University of Southern California Dissertations and Theses

The impact of adherence to guidelines on the health care expenditures of COPD patients

(USC Thesis Other)

The impact of adherence to guidelines on the health care expenditures of COPD patients

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content !"#$%&'()!$*+$(,"#-#.)#$!*$/0%,#1%.#2$*.$!"#$"#(1!"$)(-#$
#3'#.,%!0-#2$*+$)*',$'(!%#.!2
45$
+67897$#:;5<=>9?;
($,9@@ABC7C9=D$'BA@ADCAE$C=$CFA
+()01!G$*+$!"#$02)$/-(,0(!#$2)"**1
0.%H#-2%!G$*+$2*0!"#-.$)(1%+*-.%(
%D$'7BC976$+I6J966KADC$=J$CFA
-ALI9BAKADC@$J=B$CFA$,AMBAA
,*)!*-$*+$'"%1*2*'"G
N'"(-&()#0!%)(1$#)*.*&%)2$(.,$'*1%)GO
&75 PQRS
2

CONTENTS
LIST OF TABLES ................................................................................................................................ 4
LIST OF FIGURES .............................................................................................................................. 5
ABSTRACT ........................................................................................................................................ 6
DEDICATION .................................................................................................................................... 8
ACKNOWLEDGEMENTS ................................................................................................................... 9
CHAPTER 1: INTRODUCTION ......................................................................................................... 11
CHAPTER 2: BACKGROUND ........................................................................................................... 14
2.1 Epidemiology and Burden ............................................................................................... 14
2.2 Pharmacotherapy for Chronic Obstructive Pulmonary Disease ...................................... 15
2.3. GOLD guidelines ............................................................................................................. 17
2.4. COPD Severity ................................................................................................................. 19
2.5. Comorbidities ................................................................................................................. 23
2.7. Heath Care Utilization .................................................................................................... 30
2.8. Other COPD Related Topics ............................................................................................ 33
CHAPTER 3: STUDY DESIGN ........................................................................................................... 38
3.1 Motivation ........................................................................................................................... 38
3.2. Data Source ........................................................................................................................ 39
3.2.1. Data Cleaning ............................................................................................................... 41
3.3. Time Frame ......................................................................................................................... 41
3.5. Research Question.............................................................................................................. 44
CHAPTER 4: METHODOLOGY ......................................................................................................... 46
4.1. Medication Therapy ........................................................................................................... 46
4.2. Health Outcomes ................................................................................................................ 47
4.3. Independent Variables ....................................................................................................... 47
4.4. Conceptual Framework ...................................................................................................... 49
4.5. Methodological Issues ........................................................................................................ 50
4.6. Data Analysis ...................................................................................................................... 52
4.6.1. Dynamic Panel Model .................................................................................................. 55
4.6.2. Two-‐Part Model ........................................................................................................... 61
4.7. Statistical analysis ........................................................................................................... 63
3

CHAPTER 5: RESULTS ..................................................................................................................... 65
5.1 Dynamic Panel Data Results ................................................................................................ 73
5.2 Two-‐Part Model Results ...................................................................................................... 83
CHAPTER 6: DISCUSSION AND CONCLUSION ................................................................................ 87
6.1 Discussion ............................................................................................................................ 87
6.2 Limitations ........................................................................................................................... 89
6.3 Conclusion ........................................................................................................................... 90
CHAPTER 7: REFERENCES............................................................................................................... 92

4

LIST OF TABLES
Table 1: The Pharmacological Therapies available for COPD. ........................................ 16
Table 2: Classification of COPD by severity. ................................................................... 20
Table 3: Inclusion and Exclusion Criteria. ....................................................................... 44
Table 4: Group classification according to their appropriate therapy usage. ................... 46
Table 5: The description of each observable variable ݔ ௜௝
ሺ ݐ ሻ for ݆ ൌ ͳ ǡ ʹ ǡ ǥ ǡ ݇ . ............... 57
Table 6: Group classification according to their appropriate therapy usage. ................... 66
Table 7: Demographic characteristics of the patient-quarter in each group. .................... 67
Table 8: Evaluation of the factors that impact appropriate therapy usage. ....................... 70
Table 9: Unadjusted outcomes reported for each group. .................................................. 72
Table 10: Dynamic panel model including and excluding certain instruments. ............... 75
Table 11: Dynamic panel model: all patients versus non-asthmatic patients. .................. 77
Table 12: All patients vs . alive patients. ........................................................................... 79
Table 13: No oxygen therapy patients. ............................................................................. 81
Table 14: No dual eligible patients .................................................................................. 82
Table 15: Probit analysis to evaluate the probability of dying. ........................................ 84

!
!"#$%&'%'"()*+#
"#$%&'()*(+,-.(+%#/'0#1'2(3 4'5677'1/'/(89'&:;<(;:88'&1(=6&(':59(28:$'(6=(>,?.@@)A
"#$%&'(B*(C8%/<(;'/@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@DB
"#$%&'(E*( F9'(9':089(5:&'(76/'0@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@!G
"#$%&'(D*(F9'(#7;:58(6=(89'(56H:&#:8'2(61(/':89(:1/(61(89'(:/9'&'15'(86($%#/'0#1'2@@@@@@!!
"#$%&'(!*(F9'(#150%2#61(5'&#:(#1(6%&(28%/<@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@I!
"#$%&'(I*(F9'(;&'H:0'15'(6=(89'(5676&J#/#8#'2(#1(':59($&6%;@ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@IK
"#$%&'(A*(L/M%28'/(5628(H'&2%2(%1:/M%28'/(5628(;'&(N%:&8'&3;:8#'18@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@AE
"#$%&'(K*(F9'(;&'/#58'/(5628(;'&(N%:&8'&3;:8#'18(#1(89'(8O63;:&8(76/'0@@@@@@@@@@@@@@@@@@@@@@@@@@@K!
"#$%&'(P*(F68:0(5628(567;:ą(J'8O''1(8O63;:&8(76/'0(:1/(%1:/M%28'/(&'2%082@@@@@@@@@@@KI
6

ABSTRACT
OBJECTIVES: This study proposes to investigate the association among adherence to
guidelines, patients characteristics, and health care expenditures using methods for
handling endogeneity and selection bias.
METHODS: The study population consists of the random 5% COPD cohort in Medicare
for the time period between 2006 and 2008. Patients are classified in 4 groups based on
the percentage of use of appropriate therapy based on the GOLD guidelines among the
quarters: Group I) 0% - 25%; Group II) 25% - 50%; Group III) 50% - 75%; and Group
IV) 75% - 100%. In the dynamic panel data, the time unit of the panel model is set to
three months (quarter). The two part model computes the probability of a patient
surviving at a given time period and then computes the expected health care expenditures
of that patient, given that he/she has survived.
RESULTS: The unadjusted difference in cost between Group I and Group IV is $1,220
per quarter-patient. After adjusting for all covariates and also for the instrumental
variables, the difference is even more significant using the dynamic panel model. The
adjusted costs of Groups I and IV are $10,451 and $6,317 (p-value < 0.001), respectively.
Using a two-part model, the difference among the four groups is also significant (p-value
< 0.05), with Group I having a total health-care expenditure of $7,560 per patient-quarter
and Group IV having a cost of only $4,962, resulting in 34% savings per patient-quarter
(p-value < 0.01).
CONCLUSIONS: Different approaches are used to control for endogeneity and
selection bias, which show that the results of the benefits of appropriate therapy are
7

consistent. Our findings suggest that the treatment recommended by the GOLD
guidelines for moderate-to-severe COPD patients leads to a significant reduction in
health care expenditures.

8

DEDICATION

To my lovely husband
Rafael,
for his endless love and support.
No words can describe how lucky and thankful I am for having you in my life.

9

ACKNOWLEDGEMENTS
Many people were responsible for my accomplishments in these past four years. I
would like first to thank my advisor, Dr. Joel Hay, for his invaluable knowledge and for
the time he spent teaching and advising me through this thesis. I feel blessed for having
an opportunity to learn from him. I also feel lucky for having Dr. Jeffrey McCombs in
our department, a professor who is always willing to help students in every possible way.
I have always been happy to know that I could reach Dr. J. at anytime. I would also like
to thank professors Vivian Wu and John Romley for being in my qualifying exam and my
dissertation committee and I greatly appreciate their valuable feedback.
I wish I could say this in person, but unfortunately I can no longer hug her and say
how thankful I am for Dr. Kathleen Johnson being part of my life. She was a mentor and,
without a doubt, I would not be here today had she not encouraged and advised me the
way she did.
Last but not least, I would like to give a special thanks to my family and friends.
To the most important person in my life, my husband, who is the most encouraging
person through this process. Without him, I would not be here today and I would not
accomplish everything that I did. To my parents Telma and Jacob, who are always by my
side, giving me unconditional love and support. To my brother Pedro, who sometimes
forgets he is only my older brother and not my father, for always knowing what to say
and for giving me the best advices. I also have to thank him for giving me the best
nephew and niece, Breno and Luana, bringing so much joy and happiness to my life.
Additionally, I would like to give a big thanks to Janet Shin for her support, time, and
friendship. Without her, the research and the long study hours would be way less
10

pleasant. And finally, to all my friends from Brazil and US, for always being there for me
and for making this long journey much more pleasant and enjoyable.

11

CHAPTER 1: INTRODUCTION
Chronic obstructive pulmonary disease (COPD) is characterized by airflow
limitations which are not fully reversible. These limitations are usually progressive and
associated with an abnormal inflammatory response of the lungs to noxious particles or
gases (Rennard 1998). Smoking is the most significant risk factor for the development of
COPD, but other important risk factors include asthma, exposure to environmental
pollution, hereditary deficiency of antitrypsin, and lower socioeconomic status.
Individuals diagnosed with bronchitis, chronic bronchitis, chronic obstructive bronchitis,
or emphysema through the course of a lifetime are categorized as suffering from COPD
(Pauwels 2001). The clinical definition of each of these conditions is different, although
all of the above, except simple bronchitis, are characterized by airflow limitation. Even
though COPD is a preventable and treatable disease, patients still experience a
progressive decline in lung function, which worsens exercise capacity with increasing
dyspnea and frequent exacerbations (Puchelle and Vargaftig 2001).
The Global Initiative for Chronic Obstructive Lung Disease (GOLD) guidelines
were launched in 1997 as an evidence-based approach to the treatment and management
of COPD. These guidelines were developed through collaborations between leading
experts in COPD and renowned organizations, including the National Heart, Lung, and
Blood Institute (NHLBI), the National Institutes of Health (NIH), and the World Health
Organization (WHO). In 2010, a comprehensively updated version of these guidelines
was issued to include newer therapy regimes. Other COPD guidelines have also been
published by professional societies, including the American Thoracic Society and the
12

European Respiratory Society, with the common goal of improving the prevention,
diagnosis, and management of COPD.
After publishing these guidelines, it was expected that the management of COPD
would significantly improve. For instance, COPD awareness and prevention were
expected to increase, eventually leading to improvements in diagnostic abilities and
decreases in morbidity as well as mortality. However, despite the benefits of the proposed
guidelines, adherence to both guidelines and medication is still a major problem among
COPD patients. Adherence to medication is known to be around 40% (Dolce et al. 1991)
and adherence to guidelines is expected to be even lower, around 25% (Asche et al.
2008, Salinas et al. 2011). Non-adherence in COPD patients is expected to impact both
healthcare expenditure and utilization; however, to our knowledge, there are no studies
relating guideline adherence to health outcomes in moderate-to-severe COPD in
Medicare patients. Recently, it has been shown that potential cost savings can be
achieved by following the current GOLD guidelines recommendations in COPD patients
being treated with the combination of long- D F W LQJ ȕ -agonist (LABA), long-acting
muscarinic antagonist (LAMA) or inhaled corticosteroids (ICS) using the Geisinger
Health System (GHS) database (Asche et al. 2012). Using a different approach, other
studies compare the health outcomes impact of two specific therapies (Simoni-Wastila et
al. 2009, Dalal et al. 2009, Dalal, Candrilli and Davis 2011a), but not the impact of
guideline adherence on healthcare utilization and cost. Our study intends to bridge this
gap and evaluate the importance of guideline adherence for COPD patients, estimating
the cost and utilization consequences of an inadequate therapy (i.e., a therapy not
13

recommended by the guidelines). In particular, we want to evaluate how copayments
indirectly affect healthcare outcomes through adherence to guidelines.
The remainder of this dissertation is organized as following. Chapter 2 presents
the definitions, epidemiology, COPD treatments and guidelines, economic impact, health
care utilization, related comorbidities, and disease severity. Chapter 3 describes the study
design and the data source. Chapter 4 explains the conceptual framework and
econometric identification strategies. Chapter 5 presents the results, including tables and
graphs for data. Finally, chapter 6 concludes and discusses study results.

14

CHAPTER 2: BACKGROUND
This chapter describes the background on COPD, starting with epidemiology and
the burden of the disease. In the next sections, we describe the literature on COPD
severity, common comorbidities, the pharmacologic treatments available for COPD, the
related treatment costs, and health utilization.
Our literature review used the electronic databases Cochrane, Ovid, Google
Scholar, and PubMed to search papers published between 1960 and 2012. Much of the
research data regarding COPD involves primarily medical quantitative studies using
randomization or cross-sectional methods, which were completed in the United States
and in various counties worldwide. This review comprises different research topics
regarding COPD, including its economic burden, health care utilization, disease
complications, and comorbidities.

2.1 Epidemiology and Burden
Even though COPD has been treatable for a long time, it still has a high mortality
rate. It has been projected that by 2020 COPD would become the third leading cause of
mortality and the fifth leading cause of disability worldwide (Michaud, Murray and
Bloom 2001). Unfortunately, this projection already became a reality in 2008, 12 years in
advance of the projected year. In 2005, more than 126,000 deaths were attributed to
COPD in the United States (Mannino et al. 2002).
Statistics show that an estimated 13 million Americans suffer from COPD,
approximately 1 in 25 Americans (2011). Furthermore, it is also estimated that COPD
contributed to 8 million physician office and hospital outpatient visits, 1.5 million
15

emergency department visits, and 726 thousand hospitalizations in 2000. The annual cost
of COPD is estimated to be nearly $30 billion, including direct and indirect costs
(Mannino et al. 2002). Despite these large amounts, the morbidity and mortality rates as
well as the costs associated with COPD may still be underestimated, as the disease is not
usually diagnosed until clinically apparent and moderately advanced, and is more likely
to be cited as a contributory factor rather than an underlying cause of death.

2.2 Pharmacotherapy for Chronic Obs tructive Pulmonary Dis eas e
Available treatments for COPD include pharmacological therapy, acute
care/hospitalization, lifestyle modifications, and the use of invasive procedures (Pauwels
2001).
Pharmacologic therapies for different levels of COPD employ bronchodilators
(long-acting and short-acting ȕ 2-agonists, anticholinergics, methylxanthines),
glucocorticosteroids (inhaled as well as oral), and antibiotics (Pauwels et al. 2001). Table
1 summarizes the list of available therapies for COPD.
Bronchodilator medications are considered central to the management of airway
obstruction. The use of bronchodilators, especially anticholinergics and ȕ 2-agonists,
reduces airway smooth muscle tone and airflow resistance, which results in significant
improvements in several clinical outcomes, such as dyspnea, quality of life, and exercise
capacity. Current treatment guidelines recommend the use of long-acting bronchodilators,
such as long- D F WL Q J ȕ -agonist and anticholinergic, in symptomatic patients with
moderate-to-severe airflow obstruction. It is well-known that treatments with long-acting
bronchodilators, including nebulized formulations, are more effective and convenient
16

than treatment with short-acting bronchodilators (Vincken et al. 2002, Mahler et al. 1999,
Dahl et al. 2001). However, it is common to observe moderate-to-severe patients being
treated with short-acting ȕ 2-agonists instead, which is not recommended by the
guidelines.
Table 1: The pharmacological therapies available for COPD.
Anticholinergics
1) Ipratropium or
Ipratropium/Albuterol
combination
2) Tiotropium
Long-acting Beta Agonis ts
1) Arformoterol
2) Formoterol
3) Salmeterol
Fixed dos e Combination Therapy
(FDCT)
1) Fluticasone + Salmeterol

Inhaled Corticos teroids
1) Beclomethasone
2) Budesonide
3) Flunisolide
4) Fluticasone
5) Triamcinolone
Sho rt-Acting, Inhaled Beta-2
Agonis t
1) Albuterol
2) Bitolterol
3) Isoetharine
4) Isoproterenol
5) Levalbuterol
6) Metaproterenol
7) Pirbuterol
8) Terbutaline
Oral Corticos teroids
1) Betamethasone
2) Cortisone
3) Dexamethasone
4) Hydrocortisone
5) Methylprednisolone
6) Prednisolone
7) Prednisone
8) Triamcinolone
Xanthines
1) Aminophylline
2) Dyphylline
3) Oxtriphylline
4) Theophylline

Antibiotics for Res piratory Infections
1) Macrolides (azithromycin, clarithromycin,
dirithromycin, erythromycin)
2) Respiratory fluoroquinolones (ciprofloxacin,
gatifloxacin, levofloxacin, moxifloxacin)
3) Cephalosporin (cephalexin, cefaclor, cefadroxil,
cefdinir, cefditoren, cefepime, cefixime,
cefotaxime, cefpodoxime, cefprozil, ceftazidime,
ceftibuten, ceftriaxone, cefuroxime)
4) Trimethoprim-Sulfamethoxazole
5) Tetracycline derivatives (doxycycline)
6) Penicillins (amoxcillin, ampicillin)
17

2.3. GOLD guidelines
There are several therapies for COPD treatment in the GOLD guidelines. Figure 1
shows the recommended therapies, with each therapy being recommended for a different
disease stage. Several studies evaluate the difference in health outcomes among therapies.
However, most researchers compare two or three therapies in administrative claims
(Dalal et al. 2009, Akazawa et al. 2008b), and only a few compare all possible
simultaneously therapies in one study (Mannino et al. 2002, Michaud et al. 2001)

Figure 1: GOLD Guideline s - Recommended therapy pattern for each s tage of COPD.

