Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Causal estimation of the effect of medication compliance on health outcomes
(USC Thesis Other)
Causal estimation of the effect of medication compliance on health outcomes
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
CAUSAL ESTIMATION OF THE EFFECT OF MEDICATION COMPLIANCE ON
HEALTH OUTCOMES
by
Andrew Peng Yu
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF THE PHILOSOPHY
(PHARMACEUTICAL ECONOMICS AND POLICY)
May 2006
Copyright 2006 Andrew Peng Yu
UMI Number: 3237184
3237184
2007
Copyright 2006 by
Yu, Andrew Peng
UMI Microform
Copyright
All rights reserved. This microform edition is protected against
unauthorized copying under Title 17, United States Code.
ProQuest Information and Learning Company
300 North Zeeb Road
P.O. Box 1346
Ann Arbor, MI 48106-1346
All rights reserved.
by ProQuest Information and Learning Company.
ii
DEDICATION
To my wife, Yanni, for her love and support.
iii
TABLE OF CONTENTS
Dedication........................................................................................................................... ii
List of Tables ...................................................................................................................... v
List of Figures.................................................................................................................... vi
Abstract............................................................................................................................. vii
Preface................................................................................................................................ ix
Chapter 1. Introduction ....................................................................................................... 1
1.1 Causal inference and marginal structural models................................................. 1
1.2 Compliance effect, goal, hypothesis and solutions .............................................. 4
1.3 Organization of this paper .................................................................................... 6
Chapter 2. Background, Literature Critique and A Conceptual Model .............................. 8
2.1 Background on glycemic control and diabetic complications.............................. 8
2.2 Compliance research .......................................................................................... 17
2.3 Compliance-outcomes relationship .................................................................... 28
2.4 A conceptual causal model of compliance ......................................................... 38
Chapter 3. Study Design and Methods ............................................................................. 48
3.1 Data..................................................................................................................... 48
3.2 Sample selection................................................................................................. 49
3.3 Compliance, covariates and outcomes measurement ......................................... 51
3.4 Descriptive Analysis........................................................................................... 60
3.5 Cox proportional hazard models ........................................................................ 60
3.6 Marginal structural model .................................................................................. 64
Chapter 4. Results ............................................................................................................. 76
4.1 Patient characteristics ......................................................................................... 76
4.2 Compliance changing over time......................................................................... 82
4.3 Survival analysis and Cox regression models .................................................... 87
4.4 Demonstration of time-varying confounding ..................................................... 90
4.5 Causal compliance effect estimated by MSM with IPTW ................................. 94
4.6 Sensitivity analysis ............................................................................................. 98
Chapter 5. Discussion ..................................................................................................... 101
5.1 Overview .......................................................................................................... 101
5.2 The measurement of medication compliance ................................................... 103
5.3 The compliance effect ...................................................................................... 109
5.4 Intuitive understanding of the MSM with IPTW estimate ............................... 111
5.5 Advantages and disadvantages of marginal structural models......................... 113
iv
5.6 Potential impact of unmeasured confounders................................................... 114
5.7 Comparison of the five models......................................................................... 116
5.8 Limitations and future research........................................................................ 119
5.9 Implication to other health outcomes research................................................. 120
5.10 Summary......................................................................................................... 121
Bibliography ................................................................................................................... 123
v
LIST OF TABLES
Table 1. Classification of hypoglycemics......................................................................... 14
Table 2. FDA approved status for hypoglycemics combination regimens....................... 15
Table 3. ICD-9 codes for chronic microvascular complications of diabetes.................... 58
Table 4. Summary of model specification ........................................................................ 62
Table 5. Sample selection criteria and patient counts flow .............................................. 76
Table 6. Characteristics of the study population (n = 4,708)............................................ 77
Table 7. Baseline patient characteristics by one year compliance.................................... 78
Table 8. Time-varying covariates at baseline and the 4th quarter by compliance............ 80
Table 9. Hypoglycemic regimens by selected quarters .................................................... 82
Table 10. Hazard ratio of compliance variable by Cox regression for Model 1-4 ........... 88
Table 11. Time varying covariates as risk factors of survival .......................................... 90
Table 12. Time-varying covariates can predict subsequent compliance .......................... 92
Table 13. Time-varying covariates can be predicted by past compliance history............ 92
Table 14. Significance levels of time-varying confounders by three criteria................... 94
Table 15. Comparison of compliance effect estimates of Model 1-5............................... 98
Table 16. Sensitivity analysis on dichotomized compliance with different MPR cutoff
points......................................................................................................................... 99
vi
LIST OF FIGURES
Figure 1. A plot of the percentage change in outcome in relation to the compliance rate
used in 22 cost-effectiveness evaluations of drug therapies. Source: Hughes et al.
(2001)........................................................................................................................ 22
Figure 2. Hypothetical results of sequential tests in four "50% compliers". Source: Leon
Gordis (1979)............................................................................................................ 36
Figure 3. A closer look at a real-time medication compliance pattern using electronic
monitoring device. .................................................................................................... 37
Figure 4. A causal diagram for one-period conceptual model of compliance.................. 40
Figure 5. Causal diagram for a dynamic compliance model ............................................ 47
Figure 6. Diagram of study timeline................................................................................. 51
Figure 7. Diagrams of assumptions required for causal estimates by five corresponding
models....................................................................................................................... 64
Figure 8. Causal diagrams of time-varying exposure and covariates ............................... 69
Figure 9. Box plot of compliance (MPR) of hypoglycemics for each quarter. Bars
represent interquartile range; Width of the bar represents sample size for that
quarter. ...................................................................................................................... 83
Figure 10. Quarterly compliance of 10 randomly selected patients ................................. 85
Figure 11. Mean quarterly non-zero MPR by hypoglycemic classes ............................... 85
Figure 12. Mean MPR and hypoglycemic classes for all quarters with non-zero MPR.
The bubble size is proportional to the number of quarters with the hypoglycemic
class........................................................................................................................... 87
Figure 13. Unadjusted Kaplan-Meier survival curve by one-year compliance ................ 88
Figure 14. Complication-free survival curves adjusted by baseline covariates................ 89
Figure 15. Log of stabilized weights distribution by quarter............................................ 96
vii
ABSTRACT
In health outcomes research, treatment (or exposure) is often not a point
intervention, but a prolonged and time-varying process as in the case with medication
therapy for chronic conditions. After therapy initiation, many factors dynamically
interact with treatment decisions over time (i.e., time-varying confounding). For
example, intermediate outcomes that are affected by past treatment may also influence
future treatment decisions. Despite controlling for potential confounders at baseline,
standard regression models that ignore time-varying confounding will result in biased
estimate of treatment effects. Marginal structural models (MSMs) are a novel class of
causal models for estimating treatment effects of a time-varying exposure in the presence
of time-varying confounders.
In this study, a MSM was applied to estimate the causal effects of medication
compliance with hypoglycemics on the risk of microvascular complications. We
conducted a retrospective longitudinal cohort study on patients with type 2 diabetes using
the California Medicaid claims database (1995-2002). Medication compliance and
multiple time-varying confounders were measured quarterly over a maximum of 7.5
years follow up. Of 4,708 eligible patients, 2,644 (56.2%) experienced microvascular
complications during the follow up period. After controlling for baseline covariates,
standard Cox model estimated that compliance was associated with increased risk of
complication with hazard ratio of 1.09 (95% confidence interval (CI): 1.00, 1.18). After
adjusting for time-varying confounders as exogenous variables, the estimate was 0.96
(95% CI: 0.88, 1.04). In contrast, the MSM estimated the risk ratio as 0.76 (95%
viii
bootstrap CI: 0.60, 0.92), indicating a significant benefit of medication compliance with
hypoglycemics on the reduction of microvascular complications. The standard Cox
models failed to demonstrate the benefit of medication compliance because they did not
appropriately account for the dynamic interaction between time-varying confounders and
medication compliance over time.
In contrast to the past efforts of establishing compliance-outcomes association,
this study estimated the causal effects of medical compliance on health outcomes in the
presence of time-varying confounding. The findings are consistent with prior clinical
trial findings that intensive glycemic control reduces the risk of microvascular
complications in type 2 diabetic patients.
ix
PREFACE
The past decade has witnessed a rapid advancement of statistical and
epidemiological methodology in estimating the treatment effect of an intervention (or
exposure), the very question of most empirical health outcomes research. Despite the
availability of the advanced models and their superiority, standard single regression
models are still the predominant methodology when addressing treatment-outcomes
relationship, which Cook and Campbell recognized decades ago as “the error of
employing forecasting techniques for causal inference” (Cook and Campbell 1979).
A novel class of causal models, marginal structural models (MSMs), is able to
estimate the causal effect of time-varying treatment in the presence of time-varying
confounding, a ubiquitous phenomenon in health outcomes research settings. This is a
new generation of statistical model that has been presented in a solid and formal manner
with proven and evidenced validity, and efforts have been made to address its feasibility
in empirical application, often with richly collected clinical trial data.
Due to the presence of simultaneous confounders that dynamically interact with
compliance decisions over time, standard regression models cannot appropriately address
such time-varying confounding effect thus bias the estimate of compliance effect, even
with opposite direction. MSM is applied in this study to estimate the causal effect of
compliance on outcomes. In addition, this study exemplifies the use of this causal model
within commonly available administrative claims data, and shows its abundant yet
complex information that is often overlooked and can be analyzed repeatedly over time.
This paper is intended to present this model to general health outcomes
x
researchers in a less technical but more practical manner, with a special focus on its
implementation, feasibility and intuitive interpretation. As a hope, this study could evoke
the interest of empirical researchers in probing a causal, rather than “associational”
relationship between treatment and outcomes. It’s also a hope to make this model more
accessible and applicable in future empirical research. Notwithstanding the present
exercise of this model in estimating the compliance effect, its application can
significantly contribute scientific knowledge in a much broader range of research
pertaining to clinical, economic, and humanistic (quality of life) outcomes of medical or
social interventions or exposures.
This application in estimating the compliance on outcomes implies no way a
definite answer to this question of causal effect, but to present health researchers a
method that can improve the existing standard methodology.
Credit where credit is due. Special appreciations are given to Jamie Robins and
his colleagues, for their endless endeavor and unprecedented innovation and contribution
in casual effect estimation.
1
CHAPTER 1. INTRODUCTION
1.1 Causal inference and marginal structural models
The aim of treatment evaluation is to estimate the marginal (incremental) effect of
an intervention or exposure, by comparing the potential outcomes with the intervention
that had been given and had not been given to a population. Such problem can be
formulated by counterfactual (or potential outcomes) models (Rubin 1974; Holland
1986), and its estimation always presents a challenge because, for any given subject, we
can only observe the counterfactual outcomes that the subject actually receives.
Nevertheless, such effort of estimation is the core interest in economic, social and health
sciences, where the correct estimation of the direction and magnitude of the effect of a
specific treatment/exposure has important policy implication. Various terms have been
used in different fields for such effort, such as “casual estimation”, “causal inference”,
“average treatment effect (ATE)”, “treatment effect”, “identification”, “adjusting
selection bias”, “program evaluation”, “quasi-experimentation”, “internal validity”, etc.
In observational studies using longitudinal data that aim at estimating the effect of
a certain treatment on outcomes, it is a typical scenario that treatment decisions non-
randomly change over time. The changes may be triggered by observable factors such as
change of health status or patient response to the past treatment (e.g. previous outcomes
or surrogate). Even in well-controlled randomized clinical trials, patients do not always
comply with assigned therapy.
For example, filgrastim, an adjuvant therapy that boosts white blood cell and
reduces toxicity effect of chemotherapy, is often given to patients who have
2
experienced infection or had very low white blood cell (WBC) count in the past
chemotherapy cycles, and the decision of giving filgrastim is made on a cycle base. As a
more general example, the initiation of a new therapy (e.g. biologicals for rheumatoid
arthritis) is often based on the patient’s unfavorable response to the conventional therapy,
as what is usually described in a prior authorization process, and the continuation of such
therapy is influenced by responsiveness to the previous treatment. In these examples,
when accessing the treatment outcomes (e.g. hospitalizations, incremental treatment
costs, etc.) of the new (often more expensive) therapies vs. conventional therapies over
time, a standard regression model, even with attempt of adjusting a number of observed
baseline variables, may only generate “association” estimation with a sign probably
opposite to what we would reasonably expect for a causal estimate. Such phenomenon
can be explained by “selection bias” in observational studies.
Selection bias often results from the existence of unmeasured confounders, such
as severity level or health status, which influence outcomes of interest and are observable
by the clinician in treatment selection but unobserved or uncontrolled by the researcher.
In a longitudinal study where treatments, covariates, or outcomes are measured
repeatedly over time, in addition to unmeasured confounders, selection bias can also be
produced by the presence of observable time-varying confounders if their confounding
effects are ignored or not appropriately accounted for.
Simply speaking, time-varying confounders interact with treatment and correlate
with outcomes as well. Unlike a point treatment such as a surgical procedure,
medications for chronic conditions involve treatment over a period of time, and often, if
not always, the actual regimens change over time with concurrent underlying
3
factors. In this sense, time-varying confounding effect “is ubiquitous in
pharmacoepidemiology” (Hernan et al. 2005). For example, very low WBC count plays
a major role in the decision of administering filgrastim. As a result of such therapy,
WBC count may be well improved, and then the therapy may be continued or the dose
may be titrated upon the response. Obviously, WBC count is also a predictor of
outcomes of interest (e.g. hospitalization due to infection or febrile neutropenia).
Therefore, WBC count is a time-varying confounder that interacts with filgrastim
regimen over time. When estimating the effect of filgrastim on outcomes, ignoring the
interplay between filgrastim and WBC count over time and only controlling WBC count
at baseline will result in biased estimate due to residual confounding (Robins 1989).
Even adjusting WBC count as a time varying covariate in a standard longitudinal
regression, such as general estimating equation (GEE) or Cox models, the estimates can
still be biased (Robins 1997; Brumback et al. 2003) because WBC count is assumed to be
independent of treatment over time and entered as an exogenous variable in that case.
Specification of a full parametric dynamic sequential treatment model, such as
structural nested model, with presence of time-varying confounders is convoluted and
inefficient, and also requires large sample size (Robins 1999). The accuracy of the casual
estimate depends on the model specification that is usually unknown in empirical
research. Due to the complexity of model specification and method of estimation, the
application of the dynamic model has been rarely seen in health outcomes research.
Subsequent to the development of structural nested models, Robins and
colleagues developed a set of new class of causal models called marginal structural
models (MSM) that can adjust time-varying confounders and provide causal
4
estimate with weaker assumptions (Robins 1998; Robins 1999; Robins et al. 2000). It is
a semi-parametric model that uses inversed probability of treatment weight (IPTW) for
estimation. Unlike its precedent nested structural model with g-estimation formula,
MSM is a semi-parametric model and has the advantage of being more feasible and easily
applicable to health outcomes research with standard statistical software. MSM analysis
provides an asymptotically unbiased estimate of casual effect by using the inverse of the
conditional probability of receiving one’s actual treatment as weight. The details of the
model will be described in the method section.
1.2 Compliance effect, goal, hypothesis and solutions
The standard tool for evaluating the causal effect of a treatment is a randomized
controlled clinical trial. However, medication compliance, similar to many other
exposures (treatments), cannot be randomized because by definition compliance is a
patient’s subjective choice.
In observational studies, estimating the effect of medication compliance is
challenging to health outcomes researchers, because compliance is self-selected and the
decision of being compliant interact with many complex factors, and they co-evolve over
time. Standard single regression models (e.g., linear, logistic, Cox proportional hazard,
GEE models, etc.) have been widely applied to estimate such effect in almost all cases,
yet it can only estimate association (or partial correlation) between compliance and
outcomes (Cook and Campbell 1979; DiMatteo et al. 2002). With these traditional
models, the exogenous assumptions on which causal interpretation relies appear to be
violated, as compliance is not constant but inherently interact with other factors over
5
time. Consequently, resulting association estimates “will not be helpful to either
scientists, who will be unable to relate it to a mechanism, or policy makers, who will be
unable to translate it into effective interventions” (Hernan 2005).
The goal of this study is to evaluate the effect compliant use of hypoglycemic
(antidiabetics) medications on microvascular complications-free duration, among type 2
diabetes patients.
This study is also to demonstrate the merit of MSM which overcomes the
previously stated methodological problems when estimating the causal effect of
compliance on health outcomes in a longitudinal setting in comparison with conventional
models.
As generally believed, compliance (vs. noncompliance) should have beneficial
effect on health outcomes by delaying the onset of microvascular complication in diabetic
patients, which is also the hypothesis of this study.
Accounting for observable time-varying confounders, MSM with inverse
probability of treatment weights (IPTW) is applied to estimate the effect of medication
compliance on delaying the onset of microvascular complication. In contrast to
conventional approaches, compliance is treated as a dynamic process. Acknowledging the
presence of time-varying confounding, this study measured compliance repeatedly
(quarterly) over time and strictly prior to the measurement of outcomes to ensure the
erogeneity. Other covariates (e.g. concurrent medications, pill burden, comorbidities and
office/hospital visits) that change over time are also measured by quarter so that their
dynamic interaction with compliance and confounding effect could be controlled for in
the process of estimating the effect of compliance on outcomes.
6
To the best knowledge, this study is the first one that answers the question
whether medication compliance casually improves, rather than associates with, health
outcomes. This study is also the first application of marginal structural model to estimate
the causal effect in a longitudinal observational study using administrative claims data.
1.3 Organization of this paper
This paper includes five chapters, and the first chapter is to present the problem of
causal inference and introduction of marginal structural model.
Chapter 2 provides background and literature review and critiques including 1)
diabetes, treatment with hypoglycemics, glycemic control and diabetes complications, 2)
definitions and measurement of compliance, and compliance-outcomes relationship, 3)
previous studies of compliance-outcomes relationship in diabetes, and problems on
methodologies presented in these studies.
To illustrate the dynamic pattern of medication compliance and to clarify the
concept of compliance, causal conceptual diagrams of both point-treatment and dynamic
versions are presented in this chapter. The conceptual diagrams serve as a basis to
illustrate the intricate dynamic causal relationship between compliance and other factors,
thus a model appropriate for such dynamic interaction is warranted. Based on this
conceptual model, time-varying confounding effect in compliance is introduced.
Chapter 3 describes the data, patient sample selection criteria, construction of
study variables and modeling strategies including standard Cox proportional hazard
models and a marginal structural Cox model, which is more formally presented in detail
here.
7
Chapter 4 describes the results and findings, following proposed methods in
Chapter 3.
Chapter 5 discusses the compliance measure and content of compliance effects,
intuitive interpretation of MSM and its advantage and disadvantages, limitation and
future work and implication of this study. It ends with concluding remarks.
8
CHAPTER 2. BACKGROUND, LITERATURE CRITIQUE AND A
CONCEPTUAL MODEL
2.1 Background on glycemic control and diabetic complications
2.1.1. Diabetes
Diabetes mellitus (DM) is a metabolic disease characterized by hyperglycemia
(high blood glucose), resulting from impaired insulin secretion and insulin resistance.
High concentration of blood glucose (hyperglycemia) is the criterion of diagnosis
diabetes mellitus(Expert Committee on the and Classification of Diabetes 2003). In the
U.S., National Health Interview Survey (NHIS) reported that about 4.2% of U.S.
population have diagnosed diabetes from 1998-2000 (American Diabetes Association
2003). Prevalence estimates based on a larger sample telephone survey was 7.3% in
2000 (Mokdad et al. 2001), and 7.9% in 2001 (Mokdad et al. 2003). Among all diabetic
patients, about 15% have type-1 diabetes (formerly known as insulin-dependent diabetes,
or IDDM), and 85% have type 2 diabetes (formerly known as non-insulin-dependent
diabetes, or NIDDM). Total economic cost of diabetes to the US was estimated to be
$132 billion in 2002 including medical expenditures and productivity loss (American
Diabetes Association 2003).
2.1.2. Diabetes complications
Complications of diabetes are a major cause of morbidity and mortality among
diabetic patients. Chronic complications include microvascular diseases (e.g.
retinopathy, nephropathy and neuropathy) and macro-vascular diseases (e.g. coronary
9
heart disease, myocardial infarction, cerebrovascular disease or peripheral vascular
disease).
Diabetic neuropathy is the most common chronic complication in diabetes,
though it is least known among all diabetic complications (Perkins and Bril 2003). The
prevalence of diabetic neuropathy was estimated to be 66% to 75% by different
measurement criteria (Dyck and Dyck 1999; Simmons and Feldman 2002). Diabetic
neuropathy leads to visual, gastrointestinal, sexual, and peripheral vascular abnormalities,
results in pain, deteriorates patients’ quality of life, and appears to be a major reason for
physical disability in elderly diabetic patients (Resnick et al. 2002). Diabetic neuropathy
is a main indirect cause of diabetic foot problems. The incidence rate of foot ulcer in
diabetic patients is about 2-3% a year, and lifetime risk is about 15% (Reiber et al. 1998).
About 15–27% of all foot ulcers will result in some type of amputation (Jeffcoate and
Harding 2003).
Diabetic retinopathy is the major cause of blindness in the adult population. All
type 1 patients and over 60% of type 2 patients will have retinopathy within twenty years
of diabetes onset (Fong et al. 2003).
Type 2 diabetic nephropathy is the primary cause of end-stage renal disease in the
entire western world (Wolf and Ritz 2003). It is very costly in terms of morbidity,
mortality and burden to the society. It was estimated that approximately one third of type
1 and one quarter of type 2 diabetic patients would develop diabetic nephropathy
(Andersen et al. 1983; Green et al. 1985; Ballard et al. 1988). Because of the enormous
costs associated with treatment of renal disease, even interventions with modest
effectiveness in preventing nephropathy are deemed to be cost-effective
10
(Rippin et al. 2004). Therefore, prevention and treatment of diabetic nephropathy has
become a prominent goal in the treatment of patients with diabetes mellitus.
Not only do diabetic complications increase patient’s mortality, morbidity and
worsen their quality of life, complications also increase the economic burden to patients,
health care system and the society. As patients begin experiencing diabetic
complications, the costs of treatment will keep rising (Bjork 2001; O'Brien et al. 2001).
In the year of 2002, the direct costs of diabetes-attributable complications in the U.S. was
24.6 billion dollars, which was more than direct costs of diabetes care (i.e., 23.2 billion)
(American Diabetes Association 2003).
It is clear that hyperglycemia is linked to diabetic microvascular complications
(retinopathy, nephropathy, and neuropathy) in both type 1 and type 2 diabetes. However,
the relationship between hyperglycemia and macrovascular diseases were controversial
(American Association of Clinical Endocrinologists 2002). Approximately 70% of
patients with type 2 diabetes died of macrovascular diseases, and negative influence of
diabetes on cardiovascular diseases was evidenced (Julien 1997). However,
macrovascular diseases are less specific and less attributable to diabetes than
microvascular diseases. The development of macrovascular disease appears to precede
the onset of diabetes, and is associated with pre-diabetic glucose tolerance (Nathan et al.
1997; Coutinho et al. 1999). Patients with diabetes were found to be at high risk of heart
failure (Chin and Goldman 1996; Reis et al. 1997); conversely, heart failure was also an
independent risk factor of developing diabetes (Amato et al. 1997). Sharing the common
metabolic and genetic risk factors with diabetes, macrovascular diseases such as
11
abdominal obesity, hypertension, and hyperlipidemia are often viewed as comorbidities
of diabetes as well (Bell 2003).
In a retrospective study of residents of Rochester, the prevalence of retinopathy in
diabetic patients was 2.6% at the time of initial diagnosis. Among people without
retinopathy at initial diagnosis, the incidence of any retinopathy was 17.4 per 1000
person-years. After 20 years since the initial diagnosis, the cumulative incidence of
retinopathy would be over 30% (Dwyer et al. 1985). The prevalence of proteinuria was
8.2% at the diagnosis of type 2 diabetes, and subsequent incidence was 15.3 per 1000
person-years. The cumulative incidence of persistent proteinuria (an indication of
nephropathy) was 24.6% after 20 years (Ballard et al. 1988).
2.1.3. Glycemic control
Although the direct causal effect of hyperglycemia on diabetic complications was
not clearly understood or absolutely certain, the positive correlation of elevated blood
glucose levels, or its reliable marker, glycosylated hemoglobin (HbA
1c
), with
microvascular and macrovascular complications have been clearly demonstrated
(Anonymous 1993; Stratton et al. 2000).
About two decades ago, type 2 diabetes was thought to be a mild and benign
disorder and patients remained largely untreated. As a result, the traditional way of
treating type 2 diabetes was stepwise non-medication (e.g. diet) approaches followed by
oral medications (Nathan 2002).
The incidence and severity of diabetic complications can be reduced through good
glycemic control. Control of blood glycemic in both type 1 and type 2 diabetes can delay
12
or prevent the occurrence of microvascular diseases; however, the benefit in delaying or
preventing macrovascular disease is not clear with inconsistent evidence.
The Diabetes Control and Complications Trial (DCCT) has proved that intensive
glycemic control can decrease the HbA
1c
level and reduced the risk of onset and
progression of retinopathy, nephropathy, and neuropathy in patients with type 1 diabetes
(Anonymous 1993). Long-term and continuous intensive glycemic control seems to be
important because the differences in progression rates between the tight control and
standard control groups do not appear until two and a half years after the initiation.
For type 2 diabetes, two prospective studies have shown the benefit of glycemic
control in reducing diabetic complications. The Kumamoto study confirmed that tight
glycemic control can delay the onset and progression of microvascular complications,
with the risk reduction similar to the DCCT results (Ohkubo et al. 1995).
UK Prospective Diabetes Study (UKPDS) was the largest and longest clinical trial
ever conducted among diabetes patients (UKPDS Group 1998). The primary goal of the
UKPDS was to determine whether improved glucose control in type 2 diabetes would
prevent clinical complications, and whether therapy with insulin, sulphonylurea, and
metformin has specific advantages or disadvantages. Similar to the DCCT and
Kumamoto results, the UKPDS showed that intensive therapy reduced the risk of
microvascular complications (UKPDS Group 1998; UKPDS Group 1998). Moreover,
the UKPDS also showed that as long as tight glycemic control was achieved, there was
no difference in the effects of the various treatment agents on the risk of microvascular
complications (UKPDS Group 1998). However, the effect of glycemic control on
13
macrovascular disease was less certain, except for the metformin group (UKPDS Group
1998).
