Table of Contents
The CDC's Diabetes Systems Modeling Project:
Developing a New Tool for Chronic Disease Prevention and Control
Jack Homer
jack@homerconsulting.com, Homer Consulting
3618 Avalon Court, Voorhees, NJ 08043, USA
Andrew Jones
apjones@sustainer.org, Sustainability Institute
8 Lynmar Ave., Asheville, NC 28804, USA
Don Seville
dseville@sustainer.org, Sustainability Institute
3 Linden Road, Hartland, VT 05048, USA
Joyce Essien
jessien@sph.emory.edu, Rollins School of Public Health, Emory University
1518 Clifton Road NE, Atlanta, GA 30322, USA
Bobby Milstein
bmilstein@cdc.gov, Division of Adult and Community Health,
Centers for Disease Control and Prevention
4770 Buford Hwy, Atlanta, GA 30341, USA
Dara Murphy
dilm1@cdc.gov, Division of Diabetes Translation,
Centers for Disease Control and Prevention
4770 Buford Hwy, Atlanta, GA 30341, USA
The analytic tools presently used for planning and evaluating health policy are not well suited to
capture the dynamic complexity of chronic diseases, which involve long delays between causes
and health consequences. To explore the potential of system dynamics for addressing chronic
disease policy, the U.S. Centers for Disease Control and Prevention (CDC) commissioned a
dynamic model of diabetes prevalence and complications that can be used for designing and
evaluating intervention strategies. The rapid growth in diabetes in the U.S. and elsewhere has
been driven by a similarly rapid growth in obesity, which is a leading risk factor. The model was
developed through a series of meetings over several months with participants from the CDC
Division of Diabetes Translation and other Divisions with relevant expertise. The CDC and
other public health stakeholders may use the model for policy analysis at the national, state, and
local levels.
The Public Health Challenge of Chronic Disease
Chronic disease is a growing problem throughout the world, due in large part to increased
longevity and the aging of the population—one-third of the chronically ill being over 65—but
also due to numerous environmental factors. In the U.S., chronic diseases cause more illness,
disability, and death than acute diseases and accidents combined, and are responsible for at least
70% of all health care expenditures (Hoffman et al. 1996, Institute for Health and Aging 1996).
More than half of the adult population in the U.S. has at least one chronic disease, and about half
of those have more than one such affliction and are responsible for the great majority of the total
cost of chronic disease (Institute of Medicine 2001).
Most health planners who work on chronic diseases such as diabetes acknowledge the need for
ecological, system-wide interventions. However, their desire to achieve systemic change stands
in opposition to what most program staff and researchers are prepared to do. Ingrained in
conventional analytic methods and planning frameworks is the idea that each aspect of a
complicated disease control strategy can be managed and evaluated separately. Guided by these
tools, stakeholders usually act independently even though they depend on one another to achieve
comprehensive health goals for the public. In addition, many health problems, particularly
chronic diseases, involve long delays between causes and the health events that individuals
experience. Diabetes is a good example of this type of problem: it defies quick fixes and
demands highly-coordinated, system-wide interventions.
The analytic tools presently used for planning and evaluating health policy—primarily,
diagrammatic logic models and epidemiological forecasting models—are not well suited to
capture the dynamic complexity of chronic diseases. This analytic gap may be overcome by
developing system dynamics simulation models. A system dynamics model can focus on a
specific disease process, while incorporating a deeper understanding of how the surrounding
health system functions (physically, socially, and organizationally). It can also formalize the
collaborative nature of public health practice through group model building, while strengthening
the scientific foundation and active policy focus that are the hallmarks of health protection
(Homer and Milstein 2003, Homer et al. 2004).
Public health officials are frustrated with approaches that are narrow and focus on just one part
of the overall health system—say disease management, or detection, or risk factor reduction—
rather than addressing all major parts in their entirety as a system. Innovative leaders are
searching actively for an integrative, system-wide perspective that avoids this “silo” problem. In
this context, public health agencies throughout the U.S. have been encouraged to assess both
their internal and collective capacity to deliver the full spectrum of essential services related to
disease prevention and control (CDC 2003). As these assessments expose both strengths and
gaps in service capacity, questions of priority and strategic direction come to the forefront.
Integrative models could help the CDC and other stakeholders to establish priorities in a rational
way that effectively promotes health in the short- and long-term.
Growth of Diabetes in the United States and its Causes
Diabetes mellitus is a complex metabolic disorder marked by abnormally high blood glucose
levels. If left untreated, the complications of diabetes can be disabling and ultimately fatal.
Diabetes affects at least 18 million people in the U.S., a number that has been growing more
rapidly than the general population since 1990. The rapid growth has occurred among those who
have the non-insulin dependent Type 2 variety of the disease (formerly known as adult onset
diabetes), as opposed to among the one million or so who have insulin-dependent Type 1
diabetes (which almost always strikes in childhood). Total costs of diabetes in the U.S. in 2002
were estimated to be $132 billion, with $92 billion of that in direct medical expenditures and the
other $40 billion in indirect costs due to disability and premature mortality (NIDDK 2004a).
Figure | presents a graph for 1980-2000 of the fraction of the adult population (age 18 and over)
already diagnosed with Type 1 or Type 2 diabetes (CDC 2004c). The decade-plus growth trend
in prevalence, and its apparent acceleration starting at the end of the 1990s—which by all
accounts has continued into the 2000s—has led many observers to conclude that the U.S. is
experiencing a diabetes epidemic (Gorman 2003). This perception is supported by a recent
Markov modeling study that has forecasted a 120% increase in the fractional prevalence of
diagnosed diabetes during 2000 to 2050, assuming no change in rates of diabetes incidence by
age, sex, race, and ethnicity (Honeycutt et al. 2003).
7%
Diagnosed diabetes % of adults
6% + (NHIS, adjusted pre-1997)
5%
4%
3%
2%
1%
0%
1980 1985 1990 1995 2000
Figure 1. Historical trend of diagnosed diabetes in the U.S.
