University of Southern California Dissertations and Theses

Essays in health economics: evidence from Medicare

(USC Thesis Other)

Essays in health economics: evidence from Medicare

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content ESSAYS IN HEALTH ECONOMICS:
EVIDENCE FROM MEDICARE

by

Lauren Matsunaga Scarpati

A dissertation presented to the faculty of The Graduate School
of the University of Southern California
in partial fulfillment of the requirements for the degree of
Doctor of Philosophy
(Economics)

December 2015

Copyright 2015 Lauren Matsunaga Scarpati

i

Abstract
Medicare has provided health insurance for elderly Americans, as well as the
disabled and those with end-stage renal disease, since 1966. Among this population,
I investigate the demand for health care, including prescription drugs and
preventive services, as well as the relationship between wealth and health, using a
large claims dataset.
In Chapter 2, I investigate the diffusion of a new medical technology –
Zostavax, the shingles vaccine for the elderly – and investigate the effect of socio-
demographic factors, as well as supply- and demand-side social influences on its
adoption. I find that treatment patterns in small areas ultimately drive take-up.
Observed factors are unable to explain regional differences in adoption. Physician
social networks have minimal effect on Zostavax diffusion and patients’ behavior is
largely prevalence-inelastic. Patient demand is highly price-elastic.
In Chapter 3, with Dana Goldman, Florian Heiss, Daniel McFadden, Joachim
Winter, and Amelie Wuppermann, I use the housing crisis as a natural experiment to
evaluate the health-wealth gradient, wherein the affluent are healthier than those
with fewer financial resources. We use plausibly exogenous changes in five-digit
level ZIP code average home prices to evaluate the effect of wealth on health.
Beneficiaries respond to changes in wealth by decreasing their use of medical
services; we do not observe that negative wealth shocks affect health itself.
In Chapter 4, with Julie Zissimopoulos, Geoffrey Joyce, and Dana Goldman, I
assess whether the Medicare Part D prescription drug benefit reduced disparities in
access to medication. Part D’s unique and complex benefit design included a gap in
coverage after beneficiaries reached a spending threshold, which we use as a quasi-
experiment to evaluate the differential demand for drug therapies across minorities
and socioeconomic groups. Using a differences-in-differences approach, we find that
the Part D coverage gap is particularly disruptive to minorities and those living in
low-income areas. An edited version of this article, including tables and figures, was
published in the American Journal of Managed Care in February 2015.
ii

Acknowledgements
This dissertation would not have been possible without the incredible support and
guidance of my advisor, Dana Goldman. He has taught me an immeasurable amount
about health economics and conducting and presenting research. Despite his busy
schedule, he always made time for our check-ins, which helped me to make
continuous progress and complete this dissertation more expeditiously than I could
have hoped. It has been an invaluable experience to have him as a mentor and I am
so grateful.

It has been the greatest honor to work with Dan McFadden for the last few years. He
has been a continual inspiration and is one of the most fascinating people I have
met. His humility, work ethic, and pleasant attitude are qualities to which we all
should aspire.

I am especially grateful to Julie Zissimopoulos for bringing me into the Schaeffer
Center and giving me the opportunity to learn so much on my first project in health
economics. Her service on my committee and comments on my thesis manuscript
were tremendously valuable. She selflessly invested much time and effort into my
training, which I hope to be able to pay forward.

I have learned so much from the talented researchers at the Schaeffer Center and
Department of Economics. I am deeply appreciative of my other co-authors, Geoff
Joyce, Florian Heiss, Joachim Winter, and Amelie Wuppermann for sharing their
expertise and making the research process incredibly enjoyable. Thanks to Geert
Ridder and John Strauss for their insightful comments and service on my committee.

Patty St. Clair so graciously gave her time and wisdom to teach me and countless
other research assistants how to use the Medicare claims data and conduct analyses
using various statistical programs. I appreciate the programming and statistical
support from my other colleagues at the Schaeffer Center. Thanks to Darius
Lakdawalla and Neeraj Sood for use of direct-to-consumer advertising data. Morgan
Ponder, Sara Geiger, and Briana White provided excellent assistance.

Financial support from USC’s Leonard D. Schaeffer Center for Health Policy &
Economics and Dornsife College of the Art, Letters, and Sciences, and the National
Institutes of Health and National Institute on Aging (NIH/NIA R01-AG-29514,
NIH/NIA P01 AG33559-01A1, NIH 1 RC4 AG039036-01, and RCMAR Grant
P30AG043073) is gratefully acknowledged.

Finally, I am extremely thankful for the love and support of my family and friends.
Without my husband, Mike, this dissertation would not exist. His moral support and
statistical and technical expertise helped me persevere through myriad and varied
challenges. All errors are my own.
iii

Contents
Abstract ........................................................................................................................................................ i
Acknowledgements ................................................................................................................................ ii
List of Figures .......................................................................................................................................... iv
List of Tables ............................................................................................................................................ vi
CHAPTER 1. Introduction .................................................................................................................... 1
CHAPTER 2. “Diffusion of New Medical Technology: Evidence from Medicare” ........... 7
Introduction ......................................................................................................................................... 7
Background .......................................................................................................................................... 9
Data ....................................................................................................................................................... 13
Estimation .......................................................................................................................................... 22
Results ................................................................................................................................................. 25
Discussion .......................................................................................................................................... 29
Conclusion .......................................................................................................................................... 32
Tables and Figures .......................................................................................................................... 33
CHAPTER 3. “‘Healthy, wealthy, and wise?’ The effects of negative wealth shocks on
health and health care utilization” ................................................................................................ 49
Introduction ...................................................................................................................................... 49
Data ....................................................................................................................................................... 53
Estimation .......................................................................................................................................... 57
Results ................................................................................................................................................. 59
Discussion .......................................................................................................................................... 61
Conclusions ........................................................................................................................................ 62
Tables and Figures .......................................................................................................................... 64
CHAPTER 4. “Did Medicare Part D Reduce Disparities?” ..................................................... 73
Introduction ...................................................................................................................................... 73
Data ....................................................................................................................................................... 74
Statistical Analysis .......................................................................................................................... 77
Results ................................................................................................................................................. 79
Discussion .......................................................................................................................................... 82
Conclusions ........................................................................................................................................ 85
Tables and Figures .......................................................................................................................... 86
Bibliography........................................................................................................................................... 92
Appendix A. Supplemental Table and Figures for Chapter 2 ............................................ 110
Appendix B. Supplemental Table and Figures for Chapter 3 ............................................ 118
Appendix C. Supplemental Table and Figures for Chapter 4 ............................................ 119

iv

List of Figures
Chapter 2. “Diffusion of New Medical Technology: Evidence from
Medicare”
Figure 2-1: Number of Zostavax Prescriptions Filled over Time ................................... 33
Figure 2-2: Physicians Who Have Ever Prescribed Zostavax .......................................... 34
Figure 2-3: Diffusion of Zostavax Across Hospital Referral Regions and
Hospital Service Areas ...................................................................................................... 35
Figure 2-4: Hospital Service Area (HSA) Level Maps Showing Early Adoption
of Zostavax ..................................................................................................................36
Figure 2-5: Vaccination Rates over Time ................................................................................. 37
Figure 2-6: Distribution of Zostavax Prices Paid by Patients .......................................... 38
Figure 2-7: Prices Paid for Zostavax by Hospital Referral Region ................................. 39
Figure 2-8: Baseline Shingles Prevalence ................................................................................ 40
Figure 2-9: Physician Social Networks ..................................................................................... 41
Figure 2-10: Yearly Vaccination Rates by Group .................................................................. 42

Chapter 3. “‘Healthy, wealthy, and wise?’ The effects of negative wealth
shocks on health and health care utilization”
Figure 3-1: Home Prices over Time ........................................................................................... 64
Figure 3-2: Distress Rates over Time ........................................................................................ 71

Chapter 4. “Did Medicare Part D Reduce Disparities?”
Figure 4-1: Regression Adjusted Difference-in-Difference in Medication Use
(MPR), by Therapeutic Class and Race (percentage point) ................................ 87
Figure 4-2: Regression Adjusted Difference-in-Difference in Use of Generic
Substitutes (GDR), by Race (percentage point) ...................................................... 90
Figure 4-3: Regression Adjusted Difference-in-Difference in Medication Use
(MPR), by Therapeutic Class and Race for the Near-Poor Population ........... 91

Appendix A for Chapter 2
Figure A-1: Distribution of ZIP-level Population Counts over Time ............................. 110
Figure A-2: Prices Paid by Patients over Time ...................................................................... 111
Figure A-3: Geographic Variation in Patient Cost-Sharing for Zostavax ..................... 112
Figure A-4: Vaccination Rates by Income Quintile .............................................................. 113
Figure A-5: Age at the Time of Vaccination ............................................................................ 114

Appendix C for Chapter 4
Figure C-1. Cessation of Diabetes Drug Classes in the Coverage Gap in 2007 ......... 119
Figure C-2. Resumption of Diabetes Drug Classes in 2008 if Stopped in
the Coverage Gap in 2007 ............................................................................................... 119
Figure C-3. Cessation of Non-Diabetes Drug Classes in the Coverage Gap
in 2007 .................................................................................................................................... 120

v

List of Figures (continued)
Appendix C for Chapter 4 (continued)
Figure C-4. Resumption of Non-Diabetes Drug Classes in 2008 if Stopped
in the Coverage Gap in 2007 .................................................................................... 120
Figure C-5. Percent Changes in Medication Use upon Entering the Coverage
Gap: Diabetes Drugs .................................................................................................... 121
Figure C-6. Percent Changes in Medication Use upon Entering the Coverage
Gap: Non-Diabetes Drugs .......................................................................................... 121
Figure C-7. Percent Changes in Medication Use of the Near-Poor Population
upon Entering the Coverage Gap: Diabetes Drug Classes ............................. 122
Figure C-8. Percent Changes in Medication Use of the Near-Poor Population
upon Entering the Coverage Gap: Non-Diabetes Drug Classes ................... 122
Figure C-9. Percentage Point Changes in Use of Generic Drugs upon
Entering the Coverage Gap by Drug Class ........................................................... 123

vi

List of Tables
Chapter 2. “Diffusion of New Medical Technology: Evidence from
Medicare”
Table 2-1: Baseline Summary Statistics ................................................................................... 43
Table 2-2: Summary Statistics by Zostavax Adoption Rate .............................................. 44
Table 2-3: Explaining Adoption (Entire Population) .......................................................... 45
Table 2-4: Explaining Adoption (Highly-Subsidized Population) .................................. 46
Table 2-5: Explaining Adoption (65-69 Age Group) ........................................................... 47
Table 2-6: The Relationship between the Utilization of Medical Services
and Diffusion ........................................................................................................................ 48

Chapter 3. “‘Healthy, wealthy, and wise?’ The effects of negative wealth
shocks on health and health care utilization”
Table 3-1: ZIP-Year Level Descriptive Statistics ................................................................... 65
Table 3-2: Mortality & Health Outcomes ................................................................................. 67
Table 3-3: Expenditures ................................................................................................................. 68
Table 3-4: Utilization of Medical Services ............................................................................... 69
Table 3-5: Use of Preventive Services ....................................................................................... 70
Table 3-6: Selected Results Adding Distress Rates .............................................................. 72

Chapter 4. “Did Medicare Part D Reduce Disparities?”
Table 4-1: Beneficiary Characteristics by Coverage Group and Race ........................... 86
Table 4-2: Differential Stopping and Conditional Resumption Rates of the
Non-LIS Group Relative to the LIS Group .................................................................. 89

Appendix A for Chapter 2
Table A-1: Comparison of Individual-Level Summary Statistics .................................... 115
Table A-2: Unweighted Summary Statistics by Adoption Rate ....................................... 116
Table A-3: Models without Cumulative Adoption Rates .................................................... 117

Appendix B for Chapter 3
Table B-1: Preventive Services Procedure Codes................................................................. 118

Appendix C for Chapter 4
Table C-1. Beneficiaries’ Plan Switching from Year 2007 to 2008 ............................... 124

1

CHAPTER 1. Introduction
Since its inception in 1966, the Medicare program has provided health
insurance for elderly Americans, as well as the disabled and those with end-stage
renal disease. Today, over fifty-five million Americans receive benefits from
Medicare (Centers for Medicare and Medicaid Services 2015a), with fourteen
percent of federal spending in 2014 going to Medicare (Kaiser Family Foundation
2014). Massive increases in enrollment are forecasted as the Baby Boomer
generation (those born between 1946 and 1964) becomes Medicare-eligible – for a
total of sixty-four million enrollees in 2020 and eighty-two million in 2030 (Kaiser
Family Foundation 2013) – and corresponding rises in spending mean that is it even
more crucial to understand Medicare costs and the demand for medical care. The
looming insolvency of the Medicare trust fund, which is estimated to be depleted by
2030 (Boards of Trustees of the Federal Hospital Insurance and Federal
Supplementary Medical Insurance Trust Funds 2015) makes this all the more timely
and important.
Affordable prescription drug coverage for seniors became available for
purchase in 2006 through Medicare Part D. Chronic illnesses can often be managed
successfully using pharmaceutical therapies (Shrank et al. 2011), so, particularly for
the sixty-five percent of Medicare beneficiaries who have three or more chronic
conditions like diabetes (Kaiser Family Foundation 2014), Part D was very helpful in
improving access to necessary treatments. There are an estimated forty-two million
people currently enrolled in Part D (Congressional Budget Office 2015) and early
surveys showed that most individuals were pleased with their plans (KRC Research
2011).
These programs have been greatly beneficial to American seniors in
decreasing their risk of sizable financial losses due to large medical outlays and
improving their health through better access to treatments. However, there persist
geographic, socioeconomic, and racial/ethnic disparities in health outcomes and the
quality of and access to care. (Baicker, Chandra, and Skinner 2005; IOM (Institute of
2

Medicine) 2013) find that the location in which a patient resides has a large impact
on the quality and intensity of care received. There is considerable variation in
thirty-day readmission rates, which signal suboptimal coordination of care and
room for improvement in quality of care and cost reduction (Goodman, Fisher, and
Chang 2011). Some areas are much better at prescribing preventive technologies
than others (Goodney et al. 2014). These patterns are not well explained by area
demographics or other common observable factors (Goodman et al. 2010); this
thesis answers three questions which provide answers as to the underlying causes
of these patterns.
Population health has improved considerably over the last several decades,
but all groups have not benefited equally. In addition to considerable regional
variation discussed above, there also exist racial/ethnic and socioeconomic
disparities in health and access to and quality of health care among the Medicare
population. The combined costs of premature death and health inequalities were
estimated to be $1.24 trillion (Andrulis et al. 2010) between 2003 and 2006;
reducing health disparities is a clear priority. Racial and ethnic minorities have
higher rates of chronic illness than non-minorities. For example, the prevalence of
diabetes among blacks and Hispanics is considerably higher (American Diabetes
Association 2013) than white (76% and 66% higher, respectively). Black and
Hispanic enrollees report greater difficulty obtaining information and purchasing
needed medications in Part D (Haviland et al. 2012).
In this dissertation, I answer three questions in health economics that
provide insights into these geographic, racial/ethnic, and socioeconomic
inequalities. A large advantage of these empirical analyses over prior work is the
use of a large dataset comprised of medical and prescription drug claims for a
random sample of nine-million Medicare beneficiaries between 2002 and 2011. I
link these data with a variety of public use and restricted data at the ZIP code level
to incorporate information like socioeconomic status, housing wealth, and housing
distress. I exploit the longitudinal nature of these data using a variety of
econometric techniques, including differences-in-differences and panel data models.
3

Economists have long been interested in the diffusion of new technologies
(Coleman, Katz, and Menzel 1966) with good reason. New technologies are a key
driver of health care cost growth (Chandra and Skinner 2012). Further, new,
effective preventive technologies can have large effects on outcomes. In Chapter 2, I
conduct a large-scale empirical study of the adoption of a new medical technology:
Zostavax, the shingles vaccine, measured by the percent of patients within a ZIP
code who were immunized. Despite the pain and costliness of shingles and its
sequelae (Yawn et al. 2009), regional adoption rates remained low overall, with
some areas adoption much earlier and more intensely than others. Zostavax
provides an excellent test case for examining diffusion of a new pharmaceutical
technology among seniors for several reasons: (1) it was approved in the year in
which the prescription drug data become available allowing for the longest possible
study period; (2) patients are only required to receive the vaccine once, making it
simpler to model diffusion than chronic medications; and (3) Zostavax is
recommended for all elderly patients, allowing for wide sample, compared with a
therapy that is appropriate for a subset of the elderly population.
I investigate the effect of supply- and demand-side social influences on the
adoption of Zostavax. Do patient socio-demographics matter, or do information
flows on the physician side make a difference? I construct a proxy for physician
networks using the number of patients shared by pairs of physicians. Local shingles
prevalence rates capture patient-side shifters. My main finding is that treatment
patterns in small areas ultimately drive take-up. Observed factors do not explain the
large differences in adoption, and physician social networks have minimal effect on
Zostavax diffusion. Patients’ behavior is largely prevalence-inelastic and their
demand is highly price-elastic.
The health-wealth gradient, wherein the more affluent are healthier than
those with fewer financial resources, has been well-documented. However, the
causes of this relationship are not as firmly established. Understanding whether
wealth causes health or vice versa may improve public policy solutions that seek to
decrease inequalities. In Chapter 3, with Dana Goldman, Florian Heiss, Daniel
4

McFadden, Joachim Winter, and Amelie Wuppermann, I use exogenous changes in
home prices during the recent housing crisis as a natural experiment for evaluating
the effect of changes in wealth on health. Approximately 50% of the average older
American’s wealth was in the form of housing (Engelhardt, Eriksen, and Greenhalgh-
Stanley 2013) – falling home prices had large effects on the elderly’s wealth which
may have been stressful. Further, seniors may also have been affected by stress
arising from observing family, friends, and neighbors experiencing hardship or
stress arising from the uncertainty of markets and the value of their home. Chronic
exposure to the physiological mediators of stress can adversely affect the body
(McEwen 1998). Increased stress levels due to the uncertainty of housing markets
may have led to deleterious outcomes and changes in medical services use like non-
urgent hospitalizations, office visits, prescription drug use, and preventive care. We
focus on the effects among the American elderly population using a random sample
of nine-million Medicare beneficiaries, finding that negative wealth shocks are
associated with minimal effects on health outcomes. Beneficiaries respond to
changes in wealth by decreasing their use of medical services.
Both Healthy People 2020 and the 2011 Health and Human Services
Disparities Action Plan focus on improved management of chronic conditions to
decrease racial disparities in health outcomes. While pharmaceutical therapies are
effective for managing many chronic diseases, such as diabetes, such effects cannot
be fully realized without proper adherence to the prescribed regimen. Nonetheless,
non-adherence to medications is endemic (Marcum, Sevick, and Handler 2013;
Jackevicius, Mamdani, and Tu 2002; Cramer et al. 2003; Osterberg and Blaschke
2005; Haynes, McDonald, and Garg 2002), especially among ethnic and racial
minorities (Egede et al. 2011; A. S. Adams et al. 2013; Holmes et al. 2012; Gellad and
Haas 2007; Monane et al. 1996), and non-adherence costs the U.S. healthcare system
between $100 billion and $289 billion annually (Viswanathan et al. 2012).
Understanding the differential demand for prescription drugs across
races/ethnicities and levels of socioeconomic status is crucial.
5

In Chapter 4, with Julie Zissimopoulos, Geoffrey Joyce, and Dana Goldman, I
assess whether Medicare Part D reduced disparities in access to medication.
1
The
Medicare Part D benefit design is unique and complex, in that it initially contained a
gap in coverage, wherein patients would experience a sudden and temporary loss of
prescription drug coverage upon reaching a certain spending threshold. We use a
differences-in-differences design to analyze medication use of Hispanics, blacks and
whites beneficiaries with diabetes before and after reaching the Part D coverage
gap, using race-specific groups not exposed to the loss in coverage as controls. ZIP
code-level household income is used as a proxy for socioeconomic status. Hispanics
reduce use of diabetes-related medications in the coverage gap twice as much as
whites, while blacks decrease use of diabetes-related medications by thirty-three
percent more than whites. Hispanics without subsidies who live in low-income
areas reduce medication use more than similar blacks and whites. The Part D
coverage gap is thus particularly disruptive to minorities and those living in low-
income areas.
The findings in this dissertation have several policy implications. With the
elderly population growing rapidly and the Medicare cost growth between 2014 and
2024 projected to be 7.1%, understanding how to provide high-quality, lower-cost
care that improves population health is as relevant as ever. Hastening the adoption
of efficacious preventive pharmaceuticals and improving medication nonadherence
will help to improve patient health while cutting costs.
The findings in Chapter 2 yield insights as to how to increase take-up of the a
new vaccine for the elderly. While targeting specific groups (for example, those of
low socioeconomic status and racial/ethnic minorities) may help to increase the
adoption of new technologies, policymakers should understand that existence of
“high” and “low” adoption areas extends to the shingles vaccine – as in (Goodman et
al. 2010), the region in which a patient lives is the key in whether or not they
received the vaccine. Increased efforts to target areas with low propensities to

1
An edited version of this article, including all tables and figures, was published in the
American Journal of Managed Care in February 2015.
6

adopt new technologies are necessary, as well as future work more deeply
investigating the reasons underlying these patterns. Improving patient awareness of
new therapies which may dramatically improve their health and quality of life is
imperative.
To encourage the take-up of highly effective, preventive technologies,
patients should face very little to no cost-sharing, as is done with the influenza
vaccine. Policymakers should be aware that the most vulnerable individuals are
more price-sensitive, even when faced with very low levels of cost-sharing. For
chronic medications, upon experiencing price shocks, patients appear to cut back,
with Hispanic patients and the near-poor reacting most dramatically. Teaching
patients about the importance of medication adherence is a necessary step but is
not enough to decrease disparities. Increasing the use of health information
technologies that allow physicians or insurers to contact patients who do not fill or
refill a prescription may be beneficial. Finally, the successful coordination of patient
care may provide much value through better monitoring of the myriad
pharmaceuticals taken by patients, so that medications that are unnecessary or
ineffective are not taken for long.