The benefit of tight glycemic control in reducing diabetic complications appeared
to be mediated by the improvement of HbA
1c
. An observational analysis of the UKPDS
trial data found every 1% reduction of HbA
1c
was associated with 21% reduction for any
endpoint related to diabetes and death, and 37% for microvascular complications. The
risk reduction showed a gradual and log-linear relationship with no threshold of HbA
1c
for risk reduction observed for any end point (Stratton et al. 2000). The rates were
similar to the DCCT trial (Anonymous 1993). The findings of both UKPDS and DCCT
trials suggested no threshold HbA
1c
value existed for the positive benefits on
complications of diabetes, which implied that the level that the optimal control of blood
glucose could achieve should be as normal as possible (Liebl 2002). This implication
was later confirmed in a study with a nondiabetic normal population, which showed even
within the normal range of HbA
1c
, the lower HbA
1c
attained, the less risk of
cardiovascular and all cause mortality (Khaw et al. 2001). However, tight glucose
control did not achieve lower HbA
1c
without risk; both studies found that tight control
and lower HbA
1c
were associated with higher risk of hypoglycemia (Anonymous 1993;
UKPDS Group 1998).
2.1.4. Pharmacological therapy and challenges
Medication treatment for diabetes includes insulin and oral hypoglycemics. The
classes of hypoglycemics are listed below (Table 1).
14
Table 1. Classification of hypoglycemics
Mode of Action Therapeutic Class HIC3 Example
Insulin supply
Insulin (INS) C4G Insulin aspart, Human
recom. Insulin
Sulfonylureas (SUL) C4K Glyburide, Glipizide,
Glimepiride
Augment insulin supply
Nateglinide analogs
(NAT)
C4K Repaglinide,
Nateglinide
Biguanide (MET) C4L Metformin
Reduce insulin resistance Thiazolidinediones
(TZD)
C4N Rosiglitazone,
Pioglitazone
Reduce intestinal breakdown
of complex carbohydrates
Alpha-glucosidase
inhibitors (AGI)
C4M Acarbose, Miglitol
Adapted from Riddle 2000 and Van Gaal 2003
The treatment goal for diabetes is to achieve near normal glycemic control to
delay or prevent the development of diabetic complications (American Diabetes
Association 2002). According to the recommendation of American Diabetes Association,
the target for glycemic control are blood glucose levels (80-120 mg/dL preprandial, 100-
140 mg/dL bedtime) (American Diabetes Association 2002) and HbA
1c
level of less than
7%. Some more stringent guidelines recommend the target to be less than 6.5%
(American Association of Clinical Endocrinologists 2002), as that in a healthy person.
Insulin therapy is traditionally withheld as the last therapeutic approach in type 2
diabetes because of the need for administration through injection, which is presumably
associated with low compliance due to inconvenience. By the time when patients with
type 2 diabetes are treated with insulin, they usually have had diabetes for more than 10
to 15 years and have established complications (Nathan 2002), and the mean HbA1c for
type 2 diabetes patients to start insulin therapy in the US is 10.4% (Hayward et al. 1997).
However, evidence supports insulin as the most potent and durable hypoglycemic
15
intervention available (DeWitt and Hirsch 2003).
Despite initial effectiveness in reducing hyperglycemia, monotherapy of oral
hypoglycemics appears to be insufficient over the long term to reach the targeted goals
for glycemic control. For example, in UKPDS study, after an initial reduction during the
first year, HbA
1c
levels increased 0.2% to 0.3% per year on average, regardless of the
treatment patients received. By the end of fifth year, HbA
1c
deteriorated back to the
baseline level even in the tight control group (UKPDS Group 1998). As the secondary
treatment failure became evident, many patients switched to combination therapy of oral
agents or to insulin.
For type 2 diabetes, except alpha-glucosidase inhibitors (AGIs) and nateglinide,
different classes of hypoglycemic appear to be equally effective in lowering blood
glucose concentrations and HbA
1c
level (Inzucchi 2002). The effectiveness of
combination therapy generally appear to be the additive efficacy of individual
components (Inzucchi 2002). Although all classes of hypoglycemics can be used as
monotherapy in type 2 diabetes, not all possible combination regimens are approved by
FDA. The combination therapies approved by FDA are shown as follows (Table 2).
Table 2. FDA approved status for hypoglycemics combination regimens
INS SUL NAT MET AGI TZD
INS -- X X X
SUL X -- X X X
NAT -- X
MET X X X -- X
AGI X --
TZD X X X --
Adapted from Inzucchi 2002
16
A recent review of randomized clinical trials of insulin implied that combination
therapy of oral agent and insulin might be best regimen for patients with type 2 DM,
measured by glycemic control (HbA
1c
level). Among different combination regimens,
metformin with insulin appears to be the best choice for patients without
contraindications (DeWitt and Hirsch 2003). The combination of metformin and insulin
achieves the similar level of metabolic control, but with less weight gain, lower insulin
dose, and fewer hypoglycemic episodes than insulin alone, or insulin combined with
sulphonylurea (SU). Insulin/metformin combination therapy also outperforms insulin/
thiazolidinediones (TZDs) therapy in terms of safety, weight gain, and hypoglycemic
episodes (DeWitt and Hirsch 2003).
Besides direct glycemic control with medications, treatment of hypertension,
hyperlipidemia, and lifestyle modifications also benefit preventing and slowing diabetic
complication progress.
Half of people with type 2 diabetes have hypertension. As blood pressure
increases, the risk of microvascular and macrovascular diseases also increases
(Chobanian et al. 2003). Controlling blood pressure, especially with the treatment of
ACE inhibitors, can greatly help diabetic patients to reduce the risk of and slow the
progress of diabetic retinopathy, nephropathy, and certain cardiovascular diseases (Lewis
et al. 1993; UKPDS Group 1999; Podar and Tuomilehto 2002). The goal of blood
pressure control for diabetic patients is 130/80 mmHg (American Diabetes Association
2002).
Aggressive control of lipid also benefits diabetic patients in lowering risk of
macrovascular complications. Results of large clinical trials have indicted
17
that lowering low-density lipoprotein cholesterol (LDL-C) level, reducing triglyceride
levels and increasing high-density lipoprotein level (HDL) in diabetic subgroups can
decrease the risk of cardiovascular diseases at least as same as or even more than in non-
diabetic groups (UKPDS Group 1994; Sacks et al. 1996; Pyorala et al. 1997; Goldberg et
al. 1998; Rubins et al. 1999). However, the benefit of lipid control for microvascular
diseases has not been clearly evidenced.
It is highly recommended that multiple factors (glucose, lipid, blood pressure)
should be targeted at the very time when diabetes is diagnosed. A recent multifactorial
trial demonstrated that among all therapeutic strategies investigated (Gaede et al. 2003;
Pedersen and Gaede 2003), intensive and simultaneous control on multiple modifiable
factors (e.g. glucose, blood pressure, and lipid levels) demonstrated the greatest benefits
in those at high risk of type 2 diabetes.
2.2 Compliance research
2.2.1. Overview of compliance research
Myriads of effort have been made to study compliance of medical treatment since
1975 when “patient compliance” appeared as a medical subheading (MeSH) by National
Library of Medicine. A total of 24,845 articles in Medline have been listed under “patient
compliance” by April 2004, among which 2,742 were review articles. Literature
reviewers were often frustrated at the fact that little agreement had been reached in
compliance literature (Morris and Schulz 1992; Balkrishnan 1998; Brawley and Culos-
Reed 2000). As some observed, “years of research on compliance provide little
consistent information other than the fact that people do not always follow the doctor’s
18
orders” (Morris and Schulz 1992). Lack of methodological rigor in compliance literature
was an important reason for such inconsistence (Nichol et al. 1999).
The ultimate goal of understanding compliance behavior is to improve health
outcomes by identifying potential modifiable factors and implementing compliance-
promoting interventions to targeted populations. In order to achieve this goal, past
compliance literature of original and review studies have primarily focused on the
following four major areas: (1) assessment: measuring and determining the extent of
noncompliance, (2) prediction: investigating and identifying determinants, predictors, and
causes of noncompliance, (3) consequence: examining and evaluating relationship
between compliance or noncompliance and outcomes, and (4) improvement: empirically
evaluating and studying compliance-promoting interventions.
This study, with its goal of estimating the benefit of compliance, belongs to the
study area 3. However, different from all past empirical research that investigated the
compliance outcomes association, this study estimates the casual effect of compliance on
health outcomes.
2.2.2. The importance of understanding compliance-outcomes relationship
Among the four major areas in compliance literature, the relationship between
compliance and outcomes was the least investigated area. DiMatteo et al. conducted a
meta-analysis on the relation between adherence and outcomes in non-mental health
studies. In this meta-analysis, a total number of 63 articles were identified, where only 44
articles were pertaining to medication compliance (DiMatteo et al. 2002). Such under-
investigation seems to be partially because the benefit of compliance on outcomes was
often taken for granted. As DiMatteo et al. noticed, “… because, it is argued,
19
adherence to treatment improves outcomes” (DiMatteo et al. 2002). Yet, precisely
estimating the extent of effect of compliance enhancement on health outcomes is
significant for the following reasons.
First, whether higher compliance could improve health outcomes is not clear. All
past studies on compliance-outcomes relationship only identified the correlation of
compliance of outcomes (DiMatteo et al. 2002), and it is not surprising that majority of
these studies published showed positive correlation (Easterbrook et al. 1991; Dickerson et
al. 1992), and there is no statistics on the number of unpublished studies finding negative
association between compliance and outcomes. It is not known whether patients who
have or are prone to have better outcomes are more likely to comply or compliance
improves outcomes. The gold standard of causal inference is to conduct randomized
clinical trials, yet levels of compliance cannot be randomized in that compliance is a
subjective choice. The lack of application of causal modeling remains the compliance
benefit hypothetical. It is arguable whether higher compliance actually improves health
outcomes. An increasing number of articles have argued that noncompliance may be the
rational choice of patients. A school of thought views that non-compliance decision is a
patient’s rational choice involving many aspects where medication may impact, such as
personal identity, goals, and quality of life (Trostle et al. 1983; Conrad 1985; Lynn and
DeGrazia 1991; Lambert et al. 1997). Observed compliance level given each specific
individual might be his or her optimal choice, even from a health outcomes perspective.
When reasons of discontinuation are investigated, patients stop medications mostly for
legitimate reasons such as seeking alternative care, avoiding adverse effects, therapeutic
ineffectiveness, and achieving therapeutic goals (Andrade et al. 1995).
20
Acknowledging the cases where noncompliance makes sense and compliance does not,
some termed those cases as “intelligent non-compliance” or “capricious compliance”
(Weintraub 1976). Given so, do we still expect higher compliance always improves
health outcomes? Without evidence of the benefit of medication compliance, some even
called “medication compliance as an ideology” (Trostle 1988).
Despite associational design, lack of methodological rigor (Nichol et al. 1999),
and publication bias, existing literatures have not unanimously supported the positive
association between higher compliance and better health outcomes. In a review of the
effect of medication noncompliance on coronary heart disease mortality and morbidity,
out of 15 studies including 3 randomized control trials, 9 prospective and 3 retrospective
studies, only 8 showed the significant association between compliance and outcomes
(McDermott et al. 1997). Similar findings were reported in DiMatteo et al. (2002), where
a study discovered negative associations (DiMatteo et al. 2002). These revealed conflicts
may be a tip of iceberg due to the publication bias. If higher compliance always
improved health outcomes, we would expect to see that outcomes are improved when
compliance-promoting interventions significantly improve compliance. However, a
comprehensive review of adherence-improving interventions found that improved
adherence does not necessarily result in improved clinical outcomes (McDonald et al.
2002). In McDonald’s review, studies with both improved adherence and outcomes
almost always have complex interventions including combinations of convenient care,
counseling, etc. Thus, it is unidentifiable whether the improved health outcomes were
caused by improved adherence or the interventions per se. Therefore, the assumption that
21
compliance is beneficial to health outcomes remains unproven. Identifying the
compliance effect on outcomes is thus warranted.
Second, knowing the extent of the effect of compliance on health outcomes is
necessary for cost-effectiveness analysis of drug therapy in a real-world setting.
Compliance in real-world conditions marks the distinction between efficacy and
effectiveness of medication. The compliance rates seem to be lower in real-world
practice than that in the clinical trial environment (Andrade et al. 1995), partially because
of non-compliers being excluded in run-in phase of a trial (Boudes 1998), efforts to
maintain protocol adherence, simplified regimen, and access to study medications at no
cost. When assessing the “effectiveness” of drug therapy, effort should be taken to
incorporate the effect of compliance with efficacy data observed in clinical trials. In a
comprehensive review of use of compliance in cost-effectiveness analysis, Hughes et al.
(2001) found that only 22 out of approximately 3000 studies on economic evaluation of
medical treatment have considered the impact of non-compliance. He noticed that the
information of the effect of compliance on outcomes was absent in most studies. For
most evaluators who considered compliance effects on health outcomes and costs, such
effects were often derived based on assumptions or medical opinions that were “more
prone to errors.” (Hughes et al. 2001). As shown in Figure 1, different studies had taken
a great variety of assumptions to assess the percentage changes in outcomes (Hughes et
al. 2001). In addition, for some drugs with outcomes that were relatively less sensitive to
compliance, “non-compliance (in terms of premature discontinuation) might, in fact, be
cost-saving” (Brown et al. 1999; Hughes et al. 2001). The accuracy of such claims and
treatment decisions would crucially depend on the knowledge of causal effect
22
of compliance on outcomes. Due to the scarcity of methodologically-rigorous studies on
assessing the effect of noncompliance on outcomes, validity of cost-effectiveness
analyses is undermined when the therapy efficacy is extended to make real-world
decisions.
Figure 1. A plot of the percentage change in outcome in relation to the compliance rate
used in 22 cost-effectiveness evaluations of drug therapies. Source: Hughes et al. (2001)
Third, understanding the effect of medication compliance on health outcomes is
important for evaluating the potential benefit of compliance-improving interventions. In
empirical studies, strategies intended to enhance compliance were found to be
surprisingly weak in their effects, and even the most effective interventions had only
modest effect on compliance (Roter et al. 1998; McDonald et al. 2002).
23
Understanding the relationship between compliance and outcomes, and factors that
moderate the relationship between compliance and outcomes is important to guide future
endeavor in designing and assessing an effective compliance-promoting intervention
(DiMatteo et al. 2002).
Fourth, accurate estimation of the compliance effect is also important for
precisely estimating the number of patients to be treated (power calculation) in a clinical
trial. Noncompliance in clinical trials severely comprises treatment effectiveness. A
smaller number of patients would be otherwise sufficient to demonstrate the same
therapeutic effect, had patients been more compliant (Ellis et al. 2000). It was estimated
that average 50% compliance in a trial would need to increase sample size fivefold in
order to maintain the same power compared to 100% compliance (Goldsmith 1979).
In order to calculate the number of patients to be treated given a certain statistical
power, estimated effectiveness as a key parameter depends on the level of compliance
and the effect of compliance on drug effectiveness (Kastrissios and Blaschke 1997). In
addition, assessing the clinical benefits of increased compliance rates is limited by ethical
concerns of compliance-enhancing interventions (Hess 1996).
These reasons pronounce the importance of precisely estimating the effect of
compliance on health outcomes, which is the central goal of this study.
2.2.3. Definitions of compliance
It is a long and probably endless process for debating and arguing on the
definition of compliance (Steiner and Earnest 2000). As many authors have commented,
there is no agreement on a commonly accepted definition (Fawcett 1995; Miller 1997;
24
Evangelista 1999; Lutfey and Wishner 1999; Anderson and Funnell 2000; Kyngas 2000).
Alternative terminologies have been used to describe compliance-related scenarios,
including adherence, consistency, concordance, continuity, co-operation, mutuality, and
therapeutic alliance. But they were inconsistently and inappropriately defined and were
often used as synonyms (Kyngas et al. 2000; Gray et al. 2002).
The term “compliance” has been pervasive in medical literature. A frequently
quoted definition of compliance is “the extent to which a person’s behavior (in terms of
taking medications, following diets, or executing lifestyle changes) coincides with
medical or health advice” (Haynes et al. 1979). This definition was traditionally based on
the general assumption that patients should obey physician’s instructions on medication
taking, keep appointments, and follow clinician’s advice on behavioral modification such
as monitoring blood glucose, diet and lifestyle change. However, this definition has been
criticized frequently as it denies patients’ autonomy and their role in decision-making
regarding their care (Lutfey and Wishner 1999; Kyngas et al. 2000). The recent
conceptual development in compliance studies emphasized the patient’s perspective in
medical decision making and showed a paradigm shift from traditional health
professional prospective (Morris and Schulz 1992; Lutfey and Wishner 1999). From the
patient’s prospective, noncompliance may not be seen as a deviant behavior, but as a
reasoned decision making process. Incorporating the quality-of-life considerations,
noncompliance may be viewed as a rational choice of patients to assert the control over
their own disorder (Conrad 1985; Morris and Schulz 1993; Donovan 1995; Williams
1998; Lutfey and Wishner 1999). Evidence also supported a more demanding role and
responsibility of physicians in patient’s noncompliance. Based on this
25
argument, many definitions of compliance or adherence have been proposed. Hussey and
Gilliland defined compliance as “the positive behavior that patients exhibit when moving
toward mutually defined therapeutic goals” (Hussey and Gilliland 1989). A simplest
definition of noncompliance is “two people working toward different goals” (Funnell and
Anderson 2000).
Recently “adherence” has been proposed to substitute “compliance” because
“adherence” has broader implications on patient autonomy and roles of physicians, and
captures increasing complexity of health care behavior considering patients are also
decision makers (Lutfey and Wishner 1999). Proponents of using “adherence” instead of
“compliance” believe this change reflects paradigmatic shift from a medical-centered
perspective towards a more patient and social-oriented prospective (Lutfey and Wishner
1999). As a consequence, more and more literature in recent years used “adherence” in
favor of “compliance”.
Another relevant but distinct term used to describe medication-taking behavior is
“persistency”. Persistency addresses how long patients remain on a therapy (Dezii 2001).
Persistency measures the length of the period from the initiation of a medication to the
discontinuation of the medication, during which patients refill medications within a
“grace period” (Dailey et al. 2001). Technically, the grace period can be set at 50% of
days supply of the last prescription refill (Dailey et al. 2001), or a fixed number of days,
for example, 30 days (Harley et al. 2002; Larsen et al. 2002). Switching to a different
drug can also be considered as the end of persistency if persistency is intended to
measure the loyalty of patients to a given medication. Non-persistency is a type of
noncompliance in a broader sense, and it is deemed as a component of overall
26
adherence (Dezii 2001). In some cases, “persistency” is also used as a synonym of
“compliance” or “adherence” (Benner et al. 2002).
2.2.4. Measurement of compliance
Many reviewers of compliance literature have addressed the concern that
measurement of compliance has hurdled the advance of compliance research due to the
lack of valid, reliable and practical measures (McNabb 1997). Traditional measures of
compliance include administrative refill records, pill counts, drug concentration urine or
blood, questionnaires, etc. Medication event monitoring system (MEMS) sometimes is
deemed as the gold standard of measuring compliance because of its accuracy (Claxton et
al. 2001); however, the decision on how to use MEMS records to derive a compliance
measure is still subjective, and evidence showed that compliance measured by MEMS
and other traditional measures were weakly correlated (Paes et al. 1998; Vitolins et al.
2000). The expense and effort associated with using MEMS has confined its usage in
clinical trials (Farmer 1999; Vitolins et al. 2000) and limited its application in a general
compliance study.
Administrative claims databases provide a rich source of information to measure
compliance with the advantage that it is objectively measured without patient’s
awareness, and is free of bias from external influence or recall. In addition,
administrative refill records can be a relatively economic source to study the compliance
behavior together with the related factors in a large population. Because the
administrative claims database is the data source used by this study, the measurement
method of compliance associated with administrative databases will be briefly described
in this section.
27
A variety of methods of measuring medication refill compliance with such
databases have been summarized in Steiner and Prochazka (1997) (Steiner and Prochazka
1997). This article created a unique set of terminology to describe different methods
used to define compliance prior to the publication of their study: 1) continuous (C) vs.
dichotomous (D) variable; 2) single (S) or multiple (M) intervals of evaluated refill; and
3) medications during a period were available (A) or gaps between intervals (G). Three
letter combinations were used to describe a specific measurement method. Few
researchers of compliance adopted such unique set of terminology, probably because of
its unconventional approach and language. Therefore, we do not detail each combination
discussed in that article.
The combination set CMA (multiple-interval measures of medication availability)
is often applied in literature, though rarely called so. CMA is a continuous measure of
overall compliance over a period of time incorporating all refill records during such a
period. The CMA is more often called medication possession ratio (MPR) in the
empirical literature and occasionally termed as proportion days covered (PDC) (Benner et
al. 2002; Benner et al. 2004), and it’s calculated as “the sum of the days supply” obtained
over a series of intervals, divided by the total days from the beginning to the end of the
time period” (Steiner and Prochazka 1997). Because it is possible that the total days
supply can be more than the number of calendar days in the period, thus the MPR could
be greater than 1. Therefore, some versions of MPR either do not double count the
overlapped days (when some period was covered by two intervals of days supply), or
censor MPR at 1. Therefore, these versions of MPR take the value between 0 and 1.
28
2.3 Compliance-outcomes relationship
2.3.1. Studies of compliance-outcomes association in diabetes
Several articles have previously made efforts to link the compliance level of
hypoglycemic medications to outcomes of interest, with intention to show the
unfavorable outcomes due to noncompliance.
Using medication refill claims and medical records, Schectman et al. investigated
the association between medication possession ratio (MPR) of oral hypoglycemics and
HbA
1c
level as well as HbA
1c
level change using ordinary least square (OLS) regression.
Adherence was assessed during a 14-month period, and HbA
1c
levels were obtained
during a 20-month period. Attempt was made to use the most recent HbA
1c
records of
each patient; however, the outcomes measure was still overlapped with the adherence
measure, so the temporal order of adherence and outcomes was ambiguous. The results
showed an association between better glycemic control by lowering HbA
1c
0.16% and
every 10% increased in drug adherence (Schectman et al. 2002).
Using data from a longitudinal survey with 908 diabetic patients that collect four
waves of repeated measures, Kuo et al. (2003) investigated the cross-sectional
association between medication “consistency” (a dichotomous indicator) and diabetic
complications (kidney, eye, circulation problems) with logistic regressions. Poor
consistency was a dichotomous measure including discontinuity, no medication, or self-
reported inconsistency at any time during the four waves of survey. Similarly, the
dichotomous outcome variables of having complications were also self-reported at any
time during the four waves. Although the rich temporal relationship between consistency
29
and outcomes over four waves of measures was ignored, this cross-sectional logistic
regression study found a significant “relationship” between poor consistency and
increased risk of kidney problems “over a period of 7 years”, but not for eye or
circulation problems (Kuo et al. 2003). The relationship implied its causal direction to
most readers. However, though probably in a small extent, can it be possible that some
patients decreased hypoglycemics use as a result of higher demand of other medications
with the worsening kidney symptoms?
A cross-sectional survey study (Guillausseau 2003) (N=11,896) used OLS
regression to examine the association between self-reported compliance (categorical)
level with HbA
1c
lab values, among other outcomes, found 1.4% mean difference of
HbA
1c
between the best and worst compliance level.
Balkrishnan et al. (Balkrishnan et al. 2003) examined the relationship between
hypoglycemic medication adherence and total health costs in older adults with type 2
diabetes. This study used repeated measure of annual adherence up to five years, and
measured the total health care costs within each year. Using a random-effects GLS
regression, this study identified MPRs as the strongest predictors of decreased total health
care costs. This study applied longitudinal study design and used random-effects GLS
regression to adjust intra-subject variability over time. However, because of the
simultaneous measurement of compliance and health utilization, identification problem
still remained. This study, with a merit, recognized the limitation that “the observational
study design did not permit causal inference of our results”.
With a hypothesis that insulin initiation may result from poor compliance of oral
antidiabetics in type 2 diabetes, Evans et al. (Evans et al. 2002) examined the
30
difference of adherence rates between switchers and non-switchers. Using logistic
regression with switching to insulin as an outcomes measure, the study did not find any
significant association between compliance level and switching to insulin. The
compliance rate used in this study was measured by MPR.
All these five observational studies unanimously acknowledged the relationship
between compliance and better outcomes were mere “association”, with one of them
reporting no association. Nevertheless, as a common phenomenon in observational
studies, these papers also directly or indirectly implied the causal interpretation of the
estimated effect. For example, one study interpreted that “adherence to antidiabetics
medications was a greater driver of cost reduction than…”, and some other study
recommended adherence promotion intervention to improve diabetes related health
outcomes based on study findings. Some even extended such statement in conclusions
where the associational relationship could not support (Guillausseau 2003). Extending
“association” estimate to causal interpretation is especially common when the studies are
referenced by others.
2.3.2. Overview of methodological problems
It is discomposing to face the gap and conflict between the reality that studies
only provide a mere “association” on one side, and the ultimate intent for a causal effect
that can be translated directly to policy implication on the other. Such gap is often
blurred by the common desire of the readers and authors, and such causal effect of the
intervention and policy implication based on the “association” is often implied, which we
sometimes take for granted.
31
Unfortunately, the signs of association between exposure and outcomes
sometimes are contrary to our common belief about the direction of causal effect, though
such associations are truthful and cannot be denied. For example, as many empirical
outcomes researchers many have experienced, the association between compliance and
outcomes benefit are negative, i.e., higher compliance is associated with worse outcomes.
Such association is authentic also not unreasonable—sicker patients use more
medications and are more likely to experience worse outcomes. These associations
cannot be interpreted casually or provide policy implications, but do those studies with
findings of positive associations?
For obvious reasons, findings with associations close to what we would expect
about the direction of causal effect are more likely to be published. As reviewed
previously, except one without significant association, all papers reported significant
association between higher compliance and better outcomes. If compliance-outcomes
research literature is a collection of findings of associations coherent to the direction of
causal relationship, do these studies add any more knowledge beyond our existing
expectation?
These critiques as well as methodological problems to be summarized below do
not intend to discredit previous efforts on the exploration of compliance-outcomes
relationship, but only advocate causal estimation methods that have evolved to be mature
tools for outcomes researchers to probe compliance-outcomes relationship beyond
association.
Several major threats to internal validity existing in the published studies
regarding effect of compliance on outcomes are summarized. None of them is
32
a problem for “association” relationship, however, they can be critical flaws when we
extend these associations to causal interpretation, as often intended.