Concern about growing diabetes prevalence is also based on observed growth in the prevalence
of obesity, a leading risk factor for diabetes, heart disease, and several other chronic diseases
(NIDDK 2004b). If one considers that it usually takes many years for someone at high risk for
diabetes to actually develop the disease, with incidence rates being the highest for the middle
aged and the elderly (Narayan et al. 2003), then growth in obesity now may be seen as
portending continued growth in diabetes in the future. Figure 2 shows how the obese fraction of
the adult U.S. population, which grew only slightly from about 13% to 15% during the 1960s
and 1970s, rose increasingly fast during the 1980s and 1990s, and now stands at over 30% (CDC
2004b). (Obesity is defined as having a body mass index of 30 or more; BMI = kilogram weight
-3-
divided by the square of metric height.) Comparing the two trends presented in this figure, one
may say that the growth in diabetes in the 1990s was foreshadowed by similar growth in obesity
in the 1980s. The fact that obesity continued to grow rapidly through the 1990s may suggest
similar growth in diabetes in the 2000s.
50% 7%
Diagnosed diabetes % of adults 6%
NHIS, adjusted pre-1997 is
40% ( i p y
5%
9
30% 4%
9
20% 3%
Obese % of adults (NHES, NHANES)
2%
10%
1%
0% 0%
1960 1970 1980 1990 2000
Figure 2. The rise in diabetes has followed a rise in obesity
What has caused the remarkable rise in obesity in the U.S. since 1980? From a thermodynamic
standpoint, the documented increase in BMI can only have been the result of caloric intake
persistently exceeding caloric expenditures. In regard to caloric expenditures, there are three
types: physical activity, resting metabolism, and digestion (Abdel-Hamid 2002). Much has been
written about Americans having a more sedentary lifestyle than they did in decades past, but
available evidence points to no particular decline in physical activity since 1980. Indeed, the
fraction of U.S. adults who say that they get no leisure-time physical activity has been flat or
declining since 1990, though still in the 25% range as of 2001-2002 and certainly a continuing
cause for concern (CDC 2004a). Neither does available evidence point to reductions in rates of
resting metabolism (related to body weight and its fat-lean composition) nor digestion (an
essentially fixed 10% of caloric intake).!
Rather than a decline in caloric expenditure, what appears to have fueled the increase in BMI and
obesity since 1980 is an increase in caloric intake, a conclusion reached by others as well (Cutler
et al. 2003). Figure 3 presents data on caloric intake from the National Health and Nutrition
Examination Survey (NHANES; see CDC 2004b), representative of the entire U.S. adult
population of men and women. According to these data, caloric intake increased by 13% during
An analysis we have done of skinfold measurement data from NHANES (CDC 2004b), applying well-known
body density equations of Durnin and Womersley (1974), suggests a significant increase in the fat fraction during
the 1970s for both men and women, but since 1980 only a small further increase in the fat fraction for men and a
clear (and perhaps surprising) decline in the fat fraction for women.
4.
the 1980s and early 1990s, but has increased rather little since then.? The slowdown in the
growth of caloric intake may be a sign of hope that the obesity epidemic will soon reach its peak,
though what this means for future diabetes prevalence is less obvious.
50% 2,400
Mean self-reported caloric intake of adults
(NHANES, Kcal/person/day)
40%
+ 2,200
30%
+ 2,000
20%
Obese % of adults (NHES, NHANES)
+ 1,800
10% :
0% + + + + 1,600
1960-62 1971-74 1976-80 1988-94 1999-00
Figure 3. The rise in obesity has followed a rise in caloric intake
The National Diabetes Prevention and Control Program
The U.S. Government responded to the growing burden of diabetes in 1975 by establishing,
within the U.S. Centers for Disease Control and Prevention (CDC) a “National Diabetes
Prevention and Control Program” (NDPCP). As the emerging science supported new public
health responses and as resources became available, the strategy for the NDPCP has evolved
through several phases (Murphy et al. 2004).
In the late 1970s, interventions focused on educating diabetes patients and health professionals
on the proper care of people with diabetes, aiming towards reducing diabetes-related
complications such as blindness, amputations, and heart failure. In the 1980s, research supported
the addition of screening for diabetes-related complications, with a particular focus on uninsured
2 Among other time-series data we have seen, some support the idea of a very significant 1990s slowdown in the
growth of caloric intake, while others are less supportive. Data from the U.S. Department of Agriculture on calories
available and pounds of food consumed (U.S. Census Bureau 2003, Section 3, Tables 193 and 195, respectively),
suggest a slight slowdown only. More supportive, though another step or two removed in causality, are data on
cigarette smoking and income inequality. Studies show that smoking cessation leads to greater caloric intake (Flegal
et al. 1995). The prevalence of smoking in the U.S. fell rapidly in the 1980s but much less so in the 1990s (NCHS.
2003). Income inequality has been shown to be linked to emotional stress and mortality, across social and economic
classes (Wilkinson 1996). Stress, in turn, has been implicated as a driver of weight gain, most clearly through a
neuroendocrine pathway that involves the impairment of satiety perception and thereby leads to more caloric intake
(Bjérntorp 2001). The commonly used Gini Index of income concentration rose rapidly in the 1980s, but was
essentially flat for most of the 1990s (U.S. Census Bureau 2004, Table RDIS5).
people with diabetes. In the 1990s, when two landmark studies established that intensive control
of blood sugar could reduce the incidence of complications (DCCT 1993, UKPDS 1998), the
strategy shifted again. The NDPCP enhanced its patient and provider education (e.g.,
incorporating the new disease management practices) but began to minimize individual
screening activities. Around the turn of the century, the program began to focus more on
population-level or “system-based” interventions and less on direct clinical services to
individuals with diabetes. The new focus included more partnering with other organizations and
stakeholders. It also included the establishment of national objectives for testing people for
diabetes, testing for diabetes-related complications, and preventing complications such as the flu.