7

CHAPTER 2. “Diffusion of New Medical Technology:
Evidence from Medicare”
Introduction
Increasing the use of preventive health care services has been featured as a
way to improve Americans’ health and decrease spending. The 2010 Patient
Protection and Affordable Care Act includes the National Prevention Strategy, a
roadmap for improving the health in the United States for people of all ages by
expanding access to preventive services, the elimination of disparities, and
encouraging patients to live more healthful lives. Combined with the widespread
geographic variation in health outcomes, utilization and spending (Glover 1938; J.
Wennberg and Gittelsohn 1973; J. E. Wennberg and Cooper 1996; Skinner 2012;
IOM (Institute of Medicine) 2013), the diffusion of new, effective preventive
technologies is timely and important for shaping future policy. Technological change
is a key driver of health care cost growth (Newhouse 1992; Chandra and Skinner
2012). We need to better understand such adoption patterns to learn how we may
better control costs while delivering high-quality care.
I first document the diffusion of Zostavax, the vaccine used to prevent
shingles in the elderly, among elderly Americans in the six years following its
approval by the Food and Drug Administration (FDA). Then, I examine factors
associated with more rapid adoption, particularly the relationships between
diffusion and physician- and patient-side social networks. (IOM (Institute of
Medicine) 2013) emphasized the large role of the physician and physician decision-
making in driving geographic variation. However, another study found that, among
the Medicare
2
population, 40-50% of geographic variation in health care utilization
is driven by patient demand, while 50-60% can be attributed to factors on the
supply-side (Finkelstein, Gentzkow, and Williams 2014). I evaluate the effects of

2
Medicare provides elderly Americans with affordable health insurance coverage. Disabled
and end-stage renal disease patients also may be covered.
8

physician and patient-side factors, carefully adjusting for neighborhood
characteristics and changes therein, on the adoption of a new drug.
Physicians’ influence on the adoption of new drugs is not well understood –
particularly the relationship between physician social networks and take-up of new
therapies. Increased communication among physicians may help information about
a new drug to spread, for example, to discuss a patient they share. Physician
relationships may be identified using the number of patients shared (Barnett et al.
2011). Such a measure has rarely been used in economic analyses as a proxy for
social networks, but given the conclusions in (Barnett et al. 2011), appears to be
worthy of deeper study. I identify physician pairs and construct a measure for the
number of physicians with whom a given doctor shares patients, adjusted by the
sum of shared patients as a proxy for physician social networks.
Studies on measles (Philipson 1996) and the Human Immunodeficiency Virus
(HIV) (Geoffard and Philipson 1996) find that the demand for prevention is
increasing in disease prevalence. Economic epidemiology (Philipson 2000) has
dubbed the response of demand for prevention to prevalence, “prevalence-elastic
behavior.” Previous work has focused on infectious diseases among younger groups;
this paper contributes by investigating prevalence-elastic behavior among the
elderly, exploiting plausibly exogenous variation in shingles onset. While shingles
itself is not infectious, patient demand for Zostavax may respond to local shingles
prevalence rates after patients observe the associated pain and discomfort
experienced by their neighbors and seek protection. Shingles infections within a ZIP
code may be viewed as exogenous when also controlling for factors like local
mortality rates and health status with which it is likely correlated. Scientists do not
have definitive evidence for the causes which lead the virus to reactivate within the
body (Centers for Disease Control and Prevention 2014a).
This paper contributes to the literature by combining the areas of social
networks, economic epidemiology, and health policy. I estimate the effects of
physician-side social networks and shingles prevalence (as a patient-side shifter) on
diffusion. Take-up is highly persistent within ZIP-codes: neighborhood-level
9

treatment patterns drive adoption, and observable factors cannot explain much of
the geographic variation. Physician interconnectedness is not associated with
adoption. The effect of local shingles prevalence on adoption is small though
statistically significant.
Background
Shingles & Zostavax
The dormant varicella-zoster virus, chicken pox, remains in the body after
the initial infection. It can spontaneously reactivate causing herpes zoster (Centers
for Disease Control and Prevention 2014d), better known as shingles. Almost all
Americans over the age of 40 (99.5%) have had chicken pox, which means almost all
of these individuals are at risk of developing shingles, while those aged fifty and
older and/or immune-compromised are especially susceptible (Centers for Disease
Control and Prevention 2014c). A patient with shingles will develop a painful,
blistering rash, which generally subsides within thirty days after treatment, either
on its own or with the help of antiviral drugs like Valtrex (valacyclovir). Pain
medication (NSAIDs or opioid analgesics) is regularly prescribed to make patients
more comfortable. The risk of subsequent postherpetic neuralgia (PHN), irreparable
nerve damage at the infection site, is a more serious concern. In addition to being
painful, herpes zoster is also costly: (Yawn et al. 2009) report the cost of herpes
zoster between 1996-2001 to be $720 without complications or PHN and $3,998
with complications or PHN – this translates to an annual cost of $1.1 billion to treat
incident cases of shingles. Those infected with shingles may infect others who have
not had chicken pox with chicken pox, but shingles itself is not contagious to those
who have already had chicken pox (Centers for Disease Control and Prevention
2014b).
Merck developed Zostavax, which received FDA approval on May 25, 2006.
3

Clinical trials have shown that Zostavax reduces the risk of initial onset shingles by

3
This initial approval was for the sixty-and-older population. It was approved for those ages
fifty to fifty-nine five years later on May 24, 2011.
10

about 50% and, moreover, reduces the risk of PHN by 67% (Oxman et al. 2005).
Zostavax may also lessen the risk of recurring infections. The Centers for Disease
Control and Prevention (CDC) recommends that those sixty and older be vaccinated.
Patients sixty to sixty-nine years of age benefit most from the vaccine, though it also
helps older individuals (Centers for Disease Control and Prevention 2014e). There
have not been any serious problems or adverse effects reported (Centers for Disease
Control and Prevention 2014d). Zostavax remains effective for up to ten years (P. Le
and Rothberg 2015).
Previous studies have used small datasets to evaluate patterns in the uptake
of the herpes zoster vaccine among Americans (Lu et al. 2009; Lu, Euler, and Harpaz
2011; Lee et al. 2013; Lu et al. 2014).
4
These studies find that many patients were
unaware of the vaccine but would have gotten it upon a physician’s
recommendation or after learning about it from the study’s survey. Patients who
had been vaccinated reported that a doctor’s recommendation and media were most
influential in their choice to receive the drug. Whites were more likely to get the
vaccine than blacks and Hispanics, as were patients born in the U.S. compared with
foreign-borns. Immunization rates were higher among older and higher-educated
individuals. There is no evidence that Merck initially limited purchasing of Zostavax
to a limited number of areas (for example, that it targeted large, urban, metropolitan
areas).
While these studies provide an interesting foundation for this study, they
have several limitations. These use small datasets for no more than 35,000
individuals and the studies able to control for geography are only able to do so very
broadly, where regions are defined as Northeast, West, South, and Midwest. Not all
datasets used are nationally representative, and the data represent limited
timeframes. I build on these studies using panel data for more than three million

4
The 2007 National Immunization Survey-Adult (NIS-Adult; N = 3,662), the 2008 National
Health Interview Survey (NHIS; N = 5,751), a cross-sectional survey of patients sixty and
older from three primary-care clinics in North Carolina (N = 403), and the 2012 NHIS (N =
34,525; 17.8% of which are 65+ and 56.4% of which are ages 18-49) are used, respectively.
11

Medicare beneficiaries over a period of six years, carefully adjusted for narrower
regional effects at the ZIP code level.

Previous Work on the Diffusion of New Technologies
Economists have long studied the diffusion of new technologies. Studies on
agricultural technologies have been especially pervasive, starting with (Griliches
1957), and also including (Foster and Rosenzweig 1995; Foster and Rosenzweig
1996; Munshi 2004; Conley and Udry 2010). From the health economics literature,
(Goldman and Smith 2005) do not find evidence of differential diffusion of ACE
inhibitors and calcium channel blockers among different SES groups. (Skinner and
Staiger 2005) use state-level data, concluding that some states were more likely to
adopt new technologies of all sorts, whether tractors or beta-blockers, than others.
(Skinner and Staiger 2009) use macroeconomic models of productivity to look at the
relationship between survival of heart attacks and technology diffusion. They show
that the adoption rate for inexpensive and efficient technologies like aspirin explain
a considerable amount of hospital-level variation in outcomes for heart attacks.
(Chandra, Malenka, and Skinner 2013) look at the diffusion of drug-eluting stents in
the first year following their FDA approval, documenting large variations in the
initial rates of diffusion at the hospital and region levels. The data most strongly
support the hypothesis that the highest quality hospitals were fastest to adopt the
new technology and that diffusion behavior is correlated within regions.

Social Networks Literature
Physician networks are endogenous: physicians are not randomly assigned
to locations or to friends and colleagues. Those who choose to reside and practice in
a given area prefer certain things about the place that made it a better choice above
all others. Physicians within an area will be part of various groups – practices,
hospitals, friendships – some of which are formed more selectively than others. In
the economics literature, (Manski 1993) formalizes the theory of social networks
and discusses the estimation of the effects of endogenous social influence variables.
12

He describes the “reflection problem,” wherein it would be very difficult for a
bystander who does not understand mirrors to identify whether a person is causing
the movements displayed in a mirror or vice versa. Then, Manski discusses the
identification of endogenous social effects, defining three different types of effects:
correlated effects, contextual effects, and endogenous effects. Endogenous effects in
this study would mean that a larger proportion of physicians prescribing the drug in
an area would increase the likelihood that a given physician prescribes it. An
example of correlated effects would be that the prescribing behavior of doctors in a
certain area would be alike due to local institutions or being of similar types. For
example, local malpractice laws or practice norms may be such that providers in one
area are more likely than those elsewhere to seek information about new
technologies and subsequently prescribe them. Physicians’ preferences that lead
them to locate and practice in a certain area are likely to be correlated. An example
of contextual effects may be that a provider’s prescribing behavior is affected by the
socio-demographics within the area. In my analysis, I incorporate covariates which
control for these different types of effects.
Diffusion and social learning go hand-in-hand in the economics literature.
(Bikhchandani, Hirshleifer, and Welch 1992; Bikhchandani, Hirshleifer, and Welch
1998) define and examine information cascades, where individuals observe past
actions of others and follow suit, quickly converging on one action (for example, to
prescribe or not). From the marketing literature, (Iyengar, Van den Bulte, and
Valente 2011) find evidence of contagion operating through social networks.
(Goolsbee and Klenow 2002) looked at the diffusion of home computers, concluding
that individuals were more likely to purchase their first machine for home use if
they lived in areas where a larger proportion of households already had computers;
the adoption patterns were not well explained by common observed traits or by
area features, but rather stemmed from spillovers from experienced and intensive
users.
Examples of social learning and technology diffusion in the health economics
literature follow. A very early study by (Coleman, Katz, and Menzel 1966) document
13

evidence of contagion in physician prescribing of the antibiotic, tetracycline. After
conducting surveys among 126 physicians, they find that information about a new
innovation is shared within pairs of physicians leading new physicians to adopt the
drug. (Kohler 1997) looks at learning in social networks to evaluate the use of
various contraceptive methods within Korean villages: private information
transmitted between members of a social network yields different utilization
patterns of various contraceptive methods, even within social strata and villages. (E.
R. Berndt, Pindyck, and Azoulay 2003) look at the effect of herd behavior in the
diffusion of anti-ulcer drugs, finding evidence for network effects at the brand and
therapeutic class levels. (Agha and Molitor 2013) study the influence of thought
leaders on the diffusion of novel drug therapies for the treatment of cancer,
concluding that patients were more likely to be treated with the novel treatment if
they sought care in a region in which a thought leader practiced. The results of these
studies suggest that physician social networks may have non-trivial effects on the
adoption of Zostavax.
Data
I measure Zostavax adoption using a twenty-percent random sample of
Medicare beneficiaries and their prescription drug, hospital, physician, and other
medical claims. Medicare Part A covers inpatient and hospice stays. Part B provides
coverage for outpatient services and physician visits, as well as laboratory tests and
durable medical equipment, and a variety of other services and products.
5
Patients
are generally required pay some portion of these costs. The Medicare prescription

5
Per the Medicare Prescription Drug, Improvement, and Modernization Act of 2003,
vaccines for influenza, hepatitis B, and pneumococcal disease are covered under Medicare
Part B and require patients to pay almost nothing out-of-pocket. Vaccines that are
administered as part of treatment also fall under Part B: for example, a patient receiving a
tetanus shot for a dog bite or after stepping on a nail would have the vaccine covered under
Part B. As a commercially available vaccine after January 1, 2004 that is not one of the three
above, Zostavax was covered under Medicare Part D. Patients’ cost burden was thus no
different than for standard prescription drugs (Centers for Disease Control and Prevention
2008). Notably, administration of the vaccine fell under Part B in 2006 and 2007 and was
changed to be covered by Part D starting in 2008.
14

drug benefit, called Medicare Part D, was rolled out in January of 2006. Salient
variables from the drug claims data include National Drug Codes (NDC),
prescription fill dates, physician identifiers,
6
amounts paid by patients, Medicare,
and third parties, states in which prescriptions were filled, and whether the drugs
were generic or branded. I use the following information contained in the Medicare
Annual Beneficiary Summary File to construct annual five-digit ZIP-level covariates:
patient age, sex, race, mortality rates, whether the patient is in traditional fee-for-
service (FFS) Medicare or in Medicare Advantage, whether the beneficiary receives
a low-income subsidy for Part D, and whether the individual is dually eligible for
both Medicare and Medicaid.
From an initial sample of over nine million beneficiaries, I exclude
individuals who were below 65 years old in 2006. Individuals qualify for Medicare
on the basis of age at 65 and the younger group is systematically different – these
younger beneficiaries are either disabled or have end-stage renal disease (ESRD).
However, I include patients over 65 who had previously qualified for Medicare due
to their disability or ESRD. I exclude patients who did not have continuous Medicare
Part D coverage from June 2006 through the end of the data in December 2011 or
until death. This criterion ensures that vaccination is observed and that our study is
not affected by selective switching into and out of Part D plans. While individuals
with Part D at this point were the earliest enrollees and may not be representative
of the elderly population, it is crucial to capture the earliest stages of drug diffusion.
The final patient sample is comprised of an initial cohort of 3,290,456 Medicare
beneficiaries who are at least 65 years old and have continuous Part D coverage
from or until death. Patients who age into Medicare later could have been
immunized prior to entering Medicare, as could individuals who left their Part D
plan and got the vaccine while not on Part D. My sample contains 148,649

6
Physician identifiers in the Part D data are encrypted to prevent linkages with the
Medicare Parts A and B claims, including commonly used physician identifiers like the
National Provider Identifier (NPI) or Unique Physician Identification Number (UPIN).
Encryption is done by the National Council for Prescription Drug Programs (NCPDP) for its
prescriber database, which is used by CMS.
15

physicians who specialize in family medicine, general practice, or internal medicine
and appear in the Part D claims starting in 2006, ensuring that physicians are
actively practicing when it is first rolled-out. Patients are allowed to receive the
vaccine from any type of physician or in any location (for example, inpatient,
pharmacy, or primary care physician’s office), but the social network variable is
limited to primary care physicians. Geriatricians comprise a very small fraction (less
than 2%) of physicians in the Part D claims. They may differ from physicians
specializing in family practice, internal medicine, and general medicine, in that they
are largely reimbursed by Medicare and see a small, selective subset of the elderly
population. Given the lack of geriatricians in the data and that they differ
systematically from these other primary care providers, I omit them from the
physician sample.
7

323,619 patients received Zostavax between 2006 and 2011. After the
selection criteria are applied, 205,076 of those patients remain in the sample. This
noticeable drop is largely due to the strict criterion that patients be enrolled from
2006 and remain continuously enrolled in Part D plans. If patients have
supplementary insurance policies or do not file a claim for Zostavax, I will not
observe their purchase.
8

7
One may think that geriatricians prescribe Zostavax more frequently than primary care
physicians focused on all adults, given their focus on the elderly. The geriatricians in this
data actually prescribe Zostavax less frequently than physicians specializing in family
medicine, general medicine, and internal medicine.
8
The reader may be concerned about whether Zostavax administered in a hospital is
observed within the Part D claims, and thus included in the study. Per CMS guidelines,
medications administered during an inpatient stay that are part of the inpatient treatment
are covered by Part A (Centers for Medicare and Medicaid Services 2015b). The focus of
hospitals and hospitalists is to improve the patient’s health sufficiently so the patient can be
discharged. Given the ambiguity underlying the factors which cause shingles to flare up and
the lack of specificity in risk factors, it is unlikely that patients receive Zostavax as part of
inpatient care. While I observe their prescription drug claims, I do not observe hospital and
physician claims for patients in Medicare Advantage. Of those with traditional fee-for-
service (FFS), very few (5%) were hospitalized (inpatient) during the six-year study period.
In light of all of this information, I do not believe unobserved vaccinations administered in
hospitals to be a concern.
16

I define diffusion as the percent of patients within each quarter and five-digit
ZIP code that received the Zostavax vaccine. Five-digit ZIP codes are the finest
reliable geographic unit available from the claims data; in many parts of the country,
these approximate neighborhoods. ZIP codes are nested within 6,808 primary care
service areas (PCSAs), which were created by Dartmouth University and Virginia
Commonwealth University (Goodman et al. 2003) using Medicare claims data and
validated using private and Medicaid claims. They delineate local markets for
primary care. One challenge that arises when using the Part D claims is that
prescribers’ geographic information is not available below the states in which they
are licensed to practice. I assigned physicians to PCSAs following plurality rules,
wherein a physician was assigned to the region where the largest number of his
patients resided.
I conduct the following analyses at the ZIP code-level with several PCSA-level
variables. Conducting very aggregated analyses, for example at the state-level, is
suboptimal for geographic variation studies (Skinner and Staiger, 2005; Sheiner,
2014): states are incredibly broad and it is difficult to infer true diffusion patterns
without looking at more narrowly defined geographic areas. For example, Orange
County (2006 American Community Survey county-level median household income:
$70,232) and Fresno ($42,732) are both in California but are probably very different
in quality of care: the 2010 National Health Care Disparities Report (Agency for
Healthcare Research and Quality 2011) found that low SES individuals received
poorer care than those of high SES on eighty percent of core measures.
9
Further,
while county size varies widely within the United States, sweeping generalizations
about health outcomes and utilization treating counties as individual units may be
misleading. Even looking within the very large Los Angeles County, Compton is poor

9
The report defines low SES as below the Federal Poverty Line and high SES as greater than
four times the Federal Poverty Line. Core measures look at quality of and access to care.
Examples include use of preventive services and screenings, prescribing of inappropriate
medication, incident AIDS cases, and health insurance status. A full list can be found in the
report.
17

(2006 American Community Survey principal city-level median household income:
$38,682) while Santa Monica ($61,423) is wealthy.
Most maps presented are at the more aggregated hospital service area (HSA)
and hospital referral region (HRR) levels, due to data use restrictions and aesthetics.
HSAs, which define markets for care delivered in hospitals, are similar in size to
counties, with 3,436 in total. HSAs are contained within hospital referral regions
(HRRs), of which there are 306, which define markets for care typically
necessitating services from a major referral center (centers for major cardiovascular
surgical and neurosurgical procedures).
Figures 2-1 through 2-5 display initial diffusion patterns of Zostavax. Figure
2-1 presents the number of Zostavax prescriptions filled between 2006 and 2011.
Sharp dips in 2008 and 2010 reflect supply shortages, which were resolved in
2012.
10
Figure 2-2 shows the number of physicians who have ever prescribed
Zostavax increases linearly over time. This differs from the common pattern of
diffusion of new technologies, including physician prescribing of new drugs
(Coleman, Katz, and Menzel 1966), wherein adoption follows an “S” shape, with a
few early adopters prescribing, then many physicians prescribing -- a contagion
effect – with few stragglers being the last to adopt. This divergence from typical
adoption patterns reflects the low-takeup during the study period – it may be only
the very beginning of the adoption curve that is visible. Figure 2-3 shows the
diffusion of Zostavax at the HRR and HSA levels as line graphs – these show whether
any individual within the region received Zostavax. The HRR-level graph exhibits
the expected “S” shape discussed above: when looking very broadly, it is evident
that some regions and physicians are faster to adopt. Figure 2-4 is comprised of
maps at the HSA level, which highlight the regions containing the earliest adopters.
An area is shaded if at least eleven patients within the HSA received the vaccine

10
The FDA posted that Merck reported vaccine deliveries delayed three to five weeks (Food
and Drug Administration 2014), though other sources reported delays of up to fourteen
weeks (Harvard Pilgrim Health Care Network 2008). There was no evidence that suggested
that certain areas received preferential treatment in receiving expedited deliveries during
this time.
18

within the quarter.
11
Figure 2-5 shows the dispersion of HSA-level vaccination rates
by quarter: while take-up rates remained low during the study period, there is
clearly variation in adoption.
The Medicare Part D benefit design between 2006 and 2011 was unique and
complex. I describe the cost-sharing schedule faced by patients using the 2006
thresholds (CMS Office of the Actuary 2006). Following a $250 deductible, patients
faced a 25% co-insurance rate. Once reaching $2,250 out-of-pocket, individuals
reached a gap in coverage where they again paid all of their drug costs – this is also
called the “doughnut hole.” Once patients’ out-of-pocket costs exceeded $5,100, they
were only responsible for 5% of their prescription drug costs.
12
These thresholds
were adjusted yearly to account for inflation. Starting in 2010, under the Patient
Protection and Affordable Care Act, the doughnut hole began to be phased out, to be
eliminated by 2020. These changes will captured using temporal fixed effects.
Figure 2-6 provides the distribution of amounts paid by patients for Zostavax
between 2006 and 2011. Average HRR-level patient prices are presented in map
form in Figure 2-7. Most plans classified Zostavax as a brand drug, resulting in
patients paying an average of $50 for the vaccine (standard deviation of $41), with
90% paying between $3 and $162. Approximately 10% of plans provided gap
coverage for Zostavax and none treated it as a specialty drug.
I assume that the price of Zostavax is exogenous to a given consumer. The
price faced by an individual for the vaccine depended on their prescription drug
plan and the benefit phase (deductible, pre-gap, coverage gap/doughnut hole,
catastrophic) at the time of vaccination. Further, the price paid also varied based on
how the insurer classified Zostavax – as a brand, generic, or specialty drug. Patients
were required to select their drug plans during the open enrollment period at the
end of each previous year. Since Zostavax was a one-time purchase, it is highly
unlikely that patients selected plans for Zostavax coverage. The generosity of a

11
The CMS data use agreement’s cell suppression policy prohibits reporting of cells of ten or
less.
12
(G. Joyce, Zissimopoulos, and Goldman 2013) includes a particularly helpful graphical
depiction of the Medicare Part D benefit design.
19

chosen plan may be correlated with more generous coverage for Zostavax; given
how few patients knew about Zostavax (Lee et al. 2013), it is also unlikely that
patients selectively timed vaccination to limit out-of-pocket costs. Given all of this, I
assume patients to be price-takers. Appendix Figure A-2 displays prices paid by
patients over time.
Information on the expected amount of patient cost-sharing for Zostavax was
not available until 2010. Insurers may have been hesitant or slow to cover Zostavax,
and it is possible that patient cost-sharing requirements for Zostavax changed over
time between 2006 and 2010. As such, I do not use cost-sharing as a control in my
analyses, as doing so would require making too many assumptions about the wide
variety of insurance benefits. Appendix Figure A-3 contains information on
geographic variation in cost-sharing in 2010.
Information about new drugs may also spread through communication
between patients. Learning about a friend or neighbor’s experience with the painful
ailment is likely to lead an individual to seek prevention. Indeed, the physicians in
the data used by (Coleman, Katz, and Menzel 1966) that were best connected were
also the ones most likely to prescribe new drugs. I identify shingles in the Parts A
and B data using International Classification of Diseases, 9
th
Revision, Clinical
Modification (ICD-9 CM) codes.
13
Then, I divide ZIP-quarter level shingles counts by
the fee-for-service (FFS) population
14
in the quarter and ZIP. Figure 2-8 shows
baseline shingles prevalence rates by HSA in 2005, the year before Zostavax was
approved. One may be concerned that past vaccination rates affect current rates of
shingles prevalence. The data show that vaccine uptake is low and unrelated to
subsequent local shingles prevalence, thus mitigating this concern; the correlation
between the two is extremely low (1.05%).

13
The following ICD-9 codes were used to identify shingles in the claims: 053, 053.1,
053.11, 053.12, 053.13, 053.14, 053.19, 053.2, 053.21, 053.22, 053.29, 05371, 053.79, 053.8,
and 053.9. (Centers for Medicare and Medicaid Services 2015)
14
I do not require continuous coverage in FFS for this variable. This variable provides a
best-approximation of shingles prevalence within an area in a given quarter.
20

Social “distance” may play a non-trivial role in the dissemination of
information about new medical technologies among physicians. Using the claims
data, I construct a measure of physician social networks using patient sharing. I
compute the “adjusted degree” (Landon et al. 2012), the number of physicians with
whom each shares patients divided by the number of shared patients, and use this
as a proxy for physician social networks. Figure 2-9 presents histograms of the
distributions of shared patients, shared physicians, and physician-level adjusted
degrees, wherein it is evident that some patients are more “promiscuous” than
others in seeing multiple primary care providers. On average, a physician shares one
or more patients with 18 other providers. Most pairs of physicians within the data
share one patient. The measure I use in my analyses is the average of adjusted
degrees for physicians within a given PCSA and quarter.
My analyses control for the share of patients in Medicare Advantage (MA)
plans within a ZIP code. MA provides Medicare-eligible individuals with outside
options by private companies to traditional Medicare fee-for-service. Providers
serving MA patients face different financial incentives than those in traditional fee-
for-service to prescribe new technologies. Medicare pays a fixed amount each
month to MA plans for each patient, whereas fee-for-service, as its name suggests, is
reimbursed based on utilization. Thus, MA plans are more likely to encourage
patients to receive preventive therapies like Zostavax. It is also important to control
for the proportion of patients in MA within a ZIP code due to patient self-selection
into MA plans, which may be correlated with unobservable factors and could lead to
biased estimates.
I construct two PCSA-year level variables created from the Part D data to
capture Manski’s contextual effects: the percent of generalists (vs. specialists) and
the patient-to-primary care physician ratio. These are more appropriate at the
PCSA-level instead of ZIP-level, because physicians generally serve areas larger than
ZIP codes. Generalists/primary care physicians were identified using the National
Uniform Claim Committee’s taxonomy codes from the Medicare Part D Prescriber
21

file.
15
The patients-per-primary care physician captures the supply of physicians.
Less competitive areas may have lower vaccination rates due to decreased physician
incentives to maintain knowledge of and prescribe new technologies (Hamilton and
McManus 2005).
The health economics and health services research literature contain much
evidence that both race/ethnicity and socioeconomic status may drive economic
behaviors of elderly Americans with Medicare (Zissimopoulos et al. 2015). Appendix
Figure A-4 shows vaccination rates by income quintile: vaccination rates were
lowest for the poorest quintile, though the take-up does not linearly increase in ZIP
code median household income.
16
Other measures which may capture SES are the
proportion individuals in each ZIP code which are recipients of the Medicare Part D
low-income subsidy (LIS) or those eligible for Medicare and Medicaid (“dual-
eligibles”). The LIS eligibility rules require individuals to earn no more than 135% of
the Federal Poverty Line in income and hold no more than $4,000 (individual) or
$6,000 (couple) in assets (Centers for Medicare and Medicaid Services 2009). Dual-
eligibles may not earn more than 200% of the federal poverty line and may not have
more than three times the Supplemental Security Income resource limit (Centers for
Medicare and Medicaid Services 2014), which was $2,000 for individuals and $3,000
for couples in 2006 (Social Security Administration 2005). The use of a time-varying
variable to capture ZIP-level socioeconomic status would have been ideal; however,
that information was not available. The American Community Survey provides ZIP-
level measures of socioeconomic status, but as five-year estimates. Instead,
assuming minimal changes over time, socioeconomic status will be subsumed by the
ZIP fixed effects.
Table 2-1 presents weighted ZIP-level summary statistics for the initial
period, the second quarter of 2006, in which the large amount of variation across

15
A prescriber in the Part D claims was classified as a generalist/primary care physician if
his Healthcare Provider Taxonomy code was 207Q00000X (family medicine), 207R00000X
(internal medicine), or 208D00000X (general practice). He was otherwise deemed to be a
specialist if the first three characters start with 202 to 209 and the code is not one of the
three aforementioned ones.
16
The income data come from the American Community Survey.
22