Here is a prevailing scenario of estimating the compliance effect on outcomes is
described as follows, also as shown in some previously reviewed article. In an
observational study with medical and/or pharmacy records or a retrospective analysis of
clinical trial data, medication compliance is measured over a period of time of either
fixed length (e.g. one year) or varied length (e.g. from the first to the last prescription).
Health outcomes are then measured, often towards the end of, yet within, the same study
period. They can be one-year total costs (continuous), a recent clinical lab measures
(continuous), occurrence of any hospitalization or adverse event (dichotomous), time to
the first adverse event (survival), or number of visits (counts). A starndard regression
model is usually applied to investigate the “association” between compliance and
outcomes, with a hope for desirably positive coefficient in the model.
The methodological problems of this traditional approach are summarized as
follows.
2.3.3. Mere correlation between compliance and outcomes
Association (i.e. partial correlation) cannot give information on casual inference.
The ultimate answer we would seek from compliance-outcomes relationship is how much
benefit we can gain from a change in compliance level, which is termed as the treatment
effect or the causal effect, but not how compliance and outcomes correlate. In a cross-
sectional design, such problem is inevitable without the aid of a causal method such as
instrumental variable.
33
DiMatteo et al. have comprehensively reviewed published studies on the
relationship between compliance and outcomes, and made the statement as “the studies
…are all correlational” (DiMatteo et al. 2002). The effect of compliance on health
outcomes is usually established through a regression coefficient. The critique to such
method is canonical in Thomas D. Cook and Donald T. Campbell’s Quasi-
Experimentation (p. 298):
“A treatment effect is inferred if there is a statistically significant regression
coefficient (beta weight) relating to the dummy variable to the dependent variable,
after adjusting for the effect of the covariates introduced to try and correct for
selection. Unfortunately, those who use this widespread practice make the error of
employing forecasting techniques for causal inference” (Cook and Campbell
1979).
Partial correlation is sensitive to the choice of covariates controlled and is also
affected by the choice of regression functional form, so such correlation can not
contribute to the knowledge of underlying causal effect (Cook and Campbell 1979).
Empirical evidence have showed that alternate analysis of the same data can alter not
only the level of significance but also the sign of the treatment effects (Cook and
Campbell 1979). As Manski pointed out in his book Identification Problems in the
Social Sciences, “the conclusions that can be drawn from any analysis are determined by
the assumptions made and by the data brought to bear” (Manski 1995).
2.3.4. Obscured temporal relations of compliance and outcomes
The temporal order that the cause precedes the result chronologically is a self-
evident requirement for establishing causal effect (Cook and Campbell 1979; Holland
1986). This basic requirement seems not to be a barrier in most prospective studies
34
where treatments (interventions) are relatively short, and pretest and posttest measures
can be controlled in a proper time sequence by researchers.
However, in the empirical research scenario described previously, the temporal
relations between compliance and outcomes are often ambiguous. Causal inferences
strictly require measurement of cause variable to precede the measurement of effect
(outcomes) (Cook and Campbell 1979). When compliance and outcomes are measured
over the same (or similar) period of time, the effect of the compliance on outcomes and
the effect of the outcomes on compliance cannot be distinguished due to their mutual
interaction. With an econometric term, a “simultaneity” issue arises.
Unfortunately, the temporal relations are not modeled correctly in most
compliance-outcomes studies with retrospective or cross-sectional design, where
researchers are unable to identify the exact time when the outcomes are measured.
Outcomes measures can be taken in many forms, such as available lab values,
total health costs, occurrence of disease exacerbation (e.g. seizure relapse, myocardial
infarction, surgery, emergency visits, or hospitalizations), or time to some exacerbation
event. These types of outcomes cannot be measured at a fixed time point, but traditional
ways of measurement tend to violate the basic assumption of temporal precedence for
establishing causal effect.
Measuring outcomes and ongoing treatment over the same period time inevitably
confronts the simultaneity problem. Simultaneity, according to Manski, is so-called “the
identification problem” in economics and social sciences (Manski 1995). Such
measurement encounters the problem that the reciprocal effects between two measures
are undistinguishable. As noticed by DiMatteo et al., “although we assume
35
that compliance influences treatment outcomes, we acknowledge that outcomes may
influences compliance, particularly during a long course of treatment” (DiMatteo et al.
2002). In this case, we are not sure whether worsened outcomes affect compliance or
compliance affects the outcomes, especially, when the compliance is measured as a single
value over time. For example, when hospitalization was positively correlated with
noncompliance, we are not sure whether noncompliance caused worse health status and
the consequent hospitalization, or patients stopped taking medication after being
hospitalized.
2.3.5. Compliance as a static or dynamic measurement
Though often measured as a single index, compliance is not a static process, but a
process that often changes over time. Compliance often interacts with other time-varying
variables, i.e. compliance is not an independent event.
Most, if not all, studies employed a single measure of compliance, which
essentially assume that compliance is a static process over the study period. Even with
pharmacy claims data or electronic records that are longitudinal in nature, researchers
often summarize compliance level over a period of time as a single value. This is simple
and convenient, but the valuable information that could have been uncovered in a
dynamic changing process is missed. Some studies claimed to be “longitudinal cohort”
studies because of using data that span over years; however, the very term “longitudinal”
is reserved to describe repeated variable measures in epidemiology and biostatistics.
To illustrate that the same surrogate value of compliance level may possibly refer
to dramatically different behaviors, Gordis (1979) used a self-explaining diagram
reproduced in (Figure 2) (Haynes et al. 1979).
36
Patient 1 - - - - - + + + + +
Patient 4 + + - - - + + - + -
Patient 2 + + + + + - - - - -
Patient 3 + - + - + - + - + -
Figure 2. Hypothetical results of sequential tests in four "50% compliers". Source: Leon
Gordis (1979).
Compliance is not static. Studies using electronic monitoring can detect dynamic
changes of compliance (Kruse and Weber 1990). One example of showing such a
dynamic pattern was shown in (Figure 3), which was reproduced from (Dunbar-Jacob
and Mortimer-Stephens 2001). With detailed pharmacy refill records, changes in
compliance over time can also be observed when compliance is measured repeatedly
during different periods of time (Benner et al. 2002). The dynamic change of compliance
over time is not a random process. It can be correlated with other observable behaviors.
For example, it has been found that patients who used electronic monitors were more
likely to be compliant before and after a doctor visit (Cramer et al. 1990; Feinstein 1990).
In another study, patients using electronic monitors had higher compliance on blood
pressure monitoring than otherwise (Mengden et al. 1993).
37
Figure 3. A closer look at a real-time medication compliance pattern using electronic
monitoring device.
In fact, compliance is a time-varying factor that is influenced by and also
influences health outcomes. Moreover, this time-varying process is not an independent
process from health outcomes, but instead, is mediated through intermediate factors
related to outcomes. For example, patients may change the compliance level (though
altering dosage or frequency) over time because of feeling better or worse, or the change
of clinical parameters (e.g. blood sugar) (Haynes et al. 1979). This is an important issue.
Perception of disease status is one of the factors determining compliance in health belief
model (Becker and Maiman 1975), and past disease severity related variables and health
indicators are often predictors of compliance (Steiner and Prochazka 1997; Balkrishnan
1998). Failure to control such a time-varying confounding process may severely bias the
estimate of an adaptive treatment. With this respect, marginal structural models have
been developed to estimate the causal effect under weaker assumptions (Robins 1994;
Robins et al. 1999).
38
2.4 A conceptual causal model of compliance
2.4.1. Past conceptual models
Numerous factors modify the compliance-outcomes relationship, and this
phenomenon seems to be extraordinarily complicated. Some models developed to
explain general health behavior from a patient psychological perspective were also
applied to explain the medication taking behavior (compliance). Some well known
models include health belief model (Becker 1974; Rosenstock 1985; Oldridge and
Streiner 1990), theory of reasoned action (Dishman 1994; Syrjala et al. 2002), and locus
of control (Lewis et al. 1978; Wenerowicz et al. 1978), etc.
Efforts have been made to test the validity and assess usefulness of these models,
however, empirical research found utility of these models to be very limited
(Montgomery et al. 1989). Owing to the inability of conceptual model to predict
compliance behavior in real practice, some authors have cast doubt on the necessity of
such research endeavor. As some observed, “no convincing model for predicting
medication adherence has been developed. The poor and inconsistent correlation between
compliance and health beliefs suggest that the research focus should be shifted … to their
identification and the development of ways to improve compliance” (Cramer 1995).
Although these conceptual models could explain a few limited aspects of patient
behavior pertaining to medication compliance, none was developed from a perspective of
causal modeling to guide research exploring the relationship between compliance and its
outcomes.
2.4.2. One-period version and causal pathways
39
To provide a unified framework to guide compliance research with a focus on
exploring the causal relationship between compliance and outcomes and other related
factors in a comprehensive manner, the following conceptual model is contructed based
on compliance literature, causal path analysis, multiple-indictors-and-multiple-causes
(MIMIC) (Jorekog and Goldberger 1975) framework and directed acyclic graphs (DAG)
(Pearl 1995; Greenland et al. 1999). This model, tailored to the context of diabetes, is
able to clarify the intricate relationship among a large array of compliance-related
information in a concise manner, and to provide insight into the casual relationship from
compliance to outcomes. The model has a one-period version and its extension to a
multiple-period dynamic version.
The one-period version of conceptual model is presented in the form of causal
diagram shown in Figure 4. The core of this one-period version is based on the multiple-
indictors-and-multiple-causes (MIMIC) structure, which was first developed to estimate a
single latent variable not otherwise measurable (Jorekog and Goldberger 1975). The core
structure is extended with upstream causes and predictors, and downstream
consequences, to show their interactions, confounding effect, and temporal relations
within a single period.
40
Generalized Concept of Compliance
Compliance
SMBG
Medication
Refill
Keeping
Appointment
Diet
Foot care, eye
exam
Exercise
Outcomes
Multiple Indicators
(current)
Latent Construct
Individual
Level
Interpersonal
Level
Community
Level
Family and Social
Support (referent
power)
Patient-Physician
Relations (expert
power)
Insurance (copay)
Cultural norm
Social Standard
Policy (resource,
care access)
Self-efficacy
Disease Status
Beliefs
Satisfaction
Unidentified
latent constructs
Locus of control
Multiple Causes
(current or recent)
Observable Factors
(past or current)
Demographics (age,
gender, race, SEC)
Other
Comorbidities
Clinical Parameters
Personality
(forgetfulness,
responsiveness)
Health resource
utiliz ation
Physician
Variables
Drug, frequency,
pill burden, etc.
Medication
(treatment)
Level
Ease of taking,
Brand loyalty,
Side effects
Depression,
Psychological
disorders.
++
+
+
+/-
+
+
+/-
Factor level
Outcomes
(current or future)
Time
Target of compliance-
promoting intervention
Causal relation
Unidentified and weak
causal paths
strength of prediction
+/- inconsistent or
no predictive power
+ weakly predictive
++ strongly predictive
Unclear causal paths
Figure 4. A causal diagram for one-period conceptual model of compliance
This model (of both one-period and dynamic versions) is a directed acyclic graph
(DAG) because the directed arrows represent direct causal relations between variables.
There are no directed cycles, as no variable can cause itself. DAG is one of the three
equivalent representations of casual modeling, with the other two being contrafactual
approaches and structural equation models (Greenland 2000).
Description of each detailed components of this model, supporting literature, and
full model utility are beyond the scope of this paper. Only aspects pertaining to the
purpose of this study, especially those involved in the casual pathways and compliance
effect, will be accounted in this section. Nevertheless, the utility of the model can be
extended to other areas in compliance research, such as understanding the
41
determinants of compliance and stability of the relation, implication for designing more
effective compliance-promoting interventions, and so on.
The compliance construct is influenced by upstream factors at multiple levels –
medication level (e.g. dose frequency, ease of administration, side effects), individual
level (e.g. belief, self-efficacy, etc.), interpersonal level (e.g. patient-physician
interaction, family support) and community level (e.g. insurance, policy, culture norm,
etc.). Each level contains one or more factors that influence compliance. Most of these
factors are not directly observable, and they can be influenced by the past or current
factors in unknown paths. Though many may not be directly measurable or observable,
some upstream factors can be determined or reflected by other measurable factors (some
with unclear paths), usually at patient level. The observable factors may not exert direct
influence on compliance, however, they do have relations that probably are intricate. The
causal paths of these factors are not certain and their correlation with compliance depends
on the controlled covariates and interactions. This can help explain some confusion in
compliance predictor literature. Several review articles estimate over 200 factors
variables have been studied as predictors or determinants of compliance (Donovan 1995;
Anon 1997)--though they can be classified as aforementioned four levels of predictors.
Some noticed the little utility of these predictors as “none of them is consistently related
to compliance or fully predictive” (Vermeire et al. 2001), and similar frustration was
expressed elsewhere (Cramer 1995). This is not surprising given that causal models were
not applied to determine a true relationship between a predictor and outcomes.
The determinants of outcomes include compliance and other factors that also
predict compliance, which evidences the existence of confounding for the
42
compliance effect on outcomes. In the above figure, they are exemplified by the direct
arrows linking from upstream levels to the outcomes without the route of compliance.
Compliance is essentially a continuum latent construct that is not otherwise
observable. Many observable behaviors can reflect such underlying compliance
construct, such as medication refill, appointment keeping, use of self-monitoring blood
glucose, diet, exercise and foot care and eye exam. In measurement theory, an ideal
measure of the latent construct is the common variance component of all these indicators
(Campbell and Russo 2001). At the downstream of this one-period model, compliance
indication factors and many proceeding factors influence the final health outcomes. It is
important to notice the direction of time in this diagram that indicates temporal
relationship among these factors, because distinct temporal relation between cause and
effect is a necessary condition for causal inference. This is a one period causal diagram
because the arrows from outcomes (or its surrogates) to the determinants of compliance
in the future periods are not shown.
The validation of those components and paths relies on valid, reliable, and
desirably multiple measures of unobservable psychological factors (personality traits) and
all other factors, and the use of structural equation models techniques. Though such goal
of validation is not easily achievable, at a minimum, the presented model can assist us to
understand multiple factors which are possibly involved in determining compliance and
mediating compliance-outcomes relationship.
2.4.3. The content of compliance concept and compliance effect
This diagram is also intended to clarify the content of compliance concept. With
this model, we can see what is more important is the content of compliance
43
concept (the extension of a concept), rather than the literal definition of compliance (the
intension of a concept) which is often the focus of debates in the literature. In the context
of evaluating compliance effect on outcomes, the content or extension of compliance,
when measured by a proxy indicator and it becomes an explicit observable scale, includes
multiple direct causes and multiple direct indictors that are not separately measured but
coexist in the compliance latent construct.
Given such explicit clarification of the extension of the compliance concept, we
should be aware that the compliance effect on health outcomes may contain therapeutic
effect (due to additional medication intake) and, to a certain extent, non-therapeutic effect
including the intangible effect of positive belief, self-efficacy, family support and patient-
physician rapport, and/or the effect of other non-measurable downstream compliance
indicators such as the effect of unobserved self-care, regular exercise and diet control,
etc. Even for analysis with data from prospective clinical trials, these non-therapeutic
effects are inseparable from therapeutic effect of compliance because we are never able
to observe and measure all the factors that are closely related.
As described in this paper, this model is capable of defining and clarifying the
compliance content in an observational compliance-outcomes study in a gray area in the
Figure 4, which contains effect of a broader range of factors closely related to and
undividable from compliance. In addition, this more “generalized concept of
compliance” also has practical policy implication because studies on proven effective
compliance-promoting interventions rarely contained multiple interventions to enhance
compliance partially through influencing patient’s psychological factors (McDonald et al.
2002).
44
The existence of non-therapeutic effect may help explain why in some clinical
trials, better compliance even to placebo was associated with better health outcomes
(Coronary Drug Project Research Group 1980; Beta-Blocker Heart Attack Trial Research
Group 1982; Gallagher et al. 1993), though these retrospective studies of prospective
trials could not clearly distinguish the source of such benefit which could be attributed to
the non-therapeutic compliance effect and to the artifact that patients prone to good
health tend to be more compliant because compliant patients already exhibited better
health at baseline (Coronary Drug Project Research Group 1980).
Non-therapeutic aspect of the compliance effect could be attenuated or eliminated
with alternative study approach and/or certain assumptions. With retrospective analysis
of clinical trial data where compliance levels and outcomes in both treatment and placebo
arms are measured, difference-in-difference modeling with certain assumptions can factor
out placebo compliance effect from the compliance effect in the treatment group, thus
provide the therapeutic part of the compliance effect. Though it is not clear whether
studies have actually used such modeling technique, attempts have been made with
simulating studies based on pharmacodynamic parameters to estimate non-compliance
effect (Urquhart 1998), where such effect would be the pure therapeutic effect. However,
the therapeutic effect due to marginal medication intake has less policy implication
because it is unlikely to achieve such compliance state without influencing patient’s
psychological awareness and other compliance-related behavior like self-care. Therefore,
the composite effect of both therapeutic and non-therapeutic aspects of compliance is
more likely to emulate the true benefit of compliance-enhancing interventions that target
on the direct causes of compliance at different levels, as shown in the figure 4.
45
2.4.4. Dynamic version of the model
Temporal relationship between compliance and other variables and as well as
outcomes demand a dynamic model rather than the “static” cross-sectional one with a
traditional approach. Some researchers have perceived that compliance-outcomes
interaction as a dynamic process. After a detailed description of possible dynamic
behavior in compliance involved in multiple factors (domains), Ickovics and Meade
summarized –
“Adherence behaviour is determined by a matrix of interrelated factors that shift
over time as the factors and adherence itself change. The relationship between
multiple factors may be reciprocal and reinforcing. For example, better adherence
may lead to better outcome, which may increase faith in the treatment regimen
and motivation to adhere” (Ickovics and Meade 2002).
This dynamic nature of interaction of compliance and multiple factors was also
noticed by other researchers (Rudd 1993; McNabb 1997; Williams 1997; Tsasis 2001;
Spire et al. 2002). Despite such awareness, compliance studies have not been able to
account for the dynamic interaction between compliance and multiple factors, probably
because of lacking a feasible dynamic model that can be implemented by most
researchers.
It is worth noting that a single period model treating observable factors as
stationary is incapable of solving the inter-related dynamic problems. Without realizing
the possible solutions of how to use longitudinal data to solve the dynamic interrelation,
some have doubted the possibility of identifying the causal relationship. The complexity
of compliance involved in a dynamic process is vaguely expressed as “Researchers have
attempted to identify causal relationships between variables, assuming that the variables
can be treated as independent. However, the phenomenon of medication-
46
taking behavior involves variables that are inter-related with the possibility of feed-back
loops” (Morris and Schulz 1992). This process of changed health outcomes through past
compliance feedback to future compliance was also recognized by other researchers as
“reverse causality”, where “reverse causality may also play a role, so that a good outcome
may promote subsequent adherence” and “outcomes may influence adherence,
particularly during a long course of treatment” (DiMatteo et al. 2002).
By dividing a continuous time period into smaller intervals, we can adapt our one-
period compliance conceptual model to a simplified dynamic model, where determinants,
compliance and outcomes (or surrogates) can be measured over time, and their causal
relations can be modeled under certain assumptions. Many time-varying factors (e.g.
clinical status over time, outpatients visits, resource utilizations) are not exogenous, in
that they are not independent of treatment (e.g. compliance), but rather affected by
compliance in the past. Such effects, as introduced previously, are called time-
confounding effects, and those variables are called time-varying confounders.
The causal diagram of the dynamic model is depicted by Figure 5. This diagram
only depicts two periods for simplicity, and it can be extended to multiple periods as
needed. In this diagram, C(t) denotes time-varying compliance measured at time t, and
L(t) represents all causal risk factors measured at time t for compliance and outcomes Y,
which in this study are survival outcomes. L(t) includes both time-varying confounders
as well as time-invariant baseline variables. U(t) denotes unobserved factors.
47
C
t
L
t
C
t+1
Y
L
t+1
U
t U
t+1
Time
Figure 5. Causal diagram for a dynamic compliance model
In this dynamic version of the conceptual causal model, C(t) proceeds and
influences L(t+1), and L(t) predicts C(t), where we assume that L(t) is temporarily prior
to C(t), given that the time interval is small. The dynamic interaction between L(t) and
C(t) implies L(t) as time-varying confounders. Such model with the specification of the
dynamic interaction between C(t) and L(t) fully accounts for the phenomenon of “feed-
back loops” and the “reciprocal and reinforcing (relationship)” as previously described.
It is important to notice when there are no direct arrows from unmeasured risk factors
U(t) to treatment variable C(t), and we say there are no unmeasured confounders
conditioning on observed confounders L(t), for every period t. Under the assumption of
sequential randomization (ignorability), despite the complexity of the model and
interrelated relationship as it depicts, it has been proven that the treatment effect on
outcomes (Y) can be consistently estimated by marginal structural model developed by
Robins and colleagues (Robins 1999).
48
CHAPTER 3. STUDY DESIGN AND METHODS
This study is a longitudinal observational cohort study. Medicaid administrative
claims data of California are used to examine the causal effect of hypoglycemics
compliance on the microvascular complications-free duration among type 2 diabetic
patients. As described previously, the hypothesis is that compliance has beneficial effect
on health outcomes, i.e., leading to a longer duration free of complication.
In order to account for the dynamic process over time, compliance level and time-
varying covariates are measured repeatedly on a quarterly basis, starting from the first
day when hypoglycemic medication is filled. Estimations of treatment effects are
compared between traditional Cox proportional hazard model and marginal structural
Cox model estimated by inverse-probability-of-treatment weighting.
3.1 Data
The data source for this study is twenty percent of the fee-for-service portion of
the California Medicaid program (Medi-Cal) paid claims and eligibility files for Medi-
Cal enrollees from January 1
st
, 1995 to December 31
st
, 2002. Medi-Cal covers outpatient
care, nursing home, inpatient, prescription drugs as well as other medical services for the
poor and disabled California residents.
Medi-Cal paid claims files include institutional (facility) claims at claim level,
professional services claims at service level, and pharmacy claims at specific drug level.
Service claims include date of service, type of service, place of service, paid amount,
billed amount, co-payment, units (days) of service, provider ID, primary and secondary
diagnosis codes, procedure codes (CPT), etc. Pharmacy claims include
49
national drug code (NDC), fill date, days of supply, paid amount, billed amount, quantity,
etc. Eligibility files include the enrollment status of each month, in addition to enrollee’s
demographic information.
Given NDC codes, drug brand and generic names, strength, and therapeutic
classes can be identified using the National Drug Data Files (NDDF®) of First DataBank,
Inc. Hierarchical Specific Therapeutic Class Codes (HIC3) in NDDF are used to identify
therapeutic classes.
3.2 Sample selection
To estimate the effect of compliance on time to the onset of microvascular
complications, we followed adult type 2 diabetic patients who “newly” started on
hypoglycemic medications. The following patient sample criteria are applied.
1. Patients have at least one diabetes diagnosis 250.xx in claims history, and no
diagnosis codes with fifth digit as ‘1’ (type I, not stated as uncontrolled) or
‘3’ (type I, uncontrolled). The fifth digit ‘0’ (type II or unspecified type, not
stated as uncontrolled) or ‘2’ (type II or unspecified, uncontrolled) is not
required because there is no specific type II diabetes diagnosis.
2. Patients need to have at least two prescription fills of hypoglycemics during
the entire claims history to be included. Hypoglycemics include oral
hypoglycemics and insulin, with their specific therapeutic class (HIC3) codes
listed in Table 1. The date of first hypoglycemics prescription fill date is
defined as the index date.
3. “Newly” start on hypoglycemics requires a six-month or longer washout
50
period (hypoglycemics naïve period) prior to the index date, following
conventional criteria (Dailey et al. 2001).
4. A minimum 6-month continuous eligibility period prior to the index date
(pre-index period) and at least 12-month continuous eligibility after the index
date (post-index period) are required, because a sufficient length of follow up
is needed to observe outcomes and measure compliance in a standard
regression as comparison.
5. Because typical type 2 diabetes patients do not start on insulin monotherapy
(Hayward et al. 1997) and existing diagnosis codes are not specific to type 2
diabetes, patients started with insulin as the only hypoglycemics during the
first quarter after the index date will be excluded. Such exclusion criterion
further removes possible type 1 diabetes patients.
6. Considering that children might have a different glycemic response in
developing microvascular disease and that most evidence of therapeutic
glycemic controls is for adults, we restrict our patient sample to be aged 18
years or older as of the index date.
7. Patients are excluded if they have diagnosis of diabetes-related microvascular
complications before or within 3 months after the index date, because some
patients might have experienced complications at the time of starting
medication treatment (Hillier and Pedula 2003) and there is a delay of lab
results reporting after the initial diabetes evaluation when the tests are
ordered. It is also because patients with concurrent complications appear to
have different response to medication treatments than those
51
without. For example, an unexpected observation in the DCCT study was
that about 10% of the patients with preexisting retinopathy experienced a
transient worsening of retinopathy symptoms after the treatment of tight
blood glucose control (Chantelau and Kohner 1997; The Diabetes Control
and Complications Trial Research Group 1998).
3.3 Compliance, covariates and outcomes measurement
The final selected patient cohort are followed until the occurrence of any the
following events, whichever occurs first: 1) discontinuation of Medicaid eligibility, 2)
diagnosis of diabetes-related microvascular complications to be defined later, 3)
December 2002, which is the end of administrative study period. During the follow up
period, the following variables will be measured. The diagram of the study period is
shown in (Figure 6).
2 134 56 -1 -2
Time (Quarters)
6-month washout
Index Rx Date
X
Diagnosis of complication
or disenrollment
Time to event
0
Figure 6. Diagram of study timeline.
3.3.1. Compliance
There are several assumptions under which the compliance here is measured.
1. Pharmacy refill claims record the actual way patients take their medications.
That is, patients take all prescribed medications according to the days supply
52
schedule once the prescription is filled, and patients do not have access to
medications not recorded in the claims data.
2. Patients in this study (with type 2 diabetes) should always be on at least one
hypoglycemic drug ever since the initiation of any hypoglycemics.
3. Given a certain measurement interval (e.g. quarterly in this study), we
assume the actual regimen observed in this interval is the regimen advised by
the physician for that interval.
4. Within the measurement interval, it is ignorable whether compliance is
constant or changing.