Finally, as a new landmark study (DPP 2002) demonstrated that effective treatment could reduce
diabetes incidence in people with “prediabetes” (that is, elevated blood glucose levels that are
not yet in the range of diabetes), NDPCP began to include management of prediabetes as well.
In recent years, public health officials have looked even further “upstream” towards the factors
driving the incidence of prediabetes. While age and demographics are important factors in
determining an individual’s risk for prediabetes and diabetes, the primary changeable risk factor
is obesity. Public health officials are now paying significant attention to the obesity epidemic in
the U.S., and recognize that diabetes prevalence may grow further than previously forecasted
should the prevalence of obesity continue to increase. Such attention led the designers of the
NDPCP to consider not just reducing the incidence of diabetes by focusing on those with
prediabetes, but also reducing the incidence of prediabetes by controlling obesity.
By 2004, CDC staff in the Division of Diabetes Translation (DDT), which administers the
NDPCP, had grown to better understand the influential role that other organization’s programs
and policies have in focusing on diabetes prevention and achieving diabetes-related goals and
objectives. The Division has come to appreciate that efforts to prevent, control, and manage
diabetes are tied, in multiple ways, to most other chronic diseases and to a wide array of social
and physical processes at work in communities (i.e., food production and distribution,
transportation, education, etc.). It has become clear that efforts to address the increasing burden
of diabetes will necessarily involve expanding and reconfiguring the kinds of resources that fuel
public health work (Murphy et al. 2004).
Today, with the nation’s diabetes prevention and control program poised for further
transformation, many strategic questions remain unanswered. How can public health officials
engage new partners in redirecting the course of the diabetes epidemic? How can these diverse
actors build confidence that prevention activities in the present will yield reduced rates of disease
in the future? How can policy makers assess the inherent trade-offs among competing
strategies? And most importantly, what conditions must exist for people to take action based on
their understanding of how a dynamic diabetes system behaves?
The Modeling Project
In order to address the challenge of diabetes growth, a group of CDC leaders decided to employ
system dynamics modeling as a tool for enhancing both learning and action. Through a
partnership with system modelers at the Sustainability Institute and the Rollins School of Public
Health at Emory University, staff from the DDT and several other CDC Divisions worked
together to create an interactive, SD model-based “learning lab” that enables multiple
stakeholders to design and formally evaluate diabetes-related policies.
The learning lab technique engages multiple stakeholders in a facilitated process that is designed
to explore the structural reasons for observed health trends. The process begins by integrating
research findings and insights from practical experience into a dynamic hypothesis of the health
problem. Next, this hypothesis is translated into a formal simulation model using a system of
differential equations and specialized SD modeling software. The model creates a virtual world
for learning in which controlled experiments can be conducted as a prelude to introducing
policies in the real world. Stakeholders use the model to search for policies that can deliver
long-term health impacts, while simultaneously guarding against interventions that lead to
unintended effects. Users can also formally compare and evaluate prevention policies that would
otherwise have to be tested through trial and error in real populations.
Eventually DDT staff and their state-level colleagues, with training and support from SD
practitioners, may begin facilitating model-based learning labs that provide a forum for diverse
stakeholders to think together about the dynamics of diabetes strategy.
Developing the Model: Meetings and Participants
Between September 2003 and March 2004, the modeling team constructed a system dynamics
model focused on the drivers of increasing diabetes prevalence. The process involved a series
four meetings, each taking place over two days.
Meeting 1: Model Conceptualization
After two planning meetings with a team of CDC “project champions”, during which a
preliminary population stock-and-flow structure was developed, DDT leaders convened a two-
day conceptualization meeting in early December 2003. Approximately 45 individuals
participated, including CDC staff with a specialty in diabetes prevention and control, as well as
representatives from other CDC Divisions that focus on factors within the broader “diabetes
system’, including nutrition and physical activity, smoking, cardiovascular health, adolescent
and school health, adult and community health, and chronic disease in general. The agenda for
the first day included time for reviewing the principles of SD methodology; brainstorming
dilemmas that arise when confronting diabetes and its many drivers (from obesity to health care
access and utilization); and mapping the relationships among those drivers.
That evening, the modeling team simplified what had become a sprawling map of drivers into a
tighter, more generalized sector diagram. On the second day, the agenda included time for:
discussing the simplified sector diagram; sketching behavior-over-time graphs to explain how
various parts of the system have been changing; identifying sources of empirical data; and setting
overall priorities for this phase of the project.
Meeting 2: Interacting with a Simulation Model
One line of work undertaken to complement the development of the simulation model involved
educating DDT staff in how one can use a system dynamics model to improve strategy and spark
effective action. To this end, the modeling team developed a simple user interface for the core
stock-and-flow model, and in January 2004 gave staff members an opportunity to experiment
with it in teams of two in a computer lab. The intention was not to analyze specific policy
scenarios, but rather to acquaint public health professionals with what a dynamic simulation
model can do. On the second day, the modelers also met with several CDC scientists who
specialized in various aspects of the diabetes system in order to incorporate their ideas into the
model’s emerging structure.
Meeting 3: “Pushing on the Model”
A second line of work focused on developing a widely credible simulation model of the diabetes
system. This required intense collaborative work, among modelers and subject experts, to
articulate theories for the growth in diabetes prevalence; gather data to support or refute those
theories; translate those theories into differential equations; test the resulting model; and calibrate
the model with reference to historical trends. By February 2004, the emerging model had
progressed to the point where it could be critiqued by the project champions group. Again, the
focus was not on policy analysis but on potential improvements that would make the model more
useful in its goal of supporting effective conversations about diabetes prevention strategy.
Meeting 4: “Letting the Model Push on Us”
The fourth meeting will occur at the end of March 2004 and is designed around a second
iteration of the simulation model. On the the first day, we will observe how the model functions
in a conversation about diabetes strategy. Approximately 15 CDC staff members, whose jobs
entail a variety of roles in the diabetes system, will gather to experience for themselves how an
SD learning lab works. Two members of the modeling team will facilitate the event, while a
third observes the proceedings. With that experience in mind, the champions group will
reconvene on the second day to reflect on the model and the learning lab experience, as well as
next steps for widening the dialogue and acting on the main dynamic insights.