ZIP codes for several of the characteristics: the proportion of individuals dually
eligible for Medicare and Medicaid, the percent within the ZIP code that are black,
Hispanic, or enrolled in Medicare Advantage, and the population size.
17
Table 2-2
presents weighted ZIP-level summary statistics for the study population stratified
into two groups for the final quarter of study (fourth quarter of 2011): those with
above-median adoption rates and those below.
18
Areas with lower vaccination rates
appear to be those with more blacks, more Hispanics, and those of lower
socioeconomic status: more dual-eligibles, low-income subsidy recipients.
Individuals in areas with lower take-up were sicker on average (proxied by
hierarchical condition category (HCC) scores
19
).
Estimation
I build my model from Equation 1: take-up of Zostavax within a given quarter
is estimated as a function of adoption to the start of the quarter:
(1) 𝑣 𝑧𝑝𝑡 = 𝛾 𝑉 𝑧𝑝𝑡 −1
+ 𝜉 𝑉 𝑧𝑝𝑡 −1
2
+ 𝛼 𝑧 + 𝜆 𝑡 + 𝘀 𝑧𝑝𝑡
Specifically, 𝑣 𝑧𝑝𝑡 denotes the percent of patients within ZIP code 𝑧 within PCSA 𝑝
and quarter 𝑡 who receive Zostavax. 𝑉 𝑧𝑝𝑡 −1
denotes the cumulative adoption rate at
the start of quarter 𝑡 (or, as of the end of quarter 𝑡 − 1): the percent of patients who
have ever received the vaccine within ZIP code 𝑧 within PCSA 𝑝 by the start of
quarter 𝑡 . 𝜆 𝑡 are quarter-year fixed effects and 𝛼 𝑧 are ZIP-level fixed effects.
ZIP-level fixed effects, 𝛼 𝑧 , control for the presence of local, time-invariant
characteristics, including baseline demographic factors and health literacy, and
information like the presence of teaching hospitals within a region and whether the
ZIP code is rural or urban, assuming minimal changes over the study period. These
specifications using fixed effects require the assumption that there do not exist
time-varying ZIP-level characteristics that are omitted from the analysis. The model

17
Appendix Table A1 compares two groups of patients at the individual level: those who
were immunized and those who were not – on the characteristics available from the data.
18
Unweighted ZIP-level summary statistics, including t-tests between the two groups, can
be found in Appendix Table A2.
19
HCC scores are used for risk-adjustment and are computed using patient demographics
and diagnosis codes. I use them as a proxy for patients’ health status.
23

also assumes that there is no correlation across ZIPs, including those which are
contiguous, except through PCSA-level variables.
Quarterly fixed effects, 𝜆 𝑡 , are included to capture myriad changes between
2006 and 2011 which could affect Zostavax diffusion. The billing of Zostavax
administration was under Medicare Part B in 2006 and 2007, moving to Part D in
2008 (Centers for Disease Control and Prevention 2008), which likely facilitated
vaccination by streamlining the patient and/or provider reimbursement processes.
The rise of private fee-for-service Medicare Advantage (MA) plans and changes to
the Part D benefit design also occurred on the payer side. Zostavax supply shortages
occurred in 2008 and were not fully resolved until 2012. These delayed shipments
of the vaccine up to three months from the order date (Harvard Pilgrim Health Care
Network 2008). These delays should not affect the analysis because the backlog
appears to have been filled per the timing of orders received (Purvis 2011; Harvard
Pilgrim Health Care Network 2008).
The goal of this paper is to better understand the effects of demand- and
supply-side social influences driving Zostavax diffusion. To accomplish this, I
estimate the following enhanced model:
(2) 𝑣 𝑧𝑝𝑡 = 𝛾 𝑉 𝑧𝑝𝑡 −1
+ 𝜉 𝑉 𝑧𝑝𝑡 −1
2
+ 𝜙 𝐹 𝑝𝑡
+ 𝜋 log (𝑝𝑟𝑖𝑐𝑒 𝑧𝑝𝑡 ) + 𝛿 𝑠 ℎ𝑖𝑛𝑔𝑙 𝑒𝑠
𝑧𝑝𝑡 + 𝑿 𝒛𝒑𝒕 𝜷 + 𝝍 𝑮 𝒑𝒕
+ 𝛼 𝑧 + 𝜆 𝑡 + 𝘀 𝑧𝑝𝑡
The coefficients of interest are 𝜙 , 𝜋 , and 𝛿 , which correspond to physician social
networks, prices, and shingles prevalence rates, respectively. 𝐹 𝑝𝑡
is the physician
social network variable: the average number of physicians with whom a given
physician shares patients, adjusted by shared patients within PCSA 𝑝 . 𝑝𝑟𝑖𝑐𝑒 𝑧𝑝𝑡 is the
average price paid for Zostavax within the ZIP code and quarter. For early quarters
during which few patients got Zostavax and ZIP-level prices would otherwise be
missing, I use the quarter-specific average price as a proxy. 𝑠 ℎ𝑖𝑛𝑔𝑙𝑒𝑠 𝑧𝑝𝑡 is the ZIP-
quarter shingles prevalence rate.
I use a number of variables to control for changes over time within ZIP codes
which may be correlated with vaccination. 𝑿 𝒛𝒑𝒕 is a vector of time-varying ZIP-level
characteristics: average age and age-squared, percent female, percent black, percent
24

Hispanic, percent Asian, MA share, mortality rate, average “health” proxied by the
log-transformed HCC scores, percent receiving the low income subsidy, percent
dual-eligibles. Given the fixed cohort of patients for whom I analyze immunization
patterns over six years, inclusion of these variables will help to control for
differential adoption due to selective mortality (by controlling for changes in the
race composition, area health and mortality) and changes in socioeconomic status
(by controlling for the proportion that receive Part D low-income subsidies and the
fraction of Medicare/Medicaid dual-eligibles). While changes over time in these
dimensions are small compared to the cross-sectional variances, they are especially
helpful in mitigating selective attrition. Changes in the provider/reimbursement
environment, which may be related to physician quality and may affect physicians’
information, are captured by the MA share variable over time.
Social network formation is not random (𝐸 [𝘀 𝑧𝑝𝑡 𝐹 𝑝𝑡
] ≠ 0), which means that
estimates of 𝜙 may be biased in the above models. These estimates may be subject
to omitted variable bias because the data do not contain information that allow me
to control for the true factors related to the formation of social networks – for
example, personality types of physicians that would lead them to form groups as
friends or preferred colleagues, and that would be correlated with their willingness
to prescribe new drugs or preventive care. Per (Coleman, Katz, and Menzel 1966),
these are likely to be positively correlated with vaccination, which may lead to 𝜙
being overestimated. Under the assumption that these omitted factors associated
with social network formation are not changing over time within ZIP codes, the
estimates of 𝜙 are unbiased.
20
I use 𝑮 𝒑𝒕
, a vector of time-varying PCSA-level
characteristics to capture supply-side factors: patients-per-doctor and percent

20
One approach to mitigate endogeneity is to use instrumental variables. Using instruments
constructed for region-based reference groups, as in Walker et al. (2011) would have
proven problematic due to the nature of vaccination diffusion. Looking at vaccinations too
far removed spatially would have posed a weak instruments problem. Conversely,
instruments using the average of contiguous regions’ prescribing or inoculation rates would
have been correlated with the endogenous social network variable and would therefore
have violated the exclusion restriction. Given these limitations, the current approach is
more appropriate for this study.
25

generalists (versus specialists), to supplement the measure of physician
interconnectedness and capture different effects (Manski’s contextual effects).
The reported regressions and summary statistics are weighted using ZIP-
level population counts. Tests for serial correlation (Wooldridge 2010) reject the
null hypothesis of no autocorrelation when quarterly adoption rates are regressed
on lagged, cumulative adoption rates (p < 0.000). To mitigate inconsistency in the
standard errors due to serial correlation, I cluster standard errors at the ZIP code
level (Drukker 2003).
Results
Entire Population
Table 2-3 displays the main estimation results, estimated using Equation 2.
Conditional on past adoption, physician social networks are positive and statistically
significant (coefficient estimate: 0.000447). This is one example where statistical
significance does not translate into practical significance: is no evidence that
physician-side social networks affect adoption: the coefficient is positive though not
statistically significant. This means a 1% increase in average physician network size
within an area would be associated with a 0.000447 percentage point (pp) or 0.16%
increase in take-up, which is trivial. Other supply-side factors, percent generalists
(versus specialists) and patients per primary care physician do not affect take-up.
Looking at patient-side shifters, demand is highly insensitive to shingles
prevalence, which may be expected given its especially low quarterly prevalence
rate. The coefficient estimates for shingles prevalence are very similar across all
three models. There is a positive and significant relationship between local shingles
prevalence rates and Zostavax take-up (estimated coefficient: 0.0248), wherein a
0.1pp increase in shingles prevalence (25% increase within a ZIP code and quarter)
leads to a 0.0248pp or 0.90% increase in vaccination.
Changes in Medicare Advantage penetration are positive but statistically
insignificant. Thus, vaccination rates in areas with increased Medicare Advantage
penetration are not significantly different than those where more patients are in
traditional fee-for-service Medicare. While Medicare Health Maintenance
26

Organizations (HMOs) historically provided better access to preventive services
than traditional Medicare FFS, Medicare Advantage now also offers numerous
regional Preferred Provider Organization (PPO) plans in addition to the HMOs.
There is considerable heterogeneity in quality and access among Medicare
Advantage plans (Gold and Casillas 2014). These changes over time in Medicare
Advantage plans may be partially captured by quarter fixed effects. It is possible
that managed care organizations were wary of the benefits of Zostavax, given the
costs, which resulted in less of a push to vaccinate compared to more effective
products like the influenza vaccine. For example, SelectHealth, an insurer in the
intermountain west (Utah and Idaho) region, expressed these concerns and found a
twenty-percent level of patient cost-sharing to be appropriate, given the expected
benefits (Cannon 2007).
Estimates of prices are very similar across models in magnitude and
statistical significance. The coefficient on log prices suggests that patients are price-
sensitive: a 1% increase in price is associated with a 0.0241pp (8.7%) decrease in
immunization. This effect is large. While a recent study found that 65.3% of patients
who did not get Zostavax listed cost as a reason for foregoing the vaccine (Javed et
al. 2012), this estimate seems very large, that an increase in price of several dollars
would be associated with many fewer patients getting vaccinated. Evaluated at the
mean log-transformed price (3.912, which is $50) and mean quantity (0.278%), this
translates to a price elasticity of demand of -3.39, which falls well outside of the
range of demand elasticity estimates for prescription drugs (-0.10 to -0.35) and
preventive services (-0.17 to -0.43) in the empirical literature (Ringel et al. 2002). In
fact, these elasticity estimates suggest patients view Zostavax like a luxury good
(Anderson et al. 1997). (Javed et al. 2012) found that 67.8% of patients reported
that another reason they chose not to receive Zostavax was that they did not believe
they were at risk for shingles, despite 81% previous having chicken pox and 88%
knowing of shingles. If the patients in our data also believed that they were not at
risk of developing shingles, that could explain their unexpectedly elastic demand.

27

Robustness
To understand these diffusion patterns more deeply, I look at two
subsets of the Medicare Part D population: a group that pays almost nothing for
Zostavax and the group for whom the vaccine was found to be most effective.
21

Figure 2-10 compares average yearly ZIP-level vaccination rates for these two
groups with those of the entire population. Take-up is similar among the entire
population and 65-69-year-old group. Adoption is noticeably lower among the
heavily subsidized group.
Cases where physicians recommended or prescribed Zostavax, but the
patient did not fill the prescription and receive the vaccine, are not observed; in the
data, they will appear the same as patients about whom there was no discussion of
inoculation. I use beneficiaries for whom Zostavax is heavily subsidized to gain
insight into the extent of this non-compliance. This subgroup is comprised of
individuals who met one of two criteria: either they receive the most generous Part
D low-income subsidies or they have very generous coverage as a result of being
dually-eligible for Medicare and Medicaid. These groups do not include all LIS or
dual-eligible patients. Almost all patients within this subgroup paid no more than $4
for Zostavax, allowing me to evaluate adoption when the patient contribution is very
low. The average price paid by this group is $3 (standard deviation of $4) and 90%
paid between $0 and $3.30. I use two different sets of covariates with this
population: demographics for the entire ZIP code population and demographics
constructed from the data of the heavily subsidized patients within the ZIP. While
the heavily-subsidized patients pay much less for Zostavax, they are also much
poorer and typically sicker.
Table 2-4 presents the results for this subsidized group using both sets of
explanatory variables. The results are similar in sign and magnitude across the two
models using the two different sets of covariates. In both models, the estimated
coefficient of the log-transformed average price is negative and statistically
significant. Evaluated at the mean price paid for this group ($3.30, or 1.194 log-

21
This subgroup is limited to patients ages 65-69, since I do not observe those ages 60-64.
28

transformed) and mean quantity (0.118%), price elasticity estimates range between
-1.18 and -0.71. Thus, demand of the heavily subsidized patients is less elastic than
their counterparts in the general Medicare population. Physician networks do not
affect immunization rates within this population. As in the analysis for the entire
population, increased shingles prevalence within a ZIP code is associated with
marginally higher take-up.
I also look at the population for whom the vaccine was found to be most
effective: these results can be found in Table 2-5. Patients in this subsample are the
youngest within the data and Zostavax diffused similarly among this group as it did
among all beneficiaries. The effect of shingles prevalence on their take-up is similar
to the results for all elderly patients: the effect is positive and statistically
significant, but very small. While physician social network size was associated with
slightly increased vaccination for the broad elderly population, they do not affect
immunization rates for this population. Patients are less likely to get vaccinated
when faced with higher prices more than the overall older group, possibly because,
being younger, they feel healthier and less susceptible.
One concern may be that shingles prevalence at the ZIP level may be a proxy
for shingles knowledge among patients. A study found that eighty-eight percent of
patients surveyed were aware of shingles (Javed et al. 2012), though patients’
abilities to recognize shingles are not known. Some cases of shingles are resolved
without medical intervention, so the shingles prevalence rates I construct are thus
limited to occurrences where a patient sought medical care. These shingles
prevalence rates are likely to capture more severe cases, and prevalence rates are
likely to be higher in areas where patients have higher health literacy or increased
propensities to use medical care. Patient health literacy is captured by ZIP-level
fixed effects, assuming it does not vary over time within ZIP codes. I test the
relationship between increased propensities to use medical care and diffusion by
controlling for the average number of office visits and average number of
prescriptions filled within each ZIP code and quarter.
29

Patients who see their physicians more often have more opportunities
during which to discuss Zostavax. One may be concerned that such utilization of
medical services is endogenous. This analysis thus requires the assumption that the
propensity of patients within a particular ZIP code to go to the doctor is time-
invariant and thus captured by ZIP fixed effects. I interact the average number of
office visits within a ZIP code and year from the Medicare Beneficiary Summary File
(Cost & Use) with the percent in MA: this is necessary because I do not observe
physician visits for patients who are not in traditional fee-for-service, many of
which are in MA.
Table 2-6 presents results estimated using Equation 2, with the addition of
office visits. Vaccination is increasing in office visits (coefficient estimated as
0.0000888), all else equal (including area health). One additional office visit per
quarter makes a patient 8% more likely to receive Zostavax. As expected, this effect
is more muted for areas with increased MA penetration where I observe fewer office
visits. I find that the effects of shingles prevalence and prices remain unchanged.
The effect of physician size decreases slightly, remaining positive and significant,
confirming that the effect may exist but is very small.

Discussion
This paper examines the factors associated with the diffusion of a new
prescription drug. Using measures of physician social network size derived from the
claims, I do not find evidence that supply-side social influences affect take-up. While
it is possible that physician communication was trivial in the dissemination of
information about Zostavax, the quality of the proxy also merits discussion.
Measures of physician social networks using patient sharing in previous studies
(Barnett et al. 2012; Landon et al. 2012) appear to serve as sufficient proxies for
physician networks by which doctors may learn about new treatments. Nonetheless,
I do not find effects of physician social networks on Zostavax take-up. I believe that
the most plausible explanation is that more data are needed to capture physician
social networks: information networks among primary care physicians serving
30

Medicare patients are not well-captured given the available information. (Barnett et
al. 2011) use a one-hundred percent sample of Medicare Parts A and B claims data,
compared to the twenty percent sample which I use, which allows them to create a
more comprehensive web of networks. It is also possible that this measure does not
do a sufficient job as a proxy for social networks, despite the finding by (Barnett et
al. 2011). S reasons why the study may not generalize follow: eighty-eight percent of
their survey respondents practiced in academic medical centers and the study was
limited to the Boston area. Extending the study to more areas within the United
States would be costly but also could yield interesting insights.
Instead of these measures, more explicit data could be used to map physician
social networks. Within the Medicare claims, if the Part D physician identifiers were
unencrypted, it would be possible to construct better physician networks. Currently,
as a researcher with encrypted Part D physician identifiers, I am prohibited from
doing anything to link these with the National Provider Identifiers (NPIs) and
Unique Physician Identification Numbers (UPINs) the Parts A and B data. Four
suggestions for different physician-side social network measures follow. Using the
Medicare claims, researchers could (1) use inpatient & outpatient files with NPIs
and UPINs to construct geographic network measures representing physician
employment at large hospitals and in large health care systems; (2) construct
similar patient-sharing measures from NPIs or UPINs using the Parts A and B data –
unlike prescription drug data, these might be less confounded, for example, by
physician assistants writing prescriptions; and (3) the claims could be linked with
the costly American Medical Association Masterfile to provide more information
about physicians’ training, which could be related to their propensity to prescribe
and know about new treatments. Finally, data from Doximity, a social media
application for physicians, could be used.
The reader may be concerned that advertising, excluded from the model,
would materially affect the results, due to omitted variable bias. Advertising is time-
varying, and selectively targeted. Companies undertake thorough market research
to determine where and when advertising would be most effective, as well as
31

whether to allocate advertising dollars to direct-to-consumer advertising versus
marketing to physicians. I use Nielsen data on direct-to-consumer advertising for
the study period to evaluate the extent to which advertising is a concern. The data
reveal that no local direct-to-consumer advertising for Zostavax was done; national
direct-to-consumer advertising was done sporadically over the study period. This
suggests that the temporal fixed effects will capture changes over time in
advertising and that patients were not selectively subject to advertising.
Consumption of media leading to differential exposure to advertising may vary
endogenously across regions; that is outside of the scope of this paper.
Physicians may also learn about Zostavax from representatives from
pharmaceutical companies. Data on advertising to physicians are not available prior
to 2013, when CMS began collecting data for their Open Payments database.
Differential advertising to physicians in different areas could potentially explain a
large amount of variation in adoption. Marketing to physicians is concerning, as it is
unobserved, time-varying, and very much non-random: it is not captured by ZIP-
level fixed effects and may cause bias.
Information on new pharmaceuticals may also be disseminated to physicians
from insurers, including guidance on target groups and anticipated benefits. One of
the largest insurers, Anthem (formerly Wellpoint, Inc.), endorsed vaccination of
patients ages sixty and older and began covering Zostavax after the Advisory
Committee on Immunization Practices formally recommended Zostavax for that
group in October of 2006 (Wellpoint 2006). However, other insurers like
SelectHealth were more hesitant, given the vaccine cost and expected benefits
(Cannon 2007).
Lastly, the one-shot nature of Zostavax affects the external validity of these
results to the diffusion of other pharmaceuticals. However, unlike chronic drugs,
adherence is guaranteed. For example, a patient may fill a 30-day prescription for
Lipitor, but I cannot know his/her adherence from the administrative data alone.
The diffusion of new chronic medications also poses an interesting area for study,
and will require careful modeling of adherence and the effects of benefit design.
32

Conclusion
Zostavax provided the elderly with protection from shingles and PHN that is
still not available from any other drug on the market.
22
I find that Zostavax diffusion
was driven by local-area variations. Physician-side social influences are unrelated to
adoption. Patients may weakly respond to higher local shingles prevalence by
seeking prevention, though these effects are very small. To hasten the adoption of
effective, preventive technologies, policymakers should be aware that individuals
are price-sensitive and may be unwilling to pay for treatments that could be
beneficial for two reasons: high out-of-pocket costs and uncertainty about their risk
of shingles. Increased subsidies for Zostavax and improved information on risks
may be helpful. While targeting specific groups (for example, those of low
socioeconomic status and racial/ethnic minorities) may help to increase diffusion,
policymakers should understand that there may exist “high” and “low” adoption
areas, and increased efforts to target areas with low propensities to adopt new
technologies may be necessary.
This study should be extended once Part D claims for 2012 and later are
available to evaluate diffusion in the absence of supply constraints. More generally,
future work in the area of diffusion using Medicare administrative data should
incorporate the Open Payments data from the Centers for Medicare and Medicaid
Service, which was collected starting in 2013, to investigate the effect of physician
compensation on the diffusion of new drugs. Expanding such studies beyond the
Medicare Part D population to investigate spillover effects of physician payments to
younger populations may yield interesting insights. Using unencrypted physician
data, this work could be extended to incorporate physician characteristics and
factors like where the physician was trained to better understand the role of the
physician in diffusion of new technologies, as well as to construct more
comprehensive social network measures.

22
It is notable that GlaxoSmithKline is currently developing a Zostavax competitor.
33

Tables and Figures
Figure 2-1: Number of Zostavax Prescriptions Filled over Time

Note: Data for all Medicare Part D claims, 2006-2011.

34

Figure 2-2: Physicians Who Have Ever Prescribed Zostavax

Note: Data for all Medicare Part D claims, 2006-2011.
35

Figure 2-3: Diffusion of Zostavax Across Hospital Referral Regions and
Hospital Service Areas

Notes: Sample is patients ages 65 and older with continuous Part D coverage from
2006-2011 or 2006 until death. There are 3,436 hospital service areas (HSAs) and
306 hospital referral regions (HRRs).

36

Figure 2-4: Hospital Service Area (HSA) Level Maps Showing Early Adoption of
Zostavax
Fourth quarter of 2006

Fourth quarter of 2008

End of study period, fourth quarter of 2011

Notes: HSAs shown with populations of 11 or greater. A region is shaded darker
blue if at least 11 patients within the region received Zostavax. The sample is
comprised of those 65+ with continuous Medicare Part D coverage between 2006-
2011 or 2006 until death. There are 3,436 hospital service areas.
37

Figure 2-5: Vaccination Rates over Time

Notes: Sample those 65+ with continuous Medicare Part D coverage during 2006-
2011 or 2006 until death.
38

Figure 2-6: Distribution of Zostavax Prices Paid by Patients

Notes: Sample is those 65+ with continuous Medicare Part D coverage during 2006-
2011 or 2006 until death. Values are censored at the 99
th
percentile. N= 205,076.

39

Figure 2-7: Prices Paid for Zostavax by Hospital Referral Region

Notes: The map reflects quartiles of average, HRR-level prices paid for Zostavax by
patients. One HSA containing less than 11 observations was dropped in compliance
with the CMS cell-suppression policy. Sample is those ages 65 with continuous Part
D coverage from 2006-2011 or 2006 until death.

40

Figure 2-8: Baseline Shingles Prevalence

Notes: This map displays baseline shingles prevalence rates at the hospital referral
region level in 2005, the year prior to the approval of Zostavax. Sample is comprised
of patients ages sixty-five and older with Medicare Fee-For-Service (FFS). There are
306 hospital referral regions.

41

Figure 2-9: Physician Social Networks

Notes: Sample is comprised of patients ages 65 and older with continuous Part D
coverage from 2006-2011 or 2006 until death. Physicians included are primary care
physicians who are in the Part D data in 2006. The adjusted physician network
measure is the quotient of the average shared patients divided by the average
number of physicians who share patients.
42

Figure 2-10: Yearly Vaccination Rates by Group

Notes: Sample for all three groups is those 65+ with continuous Medicare Part D
coverage during 2006-2011 or 2006 until death. “All” refers to entire Medicare Part
D population. Heavily subsidized population is comprised of the Medicare/Medicaid
dual-eligibles and Medicare Part D low-income subsidy recipients who have the
lowest cost-sharing. The 65-69-year-old subpopulation is the group for which
Zostavax is purported to be most effective.
43

Table 2-1: Baseline Summary Statistics

Notes: The summary statistics above are weighted by ZIP-level population.
Unweighted summary statistics can be found in the Appendix. Data are for the first
quarter of the study, the second quarter of 2006. The sample is limited to
beneficiaries ages 65 and older with continuous Medicare Part D coverage from
2006-2011 or from 2006 until death. PCP stands for primary care physician.

ZIP-Level Variables Mean (SD)
Quarterly Mortality Rate 2.6% (4.7%)
Part D Low Income Subsidy Recipients 6.2% (5.6%)
Dual-Eligibles 25% (17%)
Average Age 75.4 (2.0)
Medicare Advantage 19% (22%)
Female 62% (9.6%)
Black 8.3% (17%)
Asian 1.7% (5.6%)
Percent Hispanic 4.7% (12%)
Log(HCC Score) -0.027 (0.17)
Population Count 134 (174)
Quarterly Shingles Prevalence Rate 0.5% (0.9%)

Number of ZIP Codes 25,547

PCSA-Level Variables Mean (SD)
Physician Network Size 1.01 (0.11)
Generalists (vs. Specialists) 39% (12%)
Average number of patients per PCP 29.7 (12.5)
44

Table 2-2: Summary statistics by Zostavax adoption rate

Low Adoption High Adoption
ZIP-Level Variables Mean (SD) Mean (SD)
Cumulative Vaccination Rate (2011Q4) 1.9% (1.9%) 12% (7.0%)
Mortality Rate 1.7% (2.4%) 1.6% (2.1%)
Part D Low Income Subsidy Recipients 6.4% (5.8%) 5.3% (5.2%)
Dual-Eligibles 29% (19%) 20% (15%)
Average Age 78.1 (1.6) 78.2 (1.6)
Medicare Advantage 31% (23%) 30% (23%)
Female 63% (9.5%) 63% (8.9%)
Black 12% (21%) 4.8% (11%)
Asian 1.7% (5.4%) 1.9% (6.3%)
Hispanic 7.0% (16%) 3.2% (8.8%)
Log(Average HCC Score) 0.096 (0.14) 0.045 (0.13)
Population Count 100 (134) 110 (133)
Average Zostavax Price $56.40 ($14.50) $58.00 ($22.10)
Shingles Prevalence Rate 0.45% (0.87%) 0.47% (0.84%)

Number of ZIP Codes 12,655 12,647
PCSA-Level Variables
Physician Network Size 0.99 (0.13) 0.98 (0.13)
Generalists (vs. Specialists) 39% (11%) 40% (7.6%)
Average number of patients per PCP 32 (30) 31 (28)

Note: Data are for the fourth quarter of 2011, the final period of the study. These are
weighted by ZIP population counts. Unweighted summary statistics can be found in
Appendix Table A2. Sample is comprised of individuals ages 65 and older with
continuous Medicare Part D coverage starting in 2006. ZIPs with fewer than eleven
beneficiaries at the baseline were dropped. PCP stands for primary care physician.
HCC stands for hierarchical condition category and the HCC score is a proxy for
health status.