Assumption 1 allows us to observe the complete medication taking history of the
patients. Assumption 2 and 3 provide us with necessary information on the other part of
the compliance concept—physician’s advice, in order to evaluate compliance.
Assumption 4 is necessary when considering compliance in a dynamic environment, and
allows us to practically measure the change of compliance over time. The rationale,
necessity, variation, and implication of violating these assumptions are further discussed
in the discussion chapter.
Under these assumptions, compliance in this study is measured repeatedly for
each quarter (i.e. 90 days) from the index prescription fill date till the end of follow up
(as previously defined). Compliance is measured as MPR of hypoglycemic medications.
The MPR is defined as ratio of total number of days covered by any prescription of
hypoglycemic class within a given quarter over the number of calendar days (90 days).
Within each hypoglycemic class, though it would be rare, multiple drugs of the same
class may be filled during this interval. Whenever this happens, they are
53
treated as indifferent. Overlaped days of supplies from multiple fills of the same class
will not be double counted. In another word, there is no difference between a day
covered by one prescription supply or two prescription supplies of the same class. When
multiple classes of hypoglycemic medications are filled within a quarter, averaged MPR
for all filled hypoglycemic classes is the MPR for this quarter. As a result of these
technical specifications, the MPR value in this study for any quarter is within the range of
0 to 1.
The above describes quarterly measured MPR that allows compliance changing
over time. To be used in a traditional model, two versions of single-index MPR measures
are also defined—fixed-length single index MPR and variable-length single index MPR.
Fixed-length single index MPR measures the compliance over one year period, by
averaging the first four quarterly measured MPR described as above. Variable-length
single index MPR measures compliance over the entire follow up period (up to the event
occurrence or censoring point), by averaging MPR over all followed quarters.
Based on the convention of compliance literature (Claxton et al. 2001; DiMatteo
et al. 2002; DiMatteo 2004), compliance level measured by MPR in this study is
dichotomized to compliance and noncompliance. We adopt the traditional threshold 80%
as cutoff point for dichotomization (Haynes et al. 1976; Granstrom 1985; Stephenson et
al. 1993; Lee et al. 1996).
In order to test the sensitivity of different cutoff points during the dichotomization
process of compliance, we compare the coefficients estimates by marginal structural
model with different choice of cutoff points, with 10 percent point increment from 40%
to 90%. Lower cutoff points, though feasible in computation, are not
54
recommended because in a meta-analysis (DiMatteo 2004) that compared adherence rates
across various illnesses, the 95% confidence interval of mean adherence for diabetes was
(58.5%, 75.8%). Patients with less than 40% of MPR can be hardly justified to be
compliant.
3.3.2. Baseline covariates
Baseline variables are patient-specific variables and time-varying confounders
measured at baseline (t=0). The patient-specific variables included variables as follows.
1. Age: measured as of index date in years.
2. Gender (male=1)
3. Race: Asian, Black, Hispanic, Caucasian, and other (mostly missing value).
4. Index year: Calendar year 1995 to 2001, of the index date.
5. Prior eligibility: measured in number of months continuously eligible in the
California Medicaid program.
6. Charlson comorbidity index: measured over the 6 months pre-index baseline
period based primary and secondary ICD-9 diagnoses (Charlson et al. 1987;
Deyo et al. 1992).
7. Initial hypoglycemic regimen: measured during the first quarter since the
index date. Six hypoglycemic classes have potentially 62 possible
combination patterns (2
6
-2), excluding insulin only. Regimens with
relatively high frequency are specified, and the rest are combined together as
“other regimens” (see Chapter 4).
Besides, all time-varying confounders defined below are also reported as baseline
variables when they are measured during the baseline period (t=0), except the
55
hypoglycemic drug variables reported as initial hypoglycemic regimens measured in
quarter 1 as shown above.
3.3.3. Time-varying confounders
In addition to the above baseline covariates, many factors are measurable over
time after the initiation of hypoglycemic regimens. These factors (1) may predict future
compliance level and predict outcomes as well, and (2) may be influenced by past
compliance history. Such factors change over time and are termed as time-varying
confounders (Robins 1999). Recognizing of the existence of these confounders and
accounting for their dynamic interaction with compliance level that changes over time
distinguishes this study from all past efforts in estimating the compliance-outcomes
relationship.
Besides compliance that is measured quarterly, the following covariates are also
time varying and may interact with compliance level as time-varying confounders.
1. Number of office/outpatient visits: measured as number of office/outpatient
visits per quarter, including visits that are not emergency, inpatient, or
nursing home, with a broader range of services than the name implies.
2. Any emergency visit (a dummy variable): for each quarter.
3. Any hospitalization (a dummy variable): for each quarter.
4. Total number of unique drugs: measured at generic active ingredient
combination level (GCN_SEQNO in NDDF database) regardless whether the
drug is diabetes related or not, for each quarter.
5. Ever use of self-monitoring blood glucose (SMBG): measured
accumulatively up to this quarter as whether there is any claim of
56
SMBG pertaining to glucose monitors, lancets and test strips. Identification
criteria include HCPCS codes as E0607, E0609, E2100, E2101, A4253 and
A4772, or Medi-Cal supply codes as 9924G, 9924G, 9924N and 9924S, or
HIC3 codes as “M4A”.
6. Ever experience of hypoglycemia: ICD-9 diagnosis codes of hypoglycemia
are 250.3 (Diabetic hypoglycemic coma), 250.8x (Diabetic hypoglycemia, or
hypoglycemic shock), 251.0x (Hypoglycemic coma), 251.1x (Other specified
hypoglycemia) and 251.2x (Hypoglycemia, unspecified).
7. Ever diagnosed with one of eight included comorbidities: this is a class of
eight variables of comorbidities related to diabetes including coronary heart
disease (CHD, ICD-9: 410-414), congestive heart failure (CHF, ICD-9:
398.91, 428), depression (ICD-9: 300.4x, 301.12, 309.0, 309.1, 311),
hypertension (ICD-9: 401, 642.0), dyslipidemia (ICD-9: 272), valvular
disease (ICD-9: 093.2, 394-397, 424, 746.3-746.6, V42.2, V43.3), obesity
(ICD-9: 278.0), and hypothyroidism (ICD-9: 243, 244). These variables are
measured accumulatively up to the quarter with first such diagnoses.
8. Use of ACE inhibitors (ACEI) or Angiotensin receptor blockers (ARB): ever
filled a prescription of ACEI or ARB in the quarter.
9. Uncontrolled status of diabetes: accumulatively measured by ICD-9: 250.x2
(type 2 or unspecified diabetes, uncontrolled status) till the quarter when
patients are first diagnosed.
10. Experience of ketoacidosis or hyperosmolarity: measured by receiving
diagnosis ICD-9 diagnosis of 250.1 or 250.2 within the quarter.
57
11. Prescription of a certain hypoglycemic class: this is a class of 6 dummy
variables respectively for the 6 hypoglycemic classes, measured by any fill of
that hypoglycemic class within the quarter. These variables also imply
information on the regimen patterns and number of hypoglycemic classes.
These time-varying confounders are measured repeatedly for each quarter since
the index date up to the end of the follow up. Office visits often correlate with higher
compliance, so called “white-coat adherence” (Raynor 1992). The above list includes a
total of 23 time-varying confounders used in this study. These variables are selected
based on the possibility that they interact with compliance level and also influence final
health outcomes (duration of complication-free time). For example, ACEI and ARB are
known to reduce the incidence of diabetic complications especially nephropathy (Podar
and Tuomilehto 2002; Wolf and Ritz 2003), and may be prescribed based on health status
observed by the physician. Hypoglycemia often occurs in type 2 diabetes patients treated
with hypoglycemics, especially with tight glycemic control (Miller et al. 2001).
Therefore, patients with experience of hypoglycemia may modify their medication taking
behaviors. For comorbidities, a new diagnosis may characterize the recognition,
exacerbation, or a certain event in patient’s medical history. All included comorbidities
are measured accumulatively because they are chronic diseases.
Not all time-varying covariates are time-varying confounders. For example, age
changes over time and influence both treatment decision and outcomes, but its
progressing is independent of treatment received. Therefore, age is not a time varying
confounder. In another words, given baseline age, age over time is exogenous of time-
changing treatment.
58
3.3.4. Outcomes measures
The essential role of diabetes management is to prevent diabetic complications.
We only consider the onset (first diagnosis) of diabetic microvascular complications as
discussed in section 2.1. Microvascular complications of diabetes include neurological
disorders, diabetics foot problem, retinopathy, and nephropathy. The definitions of
diabetic complications based on ICD-9 diagnosis codes are listed in (Table 3). Because
patients with concurrent microvascular complication have been excluded from the study
sample, the date of the first claim with any of the listed diagnosis codes is considered as
the time of onset of microvascular complication. Because there is always a delay from
the onset of a certain complication to its manifestation and diagnosis, “time from index
date to the date of the first diagnosis of microvascular complication” would be a more
precise term to describe this survival endpoint. However, for the purpose of convenience,
in addition to the term “time to onset of microvascular complication”, “duration of
complication-free period” is used as an alternative.
Table 3. ICD-9 codes for chronic microvascular complications of diabetes
Chronic complications of diabetes ICD-9 codes
Neurological symptoms
Myasthetic syndromes in diseases classified elsewhere
(amyotrophy)
358.1
Other specified idiopathic peripheral neuropathy 356.8
Mononeuritis of upper and lower limbs 354, 355
Arthropathy associated w/neurological disorders
(Charcot’s arthropathy)
713.5
Peripheral autonomic neuropathy 337.1
Polyneuropathy in diabetes 357.2
Neuralgia, neuritis, and radiculitis, unspecified 729.2
Diabetes with neurological complications 250.6
Occlusion of cerebral arteries 434
Hemorrhagic stroke 430–432
59
Table 3. Continued.
Late effects of cerebrovascular disease 438
Occlusion of stenosis of pre-cerebral arteries 433
Other and ill-defined cerebrovascular disease 437
Acute, but ill-defined, cerebrovascular disease 436
Transient ischemic attack 435
Diabetic foot symptoms
Varicose veins of lower extremities 454
Gangrene and amputations 785.4, 895–897
Chronic osteomyelitis of the foot 730.17
Renal Complications
Infections of kidney 590
Other disorders of bladder 596
Cystitis 595
Renal sclerosis, unspecified 587
Glomerulonephritis, nephrotic syndrome, nephritis,
nephropathy
580–583
Proteinuria 791
Renal failure and its sequelae 584, 586, 588
Other disorders of kidney and ureter 593
Urinary tract infection 599
Diabetes and renal complications 250.4
Chronic renal failure (end-stage renal disease) 585
Ophthalmic complications
Other retinal disorders 362
Vascular disorders of the iris and ciliary body 364.0, 364.4
Disorders of the optic nerve and visual pathways 377
Diabetes with ophthalmic complications 250.5
Cataract 366
Glaucoma 365
Visual disturbance, low vision, blindness 368–369
Note: Adapted from American Diabetes Association 2003
This comprehensive list of diagnosis is provided by American Diabetes
Association, and may not be attributable to diabetes under all cases. The diagnosis codes
may be missing in many claims records. Inclusion of more comprehensive codes (e.g.
CPT-4) will increase the sensitivity of the outcomes measure, but it will drastically
reduce the specificity by giving more false positive cases.
60
3.4 Descriptive Analysis
Patient baseline characteristics and their initial hypoglycemic regimens are
described, with mean and standard deviation for continuous variable and number and
percentage for categorical variables. Mean, median and interquartile range of MPR for
each quarter are presented with box plots. Patients’ hypoglycemic regimens are
measured quarterly during the follow up period.
Stratified by the first year MPR with a cutoff point of 80%, patients are
categorized into compliance vs. noncompliance groups. We first present their
characteristics at baseline. In order to show the differences in the changes of time-
varying covariates between these two groups, time-varying covariates are demonstrated
at both first quarter and at the fourth quarter. All patients have both the first and fourth
quarters measures because they are required to have at least one-year continuous
eligibility. Between these two groups, Wilcoxon test is used to compare continuous
variables, and Chi-Square test for categorical variables. Unadjusted Kaplan-Meier curve
for complication-free survival is also presented.
3.5 Cox proportional hazard models
We first apply the conventional methodology to estimate the compliance effects
in comparison to marginal structural models. Cox proportional hazard model (Cox 1972)
is the most commonly used multivariate approach to analyzing time to event (survival
endpoint) data. Cox model treats the unknown baseline hazard function as nuisance
function, and models the effects of covariates on hazard in a multiplicative fashion. The
hazard function for the survival time T
i
of individual i is given by
61
0
( ) exp( ) ( )
ii
ht X h t β =
where the hazard function h
i
(t) is determined by a vector of covariates, X
i
, whose impact
is measured by β, a vector of coefficients to be estimated. The baseline hazard function
h
0
(t) corresponds to the hazard with X = 0 and does not need to be estimated based on
Cox’s method, because only the partial likelihood is needed for the estimation of
coefficients β. The Cox model is a semiparametric model where the survival times are
not assumed to follow a particular statistical distribution, and the model has high
efficiency and computational simplicity. However, covariates in the Cox model act
multiplicatively on the hazard, which require a key assumption that the hazard ratio of
any two groups is a constant proportion over time. Often, this proportional hazard
assumption is too strong in applications. One way to relax such assumption of time-
invariant proportionality is to use time-dependent covariates, which offer great flexibility
in model specification so that the proportional increase in hazard does not have to be
constant over time.
To test different strategies in choice of compliance measurement and modeling
methods within the standard Cox proportional hazard regression framework that is used
as comparison to the marginal structural model, we apply the following models as the
naïve methodology to estimate the “association” between compliance and survival
outcomes.
Model 1 includes a fixed-length single index compliance variable (C
1
) only
without any other covariates, which is the overall compliance measure for the first year.
This model is equivalent to testing whether there is significant difference between two
time-invariant hazard functions. Model 2 includes two models with baseline
62
covariates B and two measures of compliance respectively: fixed-length (C1) single-
index compliance for Model 2A and variable-length (C2) single-index compliance for
Model 2B. Model 2A is approximately equivalent to the conventional approach seen in
compliance-outcomes association literature. Model 3 includes baseline covariates B and
time-varying compliance C(t) which is measured repeatedly for each quarter until the end
of the follow up. Model 4 includes baseline covariates B, all time-varying covariates
L(t), and time-varying compliance measure C(t). In Model 4, time-varying covariates are
entered as exogenous variables (not determined by past treatment history) and their
dynamic interactions with time-varying compliance are ignored. The implication and
assumptions required for causal effect estimation of these models are discussed later.
At last, marginal structural model with inverse probability treatment weight
estimation to estimate causal compliance effect on outcomes will be described in detail in
the next section. In this model, time-varying covariates are not assumed to be
independent of time-varying compliance and modeled as time-varying confounders.
Following is the summary of the study models used to estimate compliance
outcomes relationship in this study (Table 4).
Table 4. Summary of model specification
Model Model specification Model Notation
1 Fixed length single index compliance C
1
2A Fixed length single index compliance + baseline covariate C
1
+ B
2B Variable length single index compliance + baseline covariate C
2
+ B
3 Time-varying compliance + baseline covariate C(t) + B
4 Time-varying compliance + time-varying covariates C(t) + B + L(t)
5 Marginal structural model with IPTW estimation MSM
63
These five models can be interpreted causally under different scenarios as
depicted in (Figure 7), where only scenario 5 resembles the reality that can be estimated
consistently by MSM. Each of the four Cox models corresponds to four scenarios in the
order they are presented. For example, scenario 1 resembles a randomized experience
where patients remain on their assigned compliance level for the rest of the study (though
it is implausible in practice), and so on. The detailed description of these assumptive
scenarios needed to support the causal estimate of each corresponding model will be
discussed in the discussion chapter.
64
C=1
C=0
Y
1
Y
0
1.
2.
C=1
C=0
Y
B
p
(c
=
1
| B
=
b
i
)
p
(c
=
0|
B
=
b
i
)
U
N
o
3.
C
t
L
t+1
Y
B
U
N
o
C
t+1
L
t+2
N
o
4.
C
t
L
t
Y
U
N
o
C
t+1
L
t+1
N
o
5.
C
t
L
t
Y
U
N
o
C
t+1
L
t+1
N
o
C
t-1
N
o
N
o
B
B
Figure 7. Diagrams of assumptions required for causal estimates by five corresponding
models
3.6 Marginal structural model
Previously described phenomenon of time-varying confounding is ubiquitous in
longitudinal studies where treatment is an adaptive and time-varying process. Formally,
65
time varying confounders are (a) time-dependent risk factors for survival, (b) which also
predict subsequent treatment and (c) also can be affected by past treatment history. A list
of possible time varying confounders have been defined and described in the previous
section. For example, comorbidities are risk factors of complication-free survival and
may predict treatment compliance by influencing patients’ awareness of the illness, and
occurrence of these comorbidities may also be affected by past compliance history.
Similarly, concurrent medications, outpatient/office visits, incidence of ER, inpatient,
hypoglycemia, and ketoacidosis/hyperosmolarity may also be time-varying confounders.
Even when there are no unmeasured confounding factors, using traditional models
such as general estimating equation (GEE) or Cox model with time-varying covariates
may still bias the estimate of treatment effect because these models treat the time-varying
covariates as exogenous variables (Robins 1999; Robins et al. 2000), as exemplified in
the Model 4. The estimates are biased because these variables are affected by past
treatment history (Robins et al. 1999).
In this study, we demonstrate that the selected time-varying covariates meet the
criteria (a), (b) and (c) as described above, thus are time-varying confounders. Model 4,
where time-varying confounders were entered in Cox regression as time-varying
covariates, can demonstrate that they are risk factors of developing complication. For
criterion (b), a GEE regression predicting dichotomous compliance level includes these
time-varying covariates and baseline covariates, which is the same regression to be used
in weight estimation. To demonstrate criterion (c), these time-varying covariates are
entered as dependent variables to be predicted by past compliance history. By the type of
the data, there are three types of time-varying covariates: continuous
66
covariates such as outpatient/office visits and number of total drugs; quarterly measured
dichotomous covariates such as ER, inpatient, and hypoglycemic regimens; and
accumulatively measured dichotomous covariates such as comorbidities, ever use of
SMBG, etc. Therefore, three types of GEE models were used to demonstrate that these
time-varying covariates can be predicted by past compliance history. Separate regression
models are fitted for each of these time-varying confounders respectively. For
continuously outcomes variables, linear GEE regression (identity link function and
normal distribution) was applied for each followed patient quarter. For quarterly
measured dichotomous outcomes, logistic GEE regressions (logit link function and
binomial distribution) were applied. For accumulative dichotomous outcomes, because
of their monotone response, logistic GEE regression was fitted with data where a patient
is only followed up to the quarter when the dependent variable is recorded as “1” for the
first time, such as receiving diagnosis of a new comorbidity that has not been previously
diagnosed. This is because, for example, we assume that once a patient receives a
diagnosis of a certain comorbidity, he always have such comorbidity for the rest of the
follow up period. So it is a monotone function over time. Two variables are used as
compliance history variables, average compliance of all past quarters and compliance of
the last quarter. In addition, all regressions also control baseline variables and all time-
varying covariates measured with 2 quarter lag L(t-2), according to (Hernan et al. 2001).
In order to account for the presence of time-varying confounders, Robins and
colleagues developed marginal structural models (MSMs) for causal inference of time-
varying treatment in observational longitudinal study (Robins 1999; Hernan et al. 2000;
Robins et al. 2000). They are a new class of causal models to estimate the
67
causal effect of a time-varying treatment with the presence of time-varying confounding
under the assumption of sequential ignorability, i.e., there is no unmeasured confounders
conditional on past treatment and covariate history. Using inverse-probability-of-
treatment weight (IPTW) is the estimation method for MSM. MSM with IPTW is a
practical method that can be implemented with common statistical packages. The
method of estimating MSMs is IPTW, where the inverse probabilities of receiving the
actual treatment given a set of previous covariate history are used as weight to estimates
the causal effect.
The assumptions, detailed proof and description, and estimation procedures of
MSMs have been presented in many original works of Robins and colleageus (Robins et
al. 1999; Hernan et al. 2000; Robins et al. 2000; Hernan et al. 2002). Marginal structural
can be used based on traditional regression framework, such as linear regression model
for continuous outcomes variable, logistic regression for dichotomous outcomes and Cox
regression for survival outcomes. Here we only provide a description of one of its
applications in survival outcomes analysis—marginal structural Cox model that is applied
in this study.
Based on the work of Robins and colleagues (Robins et al. 1999; Hernan et al.
2000; Robins et al. 2000), the marginal structural Cox model and its estimation are
briefly presented as the follows.
The problems of time-varying exposure and time-varying confounding to be
solved by marginal structural model can be demonstrated in the causal diagrams shown in
Figure 8, which is a replicate of Figure 5 presented previously as the causal diagram of
the dynamic compliance model. In this model, C(t) denotes patient’s
68
compliance level at time t (quarter). (In Robins and colleagues’ original work, A(t) was
used to denote time-varying treatment variable.) Time-varying risk factors L(t) predict
time-varying exposure C(t) and C(t) predicts L(t+1), indicating that L(t) are time-varying
confounders. We should notice that there are no direct arrows from unobservable factors
U(t) to C(t), indicating there is no unmeasured confounder at each time point t
conditioning on past covariate history. As a contrast, in Figure 8b, there are no arrows
from any non-treatment risk factors into any exposure variable, showing that there is no
confounding effect from either measured or unmeasured factors, which Robins and
colleagues termed as “causally exogenous” (Robins et al. 2000; Robins et al. 2003).
Only when the treatment is causally exogenous as presented in Figure 8b, traditional
standard regression such as Cox regression or GEE can offer unbiased estimate.
69
C
t
L
t
C
t+1
Y
L
t+1
U
t U
t+1
Time
C
t
L
t
C
t+1
Y
L
t+1
U
t U
t+1
a.
b.
Figure 8. Causal diagrams of time-varying exposure and covariates
Let T denote the time to the microvascular complication onset since the first
hypoglycemic prescription date after the initial prescription, with time measured in
quarters. In the above diagram, the outcomes variable is represented by Y. The bar over
C(t), () Ct , represents a compliance history, so () Ct ={C(u); 0 ? u ? t} is the patient
compliance history up to quarter t. B represents a vector of baseline covariates. Then the
conditional hazard function of microvascular complication onset at time t given the
baseline independent variables and treatment history can be modeled as
70
01 2
( | (), ) ()exp( () )
T
tCt B t Ct B λλ γ γ =+ .
This model is equivalent to the Model 3 of Cox models described above.
Because the hazard is a function of C(t), it can change over time when compliance
is measured as a time-varying variable. The baseline covariates (B) include variables
listed in section 3.3.2, and all time-varying confounders described in section 3.3.3
measured at baseline (t=0). Without elaborating on censoring points, this model can be
estimated by standard Cox proportional hazard estimation procedure by maximizing Cox
partial likelihood. However, the estimated
1
γ (as in Model 3) is a biased estimate of the
casual effect of compliance on the hazard with the presence of ignored time-varying
confounders, L(t). Even if L(t) are included in the model as time-varying covariates, such
as in Model 4
01 2 3
( | (), ) ()exp( () ())
T
tCt B t Ct B Lt λλ γ γγ =++ ,
it may still produce biased estimates because L(t) can be influenced by past compliance
history (-1) Ct .
Because of such time-varying confounding process, Robins and colleagues
introduced a time- and subject-specific measure of the degree to which the treatment
process is confounded by time-varying covariates (“statistically non-exogenous”) by time
t (Hernan et al. 2000; Robins et al. 2003),
0
(( ) ( )| ( 1) ( 1), )
()
(( ) ( )| ( 1) ( 1), , ( ) ( ))
t
i
ii
i
ii k
ii
pr C k c k C k c k B b
sw t
prC k c k C k c k B b L k l k
=
=−=−=
=
=−=−= =
,
71
where the product
0
()
t
k
zk
=
is the product (0) (1) ... ( ) zz zk . The denominator
(( ) ( )| ( 1) ( 1),( ) ( ), ) ii
ii
prC k c k C k c k L k l k B b =−=− = = is the conditional probability of
receiving the observed compliance level at quarter k, given past compliance and
covariates history (1) i ck − and () i lk , which the later includes baseline covariates. The
numerator is the conditional probability of receiving observed compliance level at quarter
k, given only baseline covariates and past compliance history, and the numerator serves
as a stabilizer for this inversed conditional probability of receiving actual exposure
(denominator). Robins and colleagues have shown that when these subject- and time-
specific measures are used as “stabilized” weights in an ordinary Cox model with time-
dependent covariate, the bias due to time-varying confounding can be eliminated or
removed (Robins 1999; Hernan et al. 2000; Robins et al. 2000). “Stabilized” by the
numerator, the weights can yield a more efficient and asymptotically normal estimate,
thus possess a better numerical property. This weighting method is the essential part of
the inverse-probability-of-treatment weight (IPTW) estimation method of the marginal
structural model.
Note that the numerator and denominator become equal when the time-varying
confounders () Lt become nuisance factors for all t, i.e., there is no presence of time-
varying confounding. It implies that all these stabilized weights become one, and the
treatment process is “statistically exogenous” (Hernan et al. 2001; Robins et al. 2003),
which is formally defined as () ()| ( 1) Lt C t C t − C , i.e., probability of compliance level at
time t, and does not depend on covariate history, conditioning on past compliance history.
72
Thus, the problem degenerates to the model depicted by Figure 8.b. Without additional
unmeasured confounder assumption, it can be estimated by standard regression such as
Cox model or GEE.
The estimation method based on this stabilized weights is inverse-probability-of-
treatment-weight (IPTW) partial likelihood estimate. In the expression of ) (t sw
i
, (1) C −
is defined to be zero when k=0. As an intuitive interpretation, given a risk at time t, the
purpose of weighting is to create a pseudo-population where ) (t L no longer predicts
compliance level at time t (i.e. no longer confounders) and the causal effect of
compliance remains unchanged. Thus, an otherwise complicated model can be converted
to a traditional Cox model framework with the subject and time specific weights.
Robins has proven that the IPTW can estimate the causal effect consistently under
the assumption of sequential randomization (Robins 1999). The proof process involves
two lemmas introduced by Robins regarding the contrafactual outcomes and causal
theory, which are based on the consistency of the precedent estimation method, g-
computational algorithm that was proven earlier (Robins 1986). The lemma 2 (a
corollary of lemma 1) shows that the consistency of the weight estimator was based on
the denominator, while the numerator only affects the efficiency of the estimation.