Source: GOLD Guidelines, 2010
18

Dalal et al. (Dalal et al. 2009) compare patients treated with fluticasone
SU RSLRQDWH VDOP H WHU RO ȝ J ȝ J ) 6 & ZLWK RWKHU SD WL H QWV WUH D WHG ZLWK LSUD WURSLXP bromide/albuterol (IPA), ipratropium bromide (IPR), and tiotropium bromide (TIO). The
work compares COPD-related healthcare costs for Medicare beneficiaries among the
patients, including antibiotics and OCS as confounders. Among other results, initial
maintenance treatment with FSC provided a significantly lower cost (medical +
pharmacy) than all other medications ($295 versus IPA; $1,235 versus IPR; and $110
versus TIO) over a 1-year follow-up period.
Simoni-Wastila et al. (Simoni-Wastila et al. 2009) compare the healthcare
resource utilization associated with hospitalization or emergency department (ED) visits
between inhaled corticosteroid/long-acting ȕ 2-agonist combinations (fluticasone
propionate/salmeterol combination) and anticholinergic treatments in Medicare
beneficiaries with COPD. The authors conclude that the combination therapy has a lower
risk of COPD-related exacerbation events when compared to anticholinergic treatments.

2.3.1. Adherence to Guidelines
The presence of data on COPD severity as defined by the GOLD guidelines is
important for decision makers. COPD severity is determined by pulmonary function test
results, but this information is usually not available in administrative billing data and is
therefore not available for assessment. The stage of COPD at diagnosis is very likely to
confound the assessment of cost, as someone who is diagnosed with a more advanced
stage is more likely to have higher costs.
19

Despite several studies evaluating COPD treatments, only recently a few studies
have focused on the guidelines. In 2010, Stuart et al (Stuart et al. 2010) compare the
differences between users and nonusers of maintenance medications in terms of selected
outcomes for a nationally representative sample of Medicare beneficiaries with COPD.
However, in this work, the patients are not stratified by disease severity. In multivariate
models, maintenance drug users are found to be less likely than nonusers to be
hospitalized (OR = 0.70) or rehospitalized (OR = 0.74), and have significantly lower
annual Medicare expenditures ( ± $3,916). The authors then conclude that the use of
maintenance therapy is significantly associated with lower risks of hospitalization and
rehospitalization, therefore reducing Medicare expenditures.
Asche et al. (Asche et al. 2012) investigate the impact of the use of specific
treatments recommended by GOLD guidelines on the cost of COPD patients. The authors
use a EMR database to evaluate adherence to COPD treatment in patients treated with a
combination of LABA and/or ICS. Although this work includes laboratory results to
classify disease severity, it has a few limitations. For instance, the database does not
capture the entire medical history and does not provide data about out-of-network
healthcare utilizations (e.g., hospitalizations, ER, and office visits). The authors also do
not explain how the adherence to guidelines is measured or how long patients are
observed.

2.4. COPD Severity
COPD severity is usually assessed by spirometry, which is the most reproducible,
standardized, and objective way of measuring airflow limitation. Table 2 lists the grades
20

of COPD severity according to forced expiratory volume in 1 second (FEV
1
), and the
symptoms, and complications commonly associated with each severity grade.
When spirometry is unavailable, the diagnosis of COPD can alternatively be made
based on clinical symptoms, such as abnormal shortness of breath, and clinical signs,
such as increased forced expiratory time.
Table 2: Clas s ification of COPD by s everity.

Source: NHLBI/WHO Workshop Report, Global Initiative for Chronic Obstructive Pulmonary Disease (GOLD) .
21

Hilleman et al. (Hilleman et al. 2000) show that the disease burden of COPD is
correlated with its severity. For instance, Stage-1 patients experience the lowest annual
costs ($1,681), followed by Stage-2 ($5,037), and Stage-3 patients ($10,812).
Unfortunately, clinical information is usually unavailable in administrative
claims. Therefore, researchers commonly include severity proxy measures to control for
the unavailable clinical variables. Mapel et al. (Mapel et al. 2006) study the survival of
COPD patients using inhaled corticosteroids and long-acting ȕ agonists. The authors use
claims records of COPD-related outpatient encounters and hospitalizations to develop a
proxy measure of COPD severity, since pulmonary function tests are not available for all
SD WL H QWV & 23' VH YH ULW \ D W ED VH OL QH LH WK H \ H D U SU LRU WR WKH VW D UW RI WKH S D WL H QW¶V IROORZ -
up period) is based on the number of COPD outpatient encounters, COPD emergency
department encounters, and hospital admissions with a primary discharge diagnosis of
COPD, under the assumption that a higher level of COPD-related utilization represents
greater disease severity. Respiratory admissions for conditions other than COPD or
asthma are counted using relevant discharge codes (ICD-9 460± 490, 495, 500 ± 519).
In a similar study, Halpern et al. (Halpern et al. 2006) also use surrogate measures
of disease severity, such as episodes of serious pulmonary respiratory infections (e.g.,
chronic bronchitis and pneumonia), the number of hospitalizations, and emergency room
visits for COPD in the six months prior to the index date, in comparisons of COPD
patients with versus without anemia. In their study, however, outpatient visits are not
included. Despite this, the use of ventilation support is included.
In order to address selection bias and control for severity, Dalal et al. (Dalal et al.
2009) included COPD prescription medications for antibiotics and oral corticosteroids
22

(OCS) used in the pre-index baseline period, in addition to their proxy measures of
severity. Similarly, Simoni-Wastila et al. (Simoni-Wastila et al. 2009) and Joo et al. (Joo,
Lee and Weiss 2008) also control the analysis for selected therapeutic classes frequently
used in COPD patients to adjust for severity-of-illness, including oral corticosteroids,
short-acting ȕ 2-agonists (SABA), leukotriene modifiers, theophylline and xanthine
derivatives, antibiotics frequently used in COPD treatment (e.g., including penicillins,
cephalosporins, macrolides, sulfa/trimethoprim, quinolones, tetracyclines and
lincomycins), and all other oral and intravenous antibiotics. The difference between these
two studies is the time period that the variables are measured. In (Joo et al. 2008) all of
them are measured in the year before the event date, except oral corticosteroids, which is
measured 180 days prior to the event date, and antibiotics prescribed 30 days prior to the
event date. In contrast, in (Simoni-‐Wastila et al. 2009) the variables are measured 6
months prior to the index date.
In addition to medications as an indicator of severity, Joo et al. (Joo et al. 2008)
included non-outpatient health care utilization (i.e., hospitalization and ER visits)
variables in one unique variable called acute exacerbation. The authors identify acute
exacerbations of COPD using a combination of inpatient, outpatient, and pharmacy data.
Exacerbations are defined based on the presence of an ICD-9 code related to COPD
and/or specific to an exacerbation with one of the following: (1) an inpatient
hospitalization; (2) an emergency department visit; or (3) an outpatient visit with either
an oral steroid or antibiotic prescription dispensed within five days of the visit with less
than a 30 day supply. Outpatient visits that have a diagnosis for infections other than
23

respiratory infections (e.g., cellulitis) are not included as an exacerbation. Exacerbations
are assumed to last 30 days and after 30 days a new acute exacerbation can be identified.
Wu et al. (Wu et al. 2006) develop and validate a COPD severity score for adults,
excluding individuals older than 65 years old. Their method may prove useful in
providing insights about the benefits of COPD treatments in administrative claims
analysis. The variables associated with COPD severity included hospitalization, oxygen
therapy, spirometry tests, pulmonologist visit, oral corticosteroids, and several COPD
treatments.
Although COPD severity is an important variable to explain health care expenditure
and utilization related to COPD, Blanchette et al. (Blanchette et al. 2008b) do not include
any proxy measure of severity in their model when evaluating the asthma burden in
COPD patients. This is also seen in some other studies found in the literature and may
bias the results in such studies.

2.5. Comorbidities
Along with severity, comorbidities are essential factors related to health care
utilization and expenditures (Lin, Shaya and Scharf 2010). Therefore, it is important to
PHD VXUH SD WL H QWV¶ F RPRU ELGL WL H V ZKH Q H Y D OXDWLQJ F ODLPV GD WD There are several different
methods to estimate and include them in the model. We now mention the different
methodologies used to evaluate this variable in the COPD population.
Mapel et al. (Mapel et al. 2006) adjust for comorbidities using a comorbidity
score derived for each patient, which is calculated using the Deyo adaptation of the
Charlson Comorbidity Index (CCI) with the respiratory disease diagnosis code group
24

removed from the index. This index was originally developed to predict risk of death in
hospitalized patients and includes 17 groups of comorbidities, each given a weighted
value. The authors calculate the score in two ways: (1) using hospitalization diagnosis
codes only where patients without inpatient admissions received a score of 0; and (2)
using outpatient encounter diagnosis codes. For the outpatient score, it is required that
two encounters with a given comorbidity diagnosis be present on separate dates in order
for that comorbidity to be considered present.
Similarly, Simoni-Wastila et al. (Simoni-Wastila et al. 2009) and Drescher et al.
(Drescher et al. 2008) review medical claims to assess the prevalence of comorbidities.
The authors use the method proposed by Deyo et al. (Deyo, Cherkin and Ciol 1992),
excluding COPD from the calculations of the comorbidity score, since COPD is the
disease of interest in the study.
Blanchette et al. (Blanchette et al. 2008b) also evaluate individual CCI
components using an adapted version of the clinical index developed by Charlson and her
colleagues, which is based on medical record reviews and contains 17 categories of
comorbidities. Their CCI index contains 19 categories of comorbidities, defined using
ICD-9-CM diagnosis codes (Blanchette et al. 2008b).
Lin et al. (Lin et al. 2010) characterize a comprehensive comorbidity profile to
explore the economic implications of comorbidity among patients with COPD. As in
(Blanchette et al. 2008b), the authors include more diseases in addition to the ones
included in CCI. They examine the prevalence rates of 23 comorbid conditions, including
all 17 from the CCI, and their economic implications among Medicaid COPD patients.
The results are then compared to a control group with similar demographic
25

characteristics. The authors find that COPD patients have a substantial comorbidity
burden, in terms of higher CCI scores, higher odds-ratios for 9 out of the 23 conditions
assessed, and more comorbidities than controls. The high burden of comorbidities is
associated with incremental medical utilization and cost, suggesting that disease
management interventions are needed to improve patient care and reduce the COPD
economic burden.
The CCI and its adaptations are by far the most widely used comorbidity measure
in mortality analysis because it is constructed to predict mortality outcome and because
the scores are easy to compute based on administrative data. However, using the CCI
alone to control for confounding in health utilization and cost models may not be optimal.
Other generic comorbidity measures, such as the Elixhauser comorbidity index and the
Comorbidity symptom scale, have also been used in economic evaluation of COPD.
However, all generic measures, including the CCI, are limited in their ability to capture
the comorbidity profile specific to the COPD population.
The CCI was originally developed to predict mortality in chronic disease.
However, while a few authors claim it is a reliable and valid measure of multiple
comorbidities, others show it is not a good measure for COPD patients.
For instance, Pinckney et al. (Pinckney et al. 2004) raise a few problems with
CCI. The authors test the validity of the Charlson index and another co-morbidity
instrument, the adult co-morbidity evaluation 27 (ACE-27), in patients admitted with
COPD exacerbations. Co-morbidity measured by the ACE-27 is shown to be a good
predictor of survival whereas the CCI cannot reliably predict mortality in COPD patients.
A more careful evaluation of the CCI shows mixed results; an univariate model predicts
26

mortality with an estimated 18% increase in risk of death for every increase of one in the
CCI, but this association disappears in the multivariable analysis.
Alternatively, some authors include comorbidities in their analyses, without using
a comorbidity index. For instance, Halpern et al. (Halpern et al. 2006) study different
comorbidities in each group, such as neurologic, respiratory, musculoskeletal, skin, and
GI diseases, including each variable in their model. Beauchesne et al. (Beauchesne et al.
2008) classify co-morbidities as <2 or >2, which does not consider the severity of the
diseases. In addition, they do not mention whether they include all diseases in the
analysis.
Akazawa et al. (Akazawa et al. 2008b) identify comorbid conditions from the
diagnoses in claims and group them using the Clinical Classifications Software
developed by Elixhauser and maintained by the Agency for Healthcare Research and
Quality (AHRQ). The 20 most prevalent conditions in the study cohorts are identified. In
this study, the comorbidities are measured during the same time period as the outcomes
of interest, but the proper way to do it is to measure it during the pre-index period. In
another study, Akazawa (Akazawa et al. 2008b) identifies comorbid conditions during
the 6-month baseline period (i.e., from 3 months before till 3 months after the index
date), which are usually estimated before the index period.
Joo et al. (Joo et al. 2008) determine comorbidities by a combination of
diagnostic codes and dispensing of prescriptions associated with particular disorders at
any time from a year prior to entry date till the event date. ICD-9 codes are used to
identify diabetes mellitus, hypertension, chronic kidney disease, chronic liver disease,
heart disease (i.e., heart failure and coronary artery disease), cancer, alcoholism,
27

substance abuse, depression, mental health disorders, dementia, and other lung diseases
(i.e., cystic fibrosis with pulmonary manifestations, pneumoconiosis, and bronchiectasis).
Unlike most studies, Dalal et al. (Dalal et al. 2009) do not include comorbidities,
which might introduce a bias in their results. Similarly, there are several studies that only
include specific diseases in the final model to adjust for comorbidities without additional
explanation.
One common limitation often observed in many works is the measurement of
comorbid conditions, health utilization, and costs at the same time. This methodology,
however, might introduce bias in the results, since a comorbid condition can affect the
number of visits and costs.
Lin et al. (Lin et al. 2010) find that a few comorbid conditions not included in the
CCI, such as hypertension, depression, tobacco use, and edema, may also have
substantial clinical and economic implications for COPD patients. These conditions can
subsequently trigger the development of a COPD-specific comorbidity measure and can
potentially be used as a managerial tool for targeting high-risk patients for more tailored
disease management interventions. In addition, the chronic disease score, another method
not often used in COPD studies, can provide a valid and reliable measure for
comorbidities.
Based on the aforementioned works, including each comorbidity into the equation
might be the best way avoid estimation bias in retrospective studies.

28

2.6. Health Care Cos t
COPD is one of the leading causes of mortality and morbidity in the United States
(Hurd 2000, Sullivan, Ramsey and Lee 2000). Various studies have reported the
economic burden of COPD thus far (Wilson, Devine and So 2000, Faulkner and
Hilleman 2002, Ramsey and Sullivan 2004) which ranges from $23.9 billion (Rice,
Hodgson and Kopstein 1985) to $33.2 billion dollars (Sullivan et al. 2000). These costs
include direct as well as indirect costs. Another study shows that direct costs due to
COPD are around $14 billion dollars per year (Hurd 2000).
Hospital use, medications, and oxygen therapy are the major cost drivers for
COPD. Per-capita Medicare expenditures for COPD patients are 2.4 times higher than
those of other Medicare beneficiaries ($11,841 vs. $4,901 in 2000 values) (Grasso et al.
1998). Similarly, outpatient use is 2.2 times higher and hospital expenditures, which
constitute 64% of the total expenditures, are 2.7 times higher for COPD patients.
Predictors of costs and utilization in COPD patients include age, health status, disease
severity, physician specialty, geographic location, and type of insurance coverage
(Ruchlin and Dasbach 2001).
Rascati et al. (Rascati et al. 2007) perform cost analysis including all-cause and
COPD-related medical and pharmacy costs during the 12-month post-index period. Costs
are derived from the health plan perspective and thus represent payments for services.
COPD-related medical costs include any service (e.g., inpatient, outpatient, office visit,
ED visit) in which the primary diagnosis codes of COPD (ICD-9-CM codes 490.xx,
491.xx, 492.xx, or 496.xx) are assigned or a COPD-related medical procedure (e.g.,
spirometry test, oxygen therapy, pulmonary rehabilitation) is billed. COPD-related
29

pharmacy costs include the use of the following medications: anticholinergics, SABAs,
LABAs, oral corticosteroids/ICSs, theophylline, and oral antibiotics. The authors
evaluate the effects of fluticasone propionate with those of ipratropium in patients with
COPD and find that, despite the $260 increase in COPD-related pharmacy costs, there is
no significant difference in COPD-related medical costs.
Halpern et al. (Halpern et al. 2006) stratify health care costs by the type of service
involved (e.g., durable medical equipment, inpatient facility, hospice). Costs based on
claims and actual payments (i.e., reimbursements) are analyzed separately. When
adjusting for demographics, co-morbidities, and other markers of disease severity, the
authors reveal that anemia is independently associated with a $3,582 per-patient increase
in Medicare annual reimbursements.
Menzin et al. (Menzin et al. 2008) estimate excess costs as the paired difference
in cost between the COPD and matched comparison cohorts. The cost of COPD is
calculated as the sum of all amounts paid by the health plan for medical services and
drugs. For completeness, they also report other respiratory related medical services
among patients with COPD based on having ICD-9-CM codes 460, 490, 493.x, 495.x, or
500, 519.x listed as a primary diagnosis on an inpatient claim, or as a primary or
secondary diagnosis on an outpatient claim, as long as COPD is not listed on the same
claim. COPD patients have excess total healthcare costs about $20,500 higher (p-value <
0.0001) than the comparison cohort. Comorbidities are shown to have a high impact on
cost in the COPD cohort, accounting for 46% of the observed excess cost.
Akazawa et al. (Akazawa et al. 2008a) model health care costs using the
formulation in (Blough, Madden and Hornbrook 1999), which employs a two-part cost
30

model to address skewness of cost data and the large number of $0 costs. A logistic
regression is first performed to examine the determinants and predict the probability of
any health care expenditure during the 12-month pre-index period. Costs for subjects
with nonzero healthcare expenditures are then modeled using a generalized linear model
(GLM) with a gamma distribution and log link to account for the skewed distribution of
costs. This two-part model avoids potential difficulties introduced by transforming (e.g.,
calculating the log of the costs) and retransforming the dependent variable. In the
negative binomial (NB) regression, coefficients are presented as incidence rate ratios
(IRRs). In the GLM regression, coefficients are exponentiated to represent the ratio of
expected costs in the COPD and control cohorts. Predicted costs are also estimated using
the GLM coefficients for both cohorts. Healthcare utilization and cost data for the 12
months prior to the index date are analyzed with multivariate models. Their findings
show that, compared to the control group, the average incremental costs for
inpatient/emergency department services, office visits, and medical and pharmacy
services in COPD patients are estimated at $550, $238, $1,438, and $401 per year,
respectively, after adjusting for age, gender, region, and comorbid conditions.