Model Structure
As the diabetes system model has evolved, we have continually sought to create a structure that
is
(1) generic enough to be adaptable for other chronic diseases;
(2) realistic enough to reproduce national-level historical data on the prevalence of diabetes,
prediabetes, and obesity;
(3) comprehensible enough to test practical policies without disaggregating the population into
demographic categories of age, sex, race/ethnicity, or other individual attributes;
(4) broad enough to encompass a spectrum of policy measures that are being considered; and
(5) grounded enough in empirical experience that it does not require speculation beyond what
the project participants themselves could agree upon or what credible evidence could
support.
The preliminary stock-and-flow model, which began the iterative process described above, was
based in many of its structural details and even some of its calibration, on a model of diabetes
developed for a chronic disease healthcare initiative in Whatcom County, Washington State
(Homer et al. 2004). Other system dynamics models have also examined chronic disease care
(Hirsch and Wils 1984, Hirsch and Immediato 1999), and they too informed our work.
From the outset, this project recognized a need to reach further “upstream” in the causal chain of
disease incidence and progression than the Whatcom County model had done (McKinlay and
Marceau 2000, Milstein and Homer 2003). This would mean modeling, in health policy terms,
not only disease management and control (also described as secondary and tertiary prevention)
for those already afflicted and primary prevention for the at-risk or vulnerable population, but
even actions that can protect people from becoming at-risk in the first place. Figure 4 presents
the generic view of chronic disease dynamics that came out of these discussions. Healthcare
professionals, in their roles as service providers, commonly oversee the downstream activities of
secondary and tertiary prevention that are portrayed in this figure. Such professionals may also
be involved in primary prevention when their work focuses on risk management or health
education—for example, guiding obese individuals into diet and exercise programs and regularly
monitoring their progress.
In contrast with these professionally led healthcare services, the upstream activities of health
protection focus on changing those living conditions that leave people vulnerable to disease,
either in a targeted way for people already at risk, or in a broader way for entire populations.
Efforts to improve adverse living conditions are in the natural domain of advocacy and citizen
groups, rather than healthcare professionals; they entail a level of organized public action that
downstream services do not typically require. For financial and organizational reasons,
particularly those stemming from pressures to report near-term success in the form of traditional
health outcome measures, upstream efforts typically receive only modest support from public
health agencies (see Murphy et al. 2004). The tendency to “lock in” to downstream efforts may
be shaken loose only when it finally becomes apparent that such efforts cannot, unto themselves,
be effective in protecting the public’s health (Milstein and Homer 2003).
Public Work by Advocacy & Citizen Groups Professional Work by Clinicians &
(organizing, governance, mutual accountability) Disease Experts (service provision)
a. health
response
General 4 \*
x protection
Targeted
Adverse Living protection
Conditions
Primary
prevention
Tertiary
prevention
Secondary
prevention
| Vulnerability
cessation
ean 3 Action
Vulnerability Affliction " | Death from
onset onset progression complications
Figure 4. Generic chronic disease population stocks and flows and health system responses
Figure 5 presents the population flow structure that forms the core of the diabetes system model.
This structure evolved during the course of the project to reflect the knowledge and policy
concerns of project participants. It portrays the movement of people into and out of the
following stages:
¢ Normal blood glucose (normoglycemia);
e Prediabetes, defined as having either impaired glucose tolerance (IGT, non-fasting glucose
concentration of 140-199 mg/dl) or impaired fasting glucose (IFG, 100-125 mg/dl) (ADA
2004, Benjamin et al. 2003);
¢ Uncomplicated diabetes, that is, meeting the testing criteria for diabetes (non-fasting glucose
of at least 200 mg/dl) but not yet symptomatic nor showing detectable signs of disease in the
eyes, feet, or other organs; and
¢ Complicated diabetes.
The prediabetes and diabetes (hyperglycemic) stages are further divided into stocks for the
undiagnosed and diagnosed population. Diagnosis is important dynamically because it is a
prerequisite for proper management and control of hyperglycemia; and control, in turn, strongly
reduces the rates of diabetes onset, progression, and death (DCCT 1993, UKPDS 1998, DPP
2002, ADA/NIDDK 2002, Bowman et al. 2003). Because prediabetes and uncomplicated
diabetes are not marked by symptoms, their diagnosis generally occurs only as a result of
screening that a healthcare professional might do periodically (every three years, according to
guidelines) for patients at risk for diabetes by virtue of family history, obesity, age, and other risk
factors. While there is currently still some progress to be made in getting providers to test for
uncomplicated diabetes, the testing gap for prediabetes is much larger because guidelines for the
-10-
testing and treatment of prediabetes are relatively new and still somewhat fluid. Complicated
diabetes is more readily detected than uncomplicated, because of suspicions aroused by
presented symptoms. Even so, a significant fraction of complicated diabetes is not detected
rapidly due to issues of provider inattention or the patient’s lack of insurance or convenient
access to a provider (McKinlay and Marceau 2000).
Diabetes onset
from Undx PreD Progression to
PreD onset .omplic from Undx D Undx Complic
diab deaths
Recovery from
Undx PreD pen
ss | Uncomplic diab Complic diab
diagnosis diagnosis desposs
nee)
Recovery from Diabetes onset Progression to Dx Complic
Dx PreD from Dx PreD complic from Dx D diab deaths
Figure 5. Diabetes and prediabetes population stock and flow structure as modeled
(Undx: undiagnosed; Dx: diagnosed; PreD: prediabetes; D: diabetes.)
This is a simplified view of the full stock and flow structure. The full structure also includes an
inflow into the normoglycemic stock for general population growth, and non-diabetes-related
death flows out of each population stock.