Because no standard software package can be used to estimate Cox regression
with patient and time specific weights, this difficulty can be overcome by fitting a
weighted pooled logistic regression with each person-quarter as an observation (Hernan
et al. 2000; Robins et al. 2003). Therefore, the regression model is described as
01 2
logit [ () 1| ( 1) 0, ( 1), ] () ( 1) prY t Y t C t B t C t B β ββ =−= − = + −+ ,
73
where t is integer value denoting number of quarters since the starting hypoglycemic
treatment, Y(t)=0 if the patient has not yet received any diagnosis of microvascular
complication by quarter t, and 1 if the patient received the first diagnosis, and
0
() t β is a
time-specific intercept. When the event rate at a given quarter t is small, the estimated β
with this pooled logisitc regression is closely approximate to that in the Cox regression
model (Robins et al. 2003). As previously stated, (1) Ct − represents compliance history
up to the quarter t-1. To be specific, in this study, we use two variables to represent one’s
compliance history at quarter t-1, an average MPR measured over the period from the
index date to quarter t-1 (the average compliance in the past), and the compliance
measured by MPR during the quarter t-1 (the most recent compliance). Weight ) (t sw
i
can be simply estimated by many standard statistical packages.
When patients discontinue enrollment in the Medi-Cal program during the follow
up period or the observed study period ends before we observe the first complication
diagnosis, censoring will occur. Let’s define D(t) as a censoring indicator, and it will be
the value of one if a patient is right-censored by quarter t and zero otherwise. With the
assumption of uninformative (or ignorable) censoring (i.e., loss of follow up do not
predict survival conditioning on observed covariates), Robins and colleagues showed that
using ) ( ) ( t sw t sw
i i
∗
? as weights for subject i at quarter t, where
0
[( ) 0| ( 1) 0, ( 1) ( 1), , ]
()
[ ( ) 0 | ( 1) 0, ( 1) ( 1), , ( 1) ( 1), ]
t
i
i
i
ii k
i
pr D k D k C k c k B b T k
sw t
prDk D k C k c k B b L k l k T k
∗
=
=−= −= − = >
=
=−= −= − = −= − >
, the estimated causal parameter
1
β still produces a consistent estimate of causal effect,
under the assumption that the history of observed covariates history is
74
sufficient to adjust both confounding and bias due to loss of follow-up (Robins 1999;
Robins et al. 2003).
To estimate the patient- and time-specific weights ) ( ) ( t sw t sw
i i
∗
? , we need to
estimate the numerator and denominator of the ( )
i
sw tand ()
i
sw t
∗
for each patient and
quarter. Specifically, this can be accomplished by estimating the following four GEE
(Liang and Zeger 1986) regression models with log link function and binomial
distribution, which are logistic regressions for repeated measure.
01 2
logit ( ()| ( 1), ) () ( 1) prC k C k B k C k B α αα −= + −+
01 2 3
logit ( ()| ( 1), (), ) () ( 1) () prC k C k L k B k C k B L k αα α α −= + −++
01 2
logit [ () 0| ( 1) 0, ( 1), ] () () prDk D k C k B k C k B α αα =−= − = + +
01 2 3
logit [ () 0| ( 1) 0, (1), (1), ] () () ( 1) pr Dk Dk C k Lk B k C k B L k αα α α =−= − − = + + + −
The regression coefficients will be used to estimate the predicted
µ
() /(1 )
XX
i
pke e =+ of
each numerator and denominator of weights ) ( ) ( t sw t sw
i i
∗
? for each patient i and quarter
t.
Because the regression intercept
0
() k α is quarter specific, to estimate it
consistently and “borrow strength” from other observations, we use restricted cubic
splines smoothing technique with three knots (Durrleman and Simon 1989).
Alternatively, this can also be accomplished by kernel regression, or by simply assuming
a linear response.
75
The estimated weights ) ( ) ( t sw t sw
i i
∗
? are necessary to run the pooled logistic
regression. However, because of the weighting methods, the standard errors output by
standard statistical software for logistic regression are invalid. To obtain a more robust
error estimates, the pooled logistic regression should be fitted using a generalized
estimating equation (GEE) (Liang and Zeger 1986), and the robust variance estimators of
GEE provide a conservative confidence interval for the β estimate. As an alternative of
GEE robust standard error, we also perform non-parametric bootstrap of random
sampling with replacement of 500 times to estimate the 95% confidence interval of
compliance effect on outcomes. The size of the bootstrap sample is the same as that in
the original sample, with reapplying all steps as detailed above for all these 500 samples.
Results of previously specified five Cox regression models and the causal
estimate by the marginal structural model using IPTW estimation are compared.
76
CHAPTER 4. RESULTS
4.1 Patient characteristics
A total of 4,708 patients with type 2 diabetes met previously specified sample
selection criteria, with the patient number flow show in (Table 5). These patients had a
total of 44,464 patient-quarter observations. A summary of the patient characteristics is
presented in Table 6. The mean age was 61.0 ( ?14.8) years, and 40.0% were male. At
baseline (6 months prior to the index date), there were 21.2% of the patients with
diagnosis of hypertension, 6.7% with dyslipidemia, 7.1% with coronary heart disease and
3.6% with congestive heart failure. During the first quarter, 2544 (54.0%) patients
received sulfonylureas (SUL) monotherapy and 1190 (25.3%) received metformin (MET)
monotherapy. Sulfonylureas plus metformin (SUL + MET) was the predominant
combination therapy used as the initial therapy of choice by 520 (11.1%) patients. By the
end of the follow up, 2,644 (56.2%) patients received diagnosis of microvascular
complications, and the median complication-free survival time was seven quarters.
Table 5. Sample selection criteria and patient counts flow
Sample selection criteria
Remained
(N)
Remaine
d (%)
With at least one hypoglycemics 93481 --
Age as of index date >=18 years 91247 97.6
6 months pre-index and 12 months post-index continuous eligibility 30149 32.2
At least 2 prescriptions of oral hypoglycemics 27308 29.2
Without diagnosis of type 1 diabetes 17625 18.9
At least one diagnosis of type 2 diabetes 14646 15.7
No microvascular complication before 3 months after index date 5044 5.4
With at least one oral hypoglycemics during the first quarter 4708 5.0
77
Table 6. Characteristics of the study population (n = 4,708)
Variable
Mean ? Std
N (%)
Age 61.01 ± 14.84
Male 1882 (40.0)
Race
Asian 1261 (26.8)
Black 434 (9.2)
Hispanic 637 (13.5)
Other 881 (18.7)
Caucasian 1495 (31.8)
Index year
1995 571 (12.1)
1996 1009 (21.4)
1997 616 (13.1)
1998 530 (11.3)
1999 525 (11.2)
2000 673 (14.3)
2001 784 (16.7)
Prior eligibility length in quarter 28.28 ± 20.9
Charlson comorbidity index at baseline 1.06 ± 1.11
Number of different drugs 4.22 ± 3.78
Number of office/outpatient visits 2.22 ± 3.44
Comorbidities at baseline
CHD 336 (7.1)
CHF 169 (3.6)
Depression 88 (1.9)
Hypertension 997 (21.2)
Dyslipidemia 314 (6.7)
Valvular Disease 62 (1.3)
Obesity 85 (1.8)
Hypothyroidism 66 (1.4)
Initial hypoglycemics regimen
SUL 2544 (54.0)
MET 1190 (25.3)
SUL + MET 520 (11.1)
TZD 97 (2.1)
INS + SUL 48 (1.0)
INS + MET 43 (0.9)
SUL + TZD 45 (1.0)
78
Table 6. Continued.
MET + TZD 40 (0.9)
NAT 32 (0.7)
All Other 149 (3.2)
Other time-varying covariates at baseline
Any ER visit 184 (3.9)
Any hospitalization 501 (10.6)
Use of ACEI or ARB 869 (18.5)
Uncontrolled diabetic condition 269 (5.7)
Hypoglycemia 57 (1.2)
Ketoacidosis or hyperosmolarity 43 (0.9)
The first year (fixed-length) single index MPR of these patients was 57.1%
( ?27.6%), with median of 58.8%. With 80% as a cutoff point, 1,285 (27.3%) patients
were grouped as being compliant. In order to examine the association between
compliance and baseline covariates, we compared the baseline covariates between the
compliance and noncompliance groups. As shown in Table 7, patients in the compliance
group were very different from those in the noncompliance group at baseline. They were
older (p<.0001), used more non-hypoglycemic drugs and ACEI/ARB at baseline, tended
to be treated in more recent years, and had more macrovascular complications (e.g. CAD,
CHF, dyslipidemia) at baseline.
Table 7. Baseline patient characteristics by one year compliance
Mean ± Std
Variables
N (%)
P-value
Non-Compliance Compliance
(n = 3,418) (n = 1,285)
Age 60.35 ± 14.98 62.76 ± 14.32 <.0001
Male 27.32 ± 20.33 30.84 ± 22.16 <.0001
Race
Caucasian 1012 (29.6) 483 (37.6) <.0001
79
Table 7. Continued.
Hispanic 903 (26.4) 358 (27.9)
Black 337 (9.9) 97 (7.6)
Asian 903 (26.4) 358 (27.9)
Other 637 (18.6) 637 (18.6)
Index year
1995 444 (13.0) 127 (9.9) <.0001
1996 781 (22.8) 228 (17.7)
1997 467 (13.6) 149 (11.6)
1998 386 (11.3) 144 (11.2)
1999 360 (10.5) 165 (12.8)
2000 463 (13.5) 210 (16.3)
2001 522 (15.3) 262 (20.4)
Prior eligibility length in quarter 27.32 ± 20.33 30.84 ± 22.16 <.0001
Charlson comorbidity index 1.07 ± 1.1 1.04 ± 1.04 0.8896
Number of all drugs used at baseline 4.03 ± 3.7 4.75 ± 3.85 <.0001
Number of office/outpatient visits 2.20 ± 3.2 2.27 ± 4.01 0.9143
Comorbidities
CHD 231 (6.8) 105 (8.2) 0.0912
CHF 113 (3.3) 56 (4.4) 0.0825
Depression 62 (1.8) 26 (2.0) 0.6322
Hypertension 713 (20.8) 284 (22.1) 0.3415
Dyslipidemia 213 (6.2) 101 (7.9) 0.0449
Valvular Disease 48 (1.4) 14 (1.1) 0.4017
Obesity 60 (1.8) 25 (2.0) 0.6583
Hypothyroidism 45 (1.3) 21 (1.6) 0.406
Initial hypoglycemics regimen
SUL 1860 (54.3) 684 (53.2) 0.4965
MET 815 (23.8) 375 (29.2) 0.0002
SUL + MET 401 (11.7) 119 (9.3) 0.0167
TZD 58 (1.7) 39 (3.0) 0.0039
INS + SUL 46 (1.3) 2 (0.2) 0.0003
INS + MET 35 (1.0) 8 (0.6) 0.1988
SUL + TZD 32 (0.9) 13 (1.0) 0.8093
MET + TZD 30 (0.9) 10 (0.8) 0.7436
NAT 26 (0.8) 6 (0.5) 0.2763
All Other 120 (3.5) 29 (2.3) 0.0292
80
Table 7. Continued.
Other time-varying covariates at baseline
Any ER visit 145 (4.2) 39 (3.0) 0.0582
Any hospitalization 368 (10.8) 133 (10.4) 0.6913
Use of ACEI or ARB 566 (16.5) 303 (23.6) <.0001
Uncontrolled diabetic condition 192 (5.6) 77 (6.0) 0.6139
Hypoglycemia 41 (1.2) 16 (1.3) 0.8947
Ketoacidosis or hyperosmolarity 29 (0.9) 14 (1.1) 0.4363
Use of self monitory blood glucose 93 (2.7) 35 (2.7) 0.9898
Time-varying covariates may interact with compliance and disproportionally
progress with the two groups of compliance and noncompliance, therefore we also
demonstrated the time-varying covariates at the 4
th
quarter of these two groups in
addition to their baseline values (Table 8). In this table, at the 4
th
quarter, patients in the
compliance group had significantly more visits than those in noncompliance group
(p=0.0023), but there was no difference at baseline; furthermore, fewer patients in the
compliance group had ER visits (p=0.0002) and inpatient hospitalizations (p=0.101),
whereas the baseline difference between two groups was smaller. In addition, patients in
the compliance group were more likely to receive oral hypoglycemics (e.g. MET, SUL,
TZD), but much less likely to receive insulin than the noncompliance group. It appeared
that some time varying covariates changed non-equally between compliance and
noncompliance groups, evidencing interaction between compliance and these time-
varying covariates.
Table 8. Time-varying covariates at baseline and the 4th quarter by compliance
Baseline
Mean ± Std
N (%)
4th quarter
Mean ± Std
N (%)
81
Table 8. Continued.
Noncompliance Compliance P-value Noncompliance Compliance P-value
Different drugs 4.0 ± 3.7 4.8 ± 3.9 <.0001 4.36 ± 3.66 6.06 ± 3.4 <.0001
Office/outpatient visits 2.2 ± 3.2 2.3 ± 4.0 0.9143 2.94 ± 5.02 3.05 ± 4.9 0.0023
Comorbidities
CHD 231 (6.8) 105 (8.2) 0.0912 524 (15.3) 207 (16.1) 0.4992
CHF 113 (3.3) 56 (4.4) 0.0825 251 (7.3) 109 (8.5) 0.1860
Depression 62 (1.8) 26 (2.0) 0.6322 135 (3.9) 52 (4.1) 0.8722
Hypertension 713 (20.8) 284 (22.1) 0.3415 1443 (42.2 577 (45.0) 0.0898
Dyslipidemia 213 (6.2) 101 (7.9) 0.0449 596 (17.4) 245 (19.1) 0.1867
Valvular Disease 48 (1.4) 14 (1.1) 0.4017 137 (4.0 53 (4.1) 0.8495
Obesity 60 (1.8) 25 (2.0) 0.6583 138 (4.0) 57 (4.4) 0.5352
Hypothyroidism 45 (1.3) 21 (1.6) 0.406 120 (3.5) 57 (4.4) 0.1350
Other time-varying covariates
Any ER visit 145 (4.2) 39 (3.0) 0.0582 71 (2.1) 7 (0.5) 0.0002
Hospitalization 368 (10.8) 133 (10.4) 0.6913 196 (5.7) 58 (4.5) 0.1010
Use of ACEI/ARB 566 (16.5) 303 (23.6) <.0001 823 (24.0) 487 (37.9) <.0001
Uncontrolled
diabetes 192 (5.6) 77 (6.0) 0.6139 353 (10.3) 132 (10.3) 0.9677
Hypoglycemia 41 (1.2) 16 (1.3) 0.8947 24 (0.7) 7 (0.5) 0.5545
Ketoacidosis/
hyperosmolarity 29 (0.9) 14 (1.1) 0.4363 21 (0.6) 5 (0.4) 0.3547
Use of SMBG 93 (2.7) 35 (2.7) 0.9898 397 (11.6) 171 (13.3) 0.1087
Hypoglycemics regimens
SUL 1860 (54.3) 684 (53.2) 0.4965 1381 (40.3) 853 (66.4) <.0001
MET 815 (23.8) 375 (29.2) 0.0002 1017 (29.7) 571 (44.4) <.0001
TZD 58 (1.7) 39 (3.0) 0.0039 197 (5.8) 101 (7.9) 0.0082
NAT 26 (0.8) 6 (0.5) 0.2763 49 (1.4) 11 (0.9) 0.1169
AGI -- -- -- 32 (0.9) 13 (1.0) 0.8093
INS -- -- -- 110 (3.2) 15 (1.2) 0.0001
Before introducing the patterns of compliance changing over time, we want to
demonstrate that the hypoglycemic regimens also change over time. The following table
shows the hypoglycemic combination patterns over time for patients remaining in the
study sample (Table 9). It is shown here that most regimens decreased in
82
proportion of patients over quarters because more and more patients discontinued therapy
and had “no drug” during those quarters. One exception where the regimen portion
increased over time was the “all other regimens” group, which mostly included many
rarely used combinations across classes.
Table 9. Hypoglycemic regimens by selected quarters
Q1 Q2 Q4 Q6 Q8 Q16
% % % % % %
N 4708 4261 3519 2618 2007 758
SUL 54.0 41.2 31.2 27.1 23.9 20.2
MET 25.3 20.8 18.3 17.8 15.9 12.1
SUL + MET 11.1 9.7 10.9 10.6 12.0 11.9
TZD 2.1 1.8 1.6 1.6 1.3 0.9
INS + SUL 1.0 0.4 0.5 0.4 0.4 0.5
INS + MET 0.9 0.4 0.5 0.3 0.3 0.4
SUL + TZD 1.0 1.1 1.6 2.0 1.9 2.0
MET + TZD 0.9 1.1 1.1 1.1 1.2 2.2
NAT 0.7 0.7 0.6 0.6 0.7 0.4
No drug 0.0 19.0 30.2 34.3 38.5 44.1
Other 3.2 3.7 3.7 4.3 4.0 5.3
4.2 Compliance changing over time
As this study measured compliance by MPR for every quarter until the end of the
follow up. For the entire 44,464 included patient-quarters, the mean of MPR of all
quarters was 48.4% (±39.1%) and the median was 54.4%. The distribution was not
normal, with 13,500 (30.4%) quarters having zero MPR corresponding to no drug-
covered days during the quarter, and 4,028 (9.5%) quarters having 100% MPR.
In order to show the compliance pattern over time, we first plotted the population
compliance over time using box plot that summarized mean, median and
83
interquartile range for each quarter (Figure 9). The width of the box was proportional to
the number of patients remained in that quarter. This plot resembled the long-term
compliance patterns shown in other studies (Benner et al. 2002; Benner et al. 2004),
where compliance dropped sharply during the first year, then decreased slowly (“leveled
off”) and remained relatively stable, and followed by a slight upward trend at the end due
to the survival effect. Another interesting observation was the relative skewness of the
MPR distribution reflected by the relative position of the median and the mean. During
the first seven quarters, the median was higher than the mean, indicating a left skewness
of MPR, i.e., more patients clustering on the higher end of the MPR. Afterward, the
trend started reversing to right skewness, and more patients clustered on the lower end of
the MPR especially after three years of treatment.
Figure 9. Box plot of compliance (MPR) of hypoglycemics for each quarter. Bars
represent interquartile range; Width of the bar represents sample size for that
quarter.
84
This box plot depicted the compliance distribution over time for the entire study
sample as a whole. The trend of compliance appeared to be smooth on population level.
However, it did not provide insights no how compliance level changed on the individual
level. To evidence the pattern of compliance change on a patient level, we plotted the
quarterly compliance measured by MPR for randomly selected 10 patients (Figure 10).
As we can see in this figure, patients’ compliance over time appeared to be very erratic.
Similar erratic patterns were evidenced with more randomly selected patients (figure not
shown). Apparently, compliance over time was not a static but time-varying process.
Such changes over time might not be random, but dynamically interacted with observable
time-varying covariates, as shown later. The change of quarterly MPR is associated with
the change of hypoglycemic regimens over time. For each quarter, in addition to
measured MPR, we also recorded the hypoglycemic regimens measured by whether use
of each one of six hypoglycemic classes and the number of hypoglycemic classes used in
the quarter. The mean MPR of each hypoglycemic class at each quarter (non-zero
MPRs) was plotted over time (Figure 11). This figure showed that compliance on MET,
SUL and TZD maintained relatively stable over time with slightly increases at the end of
the study period, and AGI, which was the only agent approved for combination use with
sulfonylureas, showed decreased compliance over time. Insulin, though with overall
lower compliance compared to oral hypoglycemics, was observed with an increased
compliance during the first few years of treatment. All drug classes showed higher
variation in mean compliance at the end of the study period because of decreased number
of patients available.
85
M PR
0
10
20
30
40
50
60
70
80
90
100
quar t er
010 20 30 40
Figure 10. Quarterly compliance of 10 randomly selected patients
D r ug C l ass A I M N S T
M PR
30
40
50
60
70
80
90
100
Q uar t er
0 10 20 30
Figure 11. Mean quarterly non-zero MPR by hypoglycemic classes
86
In addition to therapeutic classes, combination therapy may also associate with
level of compliance. A bubble plot was created to demonstrate the mean number of
hypoglycemic classes and non-zero MPRs of the quarters that contained each specific
hypoglycemic class (Figure 12). As expected, this plot evidenced the negative
correlation between number of classes and mean MPR, though evidenced only at
therapeutic class level. Shown in this figure, oral hypoglycemic classes including SUL,
MET and TZD were associated with quarters with higher compliance, and not
surprisingly, insulin (INS), usually reserved for uncontrolled type 2 diabetes patients, had
the lowest compliance among all therapeutic classes due to its inconvenient
administration. Two most common hypoglycemic classes of SUL and MET were more
likely to be used as monotherapy, and others in combination therapy. Thus, we have
shown that hypoglycemic regimens did correlate with compliance level, therefore it is
important to include hypoglycemic regimen variables as time-varying confounders in the
model.
87
Figure 12. Mean MPR and hypoglycemic classes for all quarters with non-zero MPR.
The bubble size is proportional to the number of quarters with the hypoglycemic class
4.3 Survival analysis and Cox regression models
As a common approach in survival analysis, we first showed the differences of
time to onset of microvascular complication by one-year single-index compliance using
Kaplan-Meier curve. Without adjusting any baseline covariate, Kaplan-Meier curve for
complication-free survival by compliance was plotted (Figure 13). As shown in this
figure, patients with non-compliance on hypoglycemics had better outcomes – less risk of
developing microvascular complications. The unadjusted log rank test has a chi-square
statistics as 3.0561 with p value 0.0804. If we do not assume that compliance on
hypoglycemics is harmful, the presence of selection bias is self-evident with the direction
unfavorable to compliance group.
88
Figure 13. Unadjusted Kaplan-Meier survival curve by one-year compliance
As proposed in the method section, we fitted the specified four different models
using Cox proportional hazard regression with the results shown in Table 10.
Table 10. Hazard ratio of compliance variable by Cox regression for Model 1-4
No Model Hazard Ratio 95% CI P-value
1 C
1
1.08 (0.99, 1.17) 0.0923
2A C
1
+ B 1.06 (0.97, 1.15) 0.235
2B C
2
+ B 1.53 (1.40, 1.67) <.0001
3 C(t) + B 1.09 (1.00, 1.18) 0.0449
4 C(t) + B + L(t) 0.96 (0.88, 1.04) 0.3089
89
The estimates of Model 1 corresponded to the unadjusted Kaplan-Meier curve
with assumption of a constant hazard ratio between compliance and noncompliance
group (Figure 13). Without adjustment of any covariate, one-year single-index
compliance was associated with higher risk (HR=1.08) of developing microvascular
complication with p=0.0923. When adjusting for baseline covariates with inverse
probability weighting method in Model 2A, one-year single index compliance was still
associated with higher risk (HR=1.06), however, this was not statistically significant
(p=0.235). We plotted the baseline covariates-adjusted survival curve (Figure 14). The
difference between compliance and noncompliance was much smaller and not significant.
This implies that part of the selection bias shown in the Model 1 can be partially
explained by differences in baseline covariates.
Figure 14. Complication-free survival curves adjusted by baseline covariates
90
Model 2B is same as Model 2A, except using variable-length single-index
compliance, which is defined over the entire follow up period. The estimated hazard
ratio of this compliance measure was 1.53 (p<0.0001), presenting a very strong bias.
Model 3 included time-varying compliance measured quarterly up to the end of
follow up and baseline covariates, and estimated the hazard ratio of compliance as 1.09
with statistical significance (p=0.0449). Including all time-varying covariates in Model
4, the selection bias appeared to be alleviated, with estimates indicating a trivial effect of
compliance (HR=0.96, p=0.3089).
4.4 Demonstration of time-varying confounding
For each time period, time-varying confounders were shown to meet the
following three conditions as described previously: (a) they are risk factors of survival,
(b) they can predict subsequent compliance, and (c) they can be predicted by past
compliance history. Condition (a) was shown using the same Cox regression as in Model
4, where these covariates were modeled as time-varying covariates along baseline
variables and time-varying compliance. Results of the Cox regression with time-varying
covariates showed that many of these time-covariates were significant risk factors for
survival (Table 11).
Table 11. Time varying covariates as risk factors of survival
Variable Hazard ratio 95% C.I. P-value
Comorbidities
CHD 1.06 0.94 1.19 0.3542
CHF 0.93 0.80 1.09 0.3724
Depression 1.10 0.87 1.38 0.4334
Hypertension 1.13 1.03 1.25 0.0105
91
Table 11 Continued.
Dyslipidemia 1.25 1.12 1.40 <.0001
Valvular Disease 0.99 0.82 1.20 0.9185
Obesity 1.25 1.00 1.56 0.0460
Hypothyroidism 1.02 0.82 1.27 0.8569
Hypoglycemics regimens
INS 0.99 0.80 1.22 0.9257
SUL 1.17 1.07 1.28 0.0005
NAT 1.27 0.91 1.77 0.1655
MET 1.17 1.06 1.29 0.0015
AGI 1.44 1.06 1.95 0.0206
TZD 1.15 0.98 1.35 0.0837
Other time-varying covariates
Any ER visit 1.57 1.26 1.96 <.0001
Any hospitalization 1.96 1.73 2.22 <.0001
Number of office/outpatient visits 1.06 1.05 1.06 <.0001
Number of all drugs used 1.07 1.06 1.08 <.0001
Use of ACEI or ARB 1.09 0.99 1.20 0.0889
Uncontrolled diabetic condition 1.16 1.03 1.31 0.0147
Hypoglycemia (ever) 1.48 1.09 2.00 0.0121
Ketoacidosis or hyperosmolarity 1.24 0.80 1.91 0.3452
Use of SMBG (ever) 1.01 0.89 1.14 0.9195
To show the condition (b), we fitted GEE logistic model using baseline covariates
and time-varying covariates to predict compliance level at time t. Because the interval
was relatively small, we assumed that these time-varying covariates were measured prior
to the measure of compliance. Therefore, we applied the same GEE regression as the
second one in the series of four GEE regressions that were used to generate weight in the
MSM. The results showed that many of these time-varying covariates significantly
predicted subsequent compliance (Table 12). The results of the regression (Table 13)
showed compliance history variables were able to independently predict time varying
covariates, so condition (c) was met.