2.7. Heath Care Utilization
As mentioned before, hospitalization is one of the drivers of COPD related health
care expenditure. Health care utilization in COPD patients is highly skewed, with the top
10% accounting for more than 50% of total expenditures (Ruchlin and Dasbach 2001).
31

Wouters et al. (Wouters 2003) show that the annual cost of health services
utilization in the United States is $4,119 per COPD patient, with additional indirect costs
accounting for $1,527 dollars.
After adjusting for age and gender, Mapel et al. (Mapel et al. 2000) show that
COPD patients in a managed care population are 2.3 times more likely to have been
hospitalized than patients with other conditions, and those admitted have significantly
longer lengths of stay (4.7 vs. 3.9 days, p-value < 0.001).
Blanchette et al. (Blanchette et al. 2008a) perform risk analysis outcomes
including the time to first all-cause hospitalization or ED visit (any diagnosis) in the
follow-up period and the time to first COPD-related hospitalization or ED visit (as
defined by the primary diagnosis using ICD-9-CM codes 490.xx, 491.xx, 492.xx, or
496.xx). Patients who do not have a hospitalization or ED visit are censored due to loss of
eligibility or to a change in the insurance plan. The authors evaluated the effects of
fluticasone propionate with those of ipratropium (IPR) in patients with COPD. They
show that the fluticasone propionate cohort was 45% less likely to have a COPD-related
exacerbation than the IPR cohort.
Mapel et al. (Mapel et al. 2006) examine the relationship between survival and
the usage of inhaled corticosteroids and/or long-acting beta agonists in COPD patients.
The authors capture exacerbations during the follow-up period in order to examine the
assumption of independence between respiratory drug use, exacerbations, and death. An
exacerbation is defined as an outpatient visit for a respiratory illness that results in a
prescription fill for an antibiotic or oral corticosteroid within 2 days, an emergency
department visit, or hospitalization due to a respiratory illness. The authors find that
32

COPD patients who use LABA alone, ICS alone, or ICS in combination with LABA have
substantially improved survival (HR 0.55, 0.59, 0.34, respectively) even after adjusting
for asthma and other confounding factors.
Menzin et al. (Menzin et al. 2008) evaluate the usage of healthcare services and
medications among the COPD and comparison cohorts without regard to diagnosis. They
considered the following types of drugs to be attributable to COPD: (1) antibiotics and
oral steroids given within 7 days of a COPD medical claim; (2) short-acting ȕ -agonists;
(3) maintenance medications, including anticholinergics, inhaled steroids, long-acting ȕ -
agonists, inhaled combination therapies; (4) xanthines; and (5) leukotriene antagonists.
They find that COPD patients are more likely to utilize healthcare services. In all services
(i.e., inpatient, outpatient, physician visits, and ER visits), the percentage of COPD
patients utilizing health care is higher than the comparison cohort (p-value < 0.001)
Simoni-Wastila et al. (Simoni-Wastila et al. 2009) study healthcare resource
utilization associated with hospitalization or emergency department (ED) visits between
FDA-approved inhaled corticosteroid/long-acting beta-agonist combinations and
anticholinergic treatments in managed-care Medicare beneficiaries with COPD. The two
clinical outcomes of interest included in their study are hospitalizations and ED visits,
which are assessed for both all-cause reasons as well as those COPD-related. Events are
categorized as COPD-related if the primary or secondary diagnosis ICD-9-CM codes of
these events are 491.xx, 492.xx, or 496.xx. The event is defined as the first occurrence of
hospitalization or ED visit after the index date. The time to event is calculated as the days
from the index drug claim date till the date of the first occurrence of a related medical
event. They found that COPD patients treated with the combination therapy have 18%
33

lower risk of a COPD-related hospitalization or ED visit than patients treated with ATC.
The results are similar for all-cause utilization.
Akazawa et al. (Akazawa et al. 2008b) also evaluate COPD patients using inpatient
and ED visits and show that COPD patients use 1.5 to 1.6 times more
inpatient/emergency department (IP/ED) services and office visits compared to control
patients.
Although the aforementioned studies provide insights into the costs of specific
COPD treatment regimens, there are no studies evaluating the impact of the guidelines on
health care outcomes.

2.8. Other COPD Related Topics
Ane mia:
Halpern et al. (Halpern et al. 2006) evaluate an interesting group of COPD patients.
Little is known about cost implications of anemia and its association with mortality in
COPD. The authors evaluate the association of anemia with survival in the study group.
The duration of survival is calculated from the index date till the time of death and it is
based on the number of quarters of Medicare enrollment until the end of either follow-up
data on December 31, 2001, termination of Medicare entitlement, or death. The authors
find that anemia is present in 21% of COPD patients. Although more prevalent in more
severely ill COPD patients, anemia significantly and independently contributes to the
costs of care for COPD and is associated with increased mortality.

34

Asthma:
While differentiating between COPD and asthma may be difficult for clinicians,
there has been increased awareness that asthma and COPD coexist in approximately 20%
of COPD patients (Beeh et al. 2004).
Dalal et al. (Dalal et al. 2009) include COPD patients with comorbid asthma in
their study. Due to the lack of reversibility and spirometry information in claims data,
asthma cannot conclusively be ruled out based on excluding ICD-9-CM diagnosis codes.
Asthma and COPD are commonly comorbid and are often treated with similar methods.
Although the two diseases can be differentiated through a systematic clinical assessment,
the authors do not separately analyze patients in order to observe whether there is a
difference between them.
In a different approach, Blanchette et al. (Blanchette et al. 2008b) (Blanchette et
al. 2009) at two different opportunities assess the cost burden of asthma in COPD
patients in a Medicare Advantage population. The authors show that patients with both
COPD and asthma actually incur higher health care costs and use more health care
services than those with COPD without asthma. After adjusting for covariates, the
authors find that patients in the COPD + asthma cohort are more likely (adjusted OR of
1.6; 95% CI, 1.4-1.7) to have at least one acute event (e.g., ER visits or hospitalizations)
than patients in the COPD cohort and have $1,931 greater adjusted respiratory-related
health care costs (p-value < 0.001). Therefore, Medicare beneficiaries with COPD and
asthma incur higher health care costs and use. Based on these studies and others
(Blanchette et al. 2009, Shaya et al. 2008), there is enough evidence showing that patients
35

with COPD and asthma should be carefully investigated and not just excluded from the
study sample.

Antibiotics:
Beauchesne et al. (Beauchesne et al. 2008) conduct a study to describe the
different antibiotics used in the home management of COPD exacerbations and to
estimate the failure rates following the initiation of the antibiotic. This project
demonstrates that a wide range of antibiotics are prescribed to moderate-to-severe COPD
patients. Around 30% of the treatment failures occurred in the 30-day period following
the initiation of the home therapy with an antibiotic. Antibiotics are also included in
many studies as indicators of disease severity (Wu et al. 2006).

S tatins:
Recently, it has been suggested that statins have anti-inflammatory effects and it
may be beneficial to treat other diseases in which inflammation plays a key role (Mancini
et al. 2006, Søyseth et al. 2007).
Blamoun et at. (Blamoun et al. 2008) assess the rate of COPD exacerbation in
patients taking statins. The data suggest that the use of statins may be associated with a
lower incidence of exacerbations in COPD patients. The adjusted odds-ratio of an
exacerbation is 2.35 for non-statin users and 3.01 for long-acting b2 agonists as a
covariate. The time to outcome during the observation period was reduced by statins with
the hazard ratio (HR) for exacerbation of 0.19 (all p-value < 0.05).

36

Depression and heart failure:
Different studies evaluate the association between COPD and other chronic
diseases, such as depression and heart failure. For instance, Jennings et al. (Jennings et al.
2009) conclude that COPD patients with depressive symptoms have a significantly higher
risk for exacerbations. Similarly, Lainscak et al. (Lainscak et al. 2009) determine the
prevalence and clinical impact of COPD among patients hospitalized for heart failure.
The authors find that COPD is frequent among hospitalized patients with heart failure
and that ȕ -blockers are largely underused, which is probably a major reason for the
higher mortality observed in patients with concomitant chronic heart failure and COPD.
Inhaler Corticosteroids (ICS ):
Another association that concerns researchers is the one between COPD
exacerbation and ICS. Although recommended by the GOLD guidelines, this class of
medication has controversial benefits in the literature of COPD patients. There are a few
studies showing the benefits of ICS (Mapel et al. 2006, van Schayck and Reid 2006),
whereas others show no significant improvement (Joo et al. 2008, Yang et al. 2012).
Joo et al. (Joo et al. 2008) show that the use of ICS among patients with newly
diagnosed COPD is associated with an increased risk of hospitalization for pneumonia.
After adjustment for covariates, patients with current use of ICS are 1.38 times (95% CI,
1.31-1.45) more likely to have a hospitalization for pneumonia than those without current
use of ICS. In contrast, recent evidence also shows that, although ICS do not slow the
progression of the disease, they do reduce exacerbations and mortality in COPD patients.
In a retrospective analysis of managed care claims data, it is shown that COPD patients
receiving ICS and/or salmeterol experience 14% lower mortality rates than patients who
37

do not receive the drugs (Mapel and Pearson 2002). Van der Valk et al. (van der Valk et
al. 2002) show that COPD patients receiving fluticasone experience significantly less
acute exacerbations than patients not receiving fluticasone. Unlike in asthma where
bronchial hyperresponsiveness is related to severe inflammation, in COPD bronchial
hyperresponsiveness is a function of the degree of existing airway obstruction. The
GOLD guidelines released by the NHLBI recommend that ICS should only be prescribed
for symptomatic COPD patients with a documented spirometric response to
glucocorticosteroids, or for those with FEV1 < 50% and repeated exacerbations requiring
treatment with antibiotics and/or oral glucocorticosteroids (Pauwels et al. 2001). A
review of 14 studies conducted by van Schayck and Reid (van Schayck and Reid 2006)
indicate that the use of ICS in COPD seems to have beneficial effects on lung function.

Pollution:
Although, it is expected that pollution has an impact on COPD exacerbations in
claims data studies, it is hard to measure pollution. As a result, zip code or other
geographical location factor is usually used as a proxy for this unmeasured variable.
Zanobetti et al. (Zanobetti, Bind and Schwartz 2008) examine the effect of exposure to
particulate air pollution on discharged patients following an admission for COPD. Their
findings suggest that long-term exposure to particulate matter elevates the risk of
mortality in susceptible population defined by COPD admissions. The authors adjust for
several personal characteristics; however, they do not include comorbidities.

38

CHAPTER 3: STUDY DESIGN

3.1 Motivation
Despite providing significant clinical benefits to COPD patients, the adherence to
the GOLD guidelines is expected to be only 25%, a dramatically low rate (Asche et al.
2008, Salinas et al. 2011). The key reasons for not following these guidelines, however,
are still unknown and under study. Potential reasons for non-adherence are the physician
lack of knowledge or understanding of the guidelines, difficulties in administering the
GU XJ V RU S D WL H QW¶V E H OL H I s that medications should only be taken to relieve COPD
symptoms instead of preventing the disease progression. Regardless of the actual reason,
non-adherence to the GOLD guidelines is expected to result in poor disease management
and, consequently, in significantly higher health costs.
Whereas several works have demonstrated the benefits of guideline adherence in
the disease management (Lin et al. 2010, Asche et al. 2012, Nichols 2007), little effort
has been made to quantify the economic burden to the health care system of COPD
patients who are non-adherent to treatment guidelines. As a result, fundamental questions
regarding the additional costs associated with non-adherence to treatment guidelines are
unanswered and, without this information, decision makers have limited knowledge to
target cost-effective health policies.
For this reason, we set out to quantify the impact of adherence to the GOLD
guidelines on the health costs of COPD patients. Specifically, our purpose in this thesis is
to compare the health expenditure of patients using appropriate therapy with the health
expenditures of patients using inappropriate therapy. We define appropriate therapy as
39

utilization of long-term bronchodilators or inhaler corticosteroids (ICS), as recommended
by the GOLD guidelines, while inappropriate therapy does not. Our work has two goals:
to identify any patient characteristics associated with medication appropriate use, and; to
assess the impact of appropriate COPD medications on healthcare expenditure.
In the next sections, we first discuss the source of our data, and define the study
population from the data. We then describe our approach to processing the data and
extracting the relevant information for our study.
3.2. Data Source
In this study, we use the Medicare 5% COPD Sample Standard Analytical Files
data available from the Centers for Medicare & Medicaid Services (CMS). Medicare is
the U.S. government's health insurance program for people age 65 or older. Certain
people under age 65 can qualify for Medicare too, including those with disabilities,
permanent kidney failure, or amyotrophic lateral sclerosis. Medicare provides assistance
with health care expenses; however, it does not cover all medical expenses or the cost of
most long-term care. The program is divided in four different parts:
x Part A is the hospital insurance, covering inpatient costs;
x Part B pays for medical services that Part A does not cover, such as therapies
administered in the physician office;
x Part C is called Medicare Advantage and, if you have Parts A and B, you can
choose this option to receive all of your health care through a provider
organization, such as an HMO;
x Part D is the prescription drug coverage, paying for multiple medications.
40

In our database, the Medicare administrative claims database consists of
information on all aspects of member healthcare utilization, including medical and
pharmacy claims. This database enables researchers to identify a specific diagnosis,
monitor the consequences of specific treatments or diseases, and even evaluate drug use
patterns across patients over time or cross-sectionally. One can also examine the charges
associated with a disease and its treatments.
With regard to the patient information, access to death date, birth date, as well as
demographic information, such as zip code, ethnicity and gender is available. Data
regarding provider information include the provider ID and its specialty. The claims data
includes the benefit year, the patient admission and discharge dates, the claim entry date,
and the claim received date. Billing information is also available and consists of the
charged amount, the paid amount, deductibles and coinsurance, any out-of-pockets costs,
as well as any amounts paid by Medicare or other insurance. Medical claims also include
a place of service field, which indicates whether the service took place in an inpatient or
outpatient setting, physician office, urgent care center, or surgical center. Additionally,
information about Current Procedure Terminology ± 4 (CPT4) procedure codes, as well
as International Classification of Diseases ± 9th Edition, Clinical Modification
(ICD-9-CM) diagnosis codes are available. Each claim can have up to 10 ICD-9-CM
diagnosis codes listed. The pharmacy records include information regarding the date the
prescription was written, the service date, days of supply, quantity, NDC drug code, drug
name, drug category code, as well as dosage description. Medical and pharmacy claims
are two different files linked by the beneficiary ID numbers.
41

3.2.1. Data Cleaning
The data obtained is divided into two files ± medical and pharmacy. The medical
claims files contain claims for hospitalization, ED claims, and outpatient claims (e.g.,
physician visits). Elaborate criteria are used to separate the three types of claims. In
addition, strict criteria are used to verify individual episodes. For instance, in order to
separate the hospitalization claims from the ED claims, diagnosis related groups (DRG)
and Current Procedure Terminology (CPT) codes are used along with the length of stay
(LOS). The LOS variable is identified as the number of days between the first and the
last date of service. The criterion used to separate these claims is:
Claims with LOS = 0 days and a COPD-related CPT code are classified as ED claims;
7K e remaining claims with LOS = 0 days are classified as outpatient claims;
$OOFOD LP V with LOS > 1 days are classified as hospitalization claims.
In addition, outpatient claims related to laboratory tests are removed from the outpatient
files because they are not accurate.

3.3. Time Frame
The time frame of the study is three years. Since the study objective is to follow patients
taking prophylactic pharmacotherapy and to examine the impact of using prophylactic
medications on healthcare costs, a 1-year follow-up period was required to allow enough
time to examine the effects of pharmacotherapy to result in specific outcomes.
Demographic (e.g., age, gender, race), clinical (e.g., diagnoses), resource utilization (e.g.,
outpatient, inpatient, prescription, and emergency room use), and economic variables
42

(e.g., amount billed to Medicare for any type of medical service rendered) are used to
achieve the study objective.
The GOLD guidelines (2010) recommend that moderate-to-severe COPD patients
be continuously on long-term bronchodilators and/or inhaler corticosteroids. From the
data, some prescriptions are dispensed for 90 days of supply. Therefore, the time unit of
the panel model is set to three months. We measure the time-varying variables every
three months during a period of 36 months, for a total of 12 time periods during the
baseline and the follow-up period is, as illustrated in Figure 2. The index-date is set to
January 2007, after all patients had been observed for one year.
Figure 2: The s tudy period.

43

3.4. Study Population
The study population consists of a random sample of 5% COPD cohort in
Medicare for the time period between 2006 and 2008. The data is composed of patients
who have a diagnosis of COPD, chronic bronchitis, chronic airway obstruction, or
emphysema. Diagnosis is based on the International Classification of Diseases, 9th
Edition, Clinical Modification (ICD-9) codes. Patients are then classified as COPD
patients if they have at least one of the following 4 ICD-9 codes 491.x, 492.x, 494.x, or
496.x. Medicare maintains both pharmacy and medical claims data for these patients.
Information about utilization and expenditures for services can be extracted for various
time periods from the claim data.
The following steps are followed in selecting the patients eligible for this study.
First, the patients have to be continuously enrolled in Medicare during the study period
or, if they die during this period, we must have observed them for at least one year
(baseline period). All recipients having at least 2 medical claims with a primary ICD-9-
CM code for chronic bronchitis (ICD-9-CM 491.xx), emphysema (ICD-9-CM 492.xx),
bronchiectasis (ICD-9-CM 494.xx), and chronic airway obstruction (ICD-9-CM 496.xx)
are included in the cohort. Patients who are dual eligible, who are under oxygen therapy,
who survived the entire study period, or who have both asthma (ICD-9-CM 493.xx) and
COPD are analyzed separately. We exclude enrollees less than 65 years of age and also
beneficiaries enrolled in the Medicare Advantage (MA) plans, because MA beneficiaries
generate no Medicare claims and hence have no diagnostic information. The eligibility
criteria for this study are listed in Table 3.
44

Table 3: Inclus ion and exclus ion criteria for eligibility in our s tudy.
Inclusion Criteria Exclusion Crite ria
Ɣ At least two COPD claims; Ɣ Age < 65 years old;
Ɣ At least one prescription recommended
by GOLD guidelines for moderate-to-
severe patients;
Ɣ Beneficiaries enrolled in a Medicare
managed care plan during anytime of the
study period.
Ɣ One year of follow up;
Ɣ Age > 65;
Ɣ Enrollment at least 1 year prior to the
index date.

3.5. Res earch Ques tion
To the best of our knowledge, there are no studies yet showing the impact of
adherence to guidelines on health care cost in moderate-to-severe COPD Medicare
population. The two studies described in Section 2.3.1 (Asche et al. 2012, Stuart et al.
2010) with similar objectives differ in either their study population and/or their
methodology.
Treatment effectiveness estimated via observational studies is useful for policy
makers to make a healthcare decision related to drug utilization. Therefore, based on the
lack of prior evidence, this study proposes to investigate the association among adherence
to guidelines, patients characteristics, and health outcomes using an innovative method
for handling death and selection bias.