Figure 6 presents an overview of the full model, which includes the population stock-and-flow
structure as well as the main influences that affect the rates of population flow throughout the
system. The diagram also highlights (in red) those inputs that are seen as major influences
affecting the upstream and downstream flow-rate drivers, and which may be amenable to policy
intervention. Upstream, these exogenous policy-related variables include caloric intake and
physical activity as drivers of obesity. Downstream, they include testing rates and healthcare
access as drivers of disease detection; and clinical management, self-management, and drug
affordability as drivers of disease control.
-I1-
Caloric intake &
a Physical activity Access to
Testing for preventive Testing for
Testing f
Caloric ee uncomplicated D health care complicated D
Resting PreD
balance “—~ ;
i metabolism %
Mean body | 7 :
weight/BMI Porte Detection Detection of
J OF Ere! of uncomplicated D complicated D
Diabetes onset
Progression to
—pomplic from Und D p- Undx Complic;
Obese from Undx PreD
fraction of = diab deaths
popn
Recovery from
Undx PreD
| Uncomplié diab Complic diab
; diagnosis diagnosis
Risk for
diabetes
ann)
Diabetes onset Progression to Dx Complic
from Dx PreD ‘complic from Dx D diab deaths
Control of Control of Dx Control of
Undiagnosed Px Ped uncomplic D Dx complic D
D = Diabetes
PreD = Prediabetes
Sine om of Self-mgmt & Clinical mgmt of
aii Drug affordability DxD
Recovery from
Dx PreD
Figure 6. Overview of full model, showing key drivers of population flows
and inputs amenable to policy intervention (in red)
It is notable that the model does not explicitly address the social, economic, and personal
determinants behind such factors as caloric intake and the ability to self-manage, but instead
leaves such forces outside the model’s boundary. Our conceptualization meeting in December
2003 involved extensive discussion of such factors as well as the intricate interactions among
them. Despite intense interest in these drivers, the decision to keep them outside the model
boundary was made for two reasons.
(1) There was a lack of supporting evidence for hypothesized relationships. Following the
December meeting, an effort was made to gather historical evidence and literature on as
many of the determinants that had been discussed as possible. In many cases, the available
evidence either did not support the hypotheses that had been presented, or was too sparse or
full of qualifiers to allow for any firm conclusion. We ultimately concluded that we were
delving into areas of speculation that would not enjoy general acceptance among the many
stakeholders, both scientists and policymakers, identified as audiences for the initial model.
Quantitative modeling of such speculative relationships might be acceptable under some
circumstances, but not given the goals and needs of this particular project in its initial phase.
(2) Powerful insights could be gained from the stock-and-flow structure alone. At the December
meeting, and then again in January 2004, project participants commented that the core stock-
-12-
and-flow model by itself revealed counterintuitive dynamics that could have important
implications for policy selection. The value to real-world clients of dynamically interesting
stock-and-flow structures—perhaps as a first step of analysis before delving into behavioral
feedback loops—has been discussed elsewhere, first as a general proposition with examples
from various subject domains (Homer 1997), and more recently in connection with the
modeling of diabetes for Whatcom County (Homer et al. 2004).
Baseline Model Behavior
The model has been calibrated based on historical data available for the U.S. adult population, as
well as estimates from the scientific literature, and from diabetes clinicians who were
interviewed as part of the Whatcom County study.> These data address the:
e Prevalence of diagnosed diabetes, prediabetes, and obesity;
e Fraction of diabetes undiagnosed;
¢ Complicated fractions of undiagnosed and diagnosed diabetes;
e Population growth rate;
e Death rates from diabetes complications and from all other causes;
e Effects of control (glucose, lipids, blood pressure) on rates of diabetes onset, progression,
and death;
e Caloric intake;
e Mean body weight, height, and BMI;
e Fat fraction of body weight;
e Fraction of population with health insurance;
e Fraction of people with diagnosed diabetes receiving annual dilated eye exam;
e Fraction of people with diagnosed diabetes doing daily glucose self-monitoring; and
e Fraction of people who can afford diabetes medications.
As described previously, the model contains a number of exogenous time series for which future
assumptions are required for policy testing. In order to establish a baseline case against which
alternative policies could be compared, we started by asking for project participants’ best
estimates of what a “status quo” future would look like, in which the NDPCP was frozen in place
as it stands currently and no new initiatives were introduced. Interestingly, the participants felt
this task was confusing and subjective, and ultimately infeasible. So, we suggested a different
sort of baseline case, in which all exogenous time series were simply held constant at their
assumed 2004 values. Despite the likely lack of realism of such a baseline for some variables
(especially in the case of prediabetes testing and clinical management, which only got their start
in the last few years and are still early in their growth path), the participants agreed to such a
baseline because of its simplicity and clarity.
3 Many of the data were obtained from online sources, including websites of the CDC (2004a, 2004b, 2004c,
2004d), the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK 2001, 2004a, 2004b), and the
U.S. Census Bureau (2003, 2004). Data and empirical equations were also culled from the following published
articles: Abdel-Hamid 2002, Benjamin et al. 2003, Durnin and Womersley 1974, and Honeycutt et al. 2003. Many
thanks also to the CDC’s Edward Gregg for extracting summary statistics from the NHANES database, spanning the
1960-2000 period, on weight, height, BMI, and skinfold measurements.
ar
Figures 7 and 8 demonstrate the model’s ability to reproduce the 1980-2000 data for diagnosed
diabetes prevalence and obesity prevalence, respectively, as well as the model’s baseline
projections for these variables through 2010. Figure 9 takes the baseline projection out to the
year 2050, illustrating the growth in the prevalence of hyperglycemia and its three major
components: complicated diabetes, uncomplicated diabetes, and prediabetes.