92
Table 12. Time-varying covariates can predict subsequent compliance
Time-varying covariates Estimate 95% C.I. P-value
Time-varying comorbidities
CHD 0.17 0.04 0.30 0.0104
CHF 0.10 -0.08 0.28 0.2752
Depression 0.29 0.05 0.54 0.0188
Hypertension 0.01 -0.08 0.10 0.8547
Dyslipidemia 0.03 -0.08 0.14 0.5980
Valvular disease -0.01 -0.22 0.21 0.9614
Obesity -0.27 -0.47 -0.06 0.0118
Hypothyroidism -0.10 -0.33 0.13 0.3906
Hypoglycemics
INS 1.33 1.01 1.64 <.0001
SUL -0.79 -0.88 -0.70 <.0001
NAT 0.34 0.03 0.65 0.0327
MET -0.34 -0.43 -0.26 <.0001
AGI 0.25 -0.06 0.56 0.1190
TZD 0.11 -0.03 0.25 0.1094
Other time-varying covariates
Any ER visit 0.23 0.01 0.46 0.0415
Any hospitalization 0.55 0.41 0.68 <.0001
Number of office/outpatient visits 0.00 -0.01 0.01 0.4461
Number of all drugs used -0.10 -0.11 -0.08 <.0001
Use of ACEI/ARB 0.08 -0.04 0.19 0.1776
Uncontrolled diabetic condition -0.11 -0.19 -0.03 0.0093
Hypoglycemia (ever) 0.55 0.22 0.88 0.0010
Ketoacidosis or hyperosmolarity -0.28 -0.67 0.10 0.1534
Use of SMBG (ever) 0.03 -0.09 0.16 0.5777
Table 13. Time-varying covariates can be predicted by past compliance history
Compliance of the
last quarter
Past average
compliance
Estimate P-value Estimate P-value
Continuous variables
Number of office/outpatient visits 0.0063 <0.0001 -0.0078 <0.0001
93
Table 13. Continued.
Number of different drugs used 0.0171 <0.0001 -0.0136 <0.0001
Quarterly measured dichotomous variables
Any hospitalization -0.0013 0.2574 0.0014 0.3522
Any ER visit -0.0009 0.6475 0.0011 0.6667
Ketoacidosis or hyperosmolarity -0.0084 0.0360 0.0078 0.1039
Use of ACEI or ARB -0.0100 <0.0001 0.0054 <0.0001
Hypoglymic regimens
INS -0.0081 0.0003 0.0107 0.0003
SUL -0.0404 <0.0001 0.0284 <0.0001
NAT -0.0128 <0.0001 0.0114 0.0003
MET -0.0299 0.0174 0.0204 0.0014
AGI -0.0170 <0.0001 0.0134 0.0001
TZD -0.0162 <0.0001 0.0144 <0.0001
Accumulatively measured dichotomous variables
Uncontrolled diabetic condition -0.0070 <0.0001 0.0041 0.0055
Hypoglycemia (ever) -0.0038 0.2439 0.0012 0.7629
Use of SMBG (ever) -0.0077 <0.0001 0.0040 0.0064
Comorbidities
CHD -0.0019 0.1279 -0.0037 0.0137
CHF -0.0014 0.3954 -0.0032 0.1206
Depression -0.0070 0.0043 -0.00001 0.9965
Hypertension -0.0023 0.0090 -0.0040 0.0002
Dyslipidemia -0.0034 0.0023 -0.0030 0.0272
Valvular Disease 0.0022 0.2836 -0.0088 0.0007
Obesity -0.0083 0.0007 -0.0011 0.7059
Hypothyroidism -0.0036 0.165 -0.0005 0.8463
Because the above three conditions, (a), (b) and (c) were all met, it was evidenced
that these time-varying covariates were time-varying confounders. A table summarized
all the values with their level of significance (Table 14). We should notice that not all
confounders were significant with these conditions. The model form and time-varying
confounders are determined as a priori in this study, following good practice
94
recommendation for retrospective database studies (Motheral et al. 2003). We did not
select time-varying confounders based on statistical significance criteria for the same
reason that standard regression should not only include significant predictors (Greenland
1989; Austin and Tu 2004).
Table 14. Significance levels of time-varying confounders by three criteria
Criteria
Time-varying confounder
a b c
Comorbidities
CHD - ** *
CHF - - -
Depression - * **
Hypertension ** - ***
Dyslipidemia *** - **
Valvular Disease - - ***
Obesity * * ***
Hypothyroidism - - -
Hypoglycemics regimens
INS - *** ***
SUL *** *** ***
NEG - * ***
MET ** *** **
AGI * - ***
TZD - - ***
Other time-varying covariates
Any ER visit *** * -
Any hospitalization *** *** -
Number of office/outpatient visits *** - ***
Number of all drugs used *** *** ***
Use of ACEI or ARB - - ***
Uncontrolled diabetes (ever) * ** ***
Hypoglycemia (ever) * ** -
Ketoacidosis or hyperosmolarity - - *
Use of SMBG (ever) - - ***
95
- not significant, * <0.05, ** <0.01, *** <0.001
Interpretation of the dynamic mediation of each time-varying confounders with
respect to the compliance outcomes relationship is not straightforward. MSM unlike its
precedent models (g-estimation and structural nested models) does not allow estimation
of the dynamic effect of these time-varying confounders directly. We should notice that
these tables and their significance level do not offer proof of time-varying confounding
effect, but a demonstration of its presence. We could also probe the story behind these
tables for three criteria that a time-varying confounder needs to meet. For example,
number of total drugs (a proxy for pill burden) is obviously a risk factor for complication
(as shown in criterion a), and number of drugs significantly predicts subsequent
noncompliance (criterion b). Criterion c uses two variables for past compliance--the
compliance of the previous quarter and average compliance of all past quarters.
Interpretation of such effects is not straightforward. With the example of number of
drugs (pill burden), controlling past average compliance level, the compliance of the
previous quarter significantly predict higher number of drugs of this quarter. This
implies that compliance to hypoglycemics also likely to signal refill and compliance of
other drugs (e.g. cardiovascular medications). This shows a negative feedback--with the
pill burden increases, compliance would likely to be negatively impacted in the
subsequent shown by criterion b. An example of positive feedback is uncontrolled status
of type 2 diabetes. Uncontrolled status predicts noncompliance (criterion b), which
further reinforces the likelihood of receiving diagnosis of uncontrolled status of the next
quarter (criterion c).
96
4.5 Causal compliance effect estimated by MSM with IPTW
Censoring-adjusted stabilized weights ) ( ) ( t sw t sw
i i
∗
? were estimated based on
the predicted probability estimated from four separate logistic regression using GEE
methods. Combining all patient quarters, the mean of the stabilized weights was 1.37
(s.d.=19.57) and median 0.812, with interquartile range 0.592-1.070. Summarizing the
weights by quarter, the box plot shows the temporal distribution of log of stabilized
weights (Figure 15). The width of the box indicates the number of patients remained by
the quarter, and the upper and lower end of the box represents interquartile range.
Wiskers corresponds to 1.5 times of interquartile range above and below the 75
th
and 25
th
percentile. The squares denote outliers. This plot shows that the mean stabilized weights
are decreasing over time, yet being “stable” and not too far from the zero line; and the
variation of stabilized weights increase over time.
Figure 15. Log of stabilized weights distribution by quarter
97
With the estimated stabilized weights, causal effect of compliance on the risk of
microvascular onset was estimated by generalized estimation equation for logistic
regression. The estimated hazard risk was 0.73, with confidence interval estimated by
robust estimator of standard error of GEE as (0.54, 0.99) and p value of 0.0441, showing
a statically significant causal benefit of compliance. The robust estimator of GEE
provided a conservative confidence interval. Using nonparametric bootstrap method with
500 samples with replacement, the estimated 95% confidence interval was (0.60, 0.92),
narrower than that estimated by GEE. Though we did not perform nonparametric
bootstrap methods for all subsequent sensitivity analysis, we should keep in mind that the
statistical significance shown by p value generated from GEE was very conservative.
This estimate was an unbiased causal effect estimate, under the assumptions of sequential
ignorability (i.e. no unmeasured confounding at each time conditioning on past covariate
history) and correct model specification.
Combining the compliance estimates from various specifications of Cox models,
here we listed the result of all models (Table 15). The table shows that only the
compliance effect estimated by the marginal structural model, the causal effect estimate,
indicates a beneficial effect of compliance on outcomes, while all Cox regression models
indicate a positive association between compliance and risk of developing complication
and they are statistically significant in Model 1, 2B, and 3.
98
Table 15. Comparison of compliance effect estimates of Model 1-5
Model No. Model Estimate Hazard Ratio 95% C.I. P-value
1 C1 0.0737 1.08 (0.99, 1.17) 0.0923
2A C1 + B 0.0532 1.06 (0.97, 1.15) 0.235
2B C2 + B 0.4266 1.53 (1.40, 1.67) <.0001
3 C(t) + B 0.0851 1.09 (1.00, 1.18) 0.0449
4 C(t) + B + L(t) -0.0447 0.96 (0.88, 1.04) 0.3089
MSM -0.3151 0.73 (0.54, 0.99) 0.0441
5
(Bootstrap) (0.60, 0.92)
4.6 Sensitivity analysis
A sensitivity analysis was conducted to determine whether using different MPR
cutoff points for dichotomizing compliance (i.e. changing compliance definition) would
influence degree of compliance effect.
Difference compliance measures were tested using various MPR cutoff points
with 10 percent point increment from 40% to 90%. For each of these alternative cutoff
points, we repeated analysis for all specified modeling strategies, i.e., Model 1-5. The
results of the estimate hazard ratios of compliance over noncompliance were shown in
(Table 16). When cutoff points changed, compliance coefficients estimated by different
models also changed, but apparently different models had different trends and sensitivity
to the cutoff points. In general, except Model 3 (measured by variable length single
index compliance and baseline covariates), estimated hazard ratios trended towards 1.
This may be because the larger the MPR cutoff points, the less distinction between the
compliance and noncompliance groups. In Cox models, the compliance positive
associations (HR>1) were even more biased if we chose to use smaller MPR as cutoff
points as shown in the table. In contrast, as an interesting finding, the benefit
99
of compliance estimated by the causal MSM (Model 5) was more pronounced when
smaller cutoff points were used, such as 40% and 50%. At 50% cutoff points, the hazard
ratio was only 0.13 yet with confidence interval that covers 1. The unknown dose-
response effect of compliance is needed to reveal this interesting observation.
Table 16. Sensitivity analysis on dichotomized compliance with different MPR cutoff
points
Model (hazard ratio, p-value)
MPR
cutoff
point
1 2A 2B 3 4 5 (MSM)
1.17 1.14 1.84 1.47 1.39 0.32
40%
0.0003 0.0024 <.0001 <.0001 <.0001 <.0001
1.18 1.14 1.84 1.41 1.29 0.13
50%
<.0001 0.0013 <.0001 <.0001 <.0001 0.0985
1.15 1.12 1.63 1.31 1.19 0.55
60%
0.0004 0.004 <.0001 <.0001 0.0001 <.0001
1.14 1.12 1.55 1.22 1.10 0.51
70%
0.001 0.0062 <.0001 <.0001 0.0293 0.0313
1.08 1.06 1.53 1.09 0.96 0.73
80%
0.0923 0.235 <.0001 0.0449 0.3089 0.0441
1.02 1.01 1.96 1.20 1.13 0.86
90%
0.7288 0.8403 <.0001 0.0001 0.0119 0.1295
It is important to notice that this sensitivity analysis is not a dose-response fashion
between MPR and hazard risk. Though it is attempting, we do not suggest plotting the
hazard ratios over MPR cutoff points, because such plot is likely to be misinterpreted as a
dose-response relationship between hazard ratio and MPR. MSM at current stage could
not be implied to estimate a dose-response effect of compliance. Each cutoff point
divided patient-quarter observations into two buckets: compliance and noncompliance.
100
The estimated hazard ratio was approximately the ratio of the risk rate between these two
groups.
101
CHAPTER 5. DISCUSSION
5.1 Overview
It is commonly believed that higher compliance improves health outcomes, which
is the rationale for numerous efforts of compliance-promoting interventions. However, it
is also argued that noncompliance may be patients’ rational choice from their own
perspective (Donovan and Blake 1992), and the level of compliance to medications may
be optimal to individual patients from their own perspective. Lack of evidence of the
benefit of better compliance makes some argued “medication compliance as an
ideology” (Trostle 1988).
Estimating the effect of medication compliance on health outcomes is thus
paramount to justify and evaluate compliance-improving interventions as well as to
account for real world compliance in cost-effectiveness analysis of drug therapy.
Conventional methods of estimating compliance-outcomes relationship are often
applied with a pre-post design, where in the post period, both compliance and outcomes
are simultaneously defined. The estimates generated by the standard regression
inevitably present a mere association between compliance and outcomes. Such
association estimated by traditional methods not only biases estimate, but also is
incapable of distinguishing the effect of outcomes on compliance and that of compliance
on outcomes due to the obscure temporal relationship between compliance and outcomes
measures. For example, a significant positive association between compliance and
outcomes is generally assumed to imply that better health outcomes are owing to higher
102
compliance, however, the converse is also possible that patients stop taking medications
for alternative therapies due to their deteriorative health.
In addition, traditional designs measure compliance as a single value over the
study period as if it was a static process, despite its time-varying nature. As a matter of
fact, many factors dynamically interact with compliance decisions over time: medication
compliance change health status or other factors, which also influences the decisions of
future compliance. Such phenomenon of time-varying confounding has bewildered
medical researchers, and been termed as “feed-back loops” (Morris and Schulz 1992) and
“reverse causality” (DiMatteo et al. 2002). Using standard regressions, although baseline
covariates can be controlled, the dynamics of compliance and simultaneous confounders
and intermediate variables cannot be appropriately accounted for. As shown in this
study, the estimate from Cox regression did not show any benefit of medication
compliance.
As a contrast, this study treated compliance as a changing process that
dynamically interact with time-varying confounders, and measured compliance
repeatedly and strictly prior to the measure of outcomes. Introduction of chronologically
distinctive order of cause and effect variables in a longitudinal analysis is necessary in
causal inference. Cox marginal structural model (MSM) was applied to estimate the
effect of higher compliance to hypoglycemics on the delay of microvascular
complications in diabetes patients. This model, developed by Robins and colleagues for
the purpose of casual estimation of time-varying exposure, was estimated by inverse-
probability-of-treatment weighting method. The causal estimate of compliance was 0.73
103
with bootstrap confidence interval as (0.60, 0.92). The benefit was more pronounced
with smaller cutoff points of MPR.
This result is consistent with the finding from UKPDS that glycemic control by
sulfonylureas or insulin compared to conventional diet could reduce microvascular
complications (UKPDS Group 1998). This study extended such findings to a strong
benefit of hypoglycemic compliance in delaying the onset of microvascular
complications.
This finding provides an evidence for future cost-effectiveness analysis where the
effect of compliance is necessary for translating medication “efficacy” information to real
world “effectiveness” (Hughes et al. 2001).
5.2 The measurement of medication compliance
Though numerous published papers used administrative claims data to assess
medication compliance/adherence, there is no consensus on how compliance should be
measured in different scenarios, and the measurement methods are as various as study
population, disease and research designs in compliance literature.
A challenge to compliance measurement is the complexity of therapeutic regimen
in the real world that’s beyond the capacity of conventional MPR. When patients with a
certain illness can only choose and should always be on one drug, measuring compliance
using MPR would be straightforward. However, with the presence of multiple
therapeutic classes available and indicated for a certain illness such as in type 2 diabetes,
the regimen pattern over time can be exceedingly complicated by medication
combination, switching, augmentation (adding a new drug), de-augmentation
104
(withdraw a drug from combination therapy). With six diabetes classes, the possible
number of patterns over a long period time is almost innumerable. Defining a feasible
yet consistent measure of compliance across all possible treatment patterns presents a big
challenge in empirical research with pharmacy refill records.
The most frequently used definition of medication compliance is the extent to
which the person medication taking behavior “coincides with medical or health advice”
(Haynes et al. 1979). Even though the compliance definition could have various
versions, and had we observed perfectly how exactly patients take their medications, we
would still need to know about the other part of the compliance concept—the medical
“advice”, or how the physician prescribes. Without knowing the true advice that is not
recorded in a claims database, measuring compliance is not possible without additional
assumptions. For example, if patients on drug therapy discontinue one or even all drugs,
we would not be able to know if it is intended by the physician, or if it’s the patient’s own
decision (noncompliance).
As previously specified, we rely on four assumptions for the compliance defined
in this study, briefly reiterated as follows: (1) pharmacy refill claims record the actual
way patients take their medications, (2) patients should always be on hypoglycemic
drugs, (3) within a small interval (i.e. quarter), we assume the actual regimen observed is
the regimen advised by the physician for that interval, and (4) within the small interval,
compliance is constant.
Assumption (1) is a default assumption in medication compliance literature using
claims database, though rarely declared. With this assumption, we also imply patients
taking medication for prescribed days. Though patients can finish all
105
prescribed drugs within a prescribed period, erratic timing is often observed and also a
source of noncompliance (Urquhart 1997; Girard et al. 1998), and may impact on health
outcomes. With pharmacy refill records, such micro-level compliance behavior cannot
be detected. Patients should take medications as prescribed on a daily basis, and should
not cease taking medications or stop filling prescriptions after medication therapy starts.
This assumption allows repeated measurement of compliance since prescription starts till
the end of the follow-up, and considers non-persistency (discontinuation) as a type of
non-compliance. Sampling may have access to medications not recorded in the claims
database, especially before the augmentation with or switching to a different
hypoglycemic medication. The presence of sampling may underestimate the measured
compliance with existing method. On the other hand, patients do not always consume all
pills of a filled prescription, thus the compliance may also be overestimated.
Assumption (2) is also a fundamental assumption in empirical research that states
patients should always be on medication during measurement period. This assumption
implies that any calendar day that is not covered by a hypoglycemic drug will potentially
compromise one’s compliance level. Assumption (3) can be reasonable when the length
of interval is appropriate, such as quarter in this study. This is a logical extension of
conventional measurement of a single drug MPR over a certain interval (e.g. a year),
where the actual observed drug is assumed to be the advised drug for the entire interval
(e.g. a year). However, this assumption can bias the measure in both ways. For example,
if physician intents to start or stop the regimen in the middle of the interval, this
assumption would downward bias the compliance measure. On the other hand, if a
patient decides to discontinue one drug in a combination therapy for a quarter
106
or longer, we would upward bias the compliance measure. However, with the limitation
that no physician’s actual advice can be observed, this assumption allows a feasible
measure of compliance when the treatment pattern is complex such as in diabetes,
hypertension or many chronic diseases where multiple medications are available.
Assumption (4) can approximately hold true or the changing within the interval is
ignorable when the measurement interval is small (e.g. by quarter). This assumption will
be stronger in a typical and longer measurement interval (i.e. one year).
These assumptions are necessary to derive a feasible longitudinal compliance
measure for complex treatment patterns over time, as that applied by this study. It is
worth noting that this compliance measure serves well for the goal of this study in terms
of distinguishing patients of different compliance level and estimating the effect of
compliance over noncompliance. This is because this study only requires a consistent
measurement method of compliance that can be applied across all patients and able to
differentiate patients with different level of compliance. This feature can be loosely
termed as “internal validity” of compliance measure. Alternate measurement methods of
compliance may also be formulated, however, it is strongly recommended that the
underlying necessary assumptions of the compliance measures should be articulated.
All above stated four assumptions are strong given the limitation that only the
pharmacy refill records are available for us to derive a compliance measure across all
classes of drugs. One important caveat is the second assumption where prescription filled
is prescription prescribed. Switching or adding a new class can be reasonably assumed
that the prescription is advised by the physicians, however, the assumption discontinuing
one class off a combination therapy is also prescribed is an arguable
107
assumption. There are certainly two possible reasons for such discontinuation — the
patient’s own choice due to noncompliance and the physician’s acknowledgement of
partial discontinuation due to multiple reasons. Literature supports the second reason as
the dominant reason for discontinuation. For example, a study using medical chart
review found most discontinuations of lipid-lowering therapy were due to legitimate
reasons granted by physicians, including controlled cholesterol level, adverse effects,
therapeutic ineffectiveness, etc., and only 6.2% observed discontinuation were attributed
to noncompliance (Andrade et al. 1995). With only claims data, we would have treated
them equally as noncompliance for statin monotherapy. Therefore, partial
discontinuation as legitimate discontinuation appears to be a not very strong assumption
for combination therapy. This assumption tends to underestimate compliance level for
hypoglycemic monotherapy with discontinuation and overestimate compliance for
hypoglycemic combination therapy with discontinuation. In future study, the reason of
discontinuation can be accounted for with medical chart review data, thus compliance
could be more objectively measured without such assumption about reason of
discontinuation.
Though numerous attempts were made to compare compliance using different
measurement methods, the problem that different methods of measurement yield different
portion of patients with noncompliance level has long been recognized (Ley 1988;
Cameron 1996). Therefore, it may not be appropriate to compare the compliance level
reported in this study with compliance level of cohorts in studies where measurement
methods were different. The first year MPR observed in this study is 57.1%, which is
substantially lower than the mean compliance (67.5%) reported in a review
108
article that summarized 23 studies of diabetes (DiMatteo 2004). This is not surprising,
because in contrast to this study where all six classes of hypoglycemic classes are
considered, most compliance literature in diabetes only measure one or two classes such
as metformin and sulfonylureas (Venturini et al. 1999; Donnan et al. 2002; Evans et al.
2002; Winkler et al. 2002; Rosen et al. 2003), and compliance level is known to
negatively correlated with number of hypoglycemic classes (Dailey et al. 2001; Dailey et
al. 2002), which is also observed in this study.
Because the assumptions that underpin different compliance measures are likely
to be violated in the real world, both systematic and random measurement errors could
occur. However, given the objective of these study to estimate compliance effect, these
measurement errors may not be a serious concern, as long as the current measurement
method can differentiate compliance vs. noncompliance with good predictability to a
certain extent. For example, even if the measured MPR had systematically
underestimated the true compliance, we would only need to lower the cutoff point of
MPR to consistently distinguish compliance from noncompliance. As shown in the
sensitivity analysis of alternative cutoff points, the benefit of compliance cannot be
reverted by changing MPR within a certain range. However, a measurement of
compliance with poorer predictability may misclassify compliance and noncompliance.
The impact of such poorer predictability will underestimate the effect of compliance on
outcomes. This topic is certainly intriguing and calls for future research, and it is needed
to evaluate the predictability of alternative measurement methods of compliance with
claims data, and the impact of different measures on the compliance effect estimation.
109
It is also noteworthy that the compliance level measured by MPR in this study is
dichotomized to compliance vs. noncompliance. Dichotomization of compliance is not
only consistent with majority of the past literature, but such simplification is also needed
for the practical application of marginal structural model where treatment variables can
only be categorical variables.
Casual effect estimation of multiple treatment levels presents a possibility using
marginal structural model, however, the implementation is not as straightforward as
dichotomous treatment. As an analogy yet a simple model for point treatment, propensity
score method was originally developed for estimating causal effect of dichotomous
treatment over two decades ago (Rosenbaum and Rubin 1983). Not until more recently,
propensity score method was later extended to be able to estimate for multiple treatment
levels (Imbens 1999), and its application has yet rarely been seen in research.
The model and estimation methods have been so far applied to mostly a
dichotomous time-varying exposure. In cases where more than two treatment levels or a
dose-response relation, the methods can be extended to estimate a dose-response effect in
an ordinal scale (Robins et al. 2000), which may be more computationally challenging.
5.3 The compliance effect
Though this study evidenced a beneficial effect of compliance, it cannot
distinguish whether such benefit is due to the higher consumption of the drug or the
“power of belief” or “self-efficacy” as psychological causes of compliance, which also
have direct effect on outcomes not via compliance. In addition, as shown in Figure 4,
besides these upstream direct causes, the downstream other non-observable
110
compliance indicators such as compliance to behavioral modification (e.g., self-care,
exercise) will also have direct effect on outcomes. All these non-observable factors, and
other direct causes and indicators, are unlikely to be observed in reality and their effects
on outcomes are inseparable from that of observed medication compliance measurement.
Acknowledging such imperfection, all these factors are included as part of the
generalized concept of compliance, shown as the gray area of the figure. The estimated
compliance effect is the composite effect of the all these intricately related and
inseparably factors. The presence of non-therapeutic effect, of both non-observed
psychological factor and more self-care, can explain the fact in the clinical trial that
patients compliant to placebo exhibited better health outcomes (Coronary Drug Project
Research Group 1980; Beta-Blocker Heart Attack Trial Research Group 1982; Gallagher
et al. 1993).
In addition, the non-therapeutic compliance effect, which we cannot measure
directly, appears to be even stronger than therapeutic effect of medication compliance. In
UKPDS trial, intensive blood glycemic control with insulin or sulfonylureas had a 25%
risk reduction in microvascular endpoints over 10 years of follow up (UKPDS Group
1998), which is very similar to the risk reduction estimate using 80% MPR cutoff point in
this study. However, compliance with a smaller MPR cutoff point was related to much
higher risk reduction than the drug effect observed in the clinical trial, implying that the
non-therapeutic compliance effect may be higher than that of compliance effect. In a
clinical trial that evaluated lipid-lowering therapy, it was observed the compliance effect
to placebo (non-therapeutic effect) is greater than that of the drug effect (Coronary Drug
111
Project Research Group 1980), showing another evidence that non-therapeutic
component of compliance plays a major role in the compliance effect.
Furthermore, the compliance effect estimated in this study is a crude measure of
the hazard rate ratio between two buckets of patient quarters –compliance over
noncompliance, with 80% MPR as a cutoff point. Besides simplicity and feasibility,
another advantage of using dichotomous measure of compliance is that it does not need
the assumption of linear effect or a parametric dose-response effect of compliance on
health outcomes. The estimated effect can be interpreted as the ratio of the average
hazard of patient-quarters with 80% or more MPR over the average hazard of patient
quarters with less than 80% MPR (including zero).
5.4 Intuitive understanding of the MSM with IPTW estimate
The MSMs with IPTW method can estimate the casual effect in a longitudinal
study with the presence of time-varying confounding. The IPTW method has been
proven to be a consistent estimate with the assumption of sequential randomization
(ignorability) (Robins 1999; Robins 2002). That is, conditioning on observed past
treatment and time-varying covariate history, there is no unmeasured confounder at each
time point.