The average quarterly vaccination rate was 0.278%.

45

Table 2-3: Explaining adoption (Entire Population)

Physician Network Size 0.000447**

(2.99)
Percent Medicare Advantage 0.000539

(1.77)
Log(Average ZIP Price) -0.000241***

(-26.25)
Shingles Prevalence 0.0248***

(12.25)
Percent Generalists -0.000197

(-0.99)
Patients Per Primary Care Physician -0.00000328

(-1.67)

Lagged Level & Squared Cumulative
Adoption Rate
YES
ZIP Fixed Effects YES
Quarter Fixed Effects YES
ZIP-Quarter Observations 552277
R-Squared 0.078
Number of ZIPs 25511

** indicates significance at the 0.1% level, ** at the 1% level, * at the 5% level. t-
statistics in parentheses.
Notes: Regressions are weighted by ZIP-level population counts. Standard errors are
clustered at the ZIP code level. Sample is limited to beneficiaries ages 65 and older
with continuous Medicare Part D coverage from 2006-2011 or from 2006 until
death. Models include ZIP and quarter fixed effects as noted, percent female, percent
black, percent Asian, percent Hispanic, average age and age-squared, log of HCC
score to proxy health, percent receiving Part D low-income subsidies (LIS), percent
dually eligible for Medicare and Medicaid, and mortality rate.

46

Table 2-4: Explaining adoption (Highly-subsidized population)

Using LIS-Only
Demographics
Using ZIP
Demographics
Physician Network Size -0.000176 -0.000121

(-0.67) (-0.48)
Percent Medicare Advantage -0.00106*** -0.00129**

(-3.50) (-3.09)
Log(Average ZIP Price) -0.000697*** -0.00117***

(-17.09) (-26.58)
Shingles Prevalence 0.0192*** 0.0155***

(3.85) (4.19)

Lagged Level & Squared
Cumulative Adoption Rate
YES YES
ZIP Fixed Effects YES YES
Quarter Fixed Effects YES YES
ZIP-Quarter Observations 327776 327776
R-Squared 0.076 0.077
Number of ZIPs 15950 15950

*** indicates significance at the 0.1% level, ** at the 1% level, * at the 5% level. t-
statistics in parentheses.
Notes: Regressions are weighted by ZIP-level population counts. Standard errors are
clustered at the ZIP code level. Sample is limited to beneficiaries ages 65 and older
with continuous Medicare Part D coverage from 2006-2011 or from 2006 until
death. Models include ZIP and quarter fixed effects as noted, percent female, percent
black, percent Asian, percent Hispanic, average age and age-squared, log of HCC
score to proxy health, percent receiving Part D low-income subsidies (LIS), percent
dually eligible for Medicare and Medicaid, mortality rate, percent generalists (versus
specialists), and patients per primary care physician.
47

Table 2-5: Explaining adoption (65-69 age group)

Physician Network Size 0.000231

(0.82)
Percent Medicare Advantage 0.00138**

(2.89)
Log(Average ZIP Price) -0.000683***

(-19.81)
Shingles Prevalence 0.0232***

(6.21)

Lagged Level & Squared
Cumulative Adoption Rate
YES
ZIP Fixed Effects YES
Quarter Fixed Effects YES
ZIP-Quarter Observations 546500
R-Squared 0.033
Number of ZIPs 25094

** indicates significance at the 0.1% level, ** at the 1% level, * at the 5% level. t-
statistics in parentheses.
Notes: The 65-69 age group is the group within the data for which Zostavax is most
efficient. Regressions are weighted by ZIP-level population counts. Standard errors
are clustered at the ZIP code level. Sample is limited to beneficiaries ages 65 and
older with continuous Medicare Part D coverage from 2006-2011 or from 2006 until
death. Models include ZIP and quarter fixed effects as noted, percent female, percent
black, percent Asian, percent Hispanic, average age and age-squared, log of HCC
score to proxy health, percent receiving Part D low-income subsidies (LIS), percent
dually eligible for Medicare and Medicaid, mortality rate, percent generalists (versus
specialists), and patients per primary care physician.
48

Table 2-6: The relationship between the utilization of medical services and
diffusion

Physician Network Size 0.000418**

(2.79)

Percent Medicare Advantage 0.00337***

(6.58)

Log(Average ZIP Price) -0.000241***

(-26.27)

Shingles Prevalence 0.0248***

(12.29)

Average Number of Office Visits 0.0000888***

(3.40)

Office Visits * Percent MA -0.000371***

(-7.09)

Lagged Level & Squared Cumulative
Adoption Rate YES
ZIP Fixed Effects YES
Quarter Fixed Effects YES
ZIP-Quarter Observations 552277
R-Squared 0.078
Number of ZIPs 25511

** indicates significance at the 0.1% level, ** at the 1% level, * at the 5% level. t-
statistics in parentheses.
Notes: Regressions are weighted by ZIP-level population counts. Standard errors are
clustered at the ZIP code level. Sample is limited to beneficiaries ages 65 and older
with continuous Medicare Part D coverage from 2006-2011 or from 2006 until
death. Models include ZIP and quarter fixed effects as noted, percent female, percent
black, percent Asian, percent Hispanic, average age and age-squared, log of HCC
score to proxy health, percent receiving Part D low-income subsidies (LIS), percent
dually eligible for Medicare and Medicaid, mortality rate, percent generalists (versus
specialists), and patients per primary care physician.
49

CHAPTER 3. “‘Healthy, wealthy, and wise?’ The effects of
negative wealth shocks on health and health care
utilization”
23

Introduction
The health-wealth gradient, wherein the affluent are healthier than those
with fewer financial resources, has been well-established in the literature (Deaton
2002; Smith 1999). However, the direction of causality is less clear: does the
direction of causality run from health to wealth or from wealth to health?
Alternatively, are there factors which are driving both wealth and health? Prior
work using a number of different methods, for example, instrumental variables
estimation and dynamic panel data models, has found mixed results in the direction
of causality of the gradient. Reverse causality makes this question particularly tricky
to answer. In looking at the causal effect of wealth on health, the wealthy are likely
to be able to afford better care and will thus be healthier. However, at the same time,
individuals that are healthier are likely to be able to work longer, which may lead to
increased wealth as well as better health.
Our contribution to this topic is to use the recent housing crisis as a natural
experiment for evaluating the causal relationship between health and
socioeconomic status (SES). We investigate whether health outcomes, such as
chronic conditions and mortality, changed due to large, rapid changes in home
prices. We also look at possible mechanisms by which wealth may affect health: the
use of medical services like hospitalizations, office visits, prescription drug use, and
preventive care.
How might home prices affect health? According to the 2010 Health and
Retirement Survey, 80% of the near-old (55-64 year olds) and elderly (65 and
older) own their homes. We focus on the elderly (those at least 65 years old). At the
time of the housing crisis, they were not directly affected by ongoing changes in
home prices unless they had been planning to sell. However, approximately 50% of

23
This paper was co-authored with Dana Goldman, Florian Heiss, Daniel McFadden, Joachim
Winter, and Amelie Wuppermann.
50

the average older American’s wealth was in the form of housing (Engelhardt,
Eriksen, and Greenhalgh-Stanley 2013) so that the observed drop in house prices
had large effects on the elderly’s wealth possibly causing stress. Elderly Americans
may also have been affected by stress arising from observing family, friends, and
neighbors experiencing hardship or stress arising from the uncertainty of markets
and the value of their home. (Seeman 1997) evaluates the effects of allostatic load –
defined as “the strain on the body produced by repeated ups and downs of
physiologic response, as well as by the elevated activity of physiologic systems
under challenge, and changes in metabolism… that can predispose the organism to
disease” – on physical and cognitive functioning, finding that individuals with
heightened allostatic loads faced elevated risks of cognitive and physical decline and
cardiovascular disease. If the housing crisis subjected individuals to excessive
stress, we may observe negative changes in health outcomes.
Grossman's model of the demand for health (1972) suggests two ways in
which the housing crisis could have affected health outcomes. First, a home price
shock is stressful, and increased stress levels lead one’s health stock to depreciate
more rapidly. Second, a direct effect on the budget constraint means fewer
resources available to invest in the production of health. However, the empirical
literature on the effect of wealth on health presents mixed results. The Whitehall
studies I and II (Marmot, Shipley, and Rose 1984; Marmot et al. 1991) were early
and incredibly thorough in collecting information on British civil servants that was
used to document the health-wealth gradient. (Marmot, Shipley, and Rose 1984)
documented the positive relationship between health and employment status
among British civil servants, while a subsequent study (Marmot et al. 1991) found
that economic factors affect health through psycho-social mechanisms, such as
stress due to work, which induce undesirable health behaviors both directly and
indirectly.
(Smith 1999) purports that households may respond to new health events by
reducing planned bequests to heirs, instead of decreasing current or future
consumption. He documents that wealth and income have statistically significant
51

positive effects on (self-reported) health, though the effects are diminished by about
33% when behavioral risk factors are added to the model (for example, smoking,
heavy alcohol consumption, physical activity, body mass index). (P. Adams et al.
2003) use tests of Granger non-causality and do not find causal evidence that SES
affects acute-onset conditions or mortality. They mention that the effect of wealth
on chronic and mental health conditions remains unresolved. (Stowasser et al.
2011) replicate the study on different data: the Health and Retirement Study (HRS),
which contains the AHEAD data used by (P. Adams et al. 2003). They find that they
cannot reject the effects of SES on health for quite a few conditions, and note that
causal inference is sensitive to the cohorts and time periods used.
(Meer, Miller, and Rosen 2003) use the Panel Survey of Income Dynamics
(PSID) to evaluate the effect of changes in wealth on one’s health, with inheritances
as instruments. Their measure of health is based on self-reports on a five-point
scale. Their primary finding is that short run changes in wealth do not drive changes
in health. (Michaud and van Soest 2008) look among a slightly younger population
and do not find evidence that wealth affects health. Using dynamic panel data
methods and exploiting variation within-couples, they do find evidence that health
affects wealth. Evidence from the United Kingdom using housing gains to evaluate
wealth effects on health finds (Fichera and Gathergood 2013) that increases in
housing wealth are associated with lower likelihoods of acute conditions among
owners, while, as expected, there was no effect among renters.
There is an extensive literature that looks at changes in health outcomes in
response to economic fluctuations that informs our work. (Ruhm 2000) found that
health is countercyclical: namely, that a negative relationship exists between the
mortality rate and state unemployment rates. Using Behavioral Risk Factor
Surveillance Survey (BRFSS) microdata, he conjectures this is due to the worsening
of unhealthy behaviors during periods of higher unemployment. (Schwandt 2011)
focuses on wealthier, elderly Americans. He constructs wealth measures of stock-
holding from the HRS, interacting stock holding with changes in the stock market.
He looks at the incidence of new chronic conditions, finding the strongest effects for
52

high blood pressure and moderate effects for cardiovascular problems. However, he
finds no effect of negative wealth shocks on the incidence of cancers, as well as
diabetes, arthritis, and lung diseases.
Two recent papers have looked at changes in the utilization of medical
services during recent crises. (Currie and Tekin 2015) investigate the relationship
between hospitalizations and emergency department (ED) visits and foreclosures,
using data from the Panel Survey of Income Dynamics (PSID), Zillow (home prices),
RealtyTrac (foreclosures), administrative medical data, and the American
Community Survey. Their analysis is done at the ZIP-code level, focusing on the
effects felt in the four hardest-hit states: Arizona, California, Florida, and New Jersey.
Part of their analysis looks at the elderly: they find that increases in foreclosures
were associated with increases in emergency room and hospital visits for a number
of conditions, including heart attack and hypertension. Their analysis that focuses
specifically on the effect of foreclosure on the elderly finds larger effects than for
younger populations, among non-elective hospitalizations, preventable
hospitalizations, and cancer, cardiovascular, and respiratory-related
hospitalizations. They attribute this to the generally more frail state of elderly
patients and their heightened vulnerability to health shocks.
While many of the existing studies have looked primarily at the effect of the
crisis on large groups, we focus specifically on the elderly, those sixty-five and older.
The elderly are sicker than the younger population and are more susceptible to
negative health shocks (Currie and Tekin 2015). Our results will not necessarily
generalize to the younger population for a few reasons. The elderly are largely
retired, while younger groups remain in the workforce. This means that fluctuations
in employment due to the financial crisis did not affect them as much. However,
(McInerney and Mellor 2012) use the Medicare Current Beneficiary Survey and find
that self-reported health among the elderly is worse and they utilize more inpatient
care when unemployment increases, in contrast with studies looking at the general
population which have found the opposite. They also conclude that mortality among
the elderly is countercyclical between 1994 and 2008.
53

Our study builds on this literature to investigate the effects of economic
distress on health outcomes and utilization among elderly Americans. We use
hospital, physician, and prescription drug claims data for a random twenty-percent
sample of Medicare beneficiaries in combination with home price data from Zillow
(Zillow 2014). Our study period begins in 2006 and lasts through 2011. An
advantage over previous studies is that the claims data are at the individual level
and do not rely on self-reports of physical health. The data contain detailed
diagnosis and procedure information, as well as information on insurance coverage
and prices paid for medical care. Home prices are available at the five-digit ZIP code
level.
To evaluate the effects of a negative wealth shock on health outcomes, we
look at ZIP code-level health outcomes and the mechanisms by which health may be
affected, including the utilization of preventive care and medical services. We focus
on the Medicare fee-for-service (FFS) population. Our analyses are comprised of
panel regressions at the individual and quarter or year level using panel linear
probability models with ZIP and year fixed effects. We find that negative wealth
shocks do not affect mortality. Analyses of chronic conditions and housing and
distress variables show that the prevalence of most conditions does not respond to
negative wealth shocks. The prevalence of ischemic heart disease and rheumatoid
arthritis/osteoarthritis decrease when prices decrease, while the prevalence of
diabetes and that of hip fractures increase in negative wealth shocks. Individuals in
areas of home prices declines are more likely to receive cancer and diabetes
screenings and wellness exams, but less likely to receive lipid screenings. Take-up of
flu shots is unrelated to changes in home prices, as are osteoporosis screenings and
depression screenings. Health care expenditures and other measures of utilization
tend to rise when home prices fall.
Data
We use a random sample of twenty percent of Medicare beneficiaries.
Medicare provides health insurance to over forty-five million elderly and disabled
individuals in the United States. The dataset links enrollment and Parts A and B
54

claims (2002-2012) for traditional fee-for-service Medicare enrollees to Part D
prescription drug claims (2006-2012). The Part A data include information about
inpatient hospital stays, including length of stay, diagnosis-related group (DRG),
department-specific charges, and up to ten individual procedure and diagnostic
codes. Part B information includes claims submitted by physicians, and other health
care providers and facilities for services reimbursed by Part B. Each claim contains
diagnostic (ICD-9-CM) and procedure (CPT-4) codes, dates of service, demographic
information on beneficiaries, and a physician identification number.
The enrollment file contains demographic information about each
beneficiary including date of birth, date of death, gender, beneficiary type (e.g.,
recipient of the low-income subsidy), and ZIP code of residence. The Medicare data
also include externally validated measures of race/ethnicity. Self-reported
measures on race/ethnicity are refined using Research Triangle Institute estimates
based on geography and first and last names.
Our main analysis is limited to patients ages sixty-seven and older with
continuous traditional fee-for-service (FFS) coverage. The FFS requirement is due to
the fact that we are unable to observe physician and hospital claims for individuals
in Medicare Advantage. While individuals become eligible for Medicare on the basis
of age at sixty-five, using the higher age threshold is necessary for our analysis of
health outcomes, due to the required look-back period for certain conditions.
24
Since
the health data for patients in Medicare Advantage are not available; patients must
remain in FFS Medicare for the duration of the study period, or until death.
To analyze mortality, we construct an indicator for whether or not the
beneficiary died between 2007 and 2011. The exact date of death is included in the
denominator file (2006-2008) and beneficiary summary files (2009-2011). We look
at the ZIP-level prevalence of a number of chronic conditions which may be affected

24
Some chronic conditions have longer look-back windows than others to ensure proper
measurement of disease. Ischemic heart disease, rheumatoid arthritis and osteoarthritis,
and diabetes have two-year look-back periods. Hip fractures, strokes/transient ischemic
attacks, hyperlipidemia, hypertension, chronic heart failure, acute myocardial infarction,
and lung, endometrial, breast, prostate cancer have one-year lookback periods.
55

by rising stress levels (Currie and Tekin 2015) – hypertension, acute myocardial
infarction, stroke/transient ischemic attack, hyperlipidemia, chronic heart failure.
We also use rheumatoid arthritis/osteoarthritis and a number of cancers
(individually: lung, endometrial, breast, and prostate) as placebos, as prevalence of
either should not respond to changes in home prices (Ruhm 2000).
To analyze morbidity, chronic conditions indicators constructed by the
Centers for Medicare and Medicaid Services (CMS) Chronic Conditions Warehouse
(CCW) are used. Diagnosis codes from hospitals, and home health and skilled
nursing facilities are used to construct yearly indicators, as well as a variable for
whether a beneficiary was ever diagnosed with the particular condition (Chronic
Conditions Data Warehouse 2014).
25
We use the yearly variable to construct ZIP-
level prevalence rates.
How large is the role of the patient in the use of medical services? If it is small
(Chandra, Cutler, and Song 2011) and patients do not use medical services
excessively, we should expect to see minimal changes when patients are faced with
a sizable wealth shock. (Finkelstein, Gentzkow, and Williams 2014) exploit moving
patterns among the elderly to find that 40-50% of geographic variation in health
care expenditures is driven by factors on the demand-side, largely variation in the
health of patients. 50-60% of geographic variation is supply-side driven. Thus, it
seems plausible that patients may respond to wealth shocks by altering their use of
medical care. In the absence of health shocks, patients may decrease their use of
medical services. However, if acute or chronic stress due to negative wealth shocks
leads to adverse health events, patients may need to use more medical services. This
may put patients in a sub-optimal financial situation. While Medicare helps to shield
patients from the full financial burden of a decline in health, beneficiaries are
generally required to bear some amount of cost-sharing. Co-pays and co-insurance

25
One may be concerned that identifying the chronically ill from claims data may be subject
to measurement error, particularly the presence of false positives, if “rule-out” diagnoses
are recorded. CCW addresses this potential problem by using stricter criteria in identifying
chronic conditions: multiple hospitalizations or doctor visits are generally required as is
appropriate for each condition.
56

can be non-trivial: in 2010, average out-of-pocket spending for patients who
reported being in poor health was $4,505, compared with $1,774 for patients who
reported being in excellent health (Cubanski et al. 2014). Increasing health costs due
to worsening health will only decrease already-stressed patients’ quality of life. We
use a number of different measures to investigate: the most aggregated measures
are of Parts A and B spending. To evaluate changes in the utilization of medical
services, we look at inpatient stays, inpatient days, outpatient visits, emergency
department (ED) visits, and office visits, which come from the Beneficiary Summary
File of Costs and Utilization.
Understanding if patients change their use of preventive services when
experiencing a wealth shock is crucial to understanding subsequent changes in
health; increasing the use of preventive health care services has been featured as a
way to improve Americans’ health and decrease spending. The 2010 Patient
Protection and Affordable Care Act (ACA) includes the National Prevention Strategy,
a roadmap for improving the health in the United States for people of all ages by
expanding access to preventive services, the elimination of disparities, and
encouraging patients to live more healthful lives. During the study period, patients
received free or highly subsidized preventive services. We focus on wellness exams,
lipid tests, diabetes tests, osteoporosis screenings, and cancer screenings. Appendix
Table 1 contains a list of procedure codes used to identify each of these preventive
services.
We use Zillow data for home prices, which are available at the ZIP-code level.
The Zillow Home Value Index (ZHVI) is the median of all Zillow “Zestimates” within
a region (here, ZIP code).
26
Annual ZIP averages of home prices are used. Figure 1
presents the median and inter-quartile range of yearly home prices over the period

26
How does the ZHVI compare to the Case-Schiller index? “The Case-Shiller index is
computed using a repeat sales methodology, which measures price change by collecting
data on homes that have been resold in a given region. Case-Shiller limits itself to homes
that have sold at least two times recently; it also excludes all new construction. Because
homes across regions may appreciate differently and those segments are not represented
proportionally in the sample limited to repeat sales, the index may suffer bias, which is
particularly acute for smaller geographic regions with lower turnover.” (Zillow 2013)
57

from 2006 to 2011, as well as a plot which groups together sets of areas
experiencing similar patterns of home price changes. The Zillow data, when merged
with the Medicare and sociodemographic data, cover about ten thousand ZIP codes.
At the early part of the study period, the mean of ZIP-level average of home prices
was $310,676 compared with $241,857 in 2011.
We control for one’s health status using the Hierarchical Condition Category
(HCC) scores. These were implemented in 2004 and were intended to be risk-
adjustment measures; these scores help to determine payments for Medicare
Advantage (MA) plans. Such plans are reimbursed a fixed amount per beneficiary.
To mitigate the incentive for MA plans to enroll the healthiest individuals, CMS
developed risk-adjustment measures that compensate plans based on enrollees’
health. (Pope et al. 2000) We use CMS’s 2010 definition of HCCs as a proxy for
health status. HCCs use diagnosis codes to determine comorbidities and health
status, using diagnostic cost groups. Summary statistics for our sample of those ages
65 or older in Part A, and alive in 2006, are displayed in Table 1.
Estimation
To evaluate whether changes in health occur when home prices change, we
undertake an intent-to-treat analysis, estimating ZIP-year level linear probability
models of mortality, chronic conditions, hospitalizations, costs, and utilization of
medical services, including preventive care on lagged home prices.
𝑦 𝑧𝑡
= 𝛾 𝑝𝑟𝑖𝑐𝑒 𝑧𝑡 −1
+ 𝑿 𝒛𝒕
𝚩 + 𝜆 𝑧 + 𝛿 𝑡 + 𝘀 𝑧𝑡

𝑦 𝑧𝑡
is the outcome variable: ZIP-year mortality rate, ZIP-year utilization of medical
and preventive services, etc. We do not observe individuals’ home prices in the
claims data and instead use ZIP-code level home prices as a proxy. 𝑿 𝒛𝒕
is a vector of
ZIP-level characteristics, including age, sex, race, the share of individuals within a
ZIP code that are in Medicare Advantage plans, and average logged HCC score. 𝜆 𝑧 are
ZIP-level fixed effects and 𝛿 𝑡 are year fixed-effects.
Several factors lead us to decide that the year level was most appropriate for
this analysis. The chronic conditions and utilization measures from the Beneficiary
Summary File are only available at the year-level. Patients are likely to use
58

preventive screenings and flu shots only once per year. Given year-to-year changes
in the Medicare benefit design, especially in 2010 and 2011 with the start of the
implementation of the ACA, year fixed effects are especially appropriate. Given that
a year is a non-trivial amount of time, home prices are lagged so that our results do
not reflect the relationship between future home prices and health and health care
use. We will refer to these simply as “home prices” for the remainder of the paper.
Moreover, it is likely that the effects of wealth shocks are not immediate:
particularly given that the housing crisis was a historical anomaly; the health effects
of chronic stress do not manifest immediately (McEwen 1998). Lags of two or more
years may be necessary to capture the true effects of wealth on mortality and
morbidity. Such results are not presented. With the currently available claims data,
which run through 2011, using lags of two or more years would not fully capture the
worst of the housing crisis and its effects on the elderly.
Identification is achieved by assuming that changes in housing prices across
ZIP codes are exogenous. The housing crash was unanticipated, and more
importantly, the extent of the crash across areas was also impossible to foresee. Our
key assumption is changes in price are exogenous (and unanticipated), though price
levels may not be. This model assumes that there is no ZIP-specific time-variant
unobserved heterogeneity after controlling for the observable time-variant
characteristics. The ZIP fixed effects capture the baseline prevalence of chronic
conditions in the area. Another necessary assumption that we find plausible is that
other SES-related factors outside of home prices are not changing during this period
in a way that affects health and the utilization of health care services. The ZIP fixed-
effects will capture SES, as well as factors like whether the ZIP code is in a rural or
urban area, assuming minimal changes over time occur within the neighborhood.
Regressions are weighted using ZIP-level population weights.
One concern may be that changes in home prices are correlated with factors
that determine home price levels: for example, homes in areas with better school
districts may retain value better through the housing market crash and also be
worth more. However, we observe that the correlation between the percent change
59

in home prices between 2006 and 2011 versus the home price levels in 2006 is 4%.
We therefore assume that changes in home prices as a result of the recession were
unrelated to past health and past SES.
Results
Table 2 presents results of analyses of the effects of home prices on mortality
and chronic conditions. Lagged home prices are not associated with any changes in
mortality. Overall, we find mixed changes in morbidity associated with changes in
home prices. The prevalences of depression, ischemic heart disease and rheumatoid
arthritis/osteoarthritis (RA/OA) are decreasing when prices decrease: one standard
deviation decrease in home prices decreases the prevalence of depression by 0.2
percentage points (pp) (or 1.7%), ischemic heart disease by 0.7pp (or 1.8%), and
the prevalence of RA/OA by 0.5 pp (or 1.6%). These are small effects, which may be
evidence of “harvesting” (Ziebarth, Schmitt, and Karlsson 2013), wherein
hospitalizations, worsening conditions, or death of those in weak health are
hastened by the increased distress. On the other hand, the prevalence of diabetes
and that of prostate cancer are increasing in negative wealth shocks. One standard
deviation decrease in home prices is associated with a 0.6pp increase in the ZIP
prevalence of diabetes (or 2.1%) and a 0.1pp increase in prostate cancer (2.3%).
The increases in diabetes and prostate cancer prevalence may be attributable to
increased surveillance – we discuss changes in preventive services use after a
negative wealth shock below. We find no relationship between the wealth shocks
and prevalence of the following chronic conditions: hip fractures, stroke and
transient ischemic attacks, hyperlipidemia, hypertension, chronic heart failure,
acute myocardial infarction, lung, endometrial, and breast cancer.
Table 3 shows the effects of home prices on Parts A and B expenditures. At
the aggregate level, we find evidence that patients may slightly increase their use of
medical services upon a negative wealth shock. A decrease in home prices of one
standard deviation ($261,178) leads to a 2.4% increase in Part A spending and a
2.2% increase in Part B spending. This is consistent with the theory that negative
wealth cause stress, which may lead to increased utilization of health care because
60

patients feel worse. It is possible that the prevalence of serious conditions may not
be affected by wealth shocks, but patients nonetheless feel a general malaise, short
of depression but that still induces use of medical care.
The effects of wealth shocks on the utilization of medical services are
displayed in Table 4. One standard deviation decrease in home prices in the prior
year is associated with a 0.057 increase in inpatient days (or 2.7%), increase of
0.016 inpatient stays (or 4%), and 0.36 increase in office visits (4.7%). However,
outpatient visits appear to decrease when home prices drop: one standard deviation
decrease in home prices is associated with a decrease of 0.45 (7.6%) visits per
patient and year on average. There do not appear to be changes in emergency
department visits after patients experience a negative wealth shock. (Currie and
Tekin 2015) combine emergency department visits with inpatient visits because, at
the margin, socioeconomic and demographic characteristics and the admission of a
patient to the hospital may be correlated. Consistent with their analysis, we combine
these two measures (“ED + IP”) and find that a decrease of one standard deviation in
ZIP-level home prices is associated with an increase of 0.017 (1.7%) hospitalizations
per patient, per year.
Table 5 presents the estimated effects of home prices on the use of
preventive services. These effects were also mixed. Increasing when home prices
decrease are cancer screenings (the following are in terms of a standard deviation
decrease in home prices: 0.75pp or an increase in use by patients within a ZIP code
of 1.7%), diabetes screenings (1.6pp or 5.1%), and wellness exams (0.67pp or
11.8%). However, lipid tests decrease when home prices decrease (0.48pp or 0.8%).
The effect of negative wealth shocks on the use of depression screenings and
osteoporosis screenings are weakly statistically significant. Upon a standard
deviation decrease in home prices, osteoporosis screenings increase 0.18pp (or
0.02%) and depression screenings increase 0.009pp (or 123%), on average. There is
also no effect of wealth shocks on flu shots.