45

Research Question 1: How patients take their COPD medications and what are the
factors associated with the adherence to guidelines?
Research Question 2: 'R H VWKHF RPSO LDQF H WR J XLG H OL QH VLP SU RYH SD WL H QWV¶ KH D OW K outcomes?
46

CHAPTER 4: METHODOLOGY
4.1. Medication Therapy
In our study, patients are classified in 4 groups based on the percentage of use of
appropriate therapy among the quarters as shown in Table 4. For each quarter-
patient ሺ ݅ ǡ ݐ ሻ , we assign a binary variable ܽ ௜ ሺ ݐ ሻ which takes the value of 1 if patient ݅
follows the appropriate therapy recommendation during quarter ݐ , and takes the value 0
otherwise. We then compute the percentage ߩ ௜ ሺ ݐ ሻ of appropriate therapy usage of patient
݅ up to quarter ݐ as ߩ ௜ ሺ ݐ ሻ ൌ ሺ ͳ Ȁ ݐ ሻ σ ܽ ௜ ሺ ߬ ሻ ௧ ఛ ୀ ଵ . Based on ߩ ௜ ሺ ݐ ሻ ǡ each quarter-patient ሺ ݅ ǡ ݐ ሻ
is then finally classified into Groups I, II, III, and IV, according to the percentage interval
where ߩ ௜ ሺ ݐ ሻ is located.
Table 4: Group clas s ification according to their appropriate therapy us age.
Group Pe rc e ntage of us e of appropriate therapy
Group I 0% - 25%
Group II 25%-50%
Group III 50%-75%
Group IV 75% -100%

We assume, based on their therapy pattern during the baseline period, that only
moderate-to-severe patients are included in the study population.

47

4.2. Health Outcomes
In this study the health outcomes evaluated are total health care expenditures. In
Medicare claims, there are different cost variables, such as the amount paid out-of-
pocket, the total charge amount, departmental total charge amount, and Medicare
payment amount. In our work, expenditure is calculated using the latter.
The Medicare payment amount is the amount of payment made from the
Medicare program for the services covered by the billing/claim record. Generally, this
represents the amount paid to the institutional provider, with some rare exceptions (e.g.,
when there is a negative claim payment amount).

4.3. Independent Variables
The following independent variables are included in the model to investigate their
association with the guidelines compliance.
As for demographic characteristics, patient age, race, gender, and geographical
location are included in the model. Comorbidities are included in the analysis as a
categorical variable for each chronic disease reported by the Agency for Healthcare
Research and Quality (AHRQ) (Elixhauser et al. 1998). The AHRQ developed a
software tool that assigns variables to identify comorbidities in hospital discharge records
using diagnosis coding from the ICD-9-CM. The diseases included in this tool are:
congestive heart failure, valvular disease, pulmonary circulation disorders, peripheral
vascular disorders, hypertension (uncomplicated and complicated), paralysis, other
neurological disorders, diabetes, hypothyroidism, renal failure, liver disease, peptic ulcer
disease excluding bleeding, AIDS, lymphoma, metastatic cancer, solid tumor without
48

metastasis, rheumatoid arthritis/collagen vascular diseases, coagulopathy, obesity, weight
loss, fluid and electrolyte disorders, blood loss anemia, deficiency anemias, and alcohol
abuse. In addition, based on a literature review for comorbidities in COPD patients,
schizophrenia, depression, and hyperlipidemia are also included in the analyses.
According to the GOLD guidelines, COPD severity is assessed based on the
SD WL H QW¶V OHYH O RI V \ PSW R PV spirometry results, and the presence of complications, such
as weight loss. Since claims data do not have information on clinical measures of disease
severity, the following proxies for disease severity are used for the purpose of this study.
Proxies for COPD severity:
1. Number of canisters of inhaled short-acting ȕ -agonists (SABA): SABA is used to
relieve COPD symptoms and thus, if a patient uses SABA often, this may be an
indicative that the disease is progressing to a more severe stage.
2. Number of canisters of inhaled ipratropium or ipratropium/albuterol: Similar to
SABA, inhaled ipratropium or ipratropium/albuterol are also reliever medications.
3. Number of sub-classes of controller medication: As opposed to reliever
medication, such as SABA and ipratropium, a few controller medications, such as
LABA and tiotropium, are recommended for long-term treatment to prevent COPD
symptoms and exacerbation. These types of therapy must only be prescribed to
moderate-to-severe patients.
4. Number of prescriptions for oral corticosteroids (OCS): OCS are usually
administered in COPD after an exacerbation.
5. Number of prescriptions for antibiotics: Antibiotics are usually used for
management of COPD exacerbations.
49

6. Use of home oxygen therapy: Oxygen should only be prescribed for very severe
patients. Although helping patients to breathe more easily, oxygen does not have
any pharmaceutical benefits, and does not improve or prevent disease clinical
outcomes.
7. Number of spirometry tests: This is a test to measure FEV1, a clinical measure for
COPD severity. Patients who have exacerbation often will be monitored more
frequently, indicating that the disease is more severe.
8. Number of hospitalizations or ER visits for COPD.
9. Presence of an intensive care unit (ICU) stay for COPD.
10. Number of physician visits for COPD.
The last three proxy measures are included because they are related to health care
utilization, and severe patients are more likely to use the health care services. Due to
these severity proxies being highly inter-correlated with the others, this might cause a
multicolinearity issue in the econometric model estimation. Therefore in sensitivity
analyses they will be excluded from the analysis.
In addition to the above listed independent variables, dual eligibility and death
information are also included in the final model. Patients are dual eligible if they are also
enrolled in Medicaid plans. For the death indicator, the Medicare dataset provides the
death date, however, there is no information about the cause of death.
4.4. Conceptual Framewo rk
As illustrated in Figure 3, a multivariate model is developed using the health service
model framework (Andersen, 1995) to explain the factors associated with adherence to
the GOLD guidelines (appropriate therapy usage) for COPD patients. Each independent
50

variable and the dependent variable are evaluated every quarter to optimize the dynamic
panel data to compute the outcomes.
Figure 3: Adherence to the GOLD guidelines mode l.

4.5. Methodological Is s ues
In real-world data, patients are not randomized to treatments. As a result, several
issues arise in most observational studies, since retrospective electronic claims databases
are designed for reimbursement purposes, not for clinical studies. The lack of
randomization might result in unbalanced characteristics (both observable and
unobservable) between the treatment and the control groups. Moreover, it is difficult to
know the causal effect of the treatment on the outcomes. This occurs, for example, when
a variable is correlated with the residual term, which captures the omitted or the
imperfectly measured variables in the outcome equation. Therefore, endogeneity might
Appropriate Therapy
Usage
Comorbidities
Patient Characteristics
$J H
*H Q G HU
5 DF H
= LS & R G H
'HD W K
Disease Severity
Previous hospitalization
or cost
COPD therapy
Antibiotics
OCS use
Spirometry tests
Physician Visits
Oxygen Use
51

exist in this type of study. There are three situations where endogeneity can be present
(Wooldridge 2001):
a) Selection on unobservables or omitted variables;
b) Measurement errors; and
c) Simultaneous (or reverse) causality.
In the current literature, few authors ignore the importance of correcting their
models for potential endogeneity bias caused by the non-randomization of the treatment
assignments. However, there are several known methods to solve this problem, such as
propensity scores, instrumental variables, panel data fixed-effects, and random effects.
The propensity score (PS) is probably the most common method used in the literature due
to its simplicity. On the other hand, the instrument variables (IV) approach adjusts for
unobservable variable confounding in a robust way, if instruments can be identified.
Basically, the IV method addresses the endogeneity problem for explanatory variables,
but only when a strong and valid instrument is available, which is not always possible
(Wooldridge 2002, Cameron 2005, Stuart, Doshi and Terza 2009).
Both the PS and IV methods can be applied in either cross-section or panel data
analyses. Panel data approaches are more complex than the cross-sectional alternatives.
However, panel data techniques also offer solutions to model estimation challenges,
which cannot be addressed in cross-sectional analysis, such as capturing repeated
observations on the same individuals, removing individual time-invariant unobservable
variables, and investigating dynamic relationships.
52

4.6. Data Analys is
This study applies advanced econometric methods to overcome two important
sources of bias: time-varying e ndoge ne ity and time-invariant individual covariate s.
Endogeneity arises when unobserved variables severely affect the estimation of
the observed variables, resulting in inconsistent results. In our study, we suspect that the
death indicator is endogenous for the following reasons. Individuals who are more
conscious about their health are more likely to seek and adhere to more effective COPD
treatments when they consult physicians about disease management. These patients are
more inclined to stop smoking and start exercising. On the other hand, the illness severity
might be captured in the error term and might be correlated with the endogenous variable,
leading to inconsistent outcomes. As a result, death endogeneity must be addressed in our
study to obtain a consistent estimator.
In our study population, death is a major issue. Patients are elderly and usually
have more than one chronic disease, leading to a high number of deaths during the study
period. As described in the results section, almost 30% of the patients died during the
study, and therefore excluding them, as most prior studies do, would potentially introduce
uncontrolled bias.
In addition to death, it could be argued that our variable of interest, appropriate
therapy usage, is also an endogenous variable. However, in our panel data model we are
evaluating how treatment usage in the previous time period impacts the outcomes in the
current time period. As a result, we do not expect appropriate therapy usage to be
endogenous. To confirm this hypothesis, we perform an auto-correlation test with the
goal of checking whether there is any autocorrelation among the time periods. Our auto-
53

correlation test rejects the null hypothesis with a p-value of 0.805, showing that the
appropriate therapy usage in the previous time period is indeed not endogenous.
Another potential source of bias is time-invariant individual covariates, which
represent personal characteristics and behavior associated with each patient, such as
smoking habits and severe pollution exposure. These confounders are usually not
reported in claims and are very hard to estimate in practice, even with good survey
instruments. However, in order to obtain unbiased results, individual covariates must be
somehow removed from the model.
In order to address both the endogeneity and the individual covariates issues, we
use the dynamic panel data model (Section 4.6.1). This model uses instrumental variables
(IVs) to adjust for endogeneity and also removes the time-invariant unobservables. The
instrumental variable estimator can then be used for the remaining time-variant variables
that influence the endogenous variable.
In order to use this model, we must find variables that meet the two key criteria
below to be considered a valid instrument:
1) It must be correlated to the endogenous right-hand-side variable(s), but not to the
outcomes (i.e., the dependent variable);
2) It must be independent of the error term; and
Once we have theorized a good argument for a valid instrument, then tests should be
performed to check whether we need to use IVs or not. First, the Wald test of exogeneity
of the instrumental variables verifies the correlation between the instruments and the
endogenous right hand side variable using the standard t-test. Then, the Hausman test is
performed to check for endogeneity. This is a common significance test for the
54

coefficient of the endogenous variable using a two-sided t-test, which checks whether
endogeneity has a significant effect on the consistency of the estimate. If the coefficient
is insignificant, the instrumental variable approach is not required, since there is no
endogeneity problem in the model.
For the potential instruments in our study, we theorize that lagged variables, such
as treatment pattern, disease severity indicators, cost, and drug copayments, are related to
death, but not related to current period health care expenditure. In fact, we believe that
copayments and other lagged variables in previous time periods (measured in ݐ െ ʹ ,
ݐ െ ͵ and ݐ െ Ͷ ) do impact the probability of dying, but the instruments are not directly
associated with the health care cost at time ݐ . Moreover, lagged variables can be used as
instruments in dynamic panel data and other researchers have also evaluated drug
copayments and coverage as potential instruments in similar circumstances (Stuart B
et.al. 2009). Figure 4 depicts the impact of the covariates on death and on the adherence
to guidelines assumed in our model. Our results in Section 5 show that these assumptions
actually hold in practice.

55

Figure 4: The health care model.

4.6.1. Dynamic Panel Model
In this study, we use the dynamic panel model to estimate the total health care
expenditures for COPD patients. Dynamic panel data (DPD) analysis has several
advantages over cross-section models. First, it increases the precision of the estimation
due to the higher number of individual observations, increasing also the degrees of
freedom and reducing the collinearity among explanatory variables. Second, panel data
models eliminate time-invariant individual covariates, and thus the effect of these omitted
variables can be used to improve estimator precision. Finally, dynamic panel data
analysis also provides information about previous time periods, making it possible to
investigate heterogeneity (Wooldridge 2002, Hsiao 2003).
56

Using this DPD approach, we are then able to (1) evaluate the relationship
between adherence to guidelines and health outcomes over several time periods, as
opposed to only one point in time; and (2) eliminate the effects of variables which are
particular to an individual and constant over time. Another advantage of dynamic panel
data is the availability of a variety of potential instruments, such as lagged variables. For
instance, the health outcomes in the current time period are likely to be influenced by past
medication use and previous health care expenditures. Using panel data, we take this time
pattern into account to generate IVs to estimate the effects of adherence to guidelines and
death on total health care expenditures.
The potential instruments available in this model handle the presence of the
endogenous RHS variables (i.e., death indicator) which, if not adjusted for, may lead to
inconsistent results. For instance, without this adjustment, the coefficient of the
endogenous variable captures the effect of omitted variables, usually caused by self-
selection or simultaneous causality, and will be correlated to the error term. The results of
the dynamic model, however, allow us to disentangle the unobserved heterogeneity from
the health outcomes variables (Arellano and Bond 1991, Roodman 2006).
Our model specification, which includes a lagged dependent variable and related
covariates, is:
ݕ ௜ ሺ ݐ ሻ ൌ ߙ ݕ ௜ ሺ ݐ െ ͳ ሻ ൅ ࢼ ᇱ ࢞ ௜ ሺ ݐ ሻ ൅ ߤ ܽ ௜ ሺ ݐ െ ͳ ሻ ൅ ߛ ݀ ௜ ሺ ݐ ሻ ൅ ߠ ௜ ሺ ݐ ሻ ൅ ߜ ௜ ൅ ߝ ௜ ሺ ݐ ሻ , (4.1)
where ݕ ௜ ሺ ݐ ሻ ݅ ݐ ǡ ߙ is
the previous-cost parameter, ࢼ ൌ ሾ ߚ ଵ ǡ ߚ ଶ ǡ ǥ ǡ ߚ ௞ ሿ Ԣ is the coefficient vector of the time
varying exogenous variables ࢞ ௜ ሺ ݐ ሻ ൌ ሾ ݔ ௜ ଵ ሺ ݐ ሻ ǡ ݔ ௜ ଶ ሺ ݐ ሻ ǡ ǥ ǡ ݔ ௜௞
ሺ ݐ ሻ ሿ Ԣ , ߤ is the coefficient of
the variable of interest, percentage of appropriate therapy in the previous time period
57

ܽ ௜ ሺ ݐ െ ͳ ሻ ǡ ߛ is the coefficient of the endogenous variable, death, ݀ ௜ ሺ ݐ ሻ , ߠ ௜ ሺ ݐ ሻ is the
dummy variable indicating the time period ݐ , ߜ ௜ is the unobserved time-invariant
individual covariate, and ߝ ௜ ሺ ݐ ሻ is the error term at time period ݐ . Table 5 below provides
the description of each observable variable ݔ ௜௝
ሺ ݐ ሻ for ݆ ൌ ͳ ǡ ʹ ǡ ǥ ǡ ݇ .

Table 5: The des cription of each obs ervable right-hand-s ide variable ࢞ ࢏࢐
ሺ ࢚ ሻ for
࢐ ൌ ૚ ǡ ૛ ǡ ǥ ǡ .
Des cription Type
Percentage of appropriate therapy at time period ࢚ െ ૚ Continuous
Age Continuous
Death Binary endogenous
Previous cost ࢚ െ ૚ Continuous
Oxygen use ࢚ െ ૚ Binary
Number of spirometry tests ࢚ െ ૚ Continuous
Number of specialist visits ࢚ െ ૚ Continuous
Antibiotic prescription ࢚ െ ૚ Binary
Chronic heart failure ࢚ െ ૚ Binary
Valvular disease ࢚ െ ૚ Binary
Pulmonary circulation disorders ࢚ െ ૚ Binary
Peripheral vascular disorders ࢚ െ ૚ Binary
Uncomplicated Hypertension ࢚ െ ૚ Binary
Complicated Hypertension ࢚ െ ૚ Binary
Other neurological disorders ࢚ െ ૚ Binary
Hypothyroidism ࢚ െ ૚ Binary
Renal Failure ࢚ െ ૚ Binary
Schizophrenia ࢚ െ ૚ Binary
Liver disease ࢚ െ ૚ Binary
Depression ࢚ െ ૚ Binary
Peptic ulcer disease ࢚ െ ૚ Binary
HIV ࢚ െ ૚ Binary
Lymphoma ࢚ െ ૚ Binary
Asthma ࢚ െ ૚ Binary
Diabetes ࢚ െ ૚ Binary
Hyperlipidemia ࢚ െ ૚ Binary
Race Categorical
Gender Categorical
Zip code Continuous

58

In order to estimate the model parameters, we use the Arellano Bond Generalized
Method of Moments (GMM) estimator (Arellano and Bond 1991). The GMM is a
convenient and widely used estimator employed to solve the problem of strict exogeneity
(Bond 2002), which usually rules out any correlation from the current and the lagged
dependent variables. The GMM estimator has the advantage that it is consistent and
asymptotically normally distributed.
There are different approaches for GMM estimation; among them the Arellano-
Bond (Arellano and Bond 1991) estimator is increasingly popular. The Arellano-Bond
estimator is designed for situations with a small number of time periods, large sample
size, dynamic dependent variables (i.e., the current variable value depends on its own
previous values), endogenous treatment effects, fixed individual effects,
heteroskedasticity, and autocorrelation within individuals but not across them. The
Arellano-Bond estimator assumes that first differences of instruments are uncorrelated
with the fixed effects, removing the individual specific heterogeneity and allowing the
introduction of more instruments, such as lagged variables, and subsequently improving
efficiency.
For this study, we are interested in estimating the health care expenditures of
COPD patients,. Health outcomes are likely to be correlated in time (e.g., a patient with a
high health care cost at time ݐ െ ͳ is likely to also have a high cost at time ݐ ) and thus the
Arellano-Bond estimator seems to be a good method to handle this dependency.
In our model, the variables representing race, gender, and zip code are time-
invariant, and thus can be addressed by taking the difference between ݕ ௜ ሺ ݐ ሻ and ݕ ௜ ሺ ݐ െ ͳ ሻ
as follows:
59