A two-step procedure was employed with this method in this study. First, for
each patient quarter observation, logistic GEE regression models were used to predict the
probability of receiving patient’s actual treatment (compliance or noncompliance) and the
probability of being uncensored, then these predicted probabilities were used to construct
inverse probability of treatment weights. The second step was to apply these
112
weights for each patient quarter in a standard regression framework, and as a result, the
time-varying confounding effects were removed.
Using inverse probability of treatment as weights (IPTW) is equivalently to create
a pseudo-population where the past treatment and covariate history no longer predicts the
treatment (and loss of follow-up) at each quarter. Assuming no unmeasured confounders
at each time t given past treatment and covariate history, the weighting with inverse
probability of treatment is essentially a sequential randomization process. This is
fundamental in casual inference, because randomized clinical trial, as the very gold
standard of causal inference, is to ensure no factor, observable or not, could predict
treatment assignment. Loosely speaking, the IPTW method is an extension propensity
score method of point treatment to longitudinal setting, where treatment and covariates
repeatedly measured and vary over time. For point treatment, conditional on propensity
score or using propensity score as weights, no observable factors can predict probability
of receiving treatment. However, this analog may underestimate the true merit of IPTW
method because it allows controlling intermediate outcomes variables while propensity
score only controls baseline covariates.
The estimation process of MSM sequentially randomize a patient to compliance
or noncompliance at each quarter, then a patient stays on her assigned level of
compliance for that quarter before another randomized assignment for the next quarter.
Though treatment changes over time, this process only requires patient to stay on her
assigned treatment (compliance) for that small time interval (quarterly), in contrast to the
standard regression that assumes patients stay on the assigned level of compliance for the
entire follow-up period with no time-varying confounding. While the time-
113
varying confounding effect is removed, the treatment in pseudo-population becomes
statistically exogenous (Hernan et al. 2001). Thus, the treatment effect remains the same
as that in the original population and can be estimated directly using standard regression
as exemplified in this study (Robins 1999).
5.5 Advantages and disadvantages of marginal structural models
To estimate the causal effect with the presence of time-varying confounding,
Robins and colleagues introduced two methods prior to this marginal structural model:
the semiparametric g-computation algorithm formula estimator (Robins 1986) and g-
estimation for structural nested models (SNM) (Robins et al. 1992; Robins 1998). MSM
and SNM are preferable to g-computation algorithm formula because the latter is rarely
efficient (Robins 1998). Estimating using these two models requires computational
techniques beyond the capability of conventional software, and they also require special
data type where treatment and time-varying covariates all need to be categorical. In
addition, a very large number of observations are needed and the estimates are not
efficient (Robins 1999).
A major advantage of MSM is that they resemble standard regression models, and
can be estimated by standard software package, whereas SNM cannot be conveniently
estimated in most cases.
As a disadvantage, MSM cannot estimate the interaction between treatment and
time-varying confounders, or the effect of dynamic treatment regimens given a patient’s
evolving covariate history. For example, though MSM provide us an estimate of the
overall compliance effect on outcomes over noncompliance, we do not have
114
much insight on how the time-varying confounders interact with compliance and their
dynamic impact on health outcomes. As a contrast, SNM can estimate the dynamic
interaction directly.
As a limitation, MSM cannot be used to estimate the treatment effect where under
a certain time point (quarter), all subjects receive an identical treatment a(t), where the
estimation of probability of receiving treatment is not feasible. In that case, g-estimation
of structural nested models can be used (Robins 1998).
The existence of unmeasured confounder would always be a possibility with no
exception in this study. In this study, clinical parameters such as fasting blood glucose
and HbA1c may influence patient behavior on compliance over time. Conditioning on all
measured covariates, we would hope, the confounding effect of these factors is small.
5.6 Unmeasured confounders and model specification
As previously discussed, MSM require two assumptions for correct estimation of
causal effect—absence of unmeasured confounders at each period, and correct model
specification.
If such unmeasured time-varying confounding effect, such as unobserved clinical
parameters (readings of SMBG, HbA1c, etc.) is not ignorable, it is not too difficult to
show that such omission results in a biased estimate which is more likely to
underestimate the benefit of compliance. Perceived worsening health status increases
medication compliance, as shown in this study and others (Alogna 1980; DiMatteo 2004)
and implied by conceptual compliance models (Cerkoney and Hart 1980; Kurtz 1990;
Budd et al. 1996). Similarly, patients with blood sugar well controlled are
115
more likely to maintain that with behavioral therapy and less likely to use medication
(given the disutility associated with medications). Thus, clinical parameters such as
blood glucose and HbA1c function as negative mediators between compliance and risk of
complication. Models without controlling such negative mediators would underestimate
the benefit of compliance. This is also why Model 2A and Model 2B with only adjusting
baseline covariate result in a harmful association of compliance on outcomes. As an
analog, omitting time-varying confounding effect of CD4 cell count, zidovudine would
have harmful association with HIV patient’s survival (Hernan et al. 2000). In this case,
low CD4 count is a predictor of initiation of zidovudine. Therefore, with the omission of
unmeasured time-varying clinical parameters, this study may have underestimated the
true benefit of compliance on complication-free duration among type 2 diabetes patients.
In addition to unmeasured confounders, MSM also need the assumption of correct
specification of the model form as required in all studies. This assumption includes
correct specification of the baseline covariates, time-varying covariates, models for
weight estimation and weighted standard regression. Such variable selection and
definition may influence the final causal estimate, and often a model is chosen based on
simplification and convention with the absence of knowledge of true underlying variable
relationship and data generation mechanism.
One example is the complexity of therapy change, switch, augmentation, and
discontinuation during the course of hypoglycemic treatment. The therapy changing
process over time can be modeled by time varying confounders. The goal is to isolate the
compliance effect from the effects due to therapy variations. This study modeled the
complexity of therapy variation over time using six quarterly measured
116
dichotomous hypoglycemic class variables. These six variables can be sufficient to
describe all possible combination pattern of therapy over time, nevertheless, there is an
assumption that the effect of each individual class is independent and additive. However,
there may be interaction effect between these classes. One way to improve the current
design in future research is to model the combination of multiple classes separately with
additional time-varying confounders.
5.7 Comparison of the five models
The five models presented in this study include four standard Cox regression
models and one marginal structural Cox model. Though the first four models are biased
models given the context of compliance measure, their estimates can be interpreted
causally under special scenarios. From Model 1 to Model 5 (MSM), the underlying
assumptions required for causal interpretation are progressively weaker. Five diagrams
are depicted for scenarios under which the five models can correctly estimate the causal
effect respectively.
Model 1 is the simple association without adjusting for baseline covariate using
one-year compliance. The estimate can be only causal when the compliance level is
randomly assigned in a clinical trial, and patients remain constant on the assigned
compliance level.
Model 2 adjusts baseline covariates only. The compliance is a dichotomous
measure over the first year after the index date (Model 2A), or over a variable length
during the entire follow up (Model 2B). The causal interpretation of estimated effect
only holds true under the following strong assumptions: (1) there is no
117
unmeasured confounder that selects patients into different compliance level at therapy
initiation, and (2) patients choose the compliance level in the beginning and remain
constant over time. Though it is obvious that patients do not remain constant on the
chosen compliance level, this approach is commonly applied in medical literature to
establish the relationship between compliance on outcomes (DiMatteo et al. 2002). It is a
fact that such measured compliance is associated with higher risk, as the results of Model
2 showed, such associational findings would neither interest physicians nor policy
makers.
The only difference between Model 2A and Model 2B is that the compliance in
Model 2B is measured over the period prior to the development of complication or the
censored point (disenrollment). As a merit of Model 2B, it ensures the measurement of
treatment is prior to that of the outcomes. However, with the fixed-length measure, event
(complication onset) could occur within the period of fixed length and obscure the
temporal relationship between the fixed-length compliance and the outcomes measure.
The result of this study showed that Model 2B was more severely biased than Model 2A,
and the reasons lie in the fact that length of the measurement period was identical to the
observed survival time and compliance was negatively correlated to the length of follow-
up. As a result, higher survival was associated with lower compliance measure, in which
was a biased estimate.
As shown in this study, patients did change compliance over time. Model 3
allows compliance to be modeled as a time varying variable, in addition to the covariates
measured at baseline. As shown in Figure 7.3, time varying covariates influence many
time-varying covariates (such as intermediate outcomes), however, none of
118
these variables can impact on future compliance. However, the causal interpretation of
compliance effect estimated from this model needs the assumption that compliance
changing over time is either at random or predetermined at baseline.
Model 4, in addition to allowing compliance level changing over time, also
incorporates other covariates that are measured repeatedly over time after the therapy
initiation. As shown in Figure 7.4, these time-varying covariates L(t) influence future
compliance but are assumed to be not affected by past compliance history. Therefore,
these covariates can be modeled as exogenous time-varying covariates in a standard Cox
regression. The causal interpretation requires the assumption that the treatment does not
interact with these variables over time, i.e., current treatment can only be predicted by
past treatment and covariate history and does not affect future time-varying covariates.
This is called “causally exogenous” in a formal term (Hernan et al. 2001; Robins et al.
2003). When the compliance changes interact with these time-varying covariates, the
estimates by Cox regression are biased.
Model 5 (MSM) is the model for causal estimation with the weakest assumption.
It allows complex interaction with time-varying covariates L(t) and C(t) over time, and
estimates the treatment effect in the presence of time-varying confounders. Its causal
interpretation only needs the assumption of sequential ignorability (randomization).
Given each time period (e.g. quarter), if there is no unmeasured confounding of
compliance selection given the history of measured time-varying factors including the
treatment history, the estimate from this marginal structural model can be interpreted as
causal effect.
119
5.8 Limitations and future research
It is a limitation that marginal structural model can only estimate the causal effect
of discrete treatment variable. Yet so far, the application of MSM in empirical research
has only been limited to dichotomous treatment, such as the dichotomized compliance
used in this study. Categorizing compliance level into compliance vs. noncompliance (or
good vs. poor compliance) based on a simple cutoff (e.g. 80 percent) point makes this
study feasible, and it is also a common practice in compliance study with claims data.
However, it would be interesting to further investigate the compliance effect in a dose-
response pattern. Within the MSM framework, it is possible to categorize compliance
into three or more levels. A multinomial logit model or more advanced discrete choice
model of repeated measure can be used to estimate the probability of receiving the actual
treatment for each patient time observation, with IPTW method to estimate the treatment
effect of compliance effect at different categorized levels. This is a feasible yet
challenging approach for future study.
Sensitivity analysis showed some interesting findings that estimated compliance
effect varied with the MPR cutoff points, also the variation revealed a nonlinear trend of
dose-response of compliance. As an analog of the propensity score method, marginal
structural models can be only applied to estimate casual effect of categorical treatment
variables. Future studies with finer categorization of compliance are needed to explore
the nonlinear effect of compliance on outcomes.
120
5.9 Implication to other health outcomes research
The MSMs have been applied in analyzing clinical trial or registry data to control
the bias due to time varying confounding effect, where time-varying confounders are all
clinical parameters. This study is the first application of MSMs in a widely accessible
administrative claims database. In claims databases, abundant information is available
over a long period of time. In this study, factors that influence treatment decision at
baseline change over the course of treatment and also influence subsequent decisions of
treatment, thus can be treated as time-varying confounders, such as comorbidities,
concomitant drugs, concurrent visits and resource utilization, ever use of SMBG. As
shown in this study, they were measured quarterly and significantly predicted by past
compliance history over time, and also played an indispensable role in subsequent
compliance decision.
Time to complication, the outcome in this study, is measured as a single variable.
In other studies where outcomes could be measured repeatedly, the previous outcome is a
strong time varying confounder for the final outcome. Very often, the past outcome is
the best predictor of the future outcome, because the past outcome contain the
information of otherwise unobservable factors such as patient severity. For example,
CD4 counts measured over time during the treatment course is a time varying confounder
for the final CD4 count. Controlling such previous outcome as a time varying
confounder in MSM can dramatically improve the estimate of the causal effect of
treatment (Hernan et al. 2002; Ko et al. 2003). This approach provides a promising
implication for pharmaceutical outcomes research where intermediate outcomes can be
measured repeatedly over time.
121
Although the marginal structural model was developed forestimating causal effect
in the presence of time-varying confounding, it also has been applied as a non-parametric
tool for multivariate standardization as an improved method over combining stratification
and parametric modeling in epidemiological causal inference (Sato and Matsuyama
2003).
In general, longitudinal data with repeated measure of covariates where treatment
or treatment decision is a function over time (time changing process), marginal structural
models can be applied because treatment decision will vary according to the time-varying
covariates. Administrative claims databases provide a natural source of such longitudinal
data and time-varying treatment, where MSM hold promise for application in solving
problems that are not possible with standard regression methods.
5.10 Summary
Standard Cox proportional hazard models can produce a biased estimate of
compliance effect because of the evidenced presence of time-varying confounders that
dynamically interact with decisions of compliance over time.
Assuming no unmeasured confounders conditional on past treatment and
covariate history (e.g. sequential randomization), marginal structural models in this study
demonstrated the statistically significant benefit of complication risk rate ratio as 0.7
(95% bootstrap CI: 0.60, 0.92). In contrast to past efforts of establishing “association”
relationship where time-varying confounding effect is ignored, and compliance and
outcomes are measured within the same period, this study is the first evidence of causal
benefit of medication compliance to health outcomes.
122
The study findings have a significant policy implication. Higher compliance
causally improves the health outcomes, which provides justification and potential
evaluation for compliance-promoting interventions, and serves as a basis for discounting
non-compliance in cost-effectiveness analysis.
123
BIBLIOGRAPHY
Alogna, M. (1980). "Perception of severity of disease and health locus of control in
compliant and noncompliant diabetic patients." Diabetes Care 3(4): 533-4.
Amato, L., G. Paolisso, et al. (1997). "Congestive heart failure predicts the development
of non-insulin-dependent diabetes mellitus in the elderly. The Osservatorio
Geriatrico Regione Campania Group." Diabetes & Metabolism 23(3): 213-8.
American Association of Clinical Endocrinologists (2002). "American College of
Endocrinology Consensus Statement on Guidelines for Glycemic Control."
ENDOCRINE PRACTICE 8(suppl 1): 5-11.
American Diabetes Association (2002). "Standards of medical care for patients with
diabetes mellitus." Diabetes Care 25(1): 213-29.
American Diabetes Association (2003). "Economic costs of diabetes in the US in 2002."
Diabetes Care 26(3): 917-32.
Andersen, A. R., J. S. Christiansen, et al. (1983). "Diabetic nephropathy in Type 1
(insulin-dependent) diabetes: an epidemiological study." Diabetologia 25(6): 496-
501.
Anderson, R. M. and M. M. Funnell (2000). "Compliance and adherence are
dysfunctional concepts in diabetes care." Diabetes Educator 26(4): 597-604.
Andrade, S. E., A. M. Walker, et al. (1995). "Discontinuation of antihyperlipidemic
drugs--do rates reported in clinical trials reflect rates in primary care settings?[see
comment]." New England Journal of Medicine 332(17): 1125-31.
Anon (1997). Working Party from Compliance to Concordance. Achieving Shared Goals
in Medicine Taking. London, Royal Pharmaceutical Society of Great Britain.
Anonymous (1993). "The effect of intensive treatment of diabetes on the development
and progression of long-term complications in insulin-dependent diabetes
mellitus. The Diabetes Control and Complications Trial Research Group.[see
comment]." New England Journal of Medicine 329(14): 977-86.
Anonymous (1993). "UK Prospective Diabetes Study (UKPDS). X. Urinary albumin
excretion over 3 years in diet-treated type 2, (non-insulin-dependent) diabetic
patients, and association with hypertension, hyperglycaemia and
hypertriglyceridaemia." Diabetologia 36(10): 1021-9.
Austin, P. C. and J. V. Tu (2004). "Automated variable selection methods for logistic
regression produced unstable models for predicting acute myocardial infarction
mortality." Journal of Clinical Epidemiology 57(11): 1138-1146.
124
Balkrishnan, R. (1998). "Predictors of medication adherence in the elderly." Clinical
Therapeutics 20(4): 764-71.
Balkrishnan, R., R. Rajagopalan, et al. (2003). "Predictors of medication adherence and
associated health care costs in an older population with type 2 diabetes mellitus:
A longitudinal cohort study." Clinical Therapeutics 25(11): 2958-2971.
Ballard, D. J., L. L. Humphrey, et al. (1988). "Epidemiology of persistent proteinuria in
type II diabetes mellitus. Population-based study in Rochester, Minnesota."
Diabetes 37(4): 405-12.
Becker, M. H. (1974). "The Health Belief Model and Personal Health Behavior." Health
Education Monographs 2(4).
Becker, M. H. and L. A. Maiman (1975). "Sociobehaviorial determinants of compliance
with health and medical care recommendations." Medical Care 13: 10-24.
Bell, D. S. (2003). "Heart failure: the frequent, forgotten, and often fatal complication of
diabetes." Diabetes Care 26(8): 2433-41.
Benner, J. S., R. J. Glynn, et al. (2002). "Long-term persistence in use of statin therapy in
elderly patients.[see comment]." JAMA 288(4): 455-61.
Benner, J. S., J. C. Tierce, et al. (2004). "Follow-up Lipid Tests and Physician Visits are
Associated with Improved Adherence to Statin Therapy." Pharmacoeconomics
22(Suppl. 3): 13-23.
Beta-Blocker Heart Attack Trial Research Group (1982). "A randomized trial of
propranolol in patients with acute myocardial infarction, I: mortality results."
JAMA 247: 1707-1714.
Bjork, S. (2001). "The cost of diabetes and diabetes care." Diabetes Research and Clinical
Practice 54(Supplement 1): 13-18.
Boudes, P. (1998). "Drug compliance in therapeutic trials: a review." Controlled Clinical
Trials 19(3): 257-68.
Brawley, L. R. and S. N. Culos-Reed (2000). "Studying Adherence to Therapeutic
Regimens: Overview, Theories, Recommendations." Controlled Clinical Trials
21(5, Supplement 1): S156-S163.
Brown, M. C., A. A. Nimmerrichter, et al. (1999). "Cost-effectiveness of mirtazapine
compared to amitriptyline and fluoxetine in the treatment of moderate and severe
depression in austria." European Psychiatry: the Journal of the Association of
European Psychiatrists 14(4): 230-44.
125
Brumback, B., S. Greenland, et al. (2003). "The Intensity-Score Approach to Adjusting
for Confounding." Biometrics 59(2): 274-285.
Budd, R. J., I. C. Hughes, et al. (1996). "Health beliefs and compliance with
antipsychotic medication." British Journal of Clinical Psychology 35(Pt 3): 393-7.
Cameron, C. (1996). "Patient compliance: recognition of factors involved and
suggestions for promoting compliance with therapeutic regimens." Journal of
Advanced Nursing 24(2): 244-50.
Campbell, D. T. and M. J. Russo (2001). Social Measurement. Thousand Oaks, CA, Sage
Publications.
Cerkoney, K. A. and L. K. Hart (1980). "The relationship between the health belief
model and compliance of persons with diabetes mellitus." Diabetes Care 3(5):
594-8.
Chantelau, E. and E. M. Kohner (1997). "Why some cases of retinopathy worsen when
diabetic control improves." BMJ 315: 1105-6.
Charlson, M., P. Pompie, et al. (1987). "A new method of classifying prognostic
comorbidity in longitudinal studies: Development and validation." Journal of
Chronic Disease 40: 373-383.
Chin, M. H. and L. Goldman (1996). "Correlates of major complications or death in
patients admitted to the hospital with congestive heart failure." Archives of
Internal Medicine 156(16): 1814-20.
Chobanian, A. V., G. L. Bakris, et al. (2003). "Seventh report of the Joint National
Committee on Prevention, Detection, Evaluation, and Treatment of High Blood
Pressure.[see comment]." Hypertension 42(6): 1206-52.
Claxton, A. J., J. Cramer, et al. (2001). "A systematic review of the associations between
dose regimens and medication compliance*1." Clinical Therapeutics 23(8): 1296-
1310.
Conrad, P. (1985). "The meaning of medications: another look at compliance." Social
Science & Medicine 20(1): 29-37.
Cook, T. D. and D. T. Campbell (1979). Quasi Experimentation: Design & Analysis
Issues for Field Settings. Boston, Houghton-Mifflin.
Coronary Drug Project Research Group (1980). "Influence of adherence to treatment and
response of cholesterol on mortality in the coronary drug project." New England
Journal of Medicine 303(18): 1038-41.
126
Coutinho, M., H. C. Gerstein, et al. (1999). "The relationship between glucose and
incident cardiovascular events. A metaregression analysis of published data from
20 studies of 95,783 individuals followed for 12.4 years.[see comment]." Diabetes
Care 22(2): 233-40.
Cox, D. R. (1972). "Regression Models and Life Tables (with discussion)." Journal of the
Royal Statistical Society, Ser. B 34: 187-220.
Cramer, J. A. (1995). "Relationship between medication compliance and medical
outcomes." American Journal of Health-System Pharmacy 52(14 Suppl 3): S27-9.
Cramer, J. A., R. D. Scheyer, et al. (1990). "Compliance declines between clinic visits."
Arch Intern Med 150: 1509-10.
Dailey, G., M. S. Kim, et al. (2001). "Patient compliance and persistence with
antihyperglycemic drug regimens: evaluation of a medicaid patient population
with type 2 diabetes mellitus." Clinical Therapeutics 23(8): 1311-20.
Dailey, G., M. S. Kim, et al. (2002). "Patient compliance and persistence with anti-
hyperglycemic therapy: evaluation of a population of type 2 diabetic patients."
Journal of International Medical Research 30(1): 71-9.
DeWitt, D. E. and I. B. Hirsch (2003). "Outpatient insulin therapy in type 1 and type 2
diabetes mellitus: scientific review." JAMA 289(17): 2254-64.
Deyo, R., D. Cherkin, et al. (1992). "Adapting a clinical comorbidity index for use with
ICD-9-CM administrative databases." Journal of Clinical Epidemiology 45: 613-
619.
Dezii, C. M. (2001). "Persistence with drug therapy: a practical approach using
administrative claims data." Managed Care 10(2): 42-5.
Dickerson, K., Y.-I. Min, et al. (1992). "Factors influencing publication of research
results: follow-up applications submitted to 2 institutional review boards." JAMA
267: 374-378.
DiMatteo, M. R. (2004). "Variations in patients' adherence to medical recommendations:
a quantitative review of 50 years of research.[see comment]." Medical Care 42(3):
200-9.
DiMatteo, M. R., P. J. Giordani, et al. (2002). "Patient adherence and medical treatment
outcomes: a meta-analysis." Med Care 40(9): 794-811.
Dishman, R. K. (1994). "The measurement conundrum in exercise adherence research."
Medicine & Science in Sports & Exercise 26(11): 1382-90.
127
Donnan, P. T., T. M. MacDonald, et al. (2002). "Adherence to prescribed oral
hypoglycaemic medication in a population of patients with Type 2 diabetes: a
retrospective cohort study.[see comment]." Diabetic Medicine 19(4): 279-84.
Donovan, J. L. (1995). "Patient decision making. The missing ingredient in compliance
research." International Journal of Technology Assessment in Health Care 11(3):
443-55.
Donovan, J. L. and D. R. Blake (1992). "Patient non-compliance: Deviance or reasoned
decision-making?" Social Science & Medicine 34(5): 507-513.
Dunbar-Jacob, J. and M. K. Mortimer-Stephens (2001). "Treatment adherence in chronic
disease." Journal of Clinical Epidemiology 54(Suppl 1): S57-60.
Durrleman, S. and R. Simon (1989). "Flexible regression models with cubic splines."
Statistics in Medicine 8: 551-561.
Dwyer, M. S., L. J. Melton, 3rd, et al. (1985). "Incidence of diabetic retinopathy and
blindness: a population-based study in Rochester, Minnesota." Diabetes Care
8(4): 316-22.
Dyck, J. B. and P. J. Dyck (1999). Diabetic polyneuropathy. Diabetic neuropathy. P. J.
Dyck and P. K. Thomas. Philadelphia, WB Saunders: 262-263.
Easterbrook, P. J., J. A. Berlin, et al. (1991). "Publication bias in clinical research."
Lancet 337(867-872).
Ellis, S., S. Shumaker, et al. (2000). "Adherence to Pharmacological Interventions:
Current Trends and Future Directions." Controlled Clinical Trials 21(5,
Supplement 1): S218-S225.
Evangelista, L. S. (1999). "Compliance: a concept analysis." Nursing Forum 34(1): 5-11.
Evans, J. M. M., P. T. Donnan, et al. (2002). "Adherence to oral hypoglycaemic agents
prior to insulin therapy in Type 2 diabetes." Diabet Med 19(8): 685-688.
Expert Committee on the, D. and M. Classification of Diabetes (2003). "Report of the
expert committee on the diagnosis and classification of diabetes mellitus."
Diabetes Care 26(Suppl 1): S5-20.
Farmer, K. C. (1999). "Methods for measuring and monitoring medication regimen
adherence in clinical trials and clinical practice." Clinical Therapeutics 21(6):
1074-90; discussion 1073.
Fawcett, J. (1995). "Compliance: definitions and key issues." Journal of Clinical
Psychiatry 56(Suppl 1): 4-8; discussion 9-10.
128
Feinstein, A. R. (1990). "On white-coat effects and the electronic monitoring of
compliance." Archives of Internal Medicine 150: 1377-8.
Fong, D. S., L. Aiello, et al. (2003). "Diabetic retinopathy." Diabetes Care 26(1): 226-9.
Funnell, M. M. and R. M. Anderson (2000). "The problem with compliance in diabetes."
JAMA 284(13): 1709.
Gaede, P., P. Vedel, et al. (2003). "Multifactorial intervention and cardiovascular disease
in patients with type 2 diabetes.[see comment]." New England Journal of
Medicine 348(5): 383-93.
Gallagher, E. J., C. M. Viscoli, et al. (1993). "The relationship of treatment adherence to
the risk of death after myocardial infarction in women." JAMA 270: 742-744.
Girard, P., T. F. Blaschke, et al. (1998). "A Markov mixed effect regression model for
drug compliance." Stat Med 17(20): 2313-33.
Goldberg, R. B., M. J. Mellies, et al. (1998). "Cardiovascular events and their reduction
with pravastatin in diabetic and glucose-intolerant myocardial infarction survivors
with average cholesterol levels: subgroup analyses in the cholesterol and recurrent
events (CARE) trial. The Care Investigators." Circulation 98(23): 2513-9.
Goldsmith, C. H. (1979). The effect of compliance distributions on therapeutic trials.
Compliance in Health Care. B. R. Haynes, D. W. Taylor and D. L. Sackett.
Baltimore, Johns Hopkins University Press: 63-77.
Granstrom, P. (1985). "Progression of visual field defects in glaucoma." Archives of
Ophthalmology 103: 529-531.
Gray, R., T. Wykes, et al. (2002). "From compliance to concordance: a review of the
literature on interventions to enhance compliance with antipsychotic medication."
J Psychiatr Ment Health Nurs 9(3): 277-284.
Green, A., K. Borch-Johnsen, et al. (1985). "Relative mortality of type 1 (insulin-
dependent) diabetes in Denmark: 1933-1981." Diabetologia 28(6): 339-42.
Greenland, S. (1989). "Modeling and variable selection in epidemiologic analysis."
American Journal of Public Health 79(3): 340-9.
Greenland, S. (2000). "Causal analysis in the health Sciences." Journal of the American
Statistical Association 95(449): 286-289.
Greenland, S., J. Pearl, et al. (1999). "Causal diagrams for epidemiologic research."
Epidemiology 10(37-48).
129
Guillausseau, P. J. (2003). "Influence of oral antidiabetic drugs compliance on metabolic
control in type 2 diabetes. A survey in general practice." Diabetes Metab 29(1):
79-81.
Harley, C. R., A. Mishra, et al. (2002). "A 'Real World' Analysis of Persistence on and
Adherence to Glipizide GITS, Glipizide IR and Glibenclamide in Managed Care
Among Patients With Type 2 Diabetes Mellitus." Clinical Drug Investigation
22(9): 575-584.
Haynes, B. R., D. L. Sackett, et al. (1976). "Improvement of medication compliance in
uncontrolled hypertension." Lancet 1: 1265-1268.
Haynes, B. R., W. R. Taylor, et al. (1979). Compliance in Health Care. Baltimore, MD,
Johns Hopkins Univ. Press.
Hayward, R. A., W. G. Manning, et al. (1997). "Starting insulin therapy in patients with
type-2 diabetes." JAMA 278: 1663-1669.
Hernan, M. A. (2005). "Invited Commentary: Hypothetical Interventions to Define
Causal Effects--Afterthought or Prerequisite?" American Journal of Epidemiology
162(7): 618-620.
Hernan, M. A., B. Brumback, et al. (2000). "Marginal Structural Models to Estimate the
Causal Effect of Zidovudine on the Survival of HIV-Positive Men."
Epidemiology 11(5): 561-570.
Hernan, M. A., B. Brumback, et al. (2001). "Marginal structural models to estimate the
joint causal effect of nonrandomized treatments." Journal of the American
Statistical Association 96(454): 440-448.
Hernan, M. A., B. A. Brumback, et al. (2002). "Estimating the causal effect of zidovudine
on CD4 count with a marginal structural model for repeated measures." Stat Med
21(12): 1689-709.
Hernan, M. A., S. R. Cole, et al. (2005). "Structural accelerated failure time models for
survival analysis in studies with time-varying treatments."
Pharmacoepidemiology and Drug Safety: In Press.
Hess, J. D. (1996). "The ethics of compliance: a dialectic." Advances in Nursing Science
19(1): 18-27.
Hillier, T. A. and K. L. Pedula (2003). "Complications in Young Adults With Early-
Onset Type 2 Diabetes: Losing the relative protection of youth." Diabetes Care
26(11): 2999-3005.
Holland, P. (1986). "Statistics and Causal Inference." Journal of the American Statistical
Association 81(396): 945-60.
130
Hughes, D. A., A. Bagust, et al. (2001). "The impact of non-compliance on the cost-
effectiveness of pharmaceuticals: a review of the literature." Health Economics
10(7): 601-15.
Hussey, L. C. and K. Gilliland (1989). "Compliance, low literacy, and locus of control."
Nursing Clinics of North America 24(3): 605-11.
Ickovics, J. R. and C. S. Meade (2002). "Adherence to HAART among patients with
HIV: breakthroughs and barriers." AIDS Care 14(3): 309-18.
Imbens, G. (1999). "The Role of Propensity Score In Estimating Dose-Response
Functions." NBER No. 237.
Inzucchi, S. E. (2002). "Oral antihyperglycemic therapy for type 2 diabetes: scientific
review.[see comment]." JAMA 287(3): 360-72.
Jeffcoate, W. J. and K. G. Harding (2003). "Diabetic foot ulcers." Lancet 361(9368):
1545-51.
Jorekog, K. G. and A. S. Goldberger (1975). "Estimation of a model with multiple
indicators and multiple causes of a single latent variable." Journal of the
American Satistical Assocation 70: 631-639.
Julien, J. (1997). "Cardiac complications in non-insulin-dependent diabetes mellitus."
Journal of Diabetes & its Complications 11(2): 123-30.
Kastrissios, H. and T. F. Blaschke (1997). "Medication compliance as a feature in drug
development." Annu Rev Pharmacol Toxicol 37: 451-75.
Khaw, K.-T., N. Wareham, et al. (2001). "Glycated haemoglobin, diabetes, and mortality
in men in Norfolk cohort of European Prospective Investigation of Cancer and
Nutrition (EPIC-Norfolk)." BMJ 322(7277): 15-.
Ko, H., J. W. Hogan, et al. (2003). "Estimating Causal Treatment Effects from
Longitudinal HIV Natural History Studies Using Marginal Structural Models."
Biometrics 59(1): 152-162.
Kruse, W. and E. Weber (1990). "Dynamics of drug regimen compliance--its assessment
by microprocessor-based monitoring." European Journal of Clinical
Pharmacology 38(6): 561-5.
Kuo, Y.-F., M. A. Raji, et al. (2003). "Inconsistent Use of Diabetes Medications,
Diabetes Complications, and Mortality in Older Mexican Americans Over a 7-
Year Period: Data from the Hispanic Established Population for the
Epidemiologic Study of the Elderly." Diabetes Care 26(11): 3054-3060.
131
Kurtz, S. M. (1990). "Adherence to diabetes regimens: empirical status and clinical
applications." Diabetes Educator 16(1): 50-9.
Kyngas, H. (2000). "Compliance of adolescents with diabetes." Journal of Pediatric
Nursing 15(4): 260-7.
Kyngas, H., M. E. Duffy, et al. (2000). "Conceptual analysis of compliance." Journal of
Clinical Nursing 9(1): 5-12.
Lambert, B. L., R. L. Street, et al. (1997). "Provider-patient communication, patient-
centered care, and the mangle of practice." Health Communication 9: 27-43.
Larsen, J., M. Andersen, et al. (2002). "High persistence of statin use in a Danish
population: compliance study 1993-1998." British Journal of Clinical
Pharmacology 53(4): 375-8.
Lee, J. Y., J. W. Kusek, et al. (1996). "Assessing medication adherence by pill count and
electronic monitoring in the African American Study of Kidney Disease and
hypertension (AASK) pilot study*1." American Journal of Hypertension 9(8):
719-725.
Lewis, E. J., L. G. Hunsicker, et al. (1993). "The effect of angiotensin-converting-enzyme
inhibition on diabetic nephropathy. The Collaborative Study Group.[see
comment][erratum appears in N Engl J Med 1993 Jan 13;330(2):152]." New
England Journal of Medicine 329(20): 1456-62.
Lewis, F. M., D. E. Morisky, et al. (1978). "A test of the construct validity of health locus
of control: effects on self-reported compliance for hypertensive patients." Health
Education Monographs 6(2): 138-48.
Ley, P. (1988). Communicating with Patients: Improving Communication, Satisfaction
and Compliance. London, Croom Helm.
Liang, K. Y. and S. L. Zeger (1986). "Longitudinal data analysis using generalized linear
models." Biometrika 73: 13-22.
Liebl, A. (2002). "Challenges in optimal metabolic control of diabetes."
Diabetes/Metabolism Research Reviews 18(Suppl 3): S36-41.
Lutfey, K. E. and W. J. Wishner (1999). "Beyond "compliance" is "adherence".
Improving the prospect of diabetes care.[see comment]." Diabetes Care 22(4):
635-9.
Lynn, J. and D. DeGrazia (1991). "An outcomes model of medical decision making."
Theoretical Medical Bioethics 12: 325-343.
132
Manski, C. F. (1995). Identification Problems in the Social Sciences. Cambridge, MA,
Havard University Press.
McDermott, M. M., B. Schmitt, et al. (1997). "Impact of medication nonadherence on
coronary heart disease outcomes. A critical review.[see comment]." Archives of
Internal Medicine 157(17): 1921-9.
McDonald, H. P., A. X. Garg, et al. (2002). "Interventions to enhance patient adherence
to medication prescriptions: scientific review.[see comment][erratum appears in
JAMA. 2003 Jun 25;289(4):3242]." JAMA 288(22): 2868-79.
McNabb, W. L. (1997). "Adherence in Diabetes: Can We Define It and Can We Measure
It?" Diabetes Care 20(2): 215-218.
Mengden, T., B. Binswanger, et al. (1993). "The use of self-measured blood pressure
determinations in assessing dynamics of drug compliance in a study with
amlodipine once a day, morning versus evening." Journal of Hypertension 11(12):
1403-11.
Miller, C. D., L. S. Phillips, et al. (2001). "Hypoglycemia in patients with type 2
diabetes." Archives of Internal Medicine 161(1653-1659).
Miller, N. H. (1997). "Compliance with treatment regimens in chronic asymptomatic
diseases." American Journal of Medicine 102(2A): 43-9.
Mokdad, A. H., B. A. Bowman, et al. (2001). "The continuing epidemics of obesity and
diabetes in the United States." JAMA 286(10): 1195-200.
Mokdad, A. H., E. S. Ford, et al. (2003). "Prevalence of obesity, diabetes, and obesity-
related health risk factors, 2001." JAMA 289(1): 76-9.
Montgomery, S. B., J. G. Joseph, et al. (1989). "The Health Belief Model in
understanding compliance with preventive recommendations for AIDS: how
useful?" AIDS Education & Prevention 1(4): 303-23.
Morris, L. S. and R. M. Schulz (1992). "Patient compliance--an overview." Journal of
Clinical Pharmacy & Therapeutics 17(5): 283-95.
Morris, L. S. and R. M. Schulz (1993). "Medication compliance: the patient's
perspective." Clinical Therapeutics 15(3): 593-606.
Motheral, B., J. Brooks, et al. (2003). "A Checklist for Retrospective Database
Studies;Report of the ISPOR Task Force on Retrospective Databases." Value in
Health 6(2): 90-97.
133
Nathan, D. M. (2002). "Clinical practice. Initial management of glycemia in type 2
diabetes mellitus.[see comment]." New England Journal of Medicine 347(17):
1342-9.
Nathan, D. M., J. Meigs, et al. (1997). "The epidemiology of cardiovascular disease in
type 2 diabetes mellitus: how sweet it is ... or is it?" Lancet 350(Suppl 1): SI4-9.
Nichol, M. B., F. Venturini, et al. (1999). "A critical evaluation of the methodology of the
literature on medication compliance." Annals of Pharmacotherapy 33(5): 531-40.
O'Brien, J. A., I. Caro, et al. (2001). "Diabetes in Canada: direct medical costs of major
macrovascular complications." Value in Health 4(3): 258-65.
Ohkubo, Y., H. Kishikawa, et al. (1995). "Intensive insulin therapy prevents the
progression of diabetic microvascular complications in Japanese patients with
non-insulin-dependent diabetes mellitus: a randomized prospective 6-year
study.[see comment]." Diabetes Research & Clinical Practice 28(2): 103-17.
Oldridge, N. B. and D. L. Streiner (1990). "The health belief model: predicting
compliance and dropout in cardiac rehabilitation." Medicine & Science in Sports
& Exercise 22(5): 678-83.
Paes, A. H., A. Bakker, et al. (1998). "Measurement of patient compliance." Pharm
World Sci 20(2): 73-7.
Pearl, J. (1995). "Causal diagrams for experimental research." Biometrika 82: 669-710.
Pedersen, O. and P. Gaede (2003). "Intensified multifactorial intervention and
cardiovascular outcome in type 2 diabetes: the Steno-2 study." Metabolism:
Clinical & Experimental 52(8 Suppl 1): 19-23.
Perkins, B. A. and V. Bril (2003). "Diabetic neuropathy: a review emphasizing diagnostic
methods." Clinical Neurophysiology 114(7): 1167-75.
Podar, T. and J. Tuomilehto (2002). "The role of angiotensin converting enzyme
inhibitors and angiotensin II receptor antagonists in the management of diabetic
complications." Drugs 62(14): 2007-12.
Pyorala, K., T. R. Pedersen, et al. (1997). "Cholesterol lowering with simvastatin
improves prognosis of diabetic patients with coronary heart disease. A subgroup
analysis of the Scandinavian Simvastatin Survival Study (4S)[see
comment][erratum appears in Diabetes Care 1997 Jun;20(6):1048]." Diabetes
Care 20(4): 614-20.
Raynor, D. K. (1992). "Patient compliance: the pharmacist's role." International Journal
of Pharmacy Practice 1: 126-135.
134
Reiber, G. E., B. A. Lipsky, et al. (1998). "The burden of diabetic foot ulcers." American
Journal of Surgery 176(2A Suppl): 5S-10S.
Reis, S. E., R. Holubkov, et al. (1997). "Treatment of patients admitted to the hospital
with congestive heart failure: specialty-related disparities in practice patterns and
outcomes." Journal of the American College of Cardiology 30(3): 733-8.
Resnick, H. E., K. B. Stansberry, et al. (2002). "Diabetes, peripheral neuropathy, and old
age disability." Muscle & Nerve 25(1): 43-50.
Rippin, J. D., A. H. Barnett, et al. (2004). "Cost-effective strategies in the prevention of
diabetic nephropathy." Pharmacoeconomics 22(1): 9-28.
Robins, J. M. (1986). "A new approach to causal inference in mortality studies with
sustained exposure periods: application to control of healthy worker survivor
effect." Mathematical Modelling 7: 1393-1512.
Robins, J. M. (1989). "The control of confounding by intermediate variables." Statistics
in Medicine 8: 679-701.
Robins, J. M. (1994). "Correcting for non-compliance in randomized trials using
structural nested mean models." communications in Statistics 23: 2379-412.
Robins, J. M. (1997). Causal inference from complex longitudinal data. Latent Variable
Modeling and Applications to Causality: Lecture Notes in Statistics 120. M.
Berkane. New York, Springer-Verlag: 69-117.
Robins, J. M. (1998). "Correction for Non-Compliance in Equivalence Trials." Statistics
in Medicine 17: 269-302.
Robins, J. M. (1998). Marginal structural models. 1997 Proceedings of the Section on
Bayesian Statistical Science. Alexandria, VA, American Statistical Association:
1-10.
Robins, J. M. (1998). Structural nested failure time models. Survival Analysis. P. K.
Andersen and N. Keiding. Chichester, UK, John Wiley and Sons: 4372-4389.
Robins, J. M. (1999). Marginal structural models versus structural nested models as tools
for causual inference. Statistical Moels in Epidemiology: The Environment and
Clinical Trials. E. Halloran and D. Berry. New York, Springer-Verlag: 95-134.
Robins, J. M. (2002). "Commentary on "Using inverse weighting and predictive
inference to estimate the effects of time-varying treatments on the discrete-time
hazard"." Statistics in Medicine 21: 1663-1680.
135
Robins, J. M., D. Blevins, et al. (1992). "G-estimation of the effect of prophylaxis
therapy for pneumocystis carinii pneumonia on the survival of AIDS patients."
Epidemiology 3: 319-336.
Robins, J. M., S. Greenland, et al. (1999). "Estimation of the Causal Effect of a Time-
Varying Exposure on the Marginal Means of a Repeated Binary Outcome."
Journal of the American Statistical Association 94: 687-712.
Robins, J. M., M. A. Hernan, et al. (2000). "Marginal structural models and causal
inference in epidemiology." Epidemiology 11(5): 550-60.
Robins, J. M., M. A. Hernan, et al. (2003). Effects of Multiple Interventions.
Comparative Quantification of Health Risks. M. Ezzati, A. D. Lopez, A. Rodgers
and C. J. L. Murray. Geneva, World Health Organization: 2199-2238.
Rosen, M. I., J. E. Beauvais, et al. (2003). "Neuropsychological correlates of suboptimal
adherence to metformin." Journal of Behavioral Medicine 26(4): 349-60.
Rosenbaum, P. and D. Rubin (1983). "The Central Role of the Propensity Score in
Observational Studies For Causal Effects." Biometrika 70(1): 41-55.
Rosenstock, I. M. (1985). "Understanding and enhancing patient compliance with
diabetic regimens." Diabetes Care 8(6): 610-6.
Roter, D. L., J. A. Hall, et al. (1998). "Effectiveness of interventions to improve patient
compliance: a meta-analysis." Med Care 36(8): 1138-61.
Rubin, D. B. (1974). "Estimating causal effects of treatments in randomized and
nonrandomized studies." Journal of Educational Psychology 66: 688-701.
Rubins, H. B., S. J. Robins, et al. (1999). "Gemfibrozil for the secondary prevention of
coronary heart disease in men with low levels of high-density lipoprotein
cholesterol. Veterans Affairs High-Density Lipoprotein Cholesterol Intervention
Trial Study Group." New England Journal of Medicine 341(6): 410-8.
Rudd, P. (1993). "Partial compliance: implications for clinical practice." Journal of
Cardiovascular Pharmacology 22(Suppl A): S1-5.
Sacks, F. M., M. A. Pfeffer, et al. (1996). "The effect of pravastatin on coronary events
after myocardial infarction in patients with average cholesterol levels. Cholesterol
and Recurrent Events Trial investigators.[see comment]." New England Journal of
Medicine 335(14): 1001-9.
Sato, T. and Y. Matsuyama (2003). "Marginal structural models as a tool for
standardization." Epidemiology 14(6): 680-6.
136
Schectman, J. M., M. M. Nadkarni, et al. (2002). "The Association Between Diabetes
Metabolic Control and Drug Adherence in an Indigent Population." Diabetes Care
25(6): 1015-1021.
Simmons, Z. and E. L. Feldman (2002). "Update on diabetic neuropathy." Current
Opinion in Neurology 15(5): 595-603.
Spire, B., S. Duran, et al. (2002). "Adherence to highly active antiretroviral therapies
(HAART) in HIV-infected patients: from a predictive to a dynamic approach."
Social Science & Medicine 54(10): 1481-96.
Steiner, J. F. and M. A. Earnest (2000). "The Language of Medication-Taking." Ann
Intern Med 132(11): 926-930.
Steiner, J. F. and A. V. Prochazka (1997). "The assessment of refill compliance using
pharmacy records: methods, validity, and applications." Journal of Clinical
Epidemiology 50(1): 105-16.
Stephenson, B. J., B. R. Rowe, et al. (1993). "Is this patient taking the treatment as
prescribed?" JAMA(269): 2779-81.
Stratton, I. M., A. I. Adler, et al. (2000). "Association of glycaemia with macrovascular
and microvascular complications of type 2 diabetes (UKPDS 35): prospective
observational study.[see comment]." BMJ 321(7258): 405-12.
Syrjala, A. M., M. C. Niskanen, et al. (2002). "The theory of reasoned action in
describing tooth brushing, dental caries and diabetes adherence among diabetic
patients." Journal of Clinical Periodontology 29(5): 427-32.
The Diabetes Control and Complications Trial Research Group (1998). "Early worsening
of diabetic retinopathy in the Diabetes Control and Complications Trial.[see
comment][erratum appears in Arch Ophthalmol 1998 Nov;116(11):1469]." Arch
Ophthalmol 116: 874-86.
Trostle, J. A. (1988). "Medical compliance as an ideology." Social Science & Medicine
27(12): 1299-308.
Trostle, J. A., W. A. Hauser, et al. (1983). "The logic of noncompliance: management of
epilepsy from the patient's point of view." Cultural and Medical Psychiatry 7: 35-
56.
Tsasis, P. (2001). "Adherence assessment to highly active antiretroviral therapy." AIDS
Patient Care & Stds 15(3): 109-15.
UKPDS Group (1994). "UK Prospective Diabetes Study (UKPDS). XI: Biochemical risk
factors in type 2 diabetic patients at diagnosis compared with age-matched normal
subjects." Diabetic Medicine 11(6): 534-44.
137
UKPDS Group (1998). "Effect of intensive blood-glucose control with metformin on
complications in overweight patients with type 2 diabetes (UKPDS 34). UK
Prospective Diabetes Study (UKPDS) Group.[see comment][erratum appears in
Lancet 1998 Nov 7;352(9139):1557]." Lancet 352(9131): 854-65.
UKPDS Group (1998). "Intensive blood-glucose control with sulphonylureas or insulin
compared with conventional treatment and risk of complications in patients with
type 2 diabetes (UKPDS 33). UK Prospective Diabetes Study (UKPDS)
Group.[see comment][erratum appears in Lancet 1999 Aug 14;354(9178):602]."
Lancet 352(9131): 837-53.
UKPDS Group (1998). "Tight blood pressure control and risk of macrovascular and
microvascular complications in type 2 diabetes: UKPDS 38. UK Prospective
Diabetes Study Group.[see comment][erratum appears in BMJ 1999 Jan
2;318(7175):29]." BMJ 317(7160): 703-13.
UKPDS Group (1998). "UKPDS 28: a randomized trial of efficacy of early addition of
metformin in sulfonylurea-treated type 2 diabetes. U.K. Prospective Diabetes
Study Group." Diabetes Care 21(1): 87-92.
UKPDS Group (1999). "Quality of life in type 2 diabetic patients is affected by
complications but not by intensive policies to improve blood glucose or blood
pressure control (UKPDS 37). U.K. Prospective Diabetes Study Group." Diabetes
Care 22(7): 1125-36.
Urquhart, J. (1997). "The electronic medication event monitor -- lessons for
pharmacotherapy." Clinical Pharmacokinetics 32: 345-356.
Urquhart, J. (1998). "Pharmacodynamics of variable patient compliance: implications for
pharmaceutical value." Advanced Drug Delivery Reviews 33(3): 207-219.
Venturini, F., M. B. Nichol, et al. (1999). "Compliance with sulfonylureas in a health
maintenance organization: a pharmacy record-based study." Annals of
Pharmacotherapy 33(3): 281-8.
Vermeire, E., H. Hearnshaw, et al. (2001). "Patient adherence to treatment: three decades
of research. A comprehensive review." Journal of Clinical Pharmacy &
Therapeutics 26(5): 331-42.
Vitolins, M. Z., C. S. Rand, et al. (2000). "Measuring Adherence to Behavioral and
Medical Interventions." Controlled Clinical Trials 21(5, Supplement 1): S188-
S194.
Weintraub, M. (1976). Intelligent non-compliance and capricious compliance. Patient
Compliance. L. L. Mt Kisco, N.Y., Futura: 39-47.
138
Wenerowicz, W. J., J. H. Riskind, et al. (1978). "Locus of control and degree of
compliance in hemodialysis patients." Journal of Dialysis 2(5-6): 495-505.
Williams, A. (1997). "Antiretroviral therapy: factors associated with adherence." Journal
of the Association of Nurses in AIDS Care 8(Suppl): 18-23.
Williams, G. H. (1998). "Assessing patient wellness: new perspectives on quality of life
and compliance." American Journal of Hypertension 11(11 Pt 2): 186S-191S.
Winkler, A., A. U. Teuscher, et al. (2002). "Monotoring adherence to prescribed
medication in type 2 diabetic patients treated with sulfonylureas." Swiss Medical
Weekly 132(27-28): 379-85.
Wolf, G. and E. Ritz (2003). "Diabetic nephropathy in type 2 diabetes prevention and
patient management." Journal of the American Society of Nephrology 14(5):
1396-405.
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Effects of a formulary expansion of the use of SSRIs and health care services by depressed patients in the California Medicaid program
PDF
The construct validity of the Health Utilities Index in patients with chronic respiratory disease in a managed care population
PDF
Compliance study of second-generation antipsychotics on patients with schizophrenia
PDF
Time dependent survival analysis of Kaiser Permanente/USC pharmacists' consultation intervention study
PDF
Controlling for biases from measurement errors in health outcomes research: A structural equation modeling approach
PDF
Assessing the cost implications of combined pharmacotherapy in the long term management of asthma: Theory and application of methods to control selection bias
PDF
Effect of vehicles and penetration enhancers on the percutaneous absorption of methotrexate and analogs
PDF
A new paradigm to evaluate quality-adjusted life years (QALY) from secondary database: Transforming health status instrument scores to health preference
PDF
A methodology to identify high -risk patients with diabetes in the California Medicaid populations (Medi -Cal)
PDF
Physician adherence to national hypertension guidelines in an elderly Medicaid population
PDF
Evidence of age-based rationing of health care to the elderly: The effect of age and death on the use of health care services by elderly enrollees in an HMO
PDF
Immunostimulating polysaccharides from Chinese medicinal herbs: Panax notoginseng and Viola yedoensis
PDF
Enhancement of transferrin receptor-mediated transcytosis: a novel approach for macromolecular drug delivery
PDF
Stereoselective metabolism and pharmacokinetics of ifosfamide in the rat.
PDF
The structure-function relationship of the human intestinal di-peptide transporter
PDF
Two essays on linear regression models: Solving shortest length least-squares problems and the determination of birth outcome
PDF
The effect of formative processes of interorganizational relationships on subsequent interaction
PDF
Essays on host country effects of capital flows: Volatility and spillovers
PDF
Ion and drug transport across cultured rabbit conjunctival epithelial barrier
PDF
Hepatitis C in the post-interferon era: selected essays in health economics
Asset Metadata
Creator
Yu, Andrew Peng
(author)
Core Title
Causal estimation of the effect of medication compliance on health outcomes
Degree
Doctor of Philosophy
Degree Program
Pharmaceutical Economics and Policy
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
health sciences, public health,OAI-PMH Harvest
Language
English
Contributor
Digitized by ProQuest
(provenance)
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-c17-111861
Unique identifier
UC11349307
Identifier
3237184.pdf (filename),usctheses-c17-111861 (legacy record id)
Legacy Identifier
3237184.pdf
Dmrecord
111861
Document Type
Dissertation
Rights
Yu, Andrew Peng
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 au...
Repository Name
University of Southern California Digital Library
Repository Location
USC Digital Library, University of Southern California, University Park Campus, Los Angeles, California 90089, USA
Tags
health sciences, public health