61

Discussion
The findings presented above are mixed. We find that patients increase their
use of medical services and select preventive services, while health itself is
minimally affected when house prices drop. One may question why increases in the
use of medical services do not lead to observed health (through more diagnoses)
worsening. A plausible explanation may be that, since these patients are older, the
chronic conditions to which they may be pre-disposed or which were acquired due
to poor health behaviors have already been realized. Patients’ minimal changes in
the use of medical care suggest that frivolous use of health services, at least from the
beneficiary perspective, was minimal.
It is possible that ZIP-level home prices may not serve as sufficient proxies
for wealth shocks. Other measures of housing distress besides home prices may
better capture the extent to which individuals were affected by the crisis. (Currie
and Tekin 2015) find that foreclosure rates, and fluctuating home prices to a lesser
extent, affect health, though they do not present results limited to the elderly
population using home prices. We test this within our data using housing distress
rates constructed from history and assessor data from DataQuick. The key variable
within those data is a flag assigned one of eight values. The values indicate that the
transaction was part of a sequence of distress events in the foreclosure process. The
foreclosure rates used by Currie and Tekin capture an earlier part of the foreclosure
process – notice of default and notice of transfer – than available in this variable. We
construct the values of our distress variable to represent two phases of the process:
the home reached (1) the pre-auction/auction or (2) the real-estate owned (REO)
phase if it was not successfully sold in an auction or prior.
Figure 2 presents the 25
th
and 75
th
percentiles and median of the pre-
auction/auction distress rate over time. Using information on the percent of homes
in the pre-auction or auction phases to construct distress rates for analyses, overall,
we find no evidence that this level of housing distress affected patients’ health
(mortality or morbidity) or their utilization of medical care. Table 6 presents the
statistically significant results when distress rates are added to the previously
62

estimated models. Increases in distress rates are related to decreases in the local
prevalence of hyperlipidemia (0.003%) and ischemic heart disease (0.002%), and
increases in heart attacks (AMI) (1.4%). Increases in distress rates are associated
with decreases in cancer screenings (0.002%). With the exceptions of AMIs, the
effect of distress rates on both of these outcomes is too small to be of any practical
significance. Effects of negative wealth shocks increase in magnitude by about 33%
for hyperlipidemia and heart attacks but remain statistically insignificant. The
coefficients of home prices on cancer screenings and ischemic heart disease remain
similar to those estimated using models with only home prices.
When the above analyses using home prices and both home prices and
distress rates are performed on the four states used by Currie and Tekin, the results
are not strengthened: the effects of home prices are generally similar in magnitude,
and some become statistically insignificant. Distress rates are not associated with
any outcome except hyperlipidemia prevalence, which is decreasing when distress
rates increase. These results differ quite starkly from the findings of Currie and
Tekin. First, as noted by Currie and Tekin, home price shocks may only minimally
affect health. It is also possible that the foreclosure rates used by Currie and Tekin
which capture the earliest signals of foreclosure better capture housing distress
than our distress rates. Second, Currie and Tekin look at the effects of foreclosure on
the entire adult population, whereas this study focuses on the elderly. Since many
seniors may not plan to move or need to access their housing wealth, and also are
largely retired, they may be differentially affected — or not—by such fluctuations.
This appears to be conclusive evidence that short-term negative wealth shocks do
not affect health, though they may change patients’ use of health care services.
Conclusions
This paper uses plausibly exogenous variation in changes in home prices due
to the housing crisis of the late 2000s as a natural experiment for evaluating
mechanisms underlying the health-wealth gradient. We find that patients increase
their use of medical services and select preventive services, while health itself is
minimally affected when house prices drop. This indicates that health may not be
63

directly affected by wealth, as also found by (Meer, Miller, and Rosen 2003; Michaud
and van Soest 2008). Instead, changes in wealth lead to suboptimal changes in the
utilization of health care services which may affect health in the longer term.
Future work should continue this analysis with a longer study period
capturing the latter part of the Great Recession and the subsequent recovery. Use of
the Health and Retirement Study-linked Medicare claims should be used to evaluate
the effects of individuals’ home prices on their health, as opposed to neighborhood
home prices.

64

Tables and Figures

Figure 3-1: Home Prices over Time

Notes: Data on ZIP-level home prices come from Zillow’s monthly home value index
(ZHVI). ZHVI_SFR stands for ZHVI – single family residence. The bottom plot groups
together ZIPs experiencing similar changes in home prices.
65

Table 3-1: ZIP-Year Level Descriptive Statistics (continued onto next page)

VARIABLES Mean Std Dev

Demographics
Average Age 79.0 1.6
Percent Female 58.9 6.5
Percent Black 6.9 15.3
Percent Asian 2.5 7.1
Percent Hispanic 4.52 10.1
Percent Other Races 0.8 1.8
Percent White 85.2 21.1
Percent LIS 2.4 2.5
Percent Dual-Eligibles 12.4 12.5
Logged HCC Score 0.27 0.15
Medicare Advantage share 21.7 15.0
ZIP-level Population (in Sample) 198 181

Take-Up of Preventive Tests & Services

Percent Getting Cancer Screenings 45 8.5
Percent Getting Lipid Tests 62 8.9
Percent Getting Diabetes Tests 31 8.4
Percent Getting Flu Shots 51 12
Percent Getting Osteoporosis Screenings 9.0 3.9
Percent Getting Depression Screenings 0.0073 0.18
Percent Getting Wellness Exams 5.7 6.5

Annual Part A Costs $8,161 $2,916
Annual Part B Costs $4,257 $1,150

Utilization of Medical Services
Hospital Days 2.15 0.88
Office Visits 7.73 1.71
Outpatient Visits 5.94 2.47
Hospital Stays 0.40 0.12
Emergency Room Visits 0.60 0.17

Annual Mean Home Price $271,347 $252,976
Distress Rate (Pre-Auction/Auction) 0.0046 0.013

Notes: The sample is comprised of patients ages 67 and older with continuous fee-
for-service coverage starting in 2006 through the end of 2011 or until death. There
are 49,799 ZIP-year level observations (10,257 ZIPs). Annual mean home price is
the yearly average of the ZIP-level Zillow Home Value Index (ZHVI).
66

Table 3-1: ZIP-Year Level Descriptive Statistics (continued from previous
page)

VARIABLES Mean Std Dev

Yearly Mortality Rate 7.0 3.2

Prevalence of Chronic Conditions
Hip Fractures 1.3 1.3
Depression 12 4.8
Stroke/Transient Ischemic Attack 5.5 2.8
Hyperlipidemia 48 10
Hypertension 64 9.3
Diabetes 28 8.0
Chronic Heart Failure 21 6.4
Ischemic Heart Disease 38 9.4
Acute Myocardial Infarction 1.2 1.3
Rheumatoid Arthritis/Osteoarthritis 32 7.7
Lung Cancer 1.3 1.3
Endometrial Cancer 0.3 0.6
Colorectal Cancer 1.7 1.5
Breast Cancer 3.4 2.1
Prostate Cancer 4.4 2.5

Notes: The sample is comprised of patients ages 67 and older with continuous fee-
for-service coverage starting in 2006 through the end of 2011 or until death. There
are 49,799 ZIP-year level observations (10,257 ZIPs).

67

Table 3-2: Mortality & Health Outcomes

Mortality Prostate Cancer
1-Period Lagged Home Prices 2.43e-09 -4.36e-09**

(1.11) (-3.02)

R-Squared 0.0965 0.0646
SD ($261,178) * home price coef 0.00148 -0.00114

Depression RA/OA Diabetes
Ischemic
Heart
Disease
1-Period Lagged 7.13e-09* 1.76e-08*** -2.10e-08*** 2.49e-08***
Home Prices (2.44) (3.54) (-4.16) (4.84)

R-Squared 0.253 0.361 0.255 0.0744
SD ($261,178) *
home price coef 0.00186 0.00460 -0.00548 0.00650
t-statistics in parentheses

* p<0.05, ** p<0.01, *** p<0.001

Notes: The sample is comprised of patients ages 67 and older with continuous fee-
for-service coverage starting in 2006 through the end of 2011 or until death.
Analysis is at the five-digit ZIP-year level for 2006-2011. Regressions are weighted
using ZIP-level population weights. Models include average age and average age-
squared, percent black, percent Hispanic, percent Asian, percent female, percent of
patients receiving the Medicare Part D low-income subsidy, percent of patients
dually eligible for Medicare and Medicaid, the share of patients in Medicare
Advantage (as opposed to traditional fee-for-service Medicare), lagged, logged HCC
score to capture health status, and ZIP and year-level fixed effects. We find no
evidence of the effect of wealth shocks on the neighborhood prevalence of the
following conditions: hip fractures, stroke and transient ischemic attacks,
hyperlipidemia, hypertension, chronic heart failure, acute myocardial infarction,
lung, endometrial, and breast cancer). N = 49,799 ZIP-year observations (10,257 ZIP
codes).

68

Table 3-3: Expenditures

Logged Part A Costs Logged Part B Costs
1-Period Lagged Home Prices -9.36e-08*** -8.26e-08***

(-3.51) (-4.09)
R-Squared 0.229 0.275
SD ($261,178) * home price coef -0.0244 -0.0216
t-statistics in parentheses. * p<0.05, ** p<0.01, *** p<0.001
Notes: The sample is comprised of patients ages 67 and older with continuous fee-
for-service coverage starting in 2006 through the end of 2011 or until death.
Analysis is at the five-digit ZIP-year level for 2006-2011. Regressions are weighted
using ZIP-level population weights. Models include average age and average age-
squared, percent black, percent Hispanic, percent Asian, percent female, percent of
patients receiving the Medicare Part D low-income subsidy, percent of patients
dually eligible for Medicare and Medicaid, the share of patients in Medicare
Advantage (as opposed to traditional fee-for-service Medicare), lagged, logged HCC
score to capture health status, and ZIP and year-level fixed effects. N = 49,799 ZIP-
year observations (10,257 ZIP codes).
69

Table 3-4: Utilization of Medical Services

Hospital IP Days Office Visits
Hospital OP
Visits
1-Period Lagged Home -0.000000219** -0.00000137*** 0.00000171***
Prices (-2.97) (-8.12) (7.27)

R-Squared 0.0282 0.552 0.107
SD ($261,178) * home
price coef -0.0571 -0.357 0.446

Hospital IP Stays ED Visits ED + IP
1-Period Lagged Home -4.45e-08*** -1.99e-08 -6.45e-08**
Prices (-4.16) (-1.60) (-2.96)

R-Squared 0.0201 0.163 0.0895
SD ($261,178) * home
price coef -0.0116 -0.00520 -0.0168
t-statistics in parentheses. * p<0.05, ** p<0.01, *** p<0.001
Notes: The sample is comprised of patients ages 67 and older with continuous fee-
for-service coverage starting in 2006 through the end of 2011 or until death.
Analysis is at the five-digit ZIP-year level for 2006-2011. ED stands for emergency
department, OP for outpatient, IP for inpatient. The ED + IP category is the sum of
emergency department visits and inpatient visits. Regressions are weighted using
ZIP-level population weights. Models include average age and average age-squared,
percent black, percent Hispanic, percent Asian, percent female, percent of patients
receiving the Medicare Part D low-income subsidy, percent of patients dually
eligible for Medicare and Medicaid, the share of patients in Medicare Advantage (as
opposed to traditional fee-for-service Medicare), lagged, logged HCC score to
capture health status, and ZIP and year-level fixed effects. N = 49,799 ZIP-year
observations (10,257 ZIP codes).
70

Table 3-5: Use of Preventive Services

Cancer
Screenings
Lipid Tests Diabetes Tests
1-Period Lagged Home -2.86e-08*** 1.84e-08*** -6.06e-08***
Prices (-5.87) (3.51) (-5.49)

R-Squared 0.384 0.130 0.250
SD ($261,178) * home
price coef -0.00746 0.00482 -0.0158

Flu Shots Wellness Exams
Depression
Screenings
1-Period Lagged Home -7.22e-09 -2.57e-08*** -3.31e-10*
Prices (-1.15) (-5.37) (-2.14)

R-Squared 0.174 0.585 0.00369
SD ($261,178) * home
price coef -0.00189 -0.00672 -0.0000864

Osteoporosis
Screenings
1-Period Lagged Home -6.92e-09*
Prices (-2.49)

R-Squared 0.0520
SD ($261,178) * home
price coef -0.00181

t statistics in parentheses
* p<0.05, ** p<0.01, *** p<0.001 Notes: The sample is comprised of patients ages 67
and older with continuous fee-for-service coverage starting in 2006 through the end
of 2011 or until death. Analysis is at the five-digit ZIP-year level for 2006-2011.
Regressions are weighted using ZIP-level population weights. Models include
average age and average age-squared, percent black, percent Hispanic, percent
Asian, percent female, percent of patients receiving the Medicare Part D low-income
subsidy, percent of patients dually eligible for Medicare and Medicaid, the share of
patients in Medicare Advantage (as opposed to traditional fee-for-service Medicare),
lagged, logged HCC score to capture health status, and ZIP and year-level fixed
effects. N = 49,799 ZIP-year observations (10,257 ZIP codes).

71

Figure 3-2: Distress rates over time
Notes: Distress rates were constructed from DataQuick transaction and assessor
data. The above plot shows the 25
th
, 50
th
, and 75
th
percentiles of the distress rates
representing the pre-auction and auction phases of the foreclosure process.
72

Table 3-6: Selected Results Adding Distress Rates

Cancer Screenings Hyperlipidemia
1-Period Lagged Home Prices -2.71e-08*** 8.88e-09

(-5.58) (1.68)

1-Period Lagged Distress Rate -0.0433* -0.0788**

(-2.09) (-3.11)

R-Squared 0.385 0.491
SD ($261,178) * Home Price
Coef -0.00707 0.00232
SD (0.02) * Distress Rate Coef -0.000770 -0.00140

Ischemic Heart
Disease
Heart Attacks
(AMI)
1-Period Lagged Home Prices 2.65e-08*** -1.40e-09

(5.00) (-1.40)

1-Period Lagged Distress Rate -0.0440* 0.00968*

(-2.43) (2.42)

R-Squared 0.0746 0.00473
SD ($261,178) * Home Price
Coef 0.00691 -0.000366
SD (0.02) * Distress Rate Coef -0.000784 0.000172

t statistics in parentheses

* p<0.05, ** p<0.01, ***p<0.001

Notes: The sample is comprised of patients ages 67 and older with continuous fee-
for-service coverage starting in 2006 through the end of 2011 or until death.
Analysis is at the five-digit ZIP-year level for 2006-2011. Regressions are weighted
using ZIP-level population weights. Models include average age and average age-
squared, percent black, percent Hispanic, percent Asian, percent female, percent of
patients receiving the Medicare Part D low-income subsidy, percent of patients
dually eligible for Medicare and Medicaid, the share of patients in Medicare
Advantage (as opposed to traditional fee-for-service Medicare), lagged, logged HCC
score to capture health status, and ZIP and year-level fixed effects. N = 49,792 ZIP-
year observations (10,257 ZIP codes). We find no evidence of other outcomes and
negative wealth shocks; estimated coefficients for home prices in models looking at
other outcomes are robust to the inclusion of distress rates.
73

CHAPTER 4. “Did Medicare Part D Reduce Disparities?
27
”
Introduction
The primary objective of the 2003 Medicare Prescription Drug,
Improvement, and Modernization Act (MMA) was to provide seniors with affordable
coverage for their prescription medications through the new Medicare Part D
prescription drug benefit. This aim has largely been achieved: more than thirty-five
million Medicare beneficiaries are now enrolled in Part D plans with approximately
nine out of ten satisfied with their plan (KRC Research 2011). While Part D has
reduced the financial burden of prescription drug spending for beneficiaries,
particularly those with low incomes or extraordinarily high out-of-pocket drug
expenses, its gap in coverage may have induced beneficiaries to change their use of
medications or discontinue use of an effective therapy altogether.
The Part D benefit has a well-known gap in coverage commonly referred to
as the “doughnut hole.” Under the standard benefit, beneficiaries who do not receive
subsidies face a deductible, followed by a 25% co-insurance rate. Once they have
spent up to a designated level on medications in a year ($2,850 in 2014), they must
start paying the full price of their drugs. Coverage resumes only after a beneficiary
reaches the “catastrophic” limit in out-of-pocket spending, with minimal cost-
sharing thereafter. This nonlinear design is more complicated than a simple
increase in patient cost-sharing, for it alters both the current and future price of a
drug. Once a non-subsidized beneficiary (non-LIS) reaches the coverage gap, each
prescription he fills is likely to cost more. Yet, simultaneously, each fill increases the
likelihood of reaching the catastrophic threshold, thus lowering the expected price
of future prescriptions that year. Further, any price change in the gap is temporary
since benefits reset at the beginning of the next calendar year. How beneficiaries,

27
This paper was co-authored with Julie Zissimopoulos, Geoffrey Joyce, and Dana Goldman.
The published, edited version, including tables and figures, was published in the American
Journal of Managed Care February 2015 issue.
74

particularly racial/ethnic minorities and those with low levels resources, respond to
changes in coverage over the course of the year is largely unknown.
Recent work finds that the Part D coverage gap induces beneficiaries to
reduce use of essential medications

(G. Joyce, Zissimopoulos, and Goldman 2013),
but does not examine the differential responses of minorities and the near-poor who
do not qualify for federal subsidies. Racial and ethnic minorities have higher rates of
chronic illness than non-minorities, and lower socioeconomic status (SES) groups
are less able to manage complex treatment regimens often required in managing a
disease

(Goldman and Smith 2002). Indeed, black and Hispanic enrollees report
greater difficulty obtaining information and purchasing needed medications in Part
D (Haviland et al. 2012).
In this paper, we examine the effects of cycling in and out of coverage on the
prescription drug use of racial and ethnic minorities and other vulnerable
subgroups of Medicare beneficiaries using a differences-in-differences approach. We
compare changes in prescription drug use of white, black and Hispanic beneficiaries
before and after reaching the coverage gap for two different groups of beneficiaries:
(1) those eligible for the full low-income subsidy (LIS) who face minimal cost
sharing and thus unaffected by the coverage gap; and (2) non-subsidized
beneficiaries who pay the full cost of medications in the coverage gap (non-LIS).
Changes in medication use after reaching the gap are estimated separately by race.
We focus on beneficiaries with diabetes because it disproportionately affects
racial/ethnic minorities and is a major risk factor for a wide range of other health
conditions. If the gap is prompting beneficiaries to use pharmaceuticals
differently—especially if it leads them to discontinue an effective therapy—it should
be evident in this sample.
Data
We use a twenty percent random sample of Medicare beneficiaries enrolled
in Part D. This dataset links enrollment and Parts A and B claims for traditional fee-
for-service Medicare enrollees (2002-2008) to Part D claims (2006-2008). The Part
A data include information about inpatient hospital stays, including length of stay,
75

diagnosis-related group (DRG), department-specific charges, and up to ten
individual procedure codes and diagnostic codes. Part B information includes claims
submitted by physicians, and other health care providers and facilities for services
reimbursed by Part B. Each claim contains diagnostic (ICD-9-CM) and procedure
(CPT-4) codes, dates of service, demographic information on beneficiaries, and a
physician identification number.
The pharmacy data include all of the key elements related to prescription
drug events (e.g., drug name; National Drug Code (NDC); dosage; supply; date of
service). Each pharmacy claim includes the amount of the low-income subsidy; the
true out-of-pocket amount; and a field that indicated in which benefit phase a claim
was made: deductible, pre-coverage gap, coverage gap, or catastrophic phase (or
whether the claim straddles two of these phases). The Part D data identify the exact
date that non-LIS beneficiaries entered and exited the coverage gap, as well as when
LIS beneficiaries – not subject to the gap -- reach the same levels of prescription
drug spending associated with entrance into and exit from the gap.
The denominator file contains demographic information about each
beneficiary including date of birth, gender, beneficiary type (e.g., recipient of the
low-income subsidy), and zip code of residence. We link five-digit zip codes to the
American Community Survey (ACS) to measure neighborhood socioeconomic status,
including education (level of schooling attained) and median household income in
the beneficiaries’ ZIP code. The Medicare data also include externally validated
measures of race/ethnicity. Self-reported measures on race/ethnicity are refined
using Research Triangle Institute estimates based on geography and first and last
names.

Sample
The study sample consists of Medicare beneficiaries aged 65 and older with
diabetes. Persons with diabetes commonly take medications for glycemic control,
hypertension, and dyslipidemia, and proper medication adherence is associated
with large reductions in both macro and microvascular complications. Clinical trials
76

consistently show that complications from this disease can be avoided or deferred
with tight glycemic control (Goldman and Smith 2002; The Diabetes Control and
Complications Trial Research Group 1993). We identify beneficiaries with diabetes
based on at least one inpatient or skilled nursing facility diagnosis, or two or more
outpatient diagnoses of diabetes. We also assume that a beneficiary with a Part D
claim for insulin has diabetes. Once identified, beneficiaries are assumed to have
diabetes in subsequent years.
We restrict our analysis to those enrolled in traditional fee-for-service
Medicare and a stand-alone Part D drug plan (PDP). The traditional fee-for-service
criterion is necessary for identifying patients with diabetes. Individuals were
required to have the same Part D contract/plan for the entire year. These inclusion
criteria ensure that the analysis is done most cleanly, though they may affect the
generalizability of the results to the general Medicare population.
Our sample includes two groups of beneficiaries: those receiving the full low-
income subsidy (LIS) and those not receiving any type of subsidy (non-LIS) and had
no gap coverage. LIS beneficiaries do not pay Part D premiums and face minimal
cost-sharing throughout the year. As a result, they are not subject to the coverage
gap even when their level of drug spending reached the coverage gap threshold (e.g.,
$2,250 in 2006) and should not change their medication use before and after
reaching the various (hypothetical) coverage thresholds. We use the LIS as controls
and compare their medication use before and after reaching the gap to that of non-
LIS beneficiaries, who face vastly different prices over the course of the year and
spending distribution.
Given that 2006 was the initial year of the program and that beneficiaries
could enroll up to May 15th, we restricted our analyses to 2007 and 2008.
Nonetheless, we used the 2006 data for risk adjustment, categorization of
beneficiaries, and to compute medication use in 2007 for medications initiated in
2006 or earlier. In 2007, the study sample included 557,756 beneficiaries: 416,495
whites, 69,947 blacks, and 71,314 Hispanics.