ݕ ௜ ሺ ݐ ሻ െ ݕ ௜ ሺ ݐ െ ͳ ሻ ൌ ߙ ሾ ݕ ௜ ሺ ݐ െ ͳ ሻ െ ݕ ௜ ሺ ݐ െ ʹ ሻ ሿ ൅ ࢼ Ԣ ሾ ࢞ ௜ ሺ ݐ ሻ െ ࢞ ௜ ሺ ݐ െ ͳ ሻ ሿ ൅
൅ ߤ ሾ ܽ ௜ ሺ ݐ െ ͳ ሻ െ ܽ ௜ ሺ ݐ െ ʹ ሻ ሿ ൅ ߛ ሾ ݀ ௜ ሺ ݐ ሻ െ ݀ ௜ ሺ ݐ െ ͳ ሻ ሿ ൅
൅ ሾ ߠ ሺ ݐ ሻ െ ߠ ሺ ݐ െ ͳ ሻ ሿ ൅ ሾ ߝ ௜ ሺ ݐ ሻ െ ߝ ௜ ሺ ݐ െ ͳ ሻ ሿ . (4.2)
As a result, the unobserved individual heterogeneity ߜ ௜ is removed from the model. Then,
we calculate the GMM estimator for ࢼ as:
ࢼ ൌ ൫ ࢄ ᇱ ࢆ ࢃ ෢ ࢆ ᇱ ࢄ ൯ ି ଵ ሺ ࢄ ᇱ ࢆ ࢃ ෢ ࢆ ᇱ ࢅ ሻ , (4.3)
where ࢄ࢙ are the endogenous RHS and exogenous variables matrix, ࢃ is the weight
matrix (consistently estimates ሾ ܸܽݎ ሺ ܼ ௜ ᇱ ߝ ௜ ሻ ሿ ି ଵ ሻ , ܼ s are the instruments, including lagged
variables of ܻ and/or other instruments, for example, drug copayments.
In addition to the GMM, there are other estimators that we could have used in our
analysis, such as the ordinary least squares (OLS), the two-stage least squares (2SLS) or
the Heckman selection model. However, all of these estimators present significant
disadvantages over GMM. For instance, the dependence of the difference in the error
terms over time (autocorrelation) implies that the OLS estimates of ߙ are inconsistent.
Moreover, in the OLS methodology there is no adjustment for the endogeneity problem,
which may also lead to biased estimators. On the other hand, the 2SLS is able to estimate
a consistent ߙ using instrumental variables. In this case, the instruments must be
correlated with the difference ο ݕ ௜ ሺ ݐ െ ͳ ሻ ൌ ݕ ௜ ሺ ݐ െ ͳ ሻ െ ݕ ௜ ሺ ݐ െ ʹ ሻ and orthogonal to
ο ߝ ௜ ሺ ݐ ሻ ൌ ߝ ௜ ሺ ݐ ሻ െ ߝ ௜ ሺ ݐ െ ͳ ሻ , assuming that the disturbances ߝ ௜ ሺ ݐ ሻ are serially uncorrelated.
However, the key disadvantage of 2SLS estimation is that the GMM is in general more
efficient than 2SLS, because of its sophisticated reweighting based on second moments.
Nevertheless, such assertions are asymptotic (Roodman 2006, Arellano and Bond 1991).
Finally, the major limitation of the Heckman selection model is that the results are
60

sensitive to distributional assumptions, which is heavily dependent on the normality
assumption.
For these reasons, the GMM estimator is chosen to evaluate the impact of
adherence to guidelines on health care expenditures. The GMM provides a convenient
framework for estimating asymptotically efficient estimators. The key assumption for the
validity of the GMM estimators is that the autocorrelation errors are independent in time,
i.e., ݕ ௜ ሺ ݐ ሻ is not correlated to ߝ ௜ ሺ ݐ ሻ for ݐ ൌ ͳ ǡ ʹ ǡ ǥ ǡ ܶ . Therefore, tests for autocorrelated
errors should be conducted. If there is autocorrelation, we should include more distant
lags to mitigate this issue. Another crucial assumption for this model is that the
instruments are exogenous, i.e., they are not correlated with the errors in the first-
differenced equation. The exogeneity of the instruments can be tested with the so-called
Sargan test. The Sargan test is possible when the model is overidentified (Wooldridge
2002) and is computed as:
ܵ ൌ ቀ ܰ ି భ మ σ ܼ ௜ ᇱ ே ௜ ୀ ଵ ߝ ௜ ቁ ᇱ ܹ ෡ ቀ ܰ ି భ మ σ ܼ ௜ ᇱ ே ௜ ୀ ଵ ߝ ௜ ቁ ᇱ ̱ ɖ ୯ ି ୩ ଶ , (4.4)

where we must assume that the GMM estimator uses ܹ as the weight matrix. Under the
null hypothesis H
0
: ܧ ሺ ܼ ௜ ᇱ ߝ ௜ ሻ ൌ Ͳ , the test statistic ܵ has a ߯ ௤ ି ௞ ଶ distribution, where q is the
total number of instruments and k is the number of explanatory variables in the model.
In addition to the health care expenditures, we also investigate the variables
associated to patient behavior related to the appropriate therapy usage recommended by
the guidelines based on a similar model:
߬ ௜ ሺ ݐ ሻ ൌ ߙ ߬ ௜ ሺ ݐ െ ͳ ሻ ൅ ࢼ Ԣ ࢞ ௜ ሺ ݐ ሻ ൅ ߠ ሺ ݐ ሻ ൅ ߜ ௜ ൅ ߝ ௜ ሺ ݐ ሻ , (4.5)
where ߬ ௜ ሺ ݐ ሻ is the percentage of adherence to GOLD guidelines and ࢞ ௜ ሺ ݐ ሻ are the
exogenous variables listed in Table 5. The other parameters have the same meaning as
61

the previous health expenditure model in Equation (4.1), but their values are computed
separately for each model.

4.6.2. Two-Part Model
In order to validate our results, we also use another econometric methodology which
is able to circumvent the problem of death endogeneity. In this model, the expected total
healthcare cost is calculated for those patients who survived, removing death from the
cost equation. As a result, the death endogeneity is no longer an issue. However, the cost
now must be weighted by the probability of surviving to keep it consistent. This
methodology is based on the well-known two-part model, which first computes the
probability of a patient surviving at a given time period and then computes the expected
health outcome of that patient, given that he/she has survived. This approach then is able
to entirely avoid endogeneity problems while still capturing the correlation between the
health care cost and the adherence to guidelines.
In this model, the conditional probability ሾ ݀ ௜ ሺ ݐ ሻ ൌ Ͳ ȁ ࢞ ௜ ሺ ݐ ሻ ሿ of patient ݅ surviving
at time ݐ is estimated by a probability model (i.e., probit) in the first part as
ሾ ݀ ௜ ሺ ݐ ሻ ൌ Ͳ ȁ ࢞ ௜ ሺ ݐ ሻ ሿ ൌ ܩ ሾ ߙ ݕ ௜ ሺ ݐ െ ͳ ሻ ൅ ࢼ ᇱ ࢞ ௜ ሺ ݐ ሻ ൅ ߤ ܽ ௜ ሺ ݐ െ ͳ ሻ ሿ (4.6)
where ݀ ௜ ሺ ݐ ሻ is the death indicator and ܩ ሺ ݖ ሻ is the cumulative distribution function of a
normal random variable ܰ ሺ Ͳ ǡ ͳ ሻ . On top of the listed exogenous variables in Table 5,
instruments, such as drug copayments, previous health care costs, and adherence to
guidelines, are also included as exogenous variables since they are directly correlated to
62

probability of dying. The corresponding health care cost ݕ ௜ ሺ ݐ ሻ for this patient is estimated
in the second part as
ݕ ௜ ሺ ݐ ሻ ൌ ߙ ݕ ௜ ሺ ݐ െ ͳ ሻ ൅ ࢼ ᇱ ࢞ ௜ ሺ ݐ ሻ ൅ ߤ ܽ ௜ ሺ ݐ െ ͳ ሻ ൅ ߠ ሺ ݐ ሻ ൅ ߜ ௜ ൅ ߝ ௜ Ǥ (4.7)
This estimation in Equation (4.7) is very similar to the previous panel data model of
Equation (4.1), with the exception that the death indicator ݀ ௜ ሺ ݐ ሻ is not present anymore,
since this cost is computed only for the surviving patients. The expected cost is then
computed as ܧ ሾ ݕ ௜ ሺ ݐ ሻ ሿ ൌ ܧ ሾ ݕ ௜ ሺ ݐ ሻ ȁ ݀ ௜ ሺ ݐ ሻ ൌ Ͳ ሿ ሾ ݀ ௜ ሺ ݐ ሻ ൌ Ͳ ሿ . A similar method based on
the differences between two consecutive costs ݕ ௜ ሺ ݐ ሻ െ ݕ ௜ ሺ ݐ െ ͳ ሻ can be used to eliminate
the unobserved individual heterogeneity ߜ ௜ . The main assumption in this method is that
the error term in the first part of the model is conditionally independent of the error term
in the second part of the model.
In both methodologies applied in this study, the dynamic panel data and the two-
part model, the total health care expenditures ݕ ௜ ሺ ݐ ሻ of an individual ݅ at time period ݐ is
log transformed since the variable is skewed. To recover the estimates of ܧ ሾ ݕ ȁ ݔ ሿ , we
assume that ݔ is lognormally distributed with constant variance parameter ߪ ଶ , given
that ݕ ൐ Ͳ . Then, ܧ ሾ ݕ ȁ ݕ ൐ Ͳ ǡ ݔ ሿ ൌ ሺ ݔ ߚ ൅ Ͳ Ǥ ͷ ߪ ଶ ሻ is consistent if the ߚ estimated
from the linear regression is also consistent. However, in most cases, the variance of the
error term is not known, making the calculation of the expected value of the error term
more difficult. To overcome this issue, Duan suggested that a robust alternative, called
smearing estimator (Duan 1983). In this approach, a consistent homoskedastic
distribution-robust retransformation factor is estimated using the residuals: ߝ ൌ ሺ ݕ ሻ െ ݔߚ Ǥ 7KH VPHD U ID F WRU ĭ LV WKHQ H VWL PDWH G D V Ȱ ൌ ܰ ି ଵ σ ሺ ߝ ሻ ௜ఢௌ
. For
instance, in the two-part model, since we are calculating the model for y > 0,
63

ܧ ሾ ݕ ȁ ݕ ൐ Ͳ ǡ ݔ ሿ ൌ Ȱ כ ሺ ݔߚ ሻ and to estimate ܧ ሾ ݕ ȁ ݔ ሿ we multiply it by the probability
of being alive ሾ ݀ ௜ ሺ ݐ ሻ ൌ Ͳ ȁ ࢞ ௜ ሺ ݐ ሻ ሿ at the current time period ݐ .

4.7. Statis tical analys is
The SAS® statistical package is used to conduct the data analyses. Simple
descriptive statistics using frequencies and means (percentage and standard deviations)
are performed. The t-test and chi-square analysis are used to examine the significance of
the relationships between the two therapy groups. The variables analyzed for the first step
are age group, gender, ethnic group, comorbidities, hospitalization, type of medication
prescribed, and compliance to guidelines.
For the statistical difference, Arellano and Bond (Arellano and Bond 1991) ran a
Monte Carlo experiment to judge the performance of the Anderson-Hsiao estimator
against various GMM estimators and found that the GMM procedures produce
substantial efficiency gains. Statistical differences in total health care expenditures in our
model are then determined by a GMM model using dynamic panel data, controlling for
some covariates.
The covariates included in the model that identify predisposing characteristics are
age, gender, and ethnicity. In addition to the independent variables, several comorbidities
and severity are also included. In this population, death rate is high and, even when
controlling for other factors, it is still endogenous. As a result, instrumental variables are
used for the endogeneity issue.
64

COPD patients with asthma as a comorbidity are analyzed separately and the results of
the two (COPD alone vs. COPD + asthma) analyses are compared. Similarly, low income
patients are also evaluated in a separate analysis to allow us to compare it to the general
study population.

65

CHAPTER 5: RESULTS
Figure 5 shows the inclusion criteria in our study. From the entire Medicare
COPD population, we only consider a sample of 5% of this population, resulting in a
total of 363,543 patients. We then exclude 40% of these patients because they do not
have at least two COPD claims during the baseline period (i.e., the 4 quarters of 2006).
Among the remaining 214,819 patients, 25% are enrolled in Medicare Advantage and
thus we only consider the 75% enrolled in fee-for-service (FFS) for the 3 years of the
study period. From the remaining 160,264 patients, 97,441 of them have at least one
claim in Part D; however, we exclude 30,509 of these patients because they are younger
than 65 years old. From the remainder, we only include the moderate-to-severe patients,
resulting in a total of 36,786 patients, which is only 10% of the original sample.
Figure 5: The inclus ion criteria in our s tudy.

5% Medicare COPD population
363,543

At least 2 COPD claims in 2006
214,819

Enrolled in FFS during study period
160,264

Enrolled in Part D
97,441

Older than 65
66,932

Moderate-to-severe COPD patients
36,786

66

In the panel data, a total of 402,707 quarter-patients are included in the analysis. The
classification of each quarter-patient into Groups I, II, III, or IV is done as explained in
Section 4.1. Table 6 shows the classification criterion and the number as well as the
percentage of quarter-patients in each group.
Table 6: Group clas s ification according to their appropriate therapy us age.
Group Percentage of appropriate therapy Number of quarter-patients
Group I 0% to 25% 208,399 (52%)
Group II 25% to 50% 41,995 (11%)
Group III 50% to 75% 53,968 (13%)
Group IV 75% to 100% 30,594 (24%)

Based on this classification, we provide a summary of demographic characteristics of
our patients in Table 7. In all groups, the average age is not very different, varying from
77.5 to 78.2 years old. Similarly, there is not a big variation in the percentage of men in
each group, around 31% to 32%. On the other hand, other characteristics show significant
differences. For instance, Group I has a higher prevalence of dual eligible patients, with
51% in the Group I versus 46% in Group IV. With regard to ethnicity, the biggest
difference is 3% among the groups, which is observed in Whites between Group II (87%)
and Group IV (90%). The opposite characteristic is observed for African Americans
patients, who are more prevalent in Group II (6.3%) than in Group IV (4.7%), and most
prevalent in Group I (6.6%).

67

Table 7: Demographic characteris tics of the patient-quarter in each group.
Group I
(N=208,399 )
Group II
(N=41,995)
Group III
(N=53,968)
Group IV
(N=98,345)
Average age 78.2 (7.3) 77.7 (7.0) 77.8 (7.1) 77.5 (7.1)
Gender (male) 67,074 (32%) 13,111 (31%) 16,866 (31%) 30,594 (31%)
Dual eligible 106,243 (51%) 20,236 (48%) 23,991 (44%) 45,729 (46%)
Race
Unknown 276 (0.1%) 52 (0.1%) 92 (0.2%) 179 (0.2%)
White 183,031 (88%) 36,639 (87%) 48,066 (89%) 88,757 (90%)
African American 13,662 (6.6%) 2,650 (6.3%) 30,10 (5.6%) 4,599 (4.7%)
Others 1,250 (0.6%) 334 (0.8%) 384 (0.7%) 624 (0.6%)
Asian 2,773 (1.3%) 816 (1.9%) 960 (1.8%) 1,882 (1.9%)
Hispanic 6,384 (3.1%) 1,363 (3.2%) 1,272 (2.4%) 1,830 (1.9%)
Native American 1,023 (0.5%) 141 (0.3%) 184 (0.3%) 474 (0.5%)

Figure 6 depicts the prevalence of all comorbidities included in this study,
analyzed for each group. From the figure, we see that lower appropriate therapy usage is
often related to higher comorbidity prevalence. For instance, several chronic diseases,
such as chronic heart failure and diabetes, are more common in Groups I and II, where a
lower appropriate therapy usage is observed. One exception, however, is asthma, where
the percentage of asthmatic patients increases with the percentage of appropriate therapy
usage, with a prevalence of 17% in Group I versus 23% in Group IV. The largest
difference (8%) among the groups is observed in diabetes, with a prevalence of 32% in
68

Group I compared to 24% in Group IV. In addition to diabetes, uncomplicated
hypertension, hyperlipidemia, and chronic heart failure are the most common diseases
observed in this sample. All of those four diseases are prevalent among over 20% of the
patients.
Figure 6: The prevalence of the comorbidities in each group.

Table 8 reports the results of the analysis performed to investigate the factors that
impact the adherence to guidelines. As expected, previous appropriate therapy usage
ߩ ௜ ሺ ݐ െ ͳ ሻ has a positive correlation with current medication usage ߩ ௜ ሺ ݐ ሻ . However, if we
look further back in time for ߩ ௜ ሺ ݐ െ ʹ ሻ and ߩ ௜ ሺ ݐ െ ͵ ሻ , the association with ߩ ௜ ሺ ݐ ሻ becomes
negative. We hypothesize that this occurs because patients who see some improvement in
a quarter tend to continue taking the medication in the next quarter; however, patients
who have the disease under control for a few months do not present any serious
0%
10%
20%
30%
40%
50%
60%
Group I
Group II
Group III
Group IV
69

symptoms and then decide to stop the medication. We also see from Table 8 that, among
the comorbidities that are significantly associated with appropriate therapy usage,
patients with chronic heart failure and complicated hypertension have a higher use of
appropriate therapy, whereas patients with ulcer and hyperlipidemia are less adherent to
guidelines. In addition, patients under oxygen therapy tend to have a lower use of
appropriate therapy. This is also shown in Table 7 and we see that, even after adjusting
for confounders, the same pattern is observed. Similarly, a lower appropriate therapy
usage is also observed among patients who died and who are dual eligible. In contrast,
young patients tend to have a higher usage of appropriate therapy.
70

Table 8: Evaluation of the factors that impact appropriate therapy us age.
Y = Appropriate Therapy Coefficient P > z
Appropriate therapy ሺ ࢚ െ ૚ ሻ 1.142 0.000
Appropriate therapy ሺ ࢚ െ ૛ ሻ -0.153 0.000
Appropriate therapy ሺ ࢚ െ ૜ ሻ -0.054 0.000
Chronic heart failure ሺ ࢚ െ ૚ ሻ 0.001 0.017
Valvular disease ሺ ࢚ െ ૚ ሻ 0.000 0.728
Pulmonary circulation disorders ሺ ࢚ െ ૚ ሻ -0.007 0.000
Peripheral vascular disorders ሺ ࢚ െ ૚ ሻ 0.000 0.715
Uncomplicated hypertension ሺ ࢚ െ ૚ ሻ 0.000 0.507
Complicated hypertension ሺ ࢚ െ ૚ ሻ 0.002 0.000
Other neurological disorders ሺ ࢚ െ ૚ ሻ 0.000 0.723
Hypothyroidism ሺ ࢚ െ ૚ ሻ 0.000 0.693
Renal failure ሺ ࢚ െ ૚ ሻ 0.000 0.483
Schizophrenia ሺ ࢚ െ ૚ ሻ -0.001 0.463
Liver disease ሺ ࢚ െ ૚ ሻ 0.000 0.634
Depression ሺ ࢚ െ ૚ ሻ -0.001 0.486
Peptic ulcer disease ሺ ࢚ െ ૚ ሻ -0.005 0.004
HIV ሺ ࢚ െ ૚ ሻ 0.001 0.559
Lymphoma ሺ ࢚ െ ૚ ሻ 0.003 0.343
Asthma ሺ ࢚ െ ૚ ሻ 0.000 0.492
Diabetes ሺ ࢚ െ ૚ ሻ 0.000 0.569
Hyperlipidemia ሺ ࢚ െ ૚ ሻ 0.000 0.036
Oxygen use ሺ ࢚ െ ૚ ሻ -0.004 0.039
Specialist visit ሺ ࢚ െ ૚ ሻ 0.001 0.000
Spirometry tests ሺ ࢚ െ ૚ ሻ 0.002 0.058
Previous cost ሺ ࢚ െ ૚ ሻ -0.245 0.002
Previous cost ሺ ࢚ െ ૛ ሻ -0.121 0.001
Copayment ሺ ࢚ െ ૚ ሻ -0.037 0.002
Copayment ሺ ࢚ െ ૛ ሻ -0.014 0.000
Death -0.009 0.000
Dual eligible 0.016 0.000
Age -0.012 0.396
Time 5 0.000 0.666
Time 6 -0.004 0.000
Time 7 -0.005 0.000
Time 8 -0.005 0.000
Time 9 -0.002 0.000
Time 10 -0.002 0.000
Time -0.001 0.000
Constant 0.491 0.000
** Significant at 99%
* Significant at 95%
71

An unadjusted comparison of health care utilization and expenditure reveals
several significant differences among groups, as seen in Table 9. For instance, Group I
presents both the highest medical cost and the lowest pharmacy cost, possibly indicating
that inappropriate medication results in a higher number of ER visits, physician visits,
and hospitalizations. With regard to the health care utilization, patients hospitalized at
least once are up to 4% higher in Group I than in other groups, and a similar difference is
observed for patients with at least one emergency visit. Among patients with at least one
hospitalization per quarter, their length of stay is 1 day longer in Group I than in Group
IV. With regard to the percentage of patients with at least one physician visit per quarter,
all groups have a similar number around 80%; however, by looking at the specialist
visits, these numbers change dramatically, with only 14% of the patients visiting a
specialist in Group I versus 22% in Group IV . An even bigger difference among the
groups is observed regarding oxygen use, with more than half (56%) of the patients in
Group I having at least one claim for oxygen therapy versus only 37% in Group IV .
Interestingly, the pattern of antibiotic use is the opposite of oxygen use, with patients in
Group IV presenting the highest percentage (34%) than Groups I (30%), II (33%), and III
(32%).