8%
Diagnosed diabetes % of adults
Simulated
6%
4%
Data (NHIS)
2%
0%
1980 1985 1990 1995 2000 2005 2010
Figure 7. Simulated diagnosed diabetes prevalence compared with historical data
40%
Obese % of adults
Simulated
30%
20%
Data (NHANES)
10%
0%
1980 1985 1990 1995 2000 2005 2010
Figure 8. Simulated obesity prevalence compared with historical data
-14-
9.3 | Fraction of adult population
Uncomplicated diabetes
1980 1990 2000» 2010S 2020» 2030» -2040~S 2050
Time (Year)
Figure 9. Baseline growth of diabetes and prediabetes prevalence through 2050
The story told by these graphs is as follows: About two years after the assumed cessation in the
growth of caloric intake as of 2004, obesity prevalence soon ceases to grow as well. This goal-
seeking behavior reflects the negative feedback loop seen in Figure 6, wherein higher body
weight leads to a higher resting metabolism. Significant growth in hyperglycemia (diabetes and
prediabetes combined) continues, in decelerating fashion, for another 25 years, reflecting the
long delays in the system connecting obesity as a risk factor to prediabetes and then to diabetes.
The prevalence of prediabetes peaks and then actually declines somewhat after 2020, due to the
fact that the normoglycemic fraction of the population (= 1 — hyperglycemic fraction), from
which prediabetes draws, is itself on the decline. The growth of uncomplicated diabetes
prevalence continues until about 2035, while the growth of complicated diabetes prevalence
continues right through 2050. From 2000 to 2050, the prevalence of diabetes (diagnosed and
undiagnosed) increases from 8.4% to 13.5%, while the prevalence of diagnosed diabetes per se
increases from 5.6% to 9.5%.4
* As with most SD models, specific output values from this model are intended to be used only for understanding
the dynamic implications of assumptions and comparing the effects of simulated policies. They do not correspond
numerically to the values that are generated from predictive models that are designed to yield specific point
estimates of diabetes prevalence, such as the Markov model by Honeycutt et al. 2003. Various effects of population
aging, for example, have been excluded from our model in an effort to improve its transparency and therefore its
value as a tool for learning about how dynamic complexity affects intervention policies in the overall diabetes
system. In future phases of model development, after users have developed a greater understanding of how and why
the dynamic model works the way it does, we may incorporate aging and other factors that would increase the
predictive accuracy of the SD model.
sen
Illustrative Policy Tests
Project participants will soon have an opportunity to do policy testing of the model in a
workshop-type environment. They will be asked first to describe a policy strategy that they feel
is feasible, and second, to specify the plausible impacts of that strategy in terms of the time-
series inputs identified in Figure 6. We expect that three broad strategy types may be explored: a
pure downstream approach, a pure upstream approach, and a balanced approach of both
upstream and downstream interventions.
We have created sample simulations of these three scenarios and compared them with the base
run in Figures 10 to 14. The runs are as follows:
Base No changes after 2004
Mgmt90 A purely downstream approach, with clinical management of diagnosed
diabetes and prediabetes ramped up to 90% during 2005-15. (Base case
management of diagnosed: 70% for diabetes, 50% for prediabetes.)
e Cal3 A purely upstream approach, with caloric intake ramped down by 3%
during 2005-15.
Mgmt85Cal2 A balanced approach, with clinical management ramped up to 85% and
caloric intake ramped down by 2% during 2005-2015.
Figures 10 and 11 illustrate graphically the assumptions we have made for the four runs.” Note
that in Figure 11 the 3% reduction of caloric intake in Ca/3 represents a mean reduction of
“only” 74 kcal/person/day, and the 2% reduction in Mgmt85Cal2 only 49 kcal/person/day,
reductions that may sound too modest to have much discernible effect on body weight. But other
analysts have concluded that further weight gain in the U.S. could be eliminated by reducing the
mean caloric balance (intake less expenditure) by 100 kcal/person/day or less, which could be
achieved by eating just a few bites less at each meal (Hill et al. 2003). Such a task sounds easy
enough for a select group of individuals in an experimental study, but to reduce the mean daily
intake by 50-100 kcal/person/day across a population of more than 200 million adults is, in fact,
a considerable challenge, and would require a society-wide reaction against those cumulative
social forces and adverse living conditions that have caused caloric intake to increase rather than
decrease for the past 20-plus years.
Public health officials want to be able to anticipate the endpoint results of their policy decisions
in terms of reduced morbidity, mortality, and healthcare costs. One of the things our project
participants—and also we—have come to realize with respect to chronic diseases is that it may
* Note that the balanced approach, Mgmt85Cal2, achieves 3/4 of the full downstream impact of Mgmt90, and 2/3 of
the full upstream impact of Ca/3. Let us call these fractions the downstream and upstream impact fractions,
respectively, for the balanced approach. Now, think of the three intervention approaches as alternative ways of
allocating a fixed public health budget for diabetes. One might then suppose that the sum of the two impact fractions
for the balanced approach should be 1, rather than a value greater than | as in our example (3/4 + 2/3 = 17/12 > 1).
The two impact fractions would, in fact, sum to | if the downstream and upstream policies had impacts that were
strictly proportional to their budgets; in other words, if they had constant returns to scale. But if the returns to scale
are diminishing, as is true of many (and, beyond some point, all) policies, then the sum of the impact fractions for a
balanced approach would be greater than one, as in our example.
=16-
be just as important for these officials to understand the intermediate effects by which they may
gauge progress during the long years that often separate policy intervention and endpoint results.
Figures 12 and 13 represent two examples of such intermediate effects, showing policy impacts
on prediabetes prevalence and complicated diabetes prevalence, respectively. These graphs have
provided insights—that is, initially unexpected or counterintuitive results—of two types. First,
Clinically managed fraction of diagnosed diabetes popn Mgmt90_| 9 g
eee Mgmta5cal2 | 0.85
0.75
Base, Cal3 Oe
0.5
0.25
0
1980 1990-2000 2010-~-2020-S 2030» -2040~S 2050
Time (Year)
Figure 10. Assumed clinical management fraction, four policy tests
2,600 | ean caloric intake (Kcal/person/day)
Base, Mgmt90_|2,465
240 2,416 (down 2%)
2,391 (down 3%)
2,300
2,150
2,000
1980 1990 2000 ~—«-2010~=—«2020~==«2030~=S—2040~—S—«2050
Time (Year)
Figure 11. Assumed caloric intake, four policy tests
iT
in regard to upstream interventions, they suggest that it will take many years for significant
benefits of prevention to be felt at the symptomatic end of the disease chain (compare Ca/3 and
Mgmt85Cal2 with Base in Figure 13). Second, in regard to downstream interventions, Figures
12 and 13 show that disease management by itself actually increases prevalence to some extent
(compare Mgmt90 with Base). This “stacking-up” phenomenon occurs wherever disease
management reduces an outflow of progression or death from a stock but does less to reduce
inflow into that stock. Because it could increase demands on healthcare and public health
capacity, stacking-up may be an important shortcoming with the downstream-only approach, and
is a matter of significance that our project participants were not focused on previously.