77

Statistical Analysis
Our strategy is to estimate the difference in medication use before and after
the coverage gap for a treatment (non-LIS) and control group (LIS), by drug class
and race/ethnicity. We estimate race-specific changes in medication use before and
after reaching the coverage gap for the non-LIS, and benchmark these changes to
race-specific changes in the medication use of LIS beneficiaries at similar levels of
drug spending, i.e. before and after reaching the “hypothetical” threshold of the
coverage gap. We use multivariate regression to control for the variation in
demographic and socioeconomic characteristics, and interact binary indicators for
each beneficiary group (LIS/non-LIS) with race/ethnicity.
Due to the correlation between race/ethnicity and income, we estimate the
models including median household income in the beneficiary’s zip code, then
predict medication use in the coverage gap by race/ethnicity holding household
income constant at $50,000 (the approximate median within the sample). Standard
errors are clustered at the individual level and are computed using bootstrapping.
Our key outcome measure is medication adherence, measured by the
Medication Possession Ratio (MPR). MPR is the fraction of days that a patient
possesses medication, as measured by prescription fills. For example, a patient who
filled a thirty-day script on April 1st and refilled the prescription on May 10th would
have an MPR of 75% for that period, since they possessed thirty pills over a forty-
day span. For each drug class, we compute the total days’ supply of medications
before and after reaching the coverage gap to compute the percentage of compliant
days for each individual in the sample. The remaining days’ supply at the end of one
year is carried over to the subsequent year. We estimate changes in the rate of
medication use (MPR) overall and by therapeutic class, as well as the proportion of
all prescriptions dispensed as generic (generic dispensing rate, GDR).
We next examine the fraction of white, black and Hispanic beneficiaries who
stopped using a class of medication after reaching the gap and the fraction that
resumed use in the first 90 days of the next year. Discontinuation is measured by
comparing medication use within a therapeutic class in the 90 days prior to a
78

beneficiary’s gap entry date and after reaching the gap. For example, a beneficiary
observed taking an oral hypoglycemic, an antihypertensive, and a statin before
reaching the gap, but only an oral hypoglycemic and an antihypertensive after
entering the gap (for the remainder of the year) would be categorized as having
discontinued one medication within the relevant classes. We also examine the
extent to which beneficiaries switched medications after reaching the gap (from
brand to generic), limited to classes that were neither brand- nor generic-
dominated.
We measure changes in medication use for the nine most prevalent drug
classes used to treat diabetes and its comorbidities (diabetes-related medications)
and the nine most common classes used by these beneficiaries for other conditions
(nondiabetes-related). Diabetes-related classes include: oral hypoglycemic agents,
ACE inhibitors, calcium channel blockers, diuretics, beta blockers, angiotensin II
receptor blockers (ARBs), statins, loop diuretics, digitalis glycosides, and
combination antihypertensives.
28
ACE inhibitors and ARBs are combined into a
single class because they are commonly considered therapeutically interchangeable.
The set of other drugs consists of the nine most prevalent nondiabetes-related
classes used by this set of beneficiaries: antidepressants, antipsychotics, central
nervous system (CNS) medications (the majority of which are Alzheimer’s
medications like Aricept, Namenda, and Razadyne, and Lyrica, which treats nerve
and muscle pain), antiasthmatics, platelet aggregation inhibitors (e.g., Plavix),
antiulcerants, anticonvulsants, opioid analgesics, and
hormones/synthetics/modifiers. Using both diabetes-related and nondiabetes-
related medications allows us to examine whether beneficiaries with diabetes are
more or less price sensitive for their disease-specific medications. In some analyses,
we report the average price of a 30-day supply of the drugs in each class. These
prices were derived empirically from the data.

28
We do not look at changes in the demand for insulin. It is covered by Medicare Part D in
certain cases; for example, it is covered under Medicare Part B as durable medical
equipment when it is administered with an insulin pump (eHealth Medicare 2015).
79

We use estimates from multivariate regression models to predict the
change in medication use by race/ethnicity for diabetes and nondiabetes-related
classes. The models control for health status using binary indicators for the most
common comorbid conditions based on ICD-9 diagnostic codes in the medical
claims. These included twenty conditions defined in the Chronic Conditions
Warehouse (CCW), as well as hypertension, hyperlipidemia, asthma, gastro-
intestinal disorders. We also adjusted for age, age-squared, gender, time indicators,
and ZIP code-level measures of income.
Finally, we compare changes in medication use for LIS and non-LIS
beneficiaries living in low-income areas to understand the relationship between
changes in medication use and income effects proxied by the median household
income in a beneficiary’s zip code. We define the “near-poor” as white, black and
Hispanic beneficiaries who reside in zip codes with a median household income
below $25,000 (the bottom income quartile of the sample of non-LIS beneficiaries).
Results
Table 4-1 shows the characteristics of the study sample by race/ethnicity
and beneficiary group. White beneficiaries were least likely and Hispanics were
most likely to receive the full low-income subsidy: more than 80 percent of
Hispanics and less than 30 percent of whites were categorized as LIS. White
beneficiaries had more years of schooling and higher incomes than Hispanics and
blacks. Regardless of race/ethnicity, the LIS were more likely to be female and have
low SES compared to non-LIS.
Although prescription drug use differed widely by race/ethnicity, it did not
differ by beneficiary group (Table 4-1) before the gap. For example, both LIS and
non-LIS whites took their medications about 80% of the time before the coverage
gap level of spending. Pre-gap adherence was lowest among Hispanics and changed
more dramatically after reaching the coverage gap. Adherence among non-LIS
Hispanics declined by 10 percentage points (from 73% to 63%) before and after
reaching the coverage gap compared to just 2 percentage points for whites (76% to
74%).
80

Because LIS beneficiaries are in worse health than the non-LIS and face
minimal cost-sharing for their medications, they were much more likely to reach the
coverage gap threshold and reach it earlier in the year than non-LIS beneficiaries.
However, within beneficiary groups, white, black and Hispanics reached the
coverage gap level of spending at about the same time (late August to early
September). Thus average duration in the gap was about 4 months for those that did
not reach the catastrophic threshold.

Multivariate Findings
Figure 4-1 displays the percentage point change in medication use of non-LIS
relative to LIS before and after the coverage gap. We present results by
race/ethnicity, adjusting for demographic, health and socioeconomic characteristics.
The top panel displays changes in medication use across nine diabetes-related
classes and the bottom panel for nondiabetes-related classes. Drug classes are
ordered from lowest to highest average price to highlight the correlation between
adherence and out-of-pocket costs during the coverage gap. For example, use of
statins ($65 per month) declines by 9 percentage points (pp) during the coverage
gap among non-LIS Hispanics (relative to LIS Hispanics). In practical terms, these
changes imply that non-LIS Hispanics take their statins as prescribed 63% of the
time after reaching the gap, compared to 72% prior to reaching the gap (tables
available upon request). Corresponding figures for blacks and whites are 7pp and
5pp, respectively.
For the nine diabetes-related drug classes combined, medication use in the
gap declines by 6pp for Hispanics, 4pp for blacks and 3pp for whites. We find a
similar pattern in the use of nondiabetes-related medications. Over these nine
classes, use in the coverage gap declines by 9pp for Hispanics, 8pp for blacks and
6pp for whites. The differential changes in medication use are even larger in
percentage terms (as opposed to percentage points) due to racial/ethnic differences
in baseline levels of adherence (see Appendix C).
81

In addition to racial differences, Figure 4-1 also highlights the correlation
between adherence and price. Use of costly, brand-dominant classes such as
antipsychotics ($213), antiplatelets ($123) and antiulcerants ($108) declines more
sharply than use of less expensive medications such as beta blockers ($27) and
diuretics ($8). For example, use of antipsychotics drops by 8pp for whites, 10pp for
blacks, and 9pp for Hispanics, while use of less costly diuretics decreases by 4pp for
both whites and blacks and 2pp for Hispanics.
Reduced medication use can reflect different behavioral responses to
the coverage gap, such as stretching a prescription over more days (e.g., pill-
splitting) or stopping a medication altogether. Table 4-2 shows differential rates of
stopping and later resuming drug therapies, by race/ethnicity. A higher percentage
of non-LIS beneficiaries discontinue use of diabetes-related and nondiabetes-related
medications after reaching the coverage gap compared to the LIS, and a larger
fraction resume use in the next year once coverage resumed. Discontinuing use is
most common among Hispanics, who stop and resume at two to three times the rate
of blacks and whites. For example, an additional 6.7% of non-LIS Hispanics
discontinue a class of diabetes-related medication after reaching the coverage gap
relative to LIS Hispanics (compared to 4.1% of blacks and 2.4% of whites). Among
those who stop, an additional 12.5% of the non-LIS Hispanics (relative to the LIS
Hispanics) resume use in the first quarter of the next year (versus 6.7% of whites
and 5.9% of blacks).
While overall medication use declines in the coverage gap, the fraction of
drugs dispensed as generic increases modestly. Figure 4-2 shows race-specific
changes in the use of generic drugs after reaching the coverage gap for diabetes-
related and nondiabetes-related classes, relative to the LIS. Among the nine
diabetes-related classes, generic use increases by 2 to 3 percentage points in the
coverage gap for each race/ethnicity. We find similar effects among the
nondiabetes-related classes, but the difference is only statistically significant for
whites.
82

Given that race/ethnicity is correlated with income, we re-estimate the
models including median household income in the beneficiary’s zip code, then
predict medication use in the coverage gap by race/ethnicity, holding household
income constant at $25,000 (compared with at $50,000 for the results in Figure 4-
1). These results are presented in Figure 4-3. For the nine diabetes-related classes
combined, low-income Hispanics decreased medication use by 9 percentage points
in the gap relative to Hispanics receiving the low-income subsidy, a larger effect
than that of Hispanics overall (6pp, Figure 4-1). Further, the effects were larger in
more expensive classes. By contrast, the reduction in medication use among lower
income blacks (5pp) and whites (3pp) was similar to that of blacks (4pp) and whites
(3pp) overall (Figure 1).
Discussion
Our findings suggest that the Part D coverage gap is disruptive to drug
therapy, particularly for minorities and those who live in lower-income areas but do
not receive subsidies. Older, unsubsidized Hispanics with diabetes reduce their use
of diabetes-related medications by 6 percentage points during the coverage gap,
compared with 4 percentage points for blacks and 3 percentage points for whites.
The reduction in medication use reflected higher rates of medication
discontinuation and only a fraction of patients who discontinue use in the coverage
gap re-initiated therapy once coverage resume the next year.
A large body of literature has demonstrated that out-of-pocket costs affect
adherence (G. F. Joyce et al. 2002; Goldman, Joyce, and Escarce 2004; Goldman,
Joyce, and Zheng 2007; Huskamp et al. 2003; T. T. Le et al. 2008). Yet, since most
claims-based datasets do not contain information on race or ethnicity, this research
has been silent as to whether minorities are more sensitive to the cost of
prescription drugs than non-minorities. Our research begins to fill that gap. Why
Hispanic and black beneficiaries have a stronger response to changes in the price of
medication remains unclear. Older minorities may perceive drug therapies as less
efficacious or essential in the treatment of chronic disease

(T. T. Le et al. 2008),
which may make them more likely to discontinue use when out-of-pocket costs
83

increase suddenly or exceed some threshold (Mann et al. 2009). We find a strong
relationship between the price of the drug and the response to the coverage gap.
Declines in medication use are larger in drug classes costing more than $60 per
month. Other studies have shown that racial/ethnic minorities are more adversely
affected by cost-related non-adherence and have poorer overall adherence to
medication in Medicare Part D (Gellad and Haas 2007; Zhang and Baik 2014; Ngo-
Metzger et al. 2012; Tseng et al. 2008; Lauffenburger et al. 2013). Unlike a change in
copayment, the coverage gap is temporary and two-fold: it increases the current
out-of-pocket cost of medication, while simultaneously lowering the expected future
out-of-pocket cost of a drug if the beneficiary reaches the catastrophic threshold
(Johnson et al. 1997).
Changes in drug benefits have been associated with substantial morbidity
and mortality in certain high-risk populations (Johnson et al. 1997; Tamblyn et al.
2001; Gaynor, Li, and Vogt 2007; Chandra, Gruber, and McKnight 2007; Lurie et al.
1986). Reductions in medication use as a result of the Part D coverage gap raise
concerns about deleterious health effects that may manifest over time (West of
Scotland Coronary Prevention Study Group 1997). Mitigating some of this concern is
the relatively short length of time most beneficiaries spend in the coverage gap. The
median beneficiary is only subject to the gap for 3-4 months. Behavioral responses
to the coverage gap may also mitigate potential health effects. Black, white and
Hispanic beneficiaries increased their use of generic medications after reaching the
coverage gap, and switched to more generous plans the next year (see Appendix C).
While minorities were more likely to stop taking a medication after reaching the gap
than white beneficiaries, they were also more likely to resume therapy once
coverage restarted in January. Earlier research has found that, at least within a one-
year period, patients do not use medical services (inpatient, outpatient, or
emergency department) differently before and after entering the coverage gap (G.
Joyce, Zissimopoulos, and Goldman 2013), which suggests that health effects, if any,
may take longer to manifest.
84

Our study has several limitations. First, our proxy for socioeconomic status
does not fully account for the variation by race and ethnicity in adherence in the
coverage gap. While near-poor Hispanics decreased medication use in the gap more
than higher income Hispanics, income had little impact on the response of white and
blacks to the coverage gap. Since our income measure is at the zip code level rather
than the individual level, we are unable to perfectly disentangle the effect of
socioeconomic status from race. Previous work using similar SES data has found
that individuals living in lower-income areas were more price sensitive than their
higher-income counterparts (Chernew et al. 2008).
Second, beneficiaries receiving the full low-income subsidy (LIS) are
obviously poorer, more likely to be female, non-white, and sicker on average than
the non-LIS. Our results may be biased if the LIS also differ in unobserved ways that
make them an inappropriate control group. Two points mitigate these concerns.
First, the LIS had a constant level of prescription drug use before and after the
coverage gap, which is consistent with them being unaffected by the gap. Further,
our empirical approach compared medication use before and after the coverage gap
within beneficiary group and race/ethnicity, thereby using each racial/ethnic group
as its own control.
Third, we identify the chronically ill from claims data. The main concern with
this approach is false positives if “rule-out” diagnoses are recorded on the claims.
We tried to minimize this error by restricting our analysis to users of disease-
specific drugs, requiring multiple physician visits or hospitalizations for the
condition, and exploiting a long panel of Parts A and B claims (2002-2008). The use
of claims data also obscures the level of disease severity, but this potential bias is
also minimized by the difference-in-differences strategy.
Lastly, our results may overstate the impact of the coverage gap on
prescription drug use if beneficiaries obtained free samples from their providers or
paid for medications in cash at discount outlets after reaching the gap (Tseng et al.
2004). An increasing number of retail pharmacies (e.g., Wal-Mart, Target) sell a
broad range of generic drugs for $4 per prescription. While there is little empirical
85

data on the extent of this behavior, a pre-Part D study found that 6% of enrollees in
a Kaiser Permanente Medicare Advantage plan purchased prescriptions outside of
their plan after reaching the annual benefit limit (Hsu et al. 2006). We observe a
substantial and rapidly increasing number of $4 claims in the Part D data, thus the
extent of bias from uncaptured claims is likely to be small. Since entry into the
catastrophic phase was based on accumulating out-of-pocket expenses, beneficiaries
had an incentive to purchase all of their medications -- even $4 scripts -- through the
Part D program.
Conclusions
Although the coverage gap is being phased out under the Affordable Care Act
(ACA), beneficiaries will continue to face a break in coverage until 2020. In addition,
like Part D, the ACA continues the trend toward “consumer-directed” health care.
While compelling patients to take a more active role in choosing a plan and
managing their health care is generally positive, protecting vulnerable groups in the
health care marketplace requires more than just premium subsidies. Patient
education is a first step, but more substantive improvements in adherence will
require changes in health care delivery. The shift from a fee-for-service model to
bundled payments under the ACA will reward providers for better patient
outcomes, of which medication adherence is critical. Similarly, new investments in
health information technology will allow more providers and health plans to contact
patients who do not fill or refill a prescription on a timely basis and discuss with
them the reasons behind their decision, and allow them intervene when applicable.
While the success of these types of changes have not been demonstrated, it is
difficult to imagine that targeted interventions would not be cost-beneficial given
the clinical and financial consequences of poor adherence among older beneficiaries
with chronic diseases.

86

Tables and Figures
Table 4-1: Beneficiary Characteristics by Coverage Group and Race

White Black Hispanic

LIS Non-LIS LIS Non-LIS LIS Non-LIS
(n=123,033) (n=293,462) (n=50,440) (n=19,507) (n=57,283) (n=14,031)
Demographics
Age in years (mean) 75.0* 75.8 74.6* 73.9 74.5* 74.2
Male (%) 28.8* 42.4 22.9* 38.0 34.2* 44.4
Socioeconomic Status
a

Median income ($) 48,697* 57,926 40,984* 46,660 45,561* 48,256
Years of education 13.3* 13.7 12.9* 13.2 12.5* 13.2
Rx Utilization Measures
(mean)
b

Pre-gap MPR 0.80* 0.80 0.76* 0.75 0.74* 0.73
Post-gap MPR 0.78* 0.73 0.74* 0.67 0.72* 0.63
Pre-gap GDR 0.49* 0.48 0.50* 0.48 0.43 0.43
Post-gap GDR 0.54* 0.55 0.56* 0.57 0.48* 0.52
Median month of coverage
gap entry
7.10* 8.70 7.80* 9.00 7.60* 8.90

White Black Hispanic
LIS Non-LIS LIS Non-LIS LIS Non-LIS
Parts A and B utilization
(mean) 2005
c
(n=109,143) (n=268,680) (n=43,930) (n=17,151) (n=44,455) (n=12,492)
No. of office visits 9.0* 9.0 7.5* 7.7 9.2* 9.4
No. of emergency
department visits
7.7* 6.3 8.7* 6.3 7.5* 4.6
No. of inpatient stays 0.5* 0.3 0.5* 0.3 0.4* 0.3
No. of inpatient days 2.6* 1.7 2.8* 2.0 2.2* 1.5
Parts A and B spending
(mean $) 2005
c

Total 9,698* 7,869 10,260* 8,160 10,581* 6,917
Inpatient 4,251* 3,152 4,265* 3,200 3,909* 2,438
Outpatient 1,455* 1,196 2,227* 1,724 1,839* 1,033
Other 3,992* 3,521 3,768* 3,236 4,833* 3,446
Notes: Sample is individuals with diabetes and ages 65 and older. Demographics,
socioeconomic status, medication possession ration and generic dispensing ratio are
measured in year 2007. Spending and utilization measured in 2005 to demonstrate pre-Part
D differences across groups.
* Indicates low-income subsidy (LIS) and non-low-income subsidy (non-LIS) values are
significantly different at 1%.
a
Socioeconomic status is measured at zip code level.
b
Medication Possession Ratio (MPR) and Generic Dispensing Ratio (GDR) are measured
before spending reaches the coverage gap level ("pre") and after spending reaches coverage
gap levels ("post") in 2007.
c
Utilization and spending for beneficiaries covered by fee-for-
service Medicare Part A & B for all 12 months of 2005.

87

Figure 4-1: Regression Adjusted Difference-in-Difference in Medication Use
(MPR), by Therapeutic Class and Race (percentage point)

-0.20 -0.15 -0.10 -0.05 0.00 0.05
Antilipemic Agents ($65)
Antihypertensive Combinations ($58)
Oral Hypoglycemic Agents ($50)
Calcium Channel Blockers ($46)
Ace/Angiotensin II Inhibitors ($31)
Beta Blockers ($27)
Loop Diuretics ($8)
Digitalis Glycosides ($7)
Diabetes-Related Drug Classes
White
Black
Hispanic
-0.20 -0.15 -0.10 -0.05 0.00 0.05
Antipsychotics ($213)
CNS Medications ($150)
Antiasthma ($127)
Platelet Aggregation Inhibitors ($123)
Antiulcerants ($108)
Hormones/Synthetics/Modifiers ($90)
Anticonvulsants ($61)
Opioid Analgesics ($53)
Antidepressants ($49)
Nondiabetes-Related Classes
White
Black
Hispanic
88

Notes:
MPR refers to the Medication Possession Ratio, which is the fraction of days that a
patient “possesses” or has access to medication, as measured by prescription fills.
“CNS medications” refers to central nervous system medications. Changes are based
on results from multivariate models which control for age, age-squared, gender,
comorbid conditions, and socioeconomic status. Prices shown reflect the average
price paid in the sample for a 30-day supply of medication in the therapeutic class.
Diabetes-related drug classes: Whites who received the low-income subsidy (LIS):
n=74,452; Whites who did not receive the low-income subsidy (non-LIS):
n=115,333; LIS blacks: n=26,140; Non-LIS blacks: n=6,131; LIS Hispanics: n=29,113;
Non-LIS Hispanics: n=4,311. Nondiabetes-related drug classes: LIS whites:
n=65,062; Non-LIS whites: n=89,927; LIS blacks: n=21,337; Non-LIS blacks:
n=4,373; LIS Hispanics: n=25,083; Non-LIS Hispanics: n=3,464.

89

Table 4-2: Differential Stopping and Conditional Resumption Rates of the Non-
LIS Group Relative to the LIS Group

Stops Drug
(% difference non-LIS and LIS)
a

White
(n=196,178)
c

Black
(n=33,061)
Hispanic
(n=34,472)
Diabetes Drug Classes 2.4 4.1 6.7
Non-Diabetes Drug Classes 2.6 3.2 4.2
Resumes Conditional on
Stopping
(% difference non-LIS and LIS)
b

White
(n=97,589)
d
Black
(n=18,257)
Hispanic
(n=20,291)
Diabetes Drug Classes 6.7 5.9 12.5
Non-Diabetes Drug Classes 4.7 4.1 11.4
Notes:
Sample is individuals with diabetes and ages 65 and older that reached the
coverage gap in 2007. Individuals had to stay in coverage gap for at least 40 days.
Resumption is defined as stopping a drug in the coverage gap in 2007 and
resuming in the first quarter of 2008. The difference between non-low-income
subsidy (non-LIS) and low-income subsidy (LIS) rates is significantly different at
1% for all cells.
a
Percent of Non-LIS stoppers less percent of LIS stoppers, in the coverage gap in
2007
b
Among stoppers, the difference in percent of Non-LIS who resume and percent of
LIS who resumed in the first quarter of 2008
c
The number of observations for stoppers is the number of individuals of the
particular race who spent at least forty days in the coverage gap in 2007.
d
The number of observations for those who resumed refers to the number of
individuals of the particular race who spent at least forty days in the coverage gap
in 2007 and stopped medications in at least one drug class.
90

Figure 4-2: Regression Adjusted Difference-in-Difference in Use of Generic
Substitutes (GDR), by Race (percentage point)

Notes:
GDR refers to the Generic Dispensing Rate. Changes are based on results from
multivariate models which control for age, age-squared, gender, comorbid
conditions, and socioeconomic status. GDR for ACE/ARB class is for ACE Inhibitors
only since ARB class is brand-dominated. This analysis is limited to therapeutic
classes which are neither brand nor generic-dominated. Diabetes-related classes
include: oral hypoglycemic agents, ACE inhibitors, calcium channel blockers,
diuretics, beta blockers, angiotensin II receptor blockers (ARBs), statins, digitalis
glycosides, and combination antihypertensives. ACE inhibitors and ARBs are
combined into a single class because they are commonly considered therapeutically
interchangeable. The set of other drugs consists of the nine most prevalent
nondiabetes-related classes used by this set of beneficiaries: platelet aggregation
inhibitors and antiulcerants. Diabetes-related drug classes: Whites who received the
low-income subsidy (LIS): n=70,284; Whites who did not receive the low-income
subsidy (non-LIS): n=104,784; LIS blacks: n=24,412; Non-LIS blacks: n=5,475; LIS
Hispanics: n=27,159; Non-LIS Hispanics: n=3,736. Nondiabetes-related drug classes:
LIS whites: n=61,860; Non-LIS whites: n=76,652; LIS blacks: n=19,054; Non-LIS
blacks: n=3,339; LIS Hispanics: n=22,485; Non-LIS Hispanics: n=2,668.
-0.03
-0.02
-0.01
0.00
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
Diabetes-Related Classes Non-Diabetes-Related Classes
White
Black
Hispanic
91

Figure 4-3: Regression Adjusted Difference-in-Difference in Medication Use
(MPR), by Therapeutic Class and Race for the Near-Poor Population

Notes:
MPR refers to the Medication Possession Ratio, which is the fraction of days that a
patient “possesses” or has access to medication, as measured by prescription fills.
Changes are based on results from multivariate models which control for age, age-
squared, gender, comorbid conditions, and socioeconomic status. Prices shown
reflect the average price paid in the sample for a 30-day supply of medication in the
therapeutic class. ACE inhibitors and ARBs are combined into a single class because
they are commonly considered therapeutically interchangeable. We defined the
“near-poor” as white, black and Hispanics beneficiaries residing in zip codes with a
median household income below $25,000 (the bottom quartile of the sample’s
income distribution). Whites who received the low-income subsidy (LIS): n=74,452;
Whites who did not receive the low-income subsidy (non-LIS): n=115,333; LIS
blacks: n=26,140; Non-LIS blacks: n=6,131; LIS Hispanics: n=29,113; Non-LIS
Hispanics: n=4,311

-0.20 -0.15 -0.10 -0.05 0.00 0.05
Antilipemic Agents ($65)
Antihypertensive Combinations ($58)
Oral Hypoglycemic Agents ($50)
Calcium Channel Blockers ($46)
Ace/Angiotensin II Inhibitors ($31)
Beta Blockers ($27)
Loop Diuretics ($8)
Digitalis Glycosides ($7)
Diabetes-Related Drug Classes
White
Black
Hispanic
92