72

Table 9: Unadjus ted outcomes reported for each group.

Group I
(N=208,399 )
Group II
(N=41,995)
Group III
(N=53,968)
Group IV
(N=98,345)
Medical Cost $5,726 (11,326) $5,797 (11,607) $5,115 (10,620) $4,283 (9,436)
Pharmacy Cost $499 (910) $563 (1,070) $586 (995) $722 (986)
Total Cost $6,226 (11,395) $6,361 (11,701) $5,702 (10,709) $5,006 (9,533)
Previous Cost $6,125 (11,206) $6,327 (11,790) $5,632 (11,790) $4,584 (8,634)
Length of Stay 3.7 (12.3) 4.0 (3.3) 3.6 (3.1) 2.7 (2.4)
At least 1 Hospitalization 46,417 (22%) 9,438 (22%) 11,027 (20%) 17,491 (18%)
At least 1 Spirometry Test
30,093 (14%) 7,099 (17%) 9,387 (17%) 17,514 (18%)
At least 1 ER Visit
54,465 (26%) 11,093 (26%) 12,998 (24%) 21,595 (22%)
At least 1 Physician Visit
166,691 (80%) 34,066 (81%) 43,993 (81%) 80,240 (81%)
At least 1 Specialist Visits
28,809 (14%) 7,493 (18%) 10,971 (20%) 21,491 (22%)
Oxygen Use 116,469 (56%) 18,570 (44%) 20,369 (38%) 37,236 (37%)
Antibiotic Use 61,712 (30%) 13,942 (33%) 17,353 (32%) 33,235 (34%)

When evaluating the total cost among the groups, the difference among groups
becomes clear, as shown in Figure 7. The difference between Group I and Group IV is
$1,220 per quarter-patient. After adjusting for all covariates and also for the instrumental
variables, the difference is even more significant. The adjusted costs of Groups I and IV
are $10,451 and $6,317, respectively, resulting in a difference of $4,134 per quarter-
patient. In addition, we see that, for each 25% increase in appropriate therapy usage, the
total adjusted cost decreases, on average, $1,378 per quarter-patient, with p-value < 0.001
73

when comparing each group to Group I. All adjusted costs are calculated using the
predicted costs and smearing factor.
Figure 7: Adjus ted cos t vers us unadjus ted cos t per quarter-patie nt.

5.1 Dynamic Panel Data Res ults
The first method evaluated in this study is the dynamic panel data, where death is
considered an endogenous variable. Our results are shown in Table 10. After controlling
for several covariates and also for the endogeneity problem, appropriate therapy
decreases the total cost for COPD patients. Another two variables that, surprisingly, have
a negative impact on the total cost are schizophrenia and hyperlipidemia. On the other
hand, several covariates increase the total cost, including previous cost, spirometry tests,
death, age, dual eligibility, and a few comorbidities (i.e., diabetes, depression, and
chronic heart failure).
As shown in Table 10, another 3 analyses are performed for the dynamic panel
data model. In the first analysis, we exclude all instrument variables (IVs), such as lagged
0
2,000
4,000
6,000
8,000
10,000
Adjusted Total Cost Unadjusted Total Cost
Group I
Group II
Group III
Group IV
**
**
**
**
*
74

variables and copayments; in the second analysis, we only exclude the lagged variables;
and in the third analysis, we only exclude copayments from the model while leaving the
lagged variables. In all models, the significance of the percentage appropriate therapy
usage does not change. This is also observed for the other variables, except age and
death. When all variables are included, age is not significant and the coefficient is
negative. In addition, death is no longer significant in all models that do not include the
IVs; however, the direction is still the same when compared to the model including all
IVs.
We also analyze all models using the Arellano Bond second autocorrelation test
and the overidentifying restrictions (i.e., Sargan test). For the first model, the null
hypothesis is not rejected, meaning that there is no first-order autocorrelation among the
residuals in the model. For the second test, the Sargan test has a p-value bigger than 0.05,
suggesting that the overidentifying restrictions are not rejected. In addition, we perform
the Hausman test and it demonstrates that the models are significantly different. When
comparing the first model that includes all instruments against the other restricted
models, all Hausman tests have a p-value < 0.05, which indicates the presence of
endogeneity and that instrumental variables correction is appropriate.

75

Table 10: Dynamic panel model including and excluding certain ins truments .
All IVs No IVs
included
No lags as
IVs
No copayments as
IVs
Previous cost ሺ ݐ െ ͳ ሻ 0.192** 0.188** 0.188** 0.191**
Previous cost ሺ ݐ െ ʹ ሻ -0.006 -0.003 -0.003 -0.006
Chronic heart failure ሺ ݐ െ ͳ ሻ 0.046** 0.057** 0.053** 0.046**
Valvular disease ሺ ݐ െ ͳ ሻ 0.020 0.021 0.021 0.019
Pulmonary circulation disorders ሺ ݐ െ ͳ ሻ 0.122 0.123 0.118 0.124
Peripheral vascular disorders ሺ ݐ െ ͳ ሻ 0.012 0.018 0.018 0.014
Uncomplicated hypertension ሺ ݐ െ ͳ ሻ 0.002 0.003 0.003 0.003
Complicated hypertension ሺ ݐ െ ͳ ሻ 0.003 0.015 0.012 0.004
Other neurological disorders ሺ ݐ െ ͳ ሻ 0.040 0.047 0.045 0.044
Hypothyroidism ሺ ݐ െ ͳ ሻ -0.009 -0.005 -0.005 -0.008
Renal failure ሺ ݐ െ ͳ ሻ 0.075** 0.067** 0.064** 0.075*
Schizophrenia ሺ ݐ െ ͳ ሻ -0.133* -0.101 -0.102 -0.134*
Liver disease ሺ ݐ െ ͳ ሻ -0.113 -0.105 -0.111 -0.115
Depression ሺ ݐ െ ͳ ሻ 0.057** 0.058** 0.058** 0.057**
Peptic ulcer disease ሺ ݐ െ ͳ ሻ 0.015 0.016 0.017 0.012
HIV ሺ ݐ െ ͳ ሻ 0.270 0.179 0.189 0.260
Lymphoma ሺ ݐ െ ͳ ሻ -0.121 -0.033 -0.033 -0.121
Asthma ሺ ݐ െ ͳ ሻ -0.011 -0.013 -0.011 -0.012
Diabetes ሺ ݐ െ ͳ ሻ 0.043** 0.051** 0.048** 0.044**
Hyperlipidemia ሺ ݐ െ ͳ ሻ -0.03** -0.027** -0.026** -0.031**
Oxygen use ሺ ݐ െ ͳ ሻ -0.003 -0.017 -0.015 -0.010
Specialist visit ሺ ݐ െ ͳ ሻ -0.011 -0.015* -0.013* -0.012
Spirometry tests ሺ ݐ െ ͳ ሻ 0.015** 0.018** 0.018** 0.015*
Death 0.126* 0.051 0.062 0.083
Appropriate therapy ሺ ݐ െ ͳ ሻ -0.38** -0.52** -0.46** -0.43**
Dual 0.270** 0.245** 0.242** 0.270**
Age 0.093** -0.035 0.133** 0.101**
Time 5 0.015* -0.224 0.071** 0.017**
Time 6 0.045** -0.160 0.094** 0.047**
Time 7 -0.08** -0.248* -0.036** -0.077**
Time 8 -0.02** -0.156* 0.012 -0.019*
Time 9 0.062** -0.037 0.089** 0.065**
Time 10 0.070** 0.003 0.088** 0.071**
Time -0.05** -0.085** -0.042** -0.051**
Constant -0.923 9.316 -4.024** -1.503
** Significant at 99%
* Significant at 95%

76

Asthmatic ve rs us Non-Asthmatic Patients
Most of the COPD studies exclude asthmatic patients since it is hard to separate
WKH RXWFRPH V D QG WKH S D WL H QWV¶ EH KD YLRU GXH WR WKH VLP LO D ULW \ R I WKH WZR GLVHD VH V , Q contrast, we evaluate two models: one including all patients and one including only non-
asthmatic patients. The key difference between the non-asthmatic patients sample and the
unrestricted sample is the distribution of patients in each group. When we are restricted to
non-asthmatic patients, Group IV is reduced from 24% to only 21% of patients whereas
Group I increases its share from 52% to 58%. Other statistical differences are also quite
noticeable. For instance, when only non-asthmatic patients are considered, the prevalence
of patients under oxygen therapy increases by 1%, the prevalence of men increases by
4%, and the dual eligibility decreases around 2% in each group. The mean age among the
groups, however, is not different between the two samples.
As reported in Table 11, most coefficients in both models are in the same
direction; however, certain coefficients are no longer significant in the non-asthmatic
model. One exception is the previous cost in time ݐ െ ʹ , which is not significant in the
main model, but it is significant in the non-asthmatic model. Once again, the percentage
of appropriate therapy usage does not change in significance and in direction when
comparing both models. The endogenous variable (i.e., death) is also not significant when
the asthmatic patients are excluded.

77

Table 11: Dynamic panel model: all patients vers us non-as thmatic patie nts .
All patients Non Asthmatic patients
Previous cost ሺ ݐ െ ͳ ሻ 0.192** 0.189**
Previous cost ሺ ݐ െ ʹ ሻ -0.006 -0.019**
Chronic heart failure ሺ ݐ െ ͳ ሻ 0.046** 0.049*
Valvular disease ሺ ݐ െ ͳ ሻ 0.020 -0.017
Pulmonary circulation disorders ሺ ݐ െ ͳ ሻ 0.122 0.148
Peripheral vascular disorders ሺ ݐ െ ͳ ሻ 0.012 0.053
Uncomplicated hypertension ሺ ݐ െ ͳ ሻ 0.002 -0.005
Complicated hypertension ሺ ݐ െ ͳ ሻ 0.003 -0.027
Other neurological disorders ሺ ݐ െ ͳ ሻ 0.040 0.052
Hypothyroidism ሺ ݐ െ ͳ ሻ -0.009 0.004
Renal failure ሺ ݐ െ ͳ ሻ 0.075** 0.021
Schizophrenia ሺ ݐ െ ͳ ሻ -0.133* 0.011
Liver disease ሺ ݐ െ ͳ ሻ -0.113 -0.005
Depression ሺ ݐ െ ͳ ሻ 0.057** 0.066**
Peptic ulcer disease ሺ ݐ െ ͳ ሻ 0.015 0.052
HIV ሺ ݐ െ ͳ ሻ 0.270 0.445
Lymphoma ሺ ݐ െ ͳ ሻ -0.121 -0.158
Asthma ሺ ݐ െ ͳ ሻ -0.011 -
Diabetes ሺ ݐ െ ͳ ሻ 0.043** 0.052*
Hyperlipidemia ሺ ݐ െ ͳ ሻ -0.032** -0.034**
Oxygen use ሺ ݐ െ ͳ ሻ -0.003 -0.022
Specialist visit ሺ ݐ െ ͳ ሻ -0.011 -0.026*
Spirometry tests ሺ ݐ െ ͳ ሻ 0.015** 0.022*
Death 0.126* 0.077
Appropriate therapy ሺ ݐ െ ͳ ሻ -0.379** -0.410**
Dual 0.270** 0.243**
Age 0.093** 0.088**
Time 5 0.015* -0.001
Time 6 0.045** 0.044**
Time 7 -0.079** -0.076**
Time 8 -0.021** -0.035**
Time 9 0.062** 0.051**
Time 10 0.070** 0.107**
Time -0.051** -0.031*
Constant -0.923 -0.475
** Significant at 99%
* Significant at 95%

78

All Patients ve rs us Alive-Only Patients
Another common methodology in health economics is to exclude patients who die
during the study period. One explanation for the exclusion of these patients is that they
are different from those who survived. However, if we just exclude them, our estimates
may be severely biased. Therefore, a sensitive analysis is performed to compare the
results between the model including all patients and the model which only considers
surviving patients.
In our sample, 28% of patients died between January 2007 and December 2008.
The majority of these patients are classified as Group I (35%) whereas a smaller fraction
(21%) is classified as Group IV. As expected, there are key differences between the two
samples. For instance, oxygen use and other health care utilization measures, such as
length of stay and ER visits, are ORZH U LQ WKH VXU YLYL QJ SD WL H QWV¶ VD PSO H 7KH UH VWULFW H G sample is younger (on average 1 year younger) than the unrestricted sample. Other
demographic characteristics are similar in both samples.
In the model including only surviving patients, several coefficients change when
deceased patients are excluded, as seen in Table 12. Most of the coefficients that are
significant in the model including all patients are not statistically significant in the alive-
only model and vice-versa. The only variable that changes direction is oxygen use, which
is negatively correlated to the total cost but not statistically significant in the original
model while positively correlated with the outcomes as well as statistically significant at
1% in the alive-only model.

79

Table 12: All patients vers us alive-only patie nts .

** Significant at 99%
* Significant at 95%

All patients Patients alive during the study period
Previous cost ሺ ݐ െ ͳ ሻ 0.192** 0.140**
Previous cost ሺ ݐ െ ʹ ሻ -0.006 -0.030**
Chronic heart failure ሺ ݐ െ ͳ ሻ 0.046** 0.032*
Valvular disease ሺ ݐ െ ͳ ሻ 0.020 0.032*
Pulmonary circulation disorders ሺ ݐ െ ͳ ሻ 0.122 -0.031
Peripheral vascular disorders ሺ ݐ െ ͳ ሻ 0.012 0.006
Uncomplicated hypertension ሺ ݐ െ ͳ ሻ 0.002 0.007
Complicated hypertension ሺ ݐ െ ͳ ሻ 0.003 0.010
Other neurological disorders ሺ ݐ െ ͳ ሻ 0.040 0.042
Hypothyroidism ሺ ݐ െ ͳ ሻ -0.009 -0.003
Renal failure ሺ ݐ െ ͳ ሻ 0.075** 0.030
Schizophrenia ሺ ݐ െ ͳ ሻ -0.133* -0.101
Liver disease ሺ ݐ െ ͳ ሻ -0.113 -0.141
Depression ሺ ݐ െ ͳ ሻ 0.057** 0.048**
Peptic ulcer disease ሺ ݐ െ ͳ ሻ 0.015 -0.012
HIV ሺ ݐ െ ͳ ሻ 0.270 0.145
Lymphoma ሺ ݐ െ ͳ ሻ -0.121 -0.090
Asthma ሺ ݐ െ ͳ ሻ -0.011 -0.002
Diabetes ሺ ݐ െ ͳ ሻ 0.043** 0.017
Hyperlipidemia ሺ ݐ െ ͳ ሻ -0.032** -0.026**
Oxygen use ሺ ݐ െ ͳ ሻ -0.003 0.043**
Specialist visit ሺ ݐ െ ͳ ሻ -0.011 -0.005
Spirometry tests ሺ ݐ െ ͳ ሻ 0.015** 0.014*
Death 0.126* -
Appropriate therapy ሺ ݐ െ ͳ ሻ -0.379** -0.520**
Dual 0.270** 0.066
Age 0.093** 0.037
Time 5 0.015* 0.056**
Time 6 0.045** -0.077**
Time 7 -0.079** -0.016
Time 8 -0.021** 0.056
Time 9 0.062** 0.084
Time 10 0.070** -0.024
Time 11 -0.051** 0.056
Constant -0.923 3.988

80

Oxyge n The rapy ve rsus No Oxyge n The rapy
Another analysis is performed comparing patients who did not use oxygen
therapy during the study period and all patients including those who did use oxygen. A
total of 13,552 patients are analyzed in this secondary analysis. In Table 13, we can
observe that comorbidities are the key difference between the two groups. When
evaluating the variables of interest, i.e., appropriate therapy and death, the results are
similar.

Dual Eligible ve rs us Not Dual Eligible
The last sensitive analysis performed is excluding patients who are also eligible for
Medicaid. From the final sample 17,511 patients are not dual eligible. Those patients
have specific characteristics that might impact the analysis, such as lower income. As
shown in Table 14, some coefficients are very different compared to the model including
all patients. For instance, the severity proxy measures that are not significant in the main
model become significant in the sub-analysis and vice-versa. In addition, the magnitude
of appropriate therapy coefficient is much larger in the new model.