0.12
Prediabetes fraction of adult popn
0.1
0.08
0.06
0.04
1980 1990 -2000-S.2010-»S2020S2030»S 2040~—S 2050
Time (Year)
Figure 12. Simulated prediabetes prevalence, four policy tests
0.06 . . Mgmt90.
Complicated diabetes fraction of adult popn
0.05
0.04
0.03
0.02
1980 1990 2000» 2010» 2020» 2030» 2040 ~—2050
Time (Year)
Figure 13. Simulated complicated diabetes prevalence, four policy tests
18
Figure 14 shows the death rate from complications of diabetes, expressed as an annual fraction of
the total adult population. This death rate represents the product of
(1) the prevalence of complicated diabetes, as seen in Figure 13, and
(2) the death rate among those with complicated diabetes, which improved clinical
management—see Figure 10—can quickly reduce.
A comparison of the policy impacts in Figure 14 shows that the purely downstream scenario has
the virtue of acting quickly, but does nothing to prevent a subsequent growth in the death rate
along a trajectory parallel to that of the base run. The purely upstream scenario has the virtue of
significantly reducing the growth trajectory of death, but does so only after a delay of 10 to 15
years. The balanced approach succeeds in both the short term and the long term, in effect paying
attention to both those people who have the misfortune of progressing to complicated diabetes,
but also to those who might otherwise ultimately progress to that state.
0.0035 | Diabetes-related death rate per year for adult popn
Base
0.003
Mgmt85Cal2
0.0025
0.002
0.0015
1980 1990 2000» 2010»: 2020S 2030-»=S 2040 = 2050
Time (Year)
Figure 14. Simulated diabetes-related death rate, four policy tests
Making It Useful
The model is intended to help improve strategies for addressing the burden of diabetes at two
levels: national and state/regional.
At the national level, the DDT may incorporate insights from this analytical work into its own
strategy and its interaction with the other Divisions within the National Center for Chronic
Disease Prevention and Health Promotion. This will likely happen through facilitated meetings
that use the model as a framework for discussion and a tool for experimentation. The DDT will
19
also consider opportunities to use the model as a catalyst for meetings with other national
organizations working on diabetes prevention and control, such as the American Diabetes
Association.
At the state/regional level, the current plan is for a small team of leaders within the DDT to learn
to facilitate a model-based workshop we are calling the “Diabetes Action Lab.” Participants in
these one-to-two day workshops may include state and local public health officials along with a
diverse group of stakeholders from non-governmental organizations such as healthcare
professionals, leaders of not-for-profit agencies, people living with diabetes, and many others.
The goal, after studying the short- and long-term dynamics of diabetes policy options, would be
to improve the overall strategic direction for redirecting the course of the diabetes epidemic, and
to spark more determined and coordinated action towards meeting explicit performance
objectives.
For the Diabetes Action Lab to be effective in a wide range of settings for diverse participants, it
must be embedded in a carefully designed learning environment, one that incorporates facilitated
presentations, small group discussions, strategy exercises, and other supplemental learning
techniques. It may also be necessary to adjust the model’s calibration so that it reflects the
unique context of the health challenge in different states, based on local demographics, diabetes
programs already in place, and other factors.
Acknowledgments
The authors would like to acknowledge those project participants who have contributed most to
this work. From the CDC: Kaetz Beartusk, Letia Boseman, Carl Caspersen, Tom Chapel,
Cynthia Clark, Clay Cooksey, Michael Engelgau, Jeff Etchason, Deborah Galuska, H. Wayne
Giles, Edward Gregg, Virginia Bales Harris, Richard Hoffman, Jim Holt, Marilyn Metzler, Julia
Mitchell, Pat Mitchell, Dawn Satterfield, Stephen Sorensen, Russell Sniegowsky, Patricia
Thompson-Reid, Frank Vinicor, and David Williamson. From outside the CDC: Kevin Klein,
and the staff at Sustainability Institute. We are also indebted to Kathy Cahill, Paul Halverson,
and Richard Jackson from the CDC Office of the Director, who took an active interest in this
work and helped us to clarify our thoughts and direction.
References
Abdel-Hamid TK. 2002. Modeling the dynamics of human energy regulation and its implications for
obesity treatment. System Dynamics Review 18(4): 431-471.
American Diabetes Association (ADA). 2004. How to Tell if You Have Pre-Diabetes. Accessed at:
http://www.diabetes.org/pre-diabetes.jsp.
American Diabetes Association and National Institute of Diabetes and Digestive and Kidney Diseases
(ADA/NIDDK). 2002. The prevention or delay of Type 2 diabetes. Diabetes Care 25:742-749.
Benjamin SM, Valdez R, Geiss LS, Rolka DB, Narayan KMV. 2003. Estimated number of adults with
prediabetes in the U.S. in 2000: Opportunities for prevention. Diabetes Care 26(3): 645-9.
20-
Bjorntorp P. 2001. Do stress reactions cause abdominal obesity and comorbidities? Obesity Reviews 2:
73-86.
Bowman BA, Gregg EW, Williams DE, Engelgau MM, Jack, L Jr. 2003. Translating the science of
primary, secondary, and tertiary prevention to inform the public health response to diabetes. J Public
Health Mgmt Practice November(Suppl): S8-S14.
Centers for Disease Control and Prevention (CDC). 2003. The Public Health Competency Handbook:
Optimizing Individual and Organizational Performance for the Public’s Health. Atlanta, GA.
Centers for Disease Control and Prevention (CDC). 2004a. Behavioral Risk Factor Surveillance System
(BRFSS). Accessed at: http://www.cdc.gov/brfss/.
Centers for Disease Control and Prevention (CDC). 2004b. National Health and Nutrition Examination
Survey (NHANES, created in 1960 as NHES). Accessed at: http://www.cdc.gov/nchs/nhanes.htm.
Centers for Disease Control and Prevention (CDC). 2004c. National Health Interview Survey (NHIS).
Accessed at: http://www.cdc.gov/nchs/nhis.htm.
Centers for Disease Control and Prevention (CDC). 2004d. Diabetes Surveillance System. Accessed at:
http://www.cde.gov/diabetes/statistics/.
Cutler DM, Glaeser EL, Shapiro JM. 2003. Why have Americans become more obese? National Bureau
of Economic Research, Working Paper 9446. Cambridge, MA. Accessed at:
http://www.nber.org/papers/w9446.
Diabetes Control and Complications Trial Research Group (DCCT). 1993. The effect of intensive
treatment of diabetes on the development and progression of long term complications in insulin
dependent diabetes mellitus. New England Journal of Medicine 329: 977-986.
Diabetes Prevention Program Research Group (DPP). 2002. Reduction in the incidence of Type 2
diabetes with lifestyle intervention or metformin. New England Journal of Medicine 346: 393-403.
Durnin JVGA, Womersley J. 1974. Body fat assessed from total body density and its estimation from
skinfold thickness: Measurements on 481 men and women aged from 16 to 72 years. British Journal
of Nutrition 32: 77-91.
Flegal KM, Troiano RP, Pamuk ER, Kuczmarski RJ, Campbell SM. 1995. The influence of smoking
cessation on the prevalence of overweight in the United States. New England Journal of Medicine
333(18): 1165-70.
Gorman C. 2003. Why so many of us are getting diabetes. TIME (8 December): 58-69.
Hill JO, Wyatt HR, Reed GW, Peters JC. 2003. Obesity and the environment: Where do we go from
here? Science 299 (7 February): 853-855.
Hirsch GB, Immediato CS. 1999. Microworlds and generic structures as resources for integrating care and
improving health. System Dynamics Review 15(3):315-330.
Hirsch GB, Wils W. 1984. Cardiovascular disease in the Dutch population: A model-based approach to
scenarios. Ministry of Health: Conference on Health Care Scenarios. The Hague, Netherlands.
Hoffman C. Rice DP, Sung H-Y. 1996. Persons with chronic conditions: Their prevalence and costs.
JAMA 276(18):1473-9.
-21-
Homer JB. 1997. Structure, data, and compelling conclusions: Notes from the field. System Dynamics
Review 13(4):293-309.
Homer J, Hirsch G, Minniti M, Pierson P. 2004. Models for collaboration: How system dynamics helped
a community organize cost-effective care for chronic illness. System Dynamics Review (in press).
Homer J, Milstein B. 2003. Planning comprehensive public health programs: A promising opportunity
for SD. Presentation to System Dynamics Winter Conference. January 9, 2003: Austin, TX.
Honeycutt AA, Boyle JP, Broglio KR, Thompson TJ, Hoerger TJ, Geiss LS, Narayan KMV. 2003. A
dynamic Markov model for forecasting diabetes prevalence in the United States through 2050.
Health Care Mgmt Science 6: 155-164.
Institute for Health and Aging, University of California, San Francisco. 1996. Chronic Care in America:
A 21" Century Challenge. Robert Wood Johnson Foundation: Princeton, New Jersey. Accessed at:
http://www.rwjf.org/publications/publicationsPdfs/Chronic Care _in_America.pdf.
Institute of Medicine Committee on Quality of Health Care in America. 2001. Crossing the Quality
Chasm: A New Health System for the 21" Century. National Academy Press: Washington, DC.
McKinlay J, Marceau L. 2000. U.S. public health and the 21" century: Diabetes mellitus. The Lancet
356(26 August): 757-761.
Milstein B, Homer J. 2003. The dynamics of upstream and downstream: Why is it so hard for the health
system to work upstream and what can be done about it? Presentation to CDC Health Futures
Workgroup. November 25, 2003: Atlanta, GA.
Murphy D, Chapel T, Clark C. 2004. Moving diabetes care from science to practice: The evolution of
the National Diabetes Prevention and Control Program. Annals of Internal Medicine (in press).
Narayan KMV, Boyle JP, Thompson TJ, Sorensen SW, Williamson DF. 2003. Lifetime risk for diabetes
mellitus in the United States. JAMA 290: 1884-90.
National Center for Health Statistics (NCHS). 2003. Health, United States 2002 (published annually).
Data on smoking prevalence accessed at: http://www.infoplease.com/ipa/A0762370.html.
National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK). 2001. Diabetes Control and
Complications Trial. Accessed at: http://www.niddk.nih.gov/health/diabetes/pubs/dect1/dect.htm.
National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK). 2004a. National Diabetes
Statistics. Accessed at: http://diabetes.niddk.nih.gov/dm/pubs/statistics/index.htm.
National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK). 2004b. Statistics related to
overweight and obesity. Accessed at:
http://www.niddk.nih.gov/health/nutrit/pubs/statobes.htm#preval.
U.K. Prospective Diabetes Study Group (UKPDS). 1998. Tight blood pressure control and risk of
macrovascular and microvascular complications in Type 2 diabetes. The Lancet 352: 703-713.
U.S. Census Bureau. 2003. Statistical Abstract of the United States: 2002. Accessed at:
http://www.census.gov/prod/www/statistical-abstract-02.html.
U.S. Census Bureau. 2004. Historical Income Tables. Accessed at:
http://www.census.gov/hhes/income/histinc.
Wilkinson RG. 1996. Unhealthy Societies: The Afflictions of Inequality. London: Routledge.
-22-
to tl
CC BY-NC-SA 4.0