pharmacogenomics to adverse drug reactions. Expert Review of Clinical Pharmacology, 1(2), 251–260. Khoury, M. J. (2003). Genetics and genomics in practice: The continuum from genetic disease to genetic information in health and disease. Genetics in Medicine, 5(4), 261–268. Kirchheiner, J., Nickchen, K., Bauer, M., Wong, M. L., Licinio, J., Roots, I., & Brockmoller, J. (2004). Pharmacogenetics of antidepressants and antipsychotics: The contribution of allelic variations to the phenotype of drug response. Molecular Psychiatry, 9(5), 442–473. Lewin Group. (2005). The value of diagnostics innovation, adoption and diffusion into health care, pp. 1–230. Available from: http://www.socalbio.org/pdfs/thevalueofdiagnostics.pdf Lord, S. J., Irwig, L., & Simes, R. J. (2006). When is measuring sensitivity and specificity sufficient to evaluate a diagnostic test, and when do we need randomized trials? Annals of Internal Medicine, 144(11), 850–855. Luftner, D., Henschke, P., Kafka, A., Anagnostopoulos, I., Wiechen, K., Geppert, R., Stein, H., Wernecke, K. D., Kreienberg, R., & Possinger, K. (2004). Discordant results obtained for different methods of HER-2/neu testing in breast cancer – a question of standardization, automation and timing. The International Journal of Biological Markers, 19(1), 1–13. Oncor. (1998). Oncor INFORM HER-2/neu Gene detection system: Procedure and interpretation guide. Ventana Medical Systems, Tucson, AZ. Available from: http:// www.ventanamed.com PathVysion Her-2 Package Insert. (1998). Vysis, Inc. Downers Grove, III. Available from: http://www.vysis.com/PathVysionHER@DNAProbeKit_35922 Phillips, K. A., & Van Bebber, S. L. (2004). Cost-effectiveness of pharmacogenomic interventions: A systematic review of the literature. Pharmacogenomics, 5(8), 1139–1149. Phillips, K. A., Veenstra, D., Van Bebber, S. L., & Sakowski, J. (2003). An introduction to costeffectiveness and cost–benefit analysis of pharmacogenomics. Pharmacogenomics, 4(3), 231–239. Phillips, K. A., Veenstra, D. L., Ramsey,, S. D, Van Bebber, S. L., & Sakowski, J. (2004). Genetic testing and pharmacogenomics: Issues for determining the impact to health care delivery and costs. The American Journal of Managed Care, 10(7 Pt 1), 425–432. Robertson, J. A., Brody, B., Buchanan, A., Kahn, J., & McPherson, E. (2002). Pharmacogenetic challenges for the health care system. Health Affairs, 21(4), 155–167. Ross, J. S., & Ginsburg, G. S. (2003). The integration of molecular diagnostics with therapeutics. American Journal of Clinical Pathology, 119(1), 26–36. Sanderson, S., Zimmern, R., Kroese, M., Higgins, J., et al. (2005). How can the evaluation of genetic tests be enhanced? Lessons learned from the ACCE framework and evaluating genetic tests in the United Kingdom. Genetics in Medicine, 7(7), 495–500. Shah, R., Darne, B., Atar, D., Abadie, E., Adams, K. F., & Zannad, F. (2004). Pharmacogenomics in cardiovascular clinical trials. Fundamental and Clinical Pharmacology, 18(6), 705–708. Stallings, S. C., Huss, D., Finkelstein, N., Crown, W. H., Witt, W. P., Maguire, J., Hiller, A. J., Sinskey, A. J., & Ginsburg, G. S. (2006). A framework to evaluate the economic impact of pharmacogenomics. Pharmacogenomics, 7(6), 853–862. Sullivan, S. D., & Weiss, K. B. (2001). Health economics of asthma and rhinitis II. Assessing the value of interventions. The Journal of Allergy and Clinical Immunology, 107(2), 203–210. Task Force on Principles for Economic Analysis of Health Care Technology. (1995). Economic analysis of health care technology: A report on principles. Annals of Internal Medicine, 123(1), 61–70.

206

AMALIA M. ISSA

Tatsioni, A., Zarin, D. A., Aronson, N., Samson, D. J., et al. (2005). Challenges in systematic reviews of diagnostic technologies. Annals of Internal Medicine, 142(12 Pt 2), 1048–1055. United States Food and Drug Administration. (2007). New labeling information for warfarin (marketed as Coumadin). August 16, 2007. Available from: http://www.fda.gov/cder/ drug/infopage/warfarin/default.htm Van den Akker-van Marle, M. E., Gurwitz, D., Detmar, S. B., et al. (2006). Cost-effectiveness of pharmacogenomics in clinical practice: A case study of thiopurine methyltransferase genotyping in acute lymphoblastic leukemia in Europe. Pharmacogenomics, 7(5), 783–792. Varian, H. R. (1993). Intermediate microeconomics. A modern approach. New York: W. W. Norton and Company. Veenstra, D. L., Higashi, M. K., & Phillips, K. A. (2000). Assessing the cost-effectiveness of pharmacogenomics. AAPS Pharmaceutical Sciences, 2(3), E29: 1–11. Vogel, C. L., Cobleigh, M. A., Tripathy, D., Gutheil, J. C., Harris, L. N., Fehrenacher, L., Slamon, D. J., Murphy, M., Novotny, W. F., Burchmore, M., Shak, S., Steart, S. J., & Press, M. (2002). Efficacy and safety of Trastuzumab as a single agent in first-line treatment of HER 2-overexpressing metastatic breast cancer. Journal of Clinical Oncology, 20(3), 719–726. Wajapeyee, N., & Somasundaram, K. (2004). Pharmacogenomics in breast cancer: Current trends and future directions. Current Opinion in Molecular Therapeutics, 6(3), 296–301. Wang, S., Saboorian, M. H., Frenkel, E., Hynan, L., Gokaslan, S. T., & Ashfaq, R. (2000). Laboratory assessment of the status of Her-2/neu protein and oncogene in breast cancer specimens: Comparison of immunohistochemistry assay with fluorescence in situ hybridization assays. Journal of Clinical Pathology, 53(5), 374–381.