81

Table 13: Oxygen therapy vers us no oxygen therapy.
All patients No oxygen Therapy
Previous cost ሺ ࢚ െ ૚ ሻ 0.192** 0.141**
Previous cost ሺ ࢚ െ ૛ ሻ -0.006 -0.042**
Chronic heart failure ሺ ࢚ െ ૚ ሻ 0.046** 0.052*
Valvular disease ሺ ࢚ െ ૚ ሻ 0.020 0.012
Pulmonary circulation disorders ሺ ࢚ െ ૚ ሻ 0.122 0.065
Peripheral vascular disorders ሺ ࢚ െ ૚ ሻ 0.012 0.082
Uncomplicated hypertension ሺ ࢚ െ ૚ ሻ 0.002 0.005
Complicated hypertension ሺ ࢚ െ ૚ ሻ 0.003 0.019
Other neurological disorders ሺ ࢚ െ ૚ ሻ 0.040 -0.005
Hypothyroidism ሺ ࢚ െ ૚ ሻ -0.009 -0.051*
Renal failure ሺ ࢚ െ ૚ ሻ 0.075** 0.039
Schizophrenia ሺ ࢚ െ ૚ ሻ -0.133* -0.094
Liver disease ሺ ࢚ െ ૚ ሻ -0.113 0.039
Depression ሺ ࢚ െ ૚ ሻ 0.057** 0.051
Peptic ulcer disease ሺ ࢚ െ ૚ ሻ 0.015 0.072
HIV ሺ ࢚ െ ૚ ሻ 0.270 0.193
Lymphoma ሺ ࢚ െ ૚ ሻ -0.121 0.046
Asthma ሺ ࢚ െ ૚ ሻ -0.011 0.005
Diabetes ሺ ࢚ െ ૚ ሻ 0.043** 0.049
Hyperlipidemia ሺ ࢚ െ ૚ ሻ -0.032** -0.057**
Oxygen use ሺ ࢚ െ ૚ ሻ -0.003 -
Specialist visit ሺ ࢚ െ ૚ ሻ -0.011 -0.010
Spirometry tests ሺ ࢚ െ ૚ ሻ 0.015** 0.017
Death 0.126* 0.271*
Appropriate therapy (t-1) -0.379** -0.369**
Dual 0.270** 0.273**
Age 0.093** 0.019
Time 5 0.015* 0.030*
Time 6 0.045** 0.072**
Time 7 -0.079** -0.067**
Time 8 -0.021** -0.014
Time 9 0.062** 0.103
Time 10 0.070** 0.141**
Time -0.051** 0.034
Constant -0.923 5.11**
** p-value < 0.01
* p-value < 0.05

82

Table 14: Dual eligible vers us not dual eligible patients .
All patients No dual eligible
Previous cost ሺ ࢚ െ ૚ ሻ 0.192** 0.191**
Previous cost ሺ ࢚ െ ૛ ሻ -0.006 -0.006
Chronic heart failure ሺ ࢚ െ ૚ ሻ 0.046** 0.031
Valvular disease ሺ ࢚ െ ૚ ሻ 0.020 0.039
Pulmonary circulation disorders ሺ ࢚ െ ૚ ሻ 0.122 0.151
Peripheral vascular disorders ሺ ࢚ െ ૚ ሻ 0.012 -0.005
Uncomplicated hypertension ሺ ࢚ െ ૚ ሻ 0.002 -0.005
Complicated hypertension ሺ ࢚ െ ૚ ሻ 0.003 0.0001
Other neurological disorders ሺ ࢚ െ ૚ ሻ 0.040 0.125**
Hypothyroidism ሺ ࢚ െ ૚ ሻ -0.009 -0.027
Renal failure ሺ ࢚ െ ૚ ሻ 0.075** 0.119**
Schizophrenia ሺ ࢚ െ ૚ ሻ -0.133* 0.222
Liver disease ሺ ࢚ െ ૚ ሻ -0.113 -0.146
Depression ሺ ࢚ െ ૚ ሻ 0.057** 0.060*
Peptic ulcer disease ሺ ࢚ െ ૚ ሻ 0.015 0.120
HIV ሺ ࢚ െ ૚ ሻ 0.270 -0.058
Lymphoma ሺ ࢚ െ ૚ ሻ -0.121 -0.046
Asthma ሺ ࢚ െ ૚ ሻ -0.011 -0.003
Diabetes ሺ ࢚ െ ૚ ሻ 0.043** 0.016
Hyperlipidemia ሺ ࢚ െ ૚ ሻ -0.032** -0.028*
Oxygen use ሺ ࢚ െ ૚ ሻ -0.003 -0.044*
Specialist visit ሺ ࢚ െ ૚ ሻ -0.011 -.023**
Spirometry tests ሺ ࢚ െ ૚ ሻ 0.015** 0.009
Death 0.126* 0.231**
Appropriate therapy (t-1) -0.379** -0.589**
Dual 0.270** -
Age 0.093** 0.095**
Time 5 0.015* 0.019
Time 6 0.045** 0.109**
Time 7 -0.079** -0.046**
Time 8 -0.021** -0.028**
Time 9 0.062** 0.089**
Time 10 0.070** 0.136**
Time 11 -0.051** -0.017
Constant -0.923 -0.938

** Significant at 99%
* Significant at 95%
83

5.2 Two-Part Model Res ults
The second methodology applied in this study is the two-part model, where in the
first part we evaluate the variables correlated to the probability of dying using the probit
model, and in the second part we compute the expected cost.
In Table 13, the factors correlated to the probability of dying are reported. We see
that the association between previous quarter costs and the probability of dying is
positive and statistically significant (p-value < 0.001). On the other hand, although most
of the lagged appropriate therapy usage coefficients are negatively correlated with death,
as expected, this correlation is not statistically significant when going further back in time
(p-value > 0.05). Only appropriate therapy usage in ݐ െ ͳ is statistically significant,
showing that the use of appropriate therapy decreases the probability of dying by 30%.
With regard to age, we observe that older patients are positively associated with the
probability of dying (p-value < 0.001).
Surprisingly, some comorbidities are negatively correlated to death in this study,
such as asthma, diabetes, and hyperlipidemia. Other comorbidities, on the other hand, are
positively associated with the outcomes, such as chronic heart failure and renal failure.
84

Table 15: Probit analys is to evaluate the probability of dying.

Coefficient P>z
Appropriate therapy ሺ ࢚ െ ૚ ሻ
-0.30 0.001
Appropriate therapy ሺ ࢚ െ ૛ ሻ
0.15 0.229
Appropriate therapy ሺ ࢚ െ ૜ ሻ
-0.01 0.883
Previous cost ሺ ࢚ െ ૚ ሻ
0.27 0.000
Previous cost ሺ ࢚ െ ૛ ሻ
0.05 0.000
Previous cost ሺ ࢚ െ ૜ ሻ
0.02 0.000
Copayments ሺ ࢚ െ ૚ ሻ
0.00 0.228
Copayments ሺ ࢚ െ ૛ ሻ
0.02 0.020
Copayments ሺ ࢚ െ ૜ ሻ
0.04 0.000
Dual eligible
-0.02 0.050
Age
0.02 0.000
Chronic heart failure ሺ ࢚ െ ૚ ሻ
0.25 0.000
Valvular disease ሺ ࢚ െ ૚ ሻ
0.07 0.000
Pulmonary circulation disorders ሺ ࢚ െ ૚ ሻ
0.16 0.430
Peripheral vascular disorders ሺ ࢚ െ ૚ ሻ
-0.06 0.222
Uncomplicated hypertension ሺ ࢚ െ ૚ ሻ
-0.03 0.004
Complicated hypertension ሺ ࢚ െ ૚ ሻ
-0.05 0.002
Other neurological disorders ሺ ࢚ െ ૚ ሻ
0.08 0.000
Hypothyroidism ሺ ࢚ െ ૚ ሻ
0.01 0.709
Renal failure ሺ ࢚ െ ૚ ሻ
0.11 0.000
Schizophrenia ሺ ࢚ െ ૚ ሻ
0.05 0.280
Liver disease ሺ ࢚ െ ૚ ሻ
0.13 0.000
Depression ሺ ࢚ െ ૚ ሻ
0.23 0.002
Peptic ulcer disease ሺ ࢚ െ ૚ ሻ
-0.09 0.411
HIV ሺ ࢚ െ ૚ ሻ
0.35 0.027
Lymphoma ሺ ࢚ െ ૚ ሻ
0.25 0.011
Asthma ሺ ࢚ െ ૚ ሻ
-0.10 0.000
Diabetes ሺ ࢚ െ ૚ ሻ
-0.02 0.045
Hyperlipidemia ሺ ࢚ െ ૚ ሻ
-0.21 0.000
Oxygen use ሺ ࢚ െ ૚ ሻ
-0.03 0.000
Specialist visit ሺ ࢚ െ ૚ ሻ
0.02 0.078
Spirometry tests ሺ ࢚ െ ૚ ሻ
0.02 0.026
Oxygen use ሺ ࢚ െ ૛ ሻ
0.05 0.000
Specialist visit ሺ ࢚ െ ૛ ሻ
0.02 0.906
Spirometry tests ሺ ࢚ െ ૛ ሻ
-0.02 0.013
Antibiotic use ሺ ࢚ െ ૚ ሻ
0.07 0.000
Time 5
0.13 0.000
Time 6
0.02 0.351
Time 7
-0.04 0.062
Time 8
0.02 0.245
Time 9
0.12 0.000
Time 10
-0.02 0.286
Time
-0.06 0.007
Constant
-3.87 0.000
** Significant at 99%
* Significant at 95%
85

By predicting the cost for each group, we can clearly observe significant
differences among them. Figure 8 shows that the cost pattern during the study period is
similar in all four groups. In particular, the cost in Groups I, II, and III decreases in the
first four observed quarters corresponding to the calendar year of 2007, then dramatically
increases in the beginning of 2008. We hypothesize that this increase is due to the
beginning of the winter season, which severely aggravates the symptoms of COPD. Right
after this, the cost slightly drops in the spring and summer, and increases a little in the
fall. At the end of the study, the predicted cost of most groups is higher than their
respective cost at the beginning of the study. The only exception is Group IV, which
finishes with a lower cost.
Figure 8: The predicted cos t per quarter-patie nt in the tw o-part model.

Figure 9 depicts the adjusted cost per quarter-patient when using the two-part
model. We see that Group IV has an average cost of only $4,962 per quarter-patient, 34%
$4,000
$4,500
$5,000
$5,500
$6,000
$6,500
$7,000
$7,500
$8,000
Group I
Group II
Group III
Group IV
**
**
**
86

less than the $7,560 cost of Group I (p-value <0.05). Even though the results from the
two-part model and the dynamic panel data model (see Figure 7) do not have the same
magnitude, both results have the same direction. As expected, patients who follow the
guidelines have a significantly lower health cost than patients who do not follow the
GOLD guidelines.

Figure 9: Total cos t comparis on between the two-part model and unadjus ted res ults .

0
1,000
2,000
3,000
4,000
5,000
6,000
7,000
8,000
Two Part Model Unadjusted Total Cost
Group I
Group II
Group III
Group IV
**
**
**
**
*
87

CHAPTER 6: DISCUSSION AND CONCLUSION
6.1 Dis cus s ion
Among Medicare beneficiaries, the estimated annual costs of COPD can range
from $21,409 (Schneider, O'Donnell and Dean 2009) to more than $45,600 per patient
(Dalal et al. 2010, Dalal et al. 2011b) depending on disease severity. However, there are
few studies that demonstrate the significant benefits of using appropriate medication
therapy.
The GOLD guidelines for COPD recommend the use of long-acting
bronchodilators and/or ICS for maintenance treatment for all patients with the condition,
except for patients with mild COPD, to whom a SABA may offer a suitable alternative
therapy. These recommendations are based on cumulative evidence base, which shows
that maintenance treatments reduce the exacerbation risk and thus the cost of health
service, notably ER visits and hospitalizations (Pauwels 2001). Our results corroborate
this assumption and show that the usage of appropriate therapy based on GOLD
guidelines is a cost-saving treatment strategy to manage COPD disease in the Medicare
population.
In this study, we compare the total health care expenditure among moderate-to-
severe COPD patients based on their behavior using appropriate therapy. Using a
dynamic panel data model, we identify moderate-to-severe COPD patients who received
at least one treatment recommended by GOLD guidelines between January 2006 and
December 2006 and followed them for up to two additional years.
From our analysis, the mean percentage of appropriate therapy during the study
period is around 35%. The majority of the patients (63%) are classified into the groups of
88

50% or lower appropriate therapy usage, indicating that patients and/or physicians are not
usually adherent to the GOLD guidelines. In particular, for the most severe patients under
oxygen therapy, the adherence to the guidelines is low. This study also shows that most
of the severe patients are not taking any treatment in addition to oxygen therapy, which is
not recommended.
Since we only include moderate-to-severe COPD patients over 65 years old, we
expected a high mortality rate, but the difference in the rate among the 4 groups is clear.
Patients in Group I using 0%-25% of appropriate therapy have an incidence of death
around 35%, which is 14% higher than the 21% mortality rate of patients in Group IV
using 75%-100% of appropriate therapy. Even after adjusting for relevant covariates, we
are still able to see an impact of appropriate therapies on the probability of dying. For
example, a moderate-to-severe patient using an appropriate therapy has less chance of
dying than a patient using an inappropriate therapy (p-value < 0.001).
In this study, we also generate an exhaustive list of covariates to cont URO V XEMHF WV¶ demographics, COPD disease severity, prior health care utilization, and comorbidity
profile. Overall, we find that the economic burden of the disease in this population is
approximately $5,872 per patient-quarter, on average. However, this cost significantly
changes according to the therapy used. Even in the unadjusted cost, patients using an
inappropriate therapy (Group I) have a 24% higher health care expenditure ($6,226) than
the expenditure ($5,006) of patients using an appropriate therapy (Group IV). This
difference increases even more when we compare the same groups after adjusting for
covariates in the panel data econometric models. In this case, patients using an
inappropriate therapy (Group I) have a cost of $10,451 per quarter, 65% higher than the
89

$6,317 spent by patients using an appropriate therapy (Group IV). Therefore, after
adjusting for relevant demographic and clinical covariates, a higher percentage of
appropriate therapy usage is correlated with lower health care cost.
Using a two-part model to control for endogeneity and selection bias, we observe
similar results. In this case, the difference among the four groups is also significant (p-
value < 0.05), with Group I having a total health-care expenditure of $7,560 per patient-
quarter and Group IV having a cost of only $4,962 in the real-world settings, resulting in
34% savings per patient-quarter (p-value < 0.01).

6.2 Limitations
Even though our analysis is able to provide a greater understanding of the role of
appropriate therapy in moderate-to-severe patients, this study still has a few limitations.
First, the acquired dataset provides restricted information about COPD patients and
services. For instance, our data only provides medication information for patients
enrolled in the Medicare Part D program. Therefore, we had to exclude other patients
from the sample and thus the results from the claims evaluation might not apply to other
COPD populations. In addition, the data also does not have information about services
which are not covered by, or billed to, Medicare. For instance, out-of-pocket expenses or
cost sharing for services not covered by Medicare are not included in our analysis.
Therefore, the costs reported here are likely to underestimate the per-patient burden of
COPD. Nonetheless, we believe that the addition of these excluded services would have a
minor overall effect in our results. If anything, including out-of-pocket expenses would
widen the cost differences between adherent and non-adherent patients.
90

Second, this is an observational study and therefore selection bias is an important
issue. For instance, smoking might have potential confounding effects in this study;
however, this information is not available. In addition, we might not capture all factors
that measure the severity of COPD, such as the result of the forced expiratory volume in
1 second (FEV1) and the peak expiratory flow (PEF) tests. Finally, it has been shown that
influenza vaccines have benefits for COPD patients; however, our dataset does not
provide accurate information about this factor.
Finally, although adherence to guidelines or to medication is widely used for
other chronic diseases, it might not be the best measure for COPD medication devices.
Basically, even though patients have the medications in their possession, it is not possible
to determine with certainty whether they are taking their medication in a proper or timely
manner. For instance, a significant factor that is usually not taken into account is the
patient's inadequate technique in using inhalation devices. Studies documented that
patients misuse their inhaler medications, and this has been shown to decrease the drug
efficacy (Batterink et al. 2012, Press et al. 2011)

6.3 Conclus ion
Despite these limitations, our study adds valuable insights to our present
knowledge-base on cost-efficient disease management strategies for COPD by
demonstrating the potential benefits of appropriate treatment on healthcare costs for
moderate-to-severe COPD patients.
Our findings suggest that the treatment recommended by the GOLD guidelines
for moderate-to-severe COPD patients leads to a significant reduction in health care
91

expenditures. This evidence support even more the benefits already known from
guidelines that lead to better clinical outcomes as well. Different approaches are used to
control for endogeneity and selection bias, which show that the results of the benefits of
appropriate therapy are consistent. These findings are able to provide insights to assist
healthcare organizations and disease management programs to better understand the
current practice patterns and to make their decisions about the intervention. Our results
further indicate that a major effort is required not only to disseminate the evidence-based
GOLD guidelines to healthcare providers and their patients, but also to educate them
about the long-term benefits of maintenance therapy for COPD management.
92

CHAPTER 7: REFERENCES
Global Initiative for Chronic Obstructive Lung Disease. Global strategy for the diagnosis,
management and prevention of chronic obstructive lung disease. NHLBI/WHO
workshop report. 2001.
National Heart, Lung, and Blood Institute Education Strategy Development Workshop:
Chronic Obstructive Pulmonary Disease, 1997.
2011. Chronic Obstructive Pulmonary Disease (COPD) Fact Sheet.
http://www.lung.org/lung-disease/copd/resources/facts-figures/COPD-Fact-
Sheet.html.
Akazawa, M., R. Halpern, A. A. Riedel, R. H. Stanford, A. Dalal & C. M. Blanchette
(2008a) Economic burden prior to COPD diagnosis: a matched case-control study
in the United States. Respir Med, 102, 1744-52.
Akazawa, M., D. C. Hayflinger, R. H. Stanford & C. M. Blanchette (2008b) Economic
assessment of initial maintenance therapy for chronic obstructive pulmonary
disease. Am J Manag Care , 14, 438-48.
Arellano, M. & S. R. Bond (1991) Some tests of specification for panel data: Monte
Carlo evidence and an application to employment equations. Review of Economic
S tudies, 58.
Asche, C., Q. Said, V. Joish, C. O. Hall & D. Brixner (2008) Assessment of COPD-
related outcomes via a national electronic medical record database. Int J Chron
Obstruct Pulmon Dis, 3, 323-6.
93

Asche, C. V., S. Leader, C. Plauschinat, S. Raparla, M. Yan, X. Ye & D. Young (2012)
Adherence to current guidelines for chronic obstructive pulmonary disease
(COPD) among patients treated with combination of long-acting bronchodilators
or inhaled corticosteroids. Int J Chron Obstruct Pulmon Dis, 7, 201-9.
Batterink, J., K. Dahri, A. Aulakh & C. Rempel (2012) Evaluation of the use of inhaled
medications by hospital inpatients with chronic obstructive pulmonary disease.
Can J Hosp Pharm, 65, 111-8.
Beauchesne, M. F., M. Julien, L. A. Julien, D. Piquette, A. Forget, M. Labrecque & L.
Blais (2008) Antibiotics used in the ambulatory management of acute COPD
exacerbations. Int J Chron Obstruct Pulmon Dis, 3, 319-22.
Beeh, K. M., O. Kornmann, J. Beier, M. Ksoll & R. Buhl (2004) Clinical application of a
simple questionnaire for the differentiation of asthma and chronic obstructive
pulmonary disease. Respir Med, 98, 591-7.
Blamoun, A. I., G. N. Batty, V. A. DeBari, A. O. Rashid, M. Sheikh & M. A. Khan
(2008) Statins may reduce episodes of exacerbation and the requirement for
intubation in patients with COPD: evidence from a retrospective cohort study. Int
J Clin Pract, 62, 1373-8.
Blanchette, C. M., M. Akazawa, A. Dalal & L. Simoni-Wastila (2008a) Risk of
hospitalizations/emergency department visits and treatment costs associated with
initial maintenance therapy using fluticasone propionate 500 microg/salmeterol 50
microg compared with ipratropium for chronic obstructive pulmonary disease in
older adults. Am J Geriatr Pharmacother, 6, 138-46.
94

Blanchette, C. M., M. Broder, C. Ory, E. Chang, M. Akazawa & A. A. Dalal (2009) Cost
and utilization of COPD and asthma among insured adults in the US. Curr Me d
Re s Opin, 25, 1385-92.
Blanchette, C. M., B. Gutierrez, C. Ory, E. Chang & M. Akazawa (2008b) Economic
burden in direct costs of concomitant chronic obstructive pulmonary disease and
asthma in a Medicare Advantage population. J Manag Care Pharm, 14, 176-85.
Blough, D. K., C. W. Madden & M. C. Hornbrook (1999) Modeling risk using
generalized linear models. J Health Econ, 18, 153-71.
Bond, S. (2002) Dynamic panel data models: a guide to micro data methods and practice.
Portuguese Economic Journal, Volume 1, 141-162.
Cameron, C. 2005. Microeconometrics: Methods and Applications.
Dahl, R., L. A. Greefhorst, D. Nowak, V. Nonikov, A. M. Byrne, M. H. Thomson, D.
Till, G. Della Cioppa & F. i. C. O. P. D. I. S. Group (2001) Inhaled formoterol dry
powder versus ipratropium bromide in chronic obstructive pulmonary disease. Am
J R e s pir Crit Care Me d, 164, 778-84.
Dalal, A. A., S. D. Candrilli & K. L. Davis (2011a) Outcomes and costs associated with
initial maintenance therapy with fluticasone propionate-salmeterol xinafoate 250
microg/50 microg combination versus tiotropium in commercially insured
patients with COPD. Manag Care, 20, 46-50, 53-5.
Dalal, A. A., L. Christensen, F. Liu & A. A. Riedel (2010) Direct costs of chronic
obstructive pulmonary disease among managed care patients. Int J Chron
Obstruct Pulmon Dis, 5, 341-9.
95

Dalal, A. A., H. Petersen, L. Simoni-Wastila & C. M. Blanchette (2009) Healthcare costs
associated with initial maintenance therapy with fluticasone propionate 250
ȝ J VDOP H WHU RO ȝ J F RP ELQDWLRQ YH UVXV D QWL F KROL QH U J LF EU RQ F KRGLO D WRUV LQ H OGHU O \ US Medicare-eligible beneficiaries with COPD. J Med Econ, 12, 339-47.
Dalal, A. A., M. Shah, A. O. D'Souza & P. Rane (2011b) Costs of COPD exacerbations
in the emergency department and inpatient setting. Re spir Me d, 105, 454-60.
Deyo, R. A., D. C. Cherkin & M. A. Ciol (1992) Adapting a clinical comorbidity index
for use with ICD-9-CM administrative databases. J Clin Epide m iol, 45, 613-9.
Dolce, J. J., C. Crisp, B. Manzella, J. M. Richards, J. M. Hardin & W. C. Bailey (1991)
Medication adherence patterns in chronic obstructive pulmonary disease. Chest,
99, 837-41.
Drescher, G. S., B. J. Carnathan, S. Imus & G. L. Colice (2008) Incorporating tiotropium
into a respiratory therapist-directed bronchodilator protocol for managing in-
patients with COPD exacerbations decreases bronchodilator costs. Re spir Care ,
53, 1678-84.
Duan, N. (1983) Sme aring es timate: a nonparametric retrans formation method. .
JAS A, 78, 605-610.
Elixhauser, A., C. Steiner, D. R. Harris & R. M. Coffey (1998) Comorbidity measures for
use with administrative data. Med Care, 36, 8-27.
Faulkner, M. A. & D. E. Hilleman (2002) The economic impact of chronic obstructive
pulmonary disease. Expert Opin Pharmacother, 3, 219-28.
96

Grasso, M. E., W. E. Weller, T. J. Shaffer, G. B. Diette & G. F. Anderson (1998)
Capitation, managed care, and chronic obstructive pulmonary disease. Am J
Re spir Crit Care Med, 158, 133-8.
Halpern, M. T., M. D. Zilberberg, J. K. Schmier, E. C. Lau & A. F. Shorr (2006) Anemia,
costs and mortality in chronic obstructive pulmonary disease. Cost Eff Resour
Alloc, 4, 17.
Hilleman, D. E., N. Dewan, M. Malesker & M. Friedman (2000) Pharmacoeconomic
evaluation of COPD. Chest, 118, 1278-85.
Hsiao, C. 2003. Analysis of panel data.
Hurd, S. (2000) The impact of COPD on lung health worldwide: epidemiology and
incidence. Chest, 117, 1S-4S.
Jennings, J. H., B. Digiovine, D. Obeid & C. Frank (2009) The association between
depressive symptoms and acute exacerbations of COPD. Lung, 187, 128-35.
Joo, M. J., T. A. Lee & K. B. Weiss (2008) Geographic variation of spirometry use in
newly diagnosed COPD. Chest, 134, 38-45.
Lainscak, M., L. M. Hodoscek, H. D. Düngen, M. Rauchhaus, W. Doehner, S. D. Anker
& S. von Haehling (2009) The burden of chronic obstructive pulmonary disease
in patients hospitalized with heart failure. Wien Klin Wochenschr, 121, 309-13.
Lin, P. J., F. T. Shaya & S. M. Scharf (2010) Economic implications of comorbid
conditions among Medicaid beneficiaries with COPD. Respir Med, 104, 697-704.
Mahler, D. A., J. F. Donohue, R. A. Barbee, M. D. Goldman, N. J. Gross, M. E.
Wisniewski, S. W. Yancey, B. A. Zakes, K. A. Rickard & W. H. Anderson (1999)
Efficacy of salmeterol xinafoate in the treatment of COPD. Chest, 115, 957-65.
97

Mancini, G. B., M. Etminan, B. Zhang, L. E. Levesque, J. M. FitzGerald & J. M. Brophy
(2006) Reduction of morbidity and mortality by statins, angiotensin-converting
enzyme inhibitors, and angiotensin receptor blockers in patients with chronic
obstructive pulmonary disease. J Am Coll Cardiol, 47, 2554-60.
Mannino, D. M., D. M. Homa, L. J. Akinbami, E. S. Ford & S. C. Redd (2002) Chronic
obstructive pulmonary disease surveillance--United States, 1971-2000. MMWR
Su r v e i l l Su m m , 51, 1-16.
Mapel, D. & M. Pearson (2002) Obtaining evidence for use by healthcare payers on the
success of chronic obstructive pulmonary disease management. Re spir Me d, 96
Suppl C, S23-30.
Mapel, D. W., J. S. Hurley, F. J. Frost, H. V. Petersen, M. A. Picchi & D. B. Coultas
(2000) Health care utilization in chronic obstructive pulmonary disease. A case-
control study in a health maintenance organization. Arch Intern Med, 160, 2653-8.
Mapel, D. W., J. S. Hurley, D. Roblin, M. Roberts, K. J. Davis, R. Schreiner & F. J. Frost
(2006) Survival of COPD patients using inhaled corticosteroids and long-acting
beta agonists. Respir Med, 100, 595-609.
Menzin, J., L. Boulanger, J. Marton, L. Guadagno, H. Dastani, R. Dirani, A. Phillips &
H. Shah (2008) The economic burden of chronic obstructive pulmonary disease
(COPD) in a U.S. Medicare population. Re spir Me d, 102, 1248-56.
Michaud, C. M., C. J. Murray & B. R. Bloom (2001) Burden of disease--implications for
future research. JAMA, 285, 535-9.
Nichols, J. (2007) Combination inhaled bronchodilator therapy in the management of
chronic obstructive pulmonary disease. Pharmacotherapy, 27, 447-54.
98

Pauwels, R. (2001) Global initiative for chronic obstructive lung diseases (GOLD): time
to act. Eur Re spir J, 18, 901-2.
Pauwels, R. A., A. S. Buist, P. Ma, C. R. Jenkins, S. S. Hurd & G. S. Committee (2001)
Global strategy for the diagnosis, management, and prevention of chronic
obstructive pulmonary disease: National Heart, Lung, and Blood Institute and
World Health Organization Global Initiative for Chronic Obstructive Lung
Disease (GOLD): executive summary. Re spir Care , 46, 798-825.
Pinckney, R. G., R. O'Brien, J. F. Piccirillo & B. Littenberg (2004) Evaluation of co-
morbidity indices in patients admitted for chronic obstructive pulmonary disease.
Monaldi Arch Chest Dis, 61, 209-12.
Press, V. G., V. M. Arora, L. M. Shah, S. L. Lewis, K. Ivy, J. Charbeneau, S. Badlani, E.
Nareckas, E. Naurekas, A. Mazurek & J. A. Krishnan (2011) Misuse of
respiratory inhalers in hospitalized patients with asthma or COPD. J Gen Intern
Med, 26, 635-42.
Puchelle, E. & B. B. Vargaftig (2001) Chronic obstructive pulmonary disease: an old
disease with novel concepts and drug strategies. Trends Pharmacol S ci, 22, 495-7.
Ramsey, S. D. & S. D. Sullivan (2004) Chronic obstructive pulmonary disease: is there a
case for early intervention? Am J Med, 117 Suppl 12A, 3S-10S.
Rascati, K. L., M. Akazawa, M. Johnsrud, R. H. Stanford & C. M. Blanchette (2007)
Comparison of hospitalizations, emergency department visits, and costs in a
historical cohort of Texas Medicaid patients with chronic obstructive pulmonary
disease, by initial medication regimen. Clin The r, 29, 1203-13.
99

Rennard, S. I. (1998) COPD: overview of definitions, epidemiology, and factors
influencing its development. Chest, 113, 235S-241S.
Rice, D. P., T. A. Hodgson & A. N. Kopstein (1985) The economic costs of illness: a
replication and update. Health Care Financ Rev, 7, 61-80.
5 RRGPDQ ' +RZ WR 'R [ WDERQG $Q , QWU RGXF WL RQ WR ³ 'LII H UH QF H ´ D QG ³ 6 \ VWHP´ GMM in Stata. Center for Global Development.
Ruchlin, H. S. & E. J. Dasbach (2001) An economic overview of chronic obstructive
pulmonary disease. Pharmacoe conomics, 19, 623-42.
Salinas, G. D., J. C. Williamson, R. Kalhan, B. Thomashow, J. L. Scheckermann, J.
Walsh, M. Abdolrasulnia & J. A. Foster (2011) Barriers to adherence to chronic
obstructive pulmonary disease guidelines by primary care physicians. Int J Chron
Obstruct Pulmon Dis, 6, 171-9.
Schneider, K. M., B. E. O'Donnell & D. Dean (2009) Prevalence of multiple chronic
conditions in the United States' Medicare population. Health Qual Life Outcomes,
7, 82.
Shaya, F. T., D. Dongyi, M. O. Akazawa, C. M. Blanchette, J. Wang, D. W. Mapel, A.
Dalal & S. M. Scharf (2008) Burden of concomitant asthma and COPD in a
Medicaid population. Chest, 134, 14-9.
Simoni-Wastila, L., H. W. Yang, C. M. Blanchette, L. Zhao, J. Qian & A. A. Dalal
(2009) Hospital and emergency department utilization associated with treatment
for chronic obstructive pulmonary disease in a managed-care Medicare
population. Curr Med Res Opin, 25, 2729-35.
100

Stuart, B. C., J. A. Doshi & J. V. Terza (2009) Assessing the impact of drug use on
hospital costs. Health S erv Res, 44, 128-44.
Stuart, B. C., L. Simoni-Wastila, I. H. Zuckerman, A. Davidoff, T. Shaffer, H. W. Yang,
J. Qian, A. A. Dalal, D. W. Mapel & L. Bryant-Comstock (2010) Impact of
maintenance therapy on hospitalization and expenditures for Medicare
beneficiaries with chronic obstructive pulmonary disease. Am J Ge riatr
Pharmacother, 8, 441-53.
Sullivan, S. D., S. D. Ramsey & T. A. Lee (2000) The economic burden of COPD. Chest,
117, 5S-9S.
Søyseth, V., P. H. Brekke, P. Smith & T. Omland (2007) Statin use is associated with
reduced mortality in COPD. Eur Re spir J, 29, 279-83.
van der Valk, P., E. Monninkhof, J. van der Palen, G. Zielhuis & C. van Herwaarden
(2002) Effect of discontinuation of inhaled corticosteroids in patients with chronic
obstructive pulmonary disease: the COPE study. Am J Re spir Crit Care Me d, 166,
1358-63.
van Schayck, C. P. & J. Reid (2006) Effective management of COPD in primary care -
the role of long-acting beta agonist/inhaled corticosteroid combination therapy.
Prim Care R e spir J, 15, 143-51.
Vincken, W., J. A. van Noord, A. P. Greefhorst, T. A. Bantje, S. Kesten, L. Korducki, P.
J. Cornelissen & D. B. T. S. Group (2002) Improved health outcomes in patients
with COPD during 1 yr's treatment with tiotropium. Eur Re spir J, 19, 209-16.
Wilson, L., E. B. Devine & K. So (2000) Direct medical costs of chronic obstructive
pulmonary disease: chronic bronchitis and emphysema. Respir Med, 94, 204-13.
101

Wooldridge, J. 2002. Econometric analysis of cross section and panel data.
Wooldridge, J. M. 2001. Econometric Analysis of Cross S ection and Panel Data.
Cambridge, Massachusetts: The MIT Press.
Wouters, E. F. (2003) Economic analysis of the Confronting COPD survey: an overview
of results. Respir Med, 97 Suppl C, S3-14.
Wu, E. Q., H. G. Birnbaum, M. Cifaldi, Y. Kang, D. Mallet & G. Colice (2006)
Development of a COPD severity score. Curr Med Res Opin, 22, 1679-87.
Yang, I. A., M. S. Clarke, E. H. Sim & K. M. Fong (2012) Inhaled corticosteroids for
stable chronic obstructive pulmonary disease. Cochrane Database S yst Rev, 7,
CD002991.
Zanobetti, A., M. A. Bind & J. Schwartz (2008) Particulate air pollution and survival in a
COPD cohort. Environ He alth, 7, 48.

Abstract (if available)

Abstract OBJECTIVES: This study proposes to investigate the association among adherence to guidelines, patients characteristics, and health care expenditures using methods for handling endogeneity and selection bias. ❧ METHODS: The study population consists of the random 5% COPD cohort in Medicare for the time period between 2006 and 2008. Patients are classified in 4 groups based on the percentage of use of appropriate therapy based on the GOLD guidelines among the quarters: Group I) 0% - 25%

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Selected papers on methods for evaluating the impact of marijuana use on BMI and other risk factors for metabolic syndrome

PDF

Estimation of heterogeneous average treatment effect-panel data correlated random coefficients model with polychotomous endogenous treatments

PDF

Three essays on comparative economic and clinical outcomes of antipsychotics using retrospective claims data

PDF

The impact of treatment decisions and adherence on outcomes in small hereditary disease populations

PDF

Selected papers on the evaluation of healthcare costs of prematurity and necrotizing enterocolitis using large retrospective databases

PDF

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

PDF

Discriminating changes in health using patient-reported outcomes

PDF

Impact of pharmacy-based transitional care on healthcare utilization and costs

PDF

Thesis on Medicare Part D

PDF

Understanding the value of hemophilia therapies

PDF

Advances and applications for economic evaluation methods in health technology assessment (HTA)

PDF

The effect of using fibrates in conjunction with statins for the management of dyslipidemia in persons with Type II diabetes mellitus

PDF

Understanding primary nonadherence to medications and its associated healthcare outcomes: a retrospective analysis of electronic medical records in an integrated healthcare setting

PDF

The impact of public expenditures on health care on total health expenditures: an exploratory analysis of selected OECD countries

PDF

Economic aspects of obesity

PDF

Assessment of the impact of second-generation antipscyhotics in Medi-Cal patients with bipolar disorder using panel data fixed effect models

PDF

Evaluating cancer treatments with health economics methods

PDF

Effects of a formulary expansion of the use of SSRIs and health care services by depressed patients in the California Medicaid program

PDF

Essays on the use of microsimulation for health and economic policy analysis

PDF

Essays in pharmaceutical and health economics

Asset Metadata

Creator Ejzykowicz, Flavia (author)

Core Title The impact of adherence to guidelines on the health care expenditures of COPD patients

School School of Pharmacy

Degree Doctor of Philosophy

Degree Program Pharmaceutical Economics and Policy

Publication Date 05/08/2013

Defense Date 10/22/2012

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

Tag COPD,Cost,dynamic panel data,expenditures,instrumental variables,OAI-PMH Harvest,two-part model

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Hay, Joel W. (committee chair), Romley, John A. (committee member), Wu, Vivian (committee member)

Creator Email ejzykowi@usc.edu,flaviaej@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c3-252682

Unique identifier UC11293341

Identifier etd-Ejzykowicz-1669.pdf (filename),usctheses-c3-252682 (legacy record id)

Legacy Identifier etd-Ejzykowicz-1669.pdf

Dmrecord 252682

Document Type Dissertation

Format application/pdf (imt)

Rights Ejzykowicz, Flavia

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