Bibliography
Adams, Alyce S, Connie Uratsu, Wendy Dyer, David Magid, Patrick O’Connor, Arne
Beck, Melissa Butler, P Michael Ho, and Julie A Schmittdiel. 2013. “Health
System Factors and Antihypertensive Adherence in a Racially and Ethnically
Diverse Cohort of New UsersAntihypertensive Adherence in New Users.” JAMA
Internal Medicine 173 (1): 54–61.
Adams, Peter, Michael D. Hurd, Daniel McFadden, Angela Merrill, and Tiago Ribeiro.
2003. “Healthy, Wealthy, and Wise? Tests for Direct Causal Paths between
Health and Socioeconomic Status.” Journal of Econometrics 112 (1): 3–56.
Agency for Healthcare Research and Quality. 2011. National Healthcare Disparities
Report, 2010. Rockville, MD.
http://www.ahrq.gov/research/findings/nhqrdr/nhdr10/nhdr10.pdf.
Agha, Leila, and David Molitor. 2013. The Local Influence of Principal Investigators on
Technology Adoption Evidence from New Cancer Drugs.
American Diabetes Association. 2013. “Economic Costs of Diabetes in the US in
2012.” Diabetes Care 36 (4): 1033–46.
Anderson, Patrick L., Richard D. McLellan, Joseph P. Overton, and Gary L. Wolfram.
1997. “Price Elasticity of Demand.”
http://scholar.harvard.edu/files/alada/files/price_elasticity_of_demand_hando
ut.pdf.
Andrulis, Dennis P., Nadia J. Siddiqui, Jonathan P. Purtle, and Lisa Duchon. 2010.
Patient Protection and Affordable Care Act of 2010: Advancing Health Equity for
Racially and Ethnically Diverse Populations. Washington, DC.
93

Baicker, Katherine, Amitabh Chandra, and Jonathan S. Skinner. 2005. “Geographic
Variation in Health Care and the Problem of Measuring Racial Disparities.”
Perspectives in Biology and Medicine 48 (1 Supplement): S42–53.
Barnett, Michael L., Nicholas A. Christakis, A. James O’Malley, Jukka-Penna Onnela,
Nancy L. Keating, and Bruce E. Landon. 2012. “Physician Patient-Sharing
Networks and the Cost and Intensity of Care in US Hospitals.” Medical Care 50
(2): 152–60.
Barnett, Michael L., Bruce E. Landon, a. James O’Malley, Nancy L. Keating, and
Nicholas a. Christakis. 2011. “Mapping Physician Networks with Self-Reported
and Administrative Data.” Health Services Research 46 (5): 1592–1609.
Berndt, Ernst R., Robert S. Pindyck, and Pierre Azoulay. 2003. “Consumption
Externalities and Diffusion in Pharmaceutical Markets: Antiulcer Drugs.” The
Journal of Industrial Economics 51 (2): 243–70.
Bikhchandani, Sushil, David Hirshleifer, and Ivo Welch. 1992. “A Theory of Fads,
Fashion, Custom, and Cultural Change as Informational Cascades.” Journal of
Political Economy 100 (5): 992–1026.
———. 1998. “Learning from the Behavior of Others: Conformity, Fads, and
Informational Cascades.” The Journal of Economic Perspectives 12 (3): 151–70.
Boards of Trustees of the Federal Hospital Insurance and Federal Supplementary
Medical Insurance Trust Funds. 2015. 2015 ANNUAL REPORT OF THE BOARDS
OF TRUSTEES OF THE FEDERAL HOSPITAL INSURANCE AND FEDERAL
SUPPLEMENTARY MEDICAL INSURANCE TRUST FUNDS.
https://www.cms.gov/Research-Statistics-Data-and-Systems/Statistics-
Trends-and-Reports/ReportsTrustFunds/Downloads/TR2015.pdf.
Cannon, H. Eric. 2007. “Pharmacy Management of Vaccines.” Journal of Managed
Care Pharmacy 13 (7, Supplement S-b): S7–11.
94

Centers for Disease Control and Prevention. 2008. CDC Seeks to Protect Older Adults
With Shingles Vaccine Message. http://www.cdc.gov/vaccines/vpd-
vac/shingles/downloads/shingles-vac-msg-2008.pdf.
———. 2014a. “About Shingles: Overview.”
http://www.cdc.gov/shingles/about/overview.html.
———. 2014b. “About Shingles: Transmission.”
http://www.cdc.gov/shingles/about/transmission.html.
———. 2014c. “Clinical Overview.” http://www.cdc.gov/shingles/hcp/clinical-
overview.html.
———. 2014d. “Shingles Vaccination: What You Need to Know.”
http://www.cdc.gov/vaccines/vpd-vac/shingles/vacc-need-know.htm.
———. 2014e. “Vaccination.” http://www.cdc.gov/shingles/vaccination.html.
Centers for Medicare and Medicaid Services. 2015. “Herpes Zoster ICD-9 Codes.”
Accessed January 1. http://www.cms.gov/medicare-coverage-
database/staticpages/icd9-code-
range.aspx?DocType=LCD&DocID=29416&Group=1&RangeStart=053.0&Range
End=053.9.
———. 2009. “Guidance to States on the Low-Income Subsidy.”
http://www.cms.gov/Medicare/Eligibility-and-
Enrollment/LowIncSubMedicarePresCov/Downloads/StateLISGuidance02100
9.pdf.
———. 2014. Dual Eligible Beneficiaries Under the Medicare and Medicaid Programs.
http://www.cms.gov/Outreach-and-Education/Medicare-Learning-Network-
MLN/MLNProducts/downloads/Medicare_Beneficiaries_Dual_Eligibles_At_a_Gl
ance.pdf.
95

———. 2015a. “On Its 50th Anniversary, More than 55 Million Americans Covered
by Medicare.”
https://www.cms.gov/Newsroom/MediaReleaseDatabase/Press-
releases/2015-Press-releases-items/2015-07-28.html.
———. 2015b. “Your Medicare Coverage: Inpatient Hospital Care.”
https://www.medicare.gov/coverage/hospital-care-inpatient.html.
Chandra, Amitabh, David Cutler, and Zirui Song. 2011. “Who Ordered That? The
Economics of Treatment Choices in Medical Care.” In Handbook of Health
Economics, Volume 2, 397–432.
Chandra, Amitabh, Jonathan Gruber, and Robin McKnight. 2007. “Patient Cost-
Sharing, Hospitalization Offsets, and the Design of Optimal Health Insurance for
the Elderly.” National Bureau of Economic Research, Inc, NBER Working
Papers: 12972.
Chandra, Amitabh, David Malenka, and Jonathan Skinner. 2013. “The Diffusion of
New Medical Technology: The Case of Drug-Eluting Stents.” Discoveries in the
Economics of Aging, 389–403. http://www.nber.org/chapters/c12974.ack.
Chandra, Amitabh, and Jonathan S. Skinner. 2012. “Technology Growth and
Expenditure Growth in Health Care.” Journal of Economic Literature, American
Economic Association 50 (3): 645–80.
Chernew, Michael, Teresa B. Gibson, Kristina Yu-Isenberg, Michael C. Sokol, Allison
B. Rosen, and A. Mark Fendrick. 2008. “Effects of Increased Patient Cost Sharing
on Socioeconomic Disparities in Health Care.” Journal of General Internal
Medicine 23 (8): 1131–36.
Chronic Conditions Data Warehouse. 2014. “Chronic Condition Warehouse
Categories.” https://www.ccwdata.org/web/guest/condition-categories.
96

CMS Office of the Actuary. 2006. “Medicare Part D Benefit Parameters for Standard
Benefit: Annual Adjustments for 2007.”
https://www.cms.gov/Medicare/Health-
Plans/MedicareAdvtgSpecRateStats/downloads/2007_Part_D_Parameter_Upda
te.pdf.
Coleman, James S, Elihu Katz, and Herbert Menzel. 1966. Medical Innovation: A
Diffusion Study. New York: Bobbs-Merrill.
Congressional Budget Office. 2015. March 2015 Medicare Baseline. Washington, DC.
http://www.cbo.gov/sites/default/files/cbofiles/attachments/44205-2015-
03-Medicare.pdf.
Conley, Timothy G., and Christopher R. Udry. 2010. “Learning About a New
Technology : Pineapple in Ghana.” American Economic Review 100 (1): 35–69.
Cramer, Joyce, Robert Rosenheck, Gail Kirk, William Krol, and John Krystal. 2003.
“Medication Compliance Feedback and Monitoring in a Clinical Trial: Predictors
and Outcomes.” Value in Health 6 (5): 566–73.
Cubanski, Juliette, Christina Swoope, Anthony Damico, and Tricia Neuman. 2014.
How Much Is Enough? Out-of-Pocket Spending Among Medicare Beneficiaries: A
Chartbook. Menlo Park, CA. http://files.kff.org/attachment/how-much-is-
enough-out-of-pocket-spending-among-medicare-beneficiaries-a-chartbook-
report.
Currie, Janet, and Erdal Tekin. 2015. “Is There a Link Between Foreclosure and
Health?” American Economic Journal: Economic Policy 7 (1): 63–94.
Deaton, Angus. 2002. “Policy Implications of the Gradient of Health and Wealth.”
Health Affairs 21 (2): 13–30.
97

Drukker, David M. 2003. “Testing for Serial Correlation in Linear Panel-Data
Models.” Stata Journal 3 (2): 168–77.
Egede, Leonard E, Mulugeta Gebregziabher, Kelly J Hunt, Robert N Axon, Carrae
Echols, Gregory E Gilbert, and Patrick D Mauldin. 2011. “Regional, Geographic,
and Racial/ethnic Variation in Glycemic Control in a National Sample of
Veterans with Diabetes.” Diabetes Care 34 (4): 938–43.
eHealth Medicare. 2015. “Medicare Diabetes Coverage.”
http://www.ehealthmedicare.com/about-medicare/diabetes/.
Engelhardt, Gary V., Michael D. Eriksen, and Nadia Greenhalgh-Stanley. 2013. A
Profile of Housing and Health Among Older Americans.
http://www.housingamerica.org/RIHA/RIHA/Publications/86310_13205_RIH
A_Senior_Housing_Paper.pdf.
Fichera, Eleonora, and John Gathergood. 2013. Do Wealth Shocks Affect Health? New
Evidence From the Housing Boom.
Finkelstein, Amy, Matthew Gentzkow, and Heidi Williams. 2014. Sources of
Geographic Variation in Health Care : Evidence from Patient Migration. 20789.
National Bureau of Economic Research Working Paper.
http://faculty.chicagobooth.edu/matthew.gentzkow/research/movers.pdf.
Food and Drug Administration. 2014. “CBER-Regulated Products: Resolved
Shortages.”
http://www.fda.gov/BiologicsBloodVaccines/SafetyAvailability/Shortages/uc
m351943.htm.
Foster, Andrew D., and Mark R. Rosenzweig. 1995. “Learning by Doing and Learning
from Others: Human Capital and Technical Change in Agriculture.” Journal of
Political Economy 103 (6): 1176–1209.
98

———. 1996. “Technical Change and Human-Capital Returns and Investments:
Evidence from the Green Revolution.” The American Economic Review 86 (4):
931–53.
Gaynor, Martin, Jian Li, and William B. Vogt. 2007. “Substitution, Spending Offsets,
and Prescription Drug Benefit Design.” Forum for Health Economics & Policy 10
(2).
Gellad, Walid F., and Jennifer S. Haas. 2007. “Race/ethnicity and Nonadherence to
Prescription Medications among Seniors: Results of a National Study.” Journal
of General Internal Medicine 22 (11): 1572–78.
Geoffard, Pierre-Yves, and Tomas J. Philipson. 1996. “Rational Epidemics and Their
Public Control.” International Economic Review 37 (3): 603–24.
Glover, J. Allison. 1938. “The Incidence of Tonsillectomy in School Children: (Section
of Epidemiology and State Medicine).” Proceedings of the Royal Society of
Medicine 31 (10): 1219–36.
Gold, Marsha, and Giselle Casillas. 2014. What Do We Know About Health Care Access
and Quality in Medicare Advantage Versus the Traditional Medicare Program?
Menlo Park, CA. http://files.kff.org/attachment/what-do-we-know-about-
health-care-access-and-quality-in-medicare-advantage-versus-the-traditional-
medicare-program-report.
Goldman, Dana P, and James P. Smith. 2002. “Can Patient Self-Management Help
Explain the SES Health Gradient?” Proceedings of the National Academy of
Sciences of the United States of America 99 (16): 10929–34.
doi:10.1073/pnas.162086599.
Goldman, Dana P., Geoffrey F. Joyce, and Jose J. Escarce. 2004. “Pharmacy Benefits
and the Use of Drugs by the Chronically Ill.” JAMA 291 (19): 2344–50.
99

Goldman, Dana P., Geoffrey F. Joyce, and Yuhui Zheng. 2007. “Prescription Drug Cost
Sharing.” Journal of the American Medical Association 298 (1): 61–69.
Goldman, Dana P., and James P. Smith. 2005. “Socioeconomic Differences in the
Adoption of New Medical Technologies.” American Economic Review 95 (2):
234–37.
Goodman, David C, Stephen S Mick, David Bott, Therese Stukel, Chiang-hua Chang,
Nancy Marth, Jim Poage, and Henry J Carretta. 2003. “Primary Care Service
Areas: A New Tool for the Evaluation of Primary Care Services.” Health Services
Research 38 (1, Part I): 287–309.
Goodman, David C., Shannon Brownlee, Chiang-Hui Chang, and Elliott S. Fisher.
2010. Regional and Racial Variation in Primary Care and the Quality of Care
among Medicare Beneficiaries.
http://www.dartmouthatlas.org/downloads/reports/Primary_care_report_090
910.pdf.
Goodman, David C., Elliott S. Fisher, and Chiang-hua Chang. 2011. After
Hospitalization: A Dartmouth Atlas Report on Post-Acute Care for Medicare
Beneficiaries.
Goodney, Philip P., Nino Dzebisashvili, David C. Goodman, and Kristen K. Bronner.
2014. Variation in the Care of Surgical Conditions: Diabetes and Peripheral
Arterial Disease.
http://www.dartmouthatlas.org/downloads/reports/Diabetes_report_10_14_1
4.pdf.
Goolsbee, Austan, and Peter J. Klenow. 2002. “Evidence on Learning and Network
Externalities in the Diffusion of Home Computers*.” The Journal of Law and
Economics 45 (2): 317–43.
100

Griliches, Zvi. 1957. “Hybrid Corn: An Exploration in the Economics of Technological
Change.” Econometrica 25 (4): 501–22.
Grossman, Michael. 1972. “On the Concept of Health Capital and the Demand for
Health.” The Journal of Political Economy 80 (2): 223–55.
Hamilton, Barton H., and Brian McManus. 2005. Technology Diffusion and Market
Structure : Evidence from Infertility Treatment Markets.
Harvard Pilgrim Health Care Network. 2008. “Shipping Delays for Zostavax Vaccine.”
Harvard Pilgrim HealthCare Network Matters.
https://www.harvardpilgrim.org/portal/page?_pageid=253,259727&_dad=por
tal&_schema=PORTAL.
Haviland, Amelia M., Marc N. Elliott, Robert Weech-Maldonado, Katrin
Hambarsoomian, Nate Orr, and Ron D. Hays. 2012. “Racial/ethnic Disparities in
Medicare Part D Experiences.” Medical Care 50 (11): S40–47.
Haynes, R Brian, Heather P McDonald, and Amit X Garg. 2002. “Helping Patients
Follow Prescribed Treatment.” JAMA: The Journal of the American Medical
Association 288 (22): 2880–83.
Holmes, Holly M, Ruili Luo, Joseph T Hanlon, Linda S Elting, Maria Suarez‐Almazor,
and James S Goodwin. 2012. “Ethnic Disparities in Adherence to
Antihypertensive Medications of Medicare Part D Beneficiaries.” Journal of the
American Geriatrics Society 60 (7): 1298–1303.
Hsu, John, Mary Price, Jie Huang, Richard Brand, Vicki Fung, Rita Hui, Bruce Fireman,
Joseph P Newhouse, and Joseph V Selby. 2006. “Unintended Consequences of
Caps on Medicare Drug Benefits.” New England Journal of Medicine 354 (22):
2349–59.
101

Huskamp, Haiden A., Patricia A. Deverka, Arnold M. Epstein, Robert S. Epstein,
Kimberly A. McGuigan, and Richard G. Frank. 2003. “The Effect of Incentive-
Based Formularies on Prescription-Drug Utilization and Spending.” New
England Journal of Medicine 349 (23): 2224–32.
IOM (Institute of Medicine). 2013. Variation in Health Care Spending: Target Decision
Making, Not Geography. Edited by Joseph P. Newhouse, Alan M. Garber, Robin P.
Graham, Margaret A. McCoy, Michelle Mancher, and Ashna Kibria. Washington,
D.C.: The National Academies Press.
Iyengar, Raghuram, Christophe Van den Bulte, and Thomas W. Valente. 2011.
“Opinion Leadership and Social Contagion in New Product Diffusion.”
Marketing Science 30 (2): 195–212.
Jackevicius, C.A., M. Mamdani, and J.V. Tu. 2002. “ADherence with Statin Therapy in
Elderly Patients with and without Acute Coronary Syndromes.” JAMA 288 (4):
462–67.
Javed, Saba, Fatima Javed, Rana M Mays, and Stephen K Tyring. 2012. “Herpes Zoster
Vaccine Awareness among People ≥ 50 Years of Age and Its Implications on
Immunization.” Dermatology Online Journal 18 (8).
http://escholarship.org/uc/item/9n03w07g.
Johnson, Richard E., Michael J. Goodman, Mark C. Eldredge, and Michael B.
Hornbrook. 1997. “The Effect of Increased Prescription Drug Cost-Sharing on
Medical Care Utilization and Expenses of Elderly Health Maintenance
Organization Members.” Medical Care 35 (11): 1119–31.
Joyce, Geoffrey F., Jose J. Escarce, Matthew D. Solomon, and Dana P. Goldman. 2002.
“Employer Drug Benefit Plans and Spending on Prescription Drugs.” JAMA 288
(14): 1733–39.
102

Joyce, Geoffrey, Julie Zissimopoulos, and Dana Goldman. 2013. “Digesting the
Doughnut Hole.” Journal of Health Economics 32 (6): 1345–55.
Kaiser Family Foundation. 2013. Projected Change in Medicare Enrollment, 2000-
2050. Menlo Park, CA. http://kff.org/medicare/slide/projected-change-in-
medicare-enrollment-2000-2050/.
———. 2014. Medicare at a Glance, August 2014. Menlo Park, CA.
https://kaiserfamilyfoundation.files.wordpress.com/2014/09/1066-17-
medicare-at-a-glance.pdf.
Kohler, Hans-Peter. 1997. “Learning in Social Networks and Contraceptive Choice.”
Demography 34 (3): 369–83.
http://link.springer.com/article/10.2307/3038290.
KRC Research. 2011. “Seniors’ Opinions About Medicare Rx: Sixth Year Update.”
Edited by K R C Research. Washington, DC: KRC Research.
Landon, Bruce E., Nancy L. Keating, Michael L. Barnett, Jukka-Penna Onnela,
Sudeshna Paul, A. James O’Malley, Thomas Keegan, and Nicholas A. Christakis.
2012. “Variation in Patient-Sharing Networks of Physicians across the United
States.” Journal of the American Medical Association 308 (3): 265–73.
Lauffenburger, Julie C., Jennifer G. Robinson, Christine Oramasionwu, and Gang
Fang. 2013. “Racial/ethnic and Gender Gaps in the Use and Adherence of
Evidence-Based Preventive Therapies among Elderly Medicare Part D
Beneficiaries after Acute Myocardial.” Circulation 129 (7): 754A – 763A.
http://circ.ahajournals.org/content/early/2013/12/10/CIRCULATIONAHA.11
3.002658.short.
Le, Phuc, and Michael B. Rothberg. 2015. “Cost-Effectiveness of Herpes Zoster
Vaccine for Persons Aged 50 Years.” Annals of Internal Medicine, [Epub ahead of
print]. doi:10.7326/M15-0093.
103

Le, Tao T., Andrew Bilderback, Bruce Bender, Frederick S. Wamboldt, Charles F.
Turner, Cynthia S. Rand, and Susan J. Bartlett. 2008. “Do Asthma Medication
Beliefs Mediate the Relationship between Minority Status and Adherence to
Therapy?” Journal of Asthma 45 (January): 33–37.
Lee, Tae Joon, Stella Hayes, Doyle M. Cummings, Qing Cao, Kristin Carpenter, Leah
Heim, and Hollie Edwards. 2013. “Herpes Zoster Knowledge, Prevalence, and
Vaccination Rate by Race.” Journal of the American Board of Family Medicine 26
(1): 45–51.
Lu, Peng-jun, Gary L. Euler, and Rafael Harpaz. 2011. “Herpes Zoster Vaccination
among Adults Aged 60 Years and Older, in the U.S., 2008.” American Journal of
Preventive Medicine 40 (2): e1–6.
Lu, Peng-jun, Gary L. Euler, Aisha O. Jumaan, and Rafael Harpaz. 2009. “Herpes
Zoster Vaccination among Adults Aged 60 Years or Older in the United States,
2007: Uptake of the First New Vaccine to Target Seniors.” Vaccine 27: 882–87.
Lu, Peng-jun, Alfonso Rodriguez-Lainz, Alissa O’Halloran, Stacie Greby, and Walter
W. Williams. 2014. “Adult Vaccination Disparities among Foreign-Born
Populations in the U.S., 2012.” American Journal of Preventive Medicine 47 (6):
722–33.
Lurie, Nicole, Nancy B. Ward, Martin F. Shapiro, Claudio Gallego, Rati Vaghaiwalla,
and Robert H. Brook. 1986. “Termination of Medi-Cal Benefits. A Follow-up
Study One Year Later.” New England Journal of Medicine 314 (19): 1266–68.
Mann, Devin M., Diego Ponieman, Howard Leventhal, and Ethan A. Halm. 2009.
“Predictors of Adherence to Diabetes Medications: The Role of Disease and
Medication Beliefs.” Journal of Behavioral Medicine 32 (3): 278–84.
Manski, Charles F. 1993. “Identification of Endogenous Social Effects: The Reflection
Problem.” The Review of Economic Studies 60 (3): 531–42.
104

Marcum, Zachary A, Mary Ann Sevick, and Steven M Handler. 2013. “Medication
Nonadherence: A Diagnosable and Treatable Medical Condition.” JAMA 309
(20): 2105–6.
Marmot, M.G., M.J. Shipley, and G. Rose. 1984. “Inequalities in Death—specific
Explanations of a General Pattern?” The Lancet 323 (8384): 1003–6.
Marmot, M.G., G.D. Smith, S. Stansfeld, C. Patel, F. North, J. Head, I. White, E. Brunner,
and A. Feeney. 1991. “Health Inequalities among British Civil Servants: The
Whitehall II Study.” The Lancet 337 (8754): 1387–93.
McEwen, Bruce S. 1998. “Protective and Damaging Effects of Stress Mediators.” New
England Journal of Medicine 338 (3): 171–79.
McInerney, Melissa, and Jennifer M. Mellor. 2012. “Recessions and Seniors’ Health,
Health Behaviors, and Healthcare Use: Analysis of the Medicare Current
Beneficiary Survey.” Journal of Health Economics 31 (5). Elsevier B.V.: 744–51.
Meer, Jonathan, Douglas L. Miller, and Harvey S. Rosen. 2003. “Exploring the Health-
Wealth Nexus.” Journal of Health Economics 22 (5): 713–30.
Michaud, Pierre-Carl, and Arthur van Soest. 2008. “Health and Wealth of Elderly
Couples: Causality Tests Using Dynamic Panel Data Models.” Journal of Health
Economics 27: 1312–25.
Monane, Mark, Rhonda L Bohn, Jerry H Gurwitz, Robert J Glynn, Raisa Levin, and
Jerry Avorn. 1996. “Compliance with Antihypertensive Therapy among Elderly
Medicaid Enrollees: The Roles of Age, Gender, and Race.” American Journal of
Public Health 86 (12): 1805–8.
Munshi, Kaivan. 2004. “Social Learning in a Heterogeneous Population: Technology
Diffusion in the Indian Green Revolution.” Journal of Development Economics
73: 185–213.
105

Newhouse, Joseph P. 1992. “Medical Care Costs: How Much Welfare Loss?” The
Journal of Economic Perspectives 6 (3): 3–21. doi:10.1257/jep.6.3.3.
Ngo-Metzger, Quyen, Dara H. Sorkin, John Billimek, Sheldon Greenfield, and Sherrie
H. Kaplan. 2012. “The Effects of Financial Pressures on Adherence and Glucose
Control among Racial/ethnically Diverse Patients with Diabetes.” Journal of
General Internal Medicine 27 (4): 432–37.
Osterberg, Lars, and Terrence Blaschke. 2005. “Adherence to Medication.” New
England Journal of Medicine 353 (5): 487–97.
Oxman, Michael N., Myron J. Levin, G. R. Johnson, K. E. Schmader, S. E. Straus, L. D.
Gelb, and R. D. Arbeit. 2005. “A Vaccine to Prevent Herpes Zoster and
Postherpetic Neuralgia in Older Adults.” New England Journal of Medicine 352
(22): 2271–84. http://www.nejm.org/doi/full/10.1056/NEJMoa051016.
Philipson, Tomas J. 1996. “Private Vaccination and Public Health : An Empirical
Examination for U . S . Measles.” The Journal of Human Resources 31 (3): 611–
30.
———. 2000. “Economic Epidemiology and Infectious Diseases.” In Handbook of
Health Economics, edited by Anthony J. Culyer and Joseph P Newhouse, 1761–
99. Elsevier B.V.
Pope, Gregory C., Randall P. Ellis, Arlene S. Ash, J. Z. Ayanian, D. W. Bates, H. Burstin,
L. I. Iezzoni, E. Marcantonio, and B. Wu. 2000. Diagnostic Cost Group
Hierarchical Condition Category Models for Medicare Risk Adjustment. Waltham,
MA. https://www.cms.gov/Research-Statistics-Data-and-Systems/Statistics-
Trends-and-Reports/Reports/Downloads/Pope_2000_2.pdf.
Purvis, Michael. 2011. “Sault Seniors Skip Long Shingles Vaccine Line in Canada and
Head to U.S.” Sault Star. November 9.
106

http://www.saultstar.com/2011/11/09/sault-seniors-skip-long-shingles-
vaccine-line-in-canada-and-head-to-us.
Ringel, Jeanne S., Susan D. Hosek, Ben A. Vollaard, and Sergej Mahnovski. 2002. The
Elasticity of Demand for Health Care: A Review of the Literature and Its
Application to the Military Health System. Santa Monica, CA.
http://www.rand.org/pubs/monograph_reports/MR1355.
Ruhm, Christopher J. 2000. “Are Recessions Good for Your Health?” The Quarterly
Journal of Economics 115 (2): 617–50.
Schwandt, Hannes. 2011. “Wealth Shocks and Health Outcomes: Evidence from
Stock Market Fluctuations.” http://fakultaetsseminar.uni-
mannheim.de/material/Hannes Schwandt Paper.pdf.
Seeman, Teresa E. 1997. “Price of Adaptation—Allostatic Load and Its Health
Consequences.” Archives of Internal Medicine 157 (19). American Medical
Association: 2259.
Sheiner, Louise. 2014. “Why the Geographic Variation in Health Care Spending
Cannot Tell Us Much about the Efficiency or Quality of Our Health Care System.”
Brookings Papers on Economic Activity, 1–67.
Shrank, William H, Niteesh K Choudhry, Joshua N Liberman, and Troyen A Brennan.
2011. “The Use of Generic Drugs in Prevention of Chronic Disease Is Far More
Cost-Effective than Thought, and May Save Money.” Health Affairs 30 (7): 1351–
57.
Skinner, Jonathan S. 2012. “Causes and Consequences of Regional Variations in
Health Care.” In Handbook of Health Economics, 2:45–93. Elsevier B.V.
Skinner, Jonathan S., and Douglas O. Staiger. 2005. Technology Adoption from Hybrid
Corn to Beta Blockers. Edited by Ernst E. Berndt and Charles R. Hulten. Hard-to-
107

Measure Goods and Services: Essays in Honor of Zvi Griliches. Cambridge, MA:
University of Chicago Press and National Bureau of Economic Research.
———. 2009. Technology Diffusion and Productivity Growth in Health Care. 14865.
National Bureau of Economic Research Working Paper Series. Cambridge, MA.
http://www.nber.org/papers/w14865.
Smith, James P. 1999. “Healthy Bodies and Thick Wallets: The Dual Relation between
Health and Economic Status.” The Journal of Economic Perspectives 13 (2): 144–
66.
Social Security Administration. 2005. “OASDI and SSI Program Rates & Limits,
2006.”
http://www.socialsecurity.gov/policy/docs/quickfacts/prog_highlights/Rates
Limits2006.html.
Stowasser, Till, Florian Heiss, Daniel McFadden, and Joachim Winter. 2011.
“‘Healthy, Wealthy and Wise?’ Revisited: An Analysis of the Causal Pathways
from Socioeconomic Status to Health".” In Investigations in the Economics of
Aging, 267–317. National Bureau of Economic Research, Inc.
Tamblyn, Robyn, Rejean Laprise, James A. Hanley, Michael Abrahamowicz, Susan
Scott, Nancy Mayo, Jerry Hurley, et al. 2001. “Adverse Events Associated with
Prescription Drug Cost-Sharing among Poor and Elderly Persons.” JAMA: The
Journal of the American Medical Association 285 (4): 421–29.
The Diabetes Control and Complications Trial Research Group. 1993. “The Effect of
Intensive Treatment of Diabetes on the Development and Progression of Long-
Term Complications in Insulin-Dependent Diabetes Mellitus.” New England
Journal of Medicine 329 (14): 977–86.
Tseng, Chien-Wen, Robert H. Brook, Emmett Keeler, W. Neil Steers, and Carol M.
Mangione. 2004. “Cost-Lowering Strategies Used by Medicare Beneficiaries
108

Who Exceed Drug Benefit Caps and Have a Gap in Drug Coverage.” JAMA: The
Journal of the American Medical Association 292 (8): 952–60.
Tseng, Chien-Wen, Edward F. Tierney, Robert B. Gerzoff, R. Adams Dudley, Beth
Waitzfelder, Ronald T. Ackermann, Andrew J. Karter, et al. 2008.
“Race/ethnicity and Economic Differences in Cost-Related Medication
Underuse among Insured Adults with Diabetes the Translating Research into
Action for Diabetes.” Diabetes Care 31 (2): 261–66.
United Healthcare. 2013. “Coding Summary for Providers.”
https://www.unitedhealthcareonline.com/ccmcontent/ProviderII/UHC/en-
US/Assets/ProviderStaticFiles/ProviderStaticFilesPdf/Tools and
Resources/Policies and Protocols/Medical Policies/Medical
Policies/preventive_care_services_coding_guideline_summary.pdf.
Viswanathan, Meera, Carol E. Golin, Christine D. Jones, Mahima Ashok, Susan J.
Blalock, Roberta C. M. Wines, Emmanuel J. L. Coker-Schwimmer, David L. Rosen,
Priyanka Sista, and Kathleen N. Lohr. 2012. “Interventions to Improve
Adherence to Self-Administered Medications for Chronic Diseases in the United
States: A Systematic Review.” Annals of Internal Medicine 157 (11). American
College of Physicians: 785–95. doi:10.7326/0003-4819-157-11-201212040-
00538.
Wellpoint. 2006. “WellPoint to Cover Varicella Zoster Virus, or Shingles Vaccine.”
http://ir.antheminc.com/phoenix.zhtml?c=130104&p=irol-
newsArticle&ID=922030.
Wennberg, John E., and Megan M. Cooper. 1996. The Dartmouth Atlas of Health Care
in the United States. Chicago, IL: American Hospital Publishing, Inc.
Wennberg, John, and Alan Gittelsohn. 1973. “Small Area Variations in Health Care
Delivery.” Science 182 (4117): 1102–8.
109

West of Scotland Coronary Prevention Study Group. 1997. “Compliance and Adverse
Event Withdrawal.” European Heart Journal 18 (11): 1718–24.
Wooldridge, Jeffrey M. 2010. Econometric Analysis of Cross Section and Panel Data.
MIT Press.
Yawn, Barbara P., Robbin F. Itzler, Peter C. Wollan, James M. Pellissier, Lina S. Sy,
and Patricia Saddier. 2009. “Health Care Utilization and Cost Burden of Herpes
Zoster in a Community Population.” Mayo Clinic Proceedings. Mayo Clinic 84 (9):
787–94. doi:10.4065/84.9.787.
Zhang, Yuting, and Seo Hyon Baik. 2014. “Race/ethnicity, Disability, and Medication
Adherence among Medicare Beneficiaries with Heart Failure.” Journal of
General Internal Medicine 29 (4): 602–7.
Ziebarth, Nicholas R, Maike Schmitt, and Martin Karlsson. 2013. “The Short-Term
Population Health Effects of Weather and Pollution: Implications of Climate
Change.” http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2377611.
Zillow. 2013. “Zillow Home Value Index.” http://cdn1.blog-
media.zillowstatic.com/3/ZHVI-InfoSheet-04ed2b.pdf.
———. 2014. “Zillow Real Estate Research: Data.” zillow.com/research/data.
Zissimopoulos, Julie, Geoffrey F. Joyce, Lauren M. Scarpati, and Dana P. Goldman.
2015. “Did Medicare Part D Reduce Disparities?” American Journal of Managed
Care 21 (2): 119–28. http://www.ajmc.com/publications/issue/2015/2015-
vol21-n2/Did-Medicare-Part-D-Reduce-Disparities.

110

Appendix A. Supplemental Table and Figures for Chapter 2
Figure A-1: Distribution of ZIP-level Population Counts over Time

Sample is limited to beneficiaries ages 65 and older with continuous Medicare Part
D coverage from 2006-2011 or from 2006 until death.

111

Figure A-2: Prices Paid by Patients over Time

Notes: Sample is limited to beneficiaries ages 65 and older with continuous
Medicare Part D coverage from 2006-2011 or from 2006 until death.

112

Figure A-3: Geographic Variation in Patient Cost-Sharing for Zostavax

Information on the expected amount of patient cost-sharing for Zostavax was not
available until 2010. Understanding geographic variation in cost-sharing
requirements is helpful in understanding resulting adoption patterns. The top figure
shows the distribution of the average expected level of patient cost-sharing within
ZIP codes for Zostavax. The bottom figure shows the average expected level of
patient cost-sharing within hospital service areas. In both images, it is clear that
there is a considerable amount of geographic variation in cost-sharing which may be
driving take-up. Both figures show cost-sharing in the pre-gap phase (during which
two-thirds of patients received Zostavax and in which most patients spend the
calendar year) and assume a total cost of Zostavax of $175 where plans specified co-
pays instead of co-insurance. Future work should study the effects of patient cost-
sharing on the diffusion of new drugs; researchers using Part D data should focus on
drugs approved in 2010 or later.
113

Figure A-4: Vaccination Rates by Income Quintile

Notes: Sample is limited to beneficiaries ages 65 and older with continuous
Medicare Part D coverage from 2006-2011 or from 2006 until death. Income
quintiles constructed from the American Community Survey median household
income ZIP-level variable.

114

Figure A-5: Age at the Time of Vaccination

Notes: Sample is limited to beneficiaries ages 65 and older with continuous
Medicare Part D coverage from 2006-2011 or from 2006 until death.

115

Table A-1: Comparison of Individual-Level Summary Statistics

Unvaccinated Vaccinated
VARIABLES Mean SD Mean SD

Average Age 75.9* 7.86 72.7 6.05
Average HCC Score 0.01* 0.63 -0.33 0.52
Fraction Female 0.63* 0.66
Fraction Black 0.10* 0.02
Fraction Asian 0.03* 0.06
Fraction Hispanic 0.08* 0.03
Fraction Medicare Advantage 0.29 0.29
Fraction LIS 0.05* 0.03
Fraction Dual-Eligibles 0.26* 0.12

Number of observations 3,126,057 164,399

VARIABLES Unvaccinated (%) Vaccinated (%)

Females 1,955,805 (94.7%) 108,750 (5.3%)
Males 1,170,252 (95.5%) 55,649 (4.5%)
Black 314,190 (99.0%) 3,265 (1.0%)
Asian 104,081 (91.4%) 9,747 (8.6%)
Hispanic 261,201 (97.9%) 5,541 (2.1%)
Traditional Fee-For Service 2,209,823 (95.0%) 116,619 (5.0%)
Medicare Advantage 916,234 (95.0%) 47,780 (5.0%)
LIS 166,685 (96.9%) 5,311 (3.1%)
Dual-Eligibles 823,039 (97.6%) 20,397 (2.4%)

Number of observations 3,126,057 (95%) 164,399 (5%)
Notes: Data for 2006. “Vaccinated” refers to whether the patient received Zostavax
at any point between FDA approval in 2006 through 2011. The top table provides
descriptive statistics for vaccinated vs. unvaccinated groups. The bottom table
presents the proportion within each sub-group (ethnic, insurance, sex, LIS/dual-
eligible status) that was vaccinated. Sample is limited to beneficiaries ages 65 and
older with continuous Medicare Part D coverage from 2006-2011 or from 2006 until
death.
*indicates difference between values is statistically significant at 1%

116

Table A-2: Unweighted Summary Statistics by Adoption Rate

Low Adoption High Adoption

Mean (SD) Mean (SD)
ZIP-Level Variables

ZIP-Level Vaccination Rate 0.59 (1.1)* 13 (12)
ZIP-Level Quarterly Mortality Rate 1.7 (5.8) 1.6 (2.7)
Percent LIS 6.7 (12)* 5.5 (6.3)
Percent Dual-Eligibles 29 (26)* 21 (16)
Average Age 77.9 (3.21)* 78.1 (1.85)
Percent Medicare Advantage 26 (27) 27 (23)
Percent Female 61 (22)* 63 (12)
Percent Black 9.5 (21)* 4.8 (12)
Percent Asian 1.1 (4.8)* 1.6 (5.4)
Percent Hispanic 5.8 (16)* 3.3 (9.2)
Average HCC Score 0.080 (0.30)* 0.045 (0.17)
Zip Population Count 50.7 (105)* 97.9 (128)
Average Zostavax Price $57 ($9)* $58 ($21)
Quarterly Shingles Prevalence Rate 0.41 (1.7)* 0.47 (1.1)

PCSA-Level Variables
Physician Social Network Size 0.98 (0.096) 0.98 (0.091)
Percent Generalists 40 (11) 40 (10)
Mean Patients per PCP 32.0 (29.1) 31.5 (28.1)

Number of ZIP Codes 10,675 10,604

*Indicates values for high- and low-vaccination groups are significantly different at
1%
**HCC stands for Hierarchical Condition Category. These measures are used for risk-
adjustment purposes. Higher scores indicate sicker individuals.
*** PCP stands for Primary Care Physician.
Notes: Summary statistics above were derived using data for 2011. Sample is
limited to beneficiaries ages 65 and older with continuous Medicare Part D coverage
from 2006-2011 or from 2006 until death.

117

Table A-3: Models without Cumulative Adoption Rates
One may be concerned that the lagged level and squared cumulative
adoption rates, 𝛾 𝑉 𝑧𝑝𝑡 −1
and 𝜉 𝑉 𝑧𝑝𝑡 −1
2
, may overemphasize persistence in adoption. I
estimate Equation 2 without these two variables:
(𝐴 1) 𝑣 𝑧𝑝𝑡 = 𝜙 𝐹 𝑝𝑡
+ 𝜋 log (𝑝𝑟𝑖𝑐𝑒 𝑧𝑝𝑡 ) + 𝛿 𝑠 ℎ𝑖𝑛𝑔𝑙𝑒𝑠 𝑧𝑝𝑡 + 𝑿 𝒛𝒑𝒕 𝜷 + 𝝍 𝑮 𝒑𝒕
+ 𝜆 𝑡 + 𝛼 𝑧 + 𝘀 𝑧𝑝𝑡
(i) (ii)

Physician Network Size 0.000447** 0.0000821

(2.99) (0.56)
Percent Medicare Advantage 0.000539 0.000988***

(1.77) (3.47)
Log(Average ZIP Price) -0.000241*** -0.000244***

(-26.25) (-26.39)
Shingles Prevalence 0.0248*** 0.0253***

(12.25) (12.61)

Lagged Level & Squared
Cumulative Adoption Rate YES NO
ZIP Fixed Effects YES YES
Quarter Fixed Effects YES YES
ZIP-Quarter Observations 552277 577550
R-Squared 0.078 0.066
Number of ZIPs 25511 25511
Column (i) contains the results from Table 3 for comparison. The results are similar,
though the estimated physician network size coefficient becomes smaller and
statistically insignificant. This does not materially change the results and
conclusions of the paper.

118

Appendix B. Supplemental Table and Figures for Chapter 3

Table B-1: Preventive Services Procedure Codes
Cancer screenings: G0101, G0123, G0124, G0141, G0143, G0144, G0145, G0147,
G0148, Q0091, P3000, P3001, 88141, 88142, 88143, 88147, 88148, 88150, 88152,
88153, 88154, 88155, 88164, 88165, 88166, 88167, 88174, 88175, G0104, G0105,
G0106, G0120, G0121, G0122, G0328, 44388, 44389, 44392, 44393, 44394, 45330,
45331, 45333, 45338, 45339, 45378, 45380, 45381, 45383, 45384, 45385, 82270,
82274, 74263, G0102, G0103, 84152, 84153, 84154, G0202, 77052, 77057

Lipid screenings: 80061, 82465, 83718, 83719, 83721, 84478

Diabetes tests: 82947, 82948, 82950, 82951, 82952, 83036

Influenza shots: 90654, 90655, 90656, 90657, 90658, 90660, 90661, 90662,
90664, 90666, 90667, 90668, 90672, 90673, 90685, 90686, 90688, Q2034, Q2035,
Q2036, Q2037, Q2038, Q2039

Osteoporosis screenings: 76977, 77078, 77080, 77081, G0130

Depression screenings: 99420, G0444

Wellness exams: G0402, G0438, G0439, G0445, S0610, S0612, S0613, 99381,
99382, 99383, 99384, 99385, 99386, 99387, 99391, 99392, 99393, 99394, 99395,
99396, 99397, 99401, 99402, 99403, 99404, 99411, 99412, 99461

Source: (United Healthcare 2013)

119

Appendix C. Supplemental Table and Figures for Chapter 4
Figure C-1. Cessation of Diabetes Drug Classes in the Coverage Gap in 2007

Figure C-2. Resumption of Diabetes Drug Classes in 2008 if Stopped in the
Coverage Gap in 2007

-0.05 0.00 0.05 0.10 0.15
Antilipemic Agents $54
Antihypertensive Combinations $58
Oral Hypoglycemic Agents $50
Calcium Channel Blockers $46
Ace/Angiotensin II Inhibitors $31
Beta Blockers $27
Loop Diuretics $8
Digitalis Glycosides $7
White
Black
Hispanic
-0.20 -0.10 0.00 0.10 0.20 0.30
Antilipemic Agents $54
Antihypertensive Combinations $58
Oral Hypoglycemic Agents $50
Calcium Channel Blockers $46
Ace/Angiotensin II Inhibitors $31
Beta Blockers $27
Loop Diuretics $8
Digitalis Glycosides $7
White
Black
Hispanic
120

Figure C-3. Cessation of Non-Diabetes Drug Classes in the Coverage Gap in
2007

Figure C-4. Resumption of Non-Diabetes Drug Classes in 2008 if Stopped in the
Coverage Gap in 2007

-0.10 -0.05 0.00 0.05 0.10 0.15
Antipsychotics $213
CNS Medications $150
Antiasthma $127
Platelet Aggregation Inhibitors $123
Antiulcerants $108
Hormones/Synthetics/Modifiers $90
Anticonvulsants $61
Opioid Analgesics $53
Antidepressants $49
White
Black
Hispanic
-0.20 -0.10 0.00 0.10 0.20 0.30 0.40
Antipsychotics $213
CNS Medications $150
Antiasthma $127
Platelet Aggregation Inhibitors $123
Antiulcerants $108
Hormones/Synthetics/Modifiers $90
Anticonvulsants $61
Opioid Analgesics $53
Antidepressants $49
White
Black
Hispanic
121

Figure C-5. Percent Changes in Medication Use upon Entering the Coverage
Gap: Diabetes Drugs

Figure C-6. Percent Changes in Medication Use upon Entering the Coverage
Gap: Non-Diabetes Drugs

-0.30 -0.25 -0.20 -0.15 -0.10 -0.05 0.00 0.05 0.10
Antilipemic Agents $65
Antihypertensive Combinations $58
Oral Hypoglycemic Agents $50
Calcium Channel Blockers $46
Ace/Angiotensin II Inhibitors $31
Beta Blockers $27
Loop Diuretics $8
Digitalis Glycosides $7
Diabetes-Related Drug Classes
White
Black
Hispanic
-0.30-0.25-0.20-0.15-0.10-0.05 0.00 0.05 0.10
Antipsychotics $213
CNS Medications $150
Antiasthma $127
Platelet Aggregation Inhibitors $123
Antiulcerants $108
Hormones/Synthetics/Modifiers $90
Anticonvulsants $61
Opioid Analgesics $53
Antidepressants $49
Nondiabetes-Related Classes
White
Black
Hispanic
122

Figure C-7. Percent Changes in Medication Use of the Near-Poor Population
upon Entering the Coverage Gap: Diabetes Drug Classes

Figure C-8. Percent Changes in Medication Use of the Near-Poor Population
upon Entering the Coverage Gap: Non-Diabetes Drug Classes

-0.40 -0.30 -0.20 -0.10 0.00 0.10 0.20
Antilipemic Agents $65
Antihypertensive Combinations $58
Oral Hypoglycemic Agents $50
Calcium Channel Blockers $46
Ace/Angiotensin II Inhibitors $31
Beta Blockers $27
Loop Diuretics $8
Digitalis Glycosides $7
Diabetes-Related Drug Classes
White
Black
Hispanic
-0.40 -0.30 -0.20 -0.10 0.00 0.10 0.20
Antipsychotics $213
CNS Medications $150
Antiasthma $127
Platelet Aggregation Inhibitors $123
Antiulcerants $108
Hormones/Synthetics/Modifiers $90
Anticonvulsants $61
Opioid Analgesics $53
Antidepressants $49
Nondiabetes-Related Classes
White
Black
Hispanic
123

Figure C-9. Percentage Point Changes in Use of Generic Drugs upon Entering
the Coverage Gap by Drug Class*

*Classes shown in the Appendix Figure are those such that the drug class is neither
brand (GDR < 20%) nor generic dominated (GDR > 80%)

-0.15
-0.10
-0.05
0.00
0.05
0.10
White
Black
Hispanic
124

Table C-1. Beneficiaries’ Plan Switching from Year 2007 to 2008

White Black Hispanic
LIS
(n=62,342)
Non-LIS
(n=102,898)
LIS
(n=21,179)
Non-LIS
(n=4,457)
LIS
(n=24,325)
Non-LIS
(n=3,688)
Switched Plans, % 27.0 14.0 28.1 11.8 39.6 11.4
Change in OOP, $ -12.65 -153.0 -7.22 -198.8 9.3 -167.4
Change in OOP, % -2.8 -7.4 -1.7 -9.6 2.7 -7.6
Do Not Switch Plans, % 73.0 86.0 71.9 88.2 60.4 88.6
Change in OOP, $ 16.67 -51.5 7.82 -98.4 16.44 -82.44
Change in OOP, % 3.4 -2.3 1.6 -4.8 4.1 -4.0
Plan Change Type
Added gap coverage, % 0.2 16.5 0.2 14.7 0.1 20.1
Lowered premium, % 32.6 25.4 47.1 25.8 28.9 18.2
Lowered deductible, % 17.2 25.8 17.8 32.8 24.6 27.0
Other, % 50.0 32.3 34.9 26.7 46.4 34.6

NOTES:
Sample is individuals with diabetes and ages 65 and older that reached coverage gap in 2007. Individuals who
switched from LIS to non-LIS status or vice versa are excluded from this analysis. Changes are from year 2007 to
year 2008. Changes in out-of-pocket costs (OOP) include premiums.

Abstract (if available)

Abstract Medicare has provided health insurance for elderly Americans, as well as the disabled and those with end-stage renal disease, since 1966. Among this population, I investigate the demand for health care, including prescription drugs and preventive services, as well as the relationship between wealth and health, using a large claims dataset. ❧ In Chapter 2, I investigate the diffusion of a new medical technology—Zostavax, the shingles vaccine for the elderly—and investigate the effect of socio-demographic factors, as well as supply- and demand-side social influences on its adoption. I find that treatment patterns in small areas ultimately drive take-up. Observed factors are unable to explain regional differences in adoption. Physician social networks have minimal effect on Zostavax diffusion and patients’ behavior is largely prevalence-inelastic. Patient demand is highly price-elastic. ❧ In Chapter 3, with Dana Goldman, Florian Heiss, Daniel McFadden, Joachim Winter, and Amelie Wuppermann, I use the housing crisis as a natural experiment to evaluate the health-wealth gradient, wherein the affluent are healthier than those with fewer financial resources. We use plausibly exogenous changes in five-digit level ZIP code average home prices to evaluate the effect of wealth on health. Beneficiaries respond to changes in wealth by decreasing their use of medical services

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Essays in health economics and provider behavior

PDF

Three essays on health economics

PDF

Three essays in health economics

PDF

Essays on development economics

PDF

Essays on health economics

PDF

Essays in the economics of education and conflict

PDF

Essays on development and health economics: social media and education policy

PDF

The value of novel antihyperlipidemic treatments in the U.S. healthcare system: Reducing the burden of cardiovascular diseases and filling the gap of low adherence in statins

PDF

Essays on health and aging with focus on the spillover of human capital

PDF

Essays on economics of education

PDF

Three essays on economics of early life health in developing countries

PDF

Panel data forecasting and application to epidemic disease

PDF

Essays on economics of education and private tutoring

PDF

Essays on family and labor economics

PDF

Long-term impacts of childhood adversity on health and human capital

PDF

Essays on competition and antitrust issues in the airline industry

PDF

Behavioral approaches to industrial organization

PDF

Three essays on behavioral economics approaches to understanding the implications of mental health stigma

PDF

Delivering better care for children with special health care needs: analyses of patient-centered medical home and types of insurance

PDF

Two essays in economic policy: The influence of perceived comparative need on financial subsidy requests; &, Unintended consequences of the FOSTA-SESTA legislation

Asset Metadata

Creator Scarpati, Lauren Matsunaga (author)

Core Title Essays in health economics: evidence from Medicare

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Publication Date 10/16/2017

Defense Date 10/01/2015

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag Economics,geographic variation,health,Health Economics,health-wealth gradient,home prices,housing crisis,Medicare,OAI-PMH Harvest,shingles,technology adoption,technology diffusion,vaccines

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Goldman, Dana P. (committee chair), McFadden, Daniel L. (committee member), Ridder, Geert (committee member), Strauss, John A. (committee member), Zissimopoulos, Julie M. (committee member)

Creator Email lauren.scarpati@gmail.com,scarpati@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c40-191055

Unique identifier UC11279360

Identifier etd-ScarpatiLa-3977.pdf (filename),usctheses-c40-191055 (legacy record id)

Legacy Identifier etd-ScarpatiLa-3977-1.pdf

Dmrecord 191055

Document Type Dissertation

Format application/pdf (imt)

Rights Scarpati, Lauren Matsunaga

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA