Beruflich Dokumente
Kultur Dokumente
by
Laura C. A. Rosella
UNIVERSITY OF TORONTO
2009
Abstract
Since the publication of the Framingham algorithm for heart disease, tools that predict disease
risk have been increasingly integrated into standards of practice. The utility of algorithms at the
population level can serve several purposes in health care decision-making and planning. A
population-based risk prediction tool for Diabetes Mellitus (DM) can be particularly valuable for
public health given the significant burden of diabetes and its projected increase in the coming
years. This thesis addresses various aspects related to diabetes risk in addition to incorporating
methodologies that advance the practice of epidemiology. The goal of this thesis is to
demonstrate and inform the methods of population-based diabetes risk prediction. This is
studied in three components: (I) development and validation of a diabetes population risk tool,
(II) measurement and (III) obesity risk. Analytic methods used include prediction survival
modeling, simulation, and multilevel growth modeling. Several types of data were analyzed
including population healthy survey, health administrative, simulation and longitudinal data.
The results from this thesis reveal several important findings relevant to diabetes, obesity,
population-based risk prediction, and measurement in the population setting. In this thesis a
model (Diabetes Population Risk Tool or DPoRT) to predict 10-year risk for diabetes, which can
be applied using commonly-collected national survey data was developed and validated.
Conclusions drawn from the measurement analysis can inform research on the influence of
measurement properties (error and type) on modeling and statistical prediction. Furthermore, the
ii
use of new modeling strategies to model change of body mass index (BMI) over time both
enhance our understanding of obesity and diabetes risk and demonstrate an important
methodology for future epidemiological studies. Epidemiologists are in need of innovative and
accessible tools to assess population risk making these types of risk algorithms an important
scientific advance. Population-based prediction models can be used to improve health planning,
explore the impact of prevention strategies, and enhance our understanding of the distribution of
diabetes in the population. This work can be extended to future studies which develop tools for
disease planning at the population level in Canada and to enrich the epidemiologic literature on
modeling strategies.
iii
Supervisors: Dr. Douglas Manuel/ Dr. Cam Mustard,
Dalla Lana School of Public Health, University of Toronto
Committee Members:
Dr. Paul Corey, Dept. of Public Health Sciences, University of Toronto
Dr. Jan Hux, Health Policy Management and Evaluatio, University of Toronto
Dr. Les Roos, Dept. of Community Health Sciences, University of Manitoba
Dr. Thérèse A. Stukel, Dept. of Health Policy Management and Evaluation, University of
Toronto
Acknowledgements
I would like to thank my supervisors and committee for their brilliant mentorship and guidance.
Each has provided a unique contribution to my training and this work.
I would like to acknowledge the valuable insight of my classmates: Jennifer Bethel, Sarah Jane
Taleski and Marcelo Urquia and my colleagues at ICES: Nick Daneman, Jeff Kwong, and Refik
Saskin.
I would like to thank my husband Luigi for his unending patience, support and love. Also, I
would like to thank my parents, and my older brothers and sisters who have encouraged and
inspired me. I am blessed to have all their support and without them this work would not be
possible.
iv
Table of Contents
1 THESIS INTRODUCTION ................................................................................................................1
1.1 Introduction/Background ....................................................................................................1
1.2 References .........................................................................................................................10
2. DEVELOPMENT AND VALIDATION OF DPORT .........................................................................16
2.1 Abstract .............................................................................................................................16
2.2 Introduction .......................................................................................................................18
2.3 Methods .............................................................................................................................21
2.4 Results ...............................................................................................................................29
2.5 Discussion .........................................................................................................................40
2.6 References .........................................................................................................................46
3. MEASUREMENT IN RISK PREDICTION MODELS .......................................................................51
3.1 The influence of measurement error on accuracy (calibration), discrimination, and
overall estimation of a risk prediction model ........................................................................51
3.1.2 Abstract .....................................................................................................................51
3.1.2 Introduction ...............................................................................................................53
3.1.3 Methods .....................................................................................................................57
3.1.4 Results ........................................................................................................................66
3.1.5 Discussion ..................................................................................................................81
3.1.6 References ..................................................................................................................87
3.2 The role of ethnicity in the population-based prediction of diabetes ............................90
3.2.2 Abstract .....................................................................................................................90
3.2.2 Introduction ...............................................................................................................92
3.2.3 Methods .....................................................................................................................95
3.2.4 Results ......................................................................................................................105
3.2.5 Discussion ................................................................................................................115
3.2.6 References ................................................................................................................121
4. OBESITY RISK ..........................................................................................................................126
4.1 Abstract ...........................................................................................................................126
4.2 Introduction .....................................................................................................................128
v
4.3 Methods ...........................................................................................................................130
4.4 Results .............................................................................................................................139
4.5 Discussion .......................................................................................................................157
4.6 References .......................................................................................................................163
5. CONCLUSION ............................................................................................................................167
6. APPENDIX .................................................................................................................................172
6.1 Glossary of frequently used terms and acronyms ...........................................................173
6.2 Ethics ..............................................................................................................................176
6.3 Study flow diagram and SAS code for simulation ..........................................................177
6.4 An example of ICC in the context of BMI distribution in the population ......................184
6.5 Detailed results from simulations ....................................................................................187
vi
1. Thesis Introduction
1.1 Introduction/Background
In many scientific disciplines, studies that predict or forecast what will happen in the future have
contributed to our understanding of the world. The value of scientific studies that provide
models to inform strategies that can modify and possibly mitigate future events is of importance
to society. Examples include estimating the impact of climate or environmental changes on the
earth‘s ecosystems or the impact of policy changes on the economy (1-3). These prediction
models have been accepted as valuable tools by scientists and have provided critical information
In the field of epidemiology, prediction models are underrepresented and the concept of
risk prediction is overshadowed by the estimation of relative risk measures to clarify etiological
perspectives of disease. Etiological models use the same estimation procedures as most
predictive modeling (i.e., regression) in order to quantify the relative risk associated with a
particular exposure on an outcome. Though regression is often used for both purposes, the way
in which the model is constructed will differ due to the goals of the model. The goal of a
multivariate etiological model is to optimize the accuracy of the relative risk estimate by
controlling for confounding. The goal of a prediction model differs in several important ways.
First, the outcome which needs to be optimized is an absolute measure of risk, often expressed as
percentage or probability (versus a risk or hazard ratio). Second, the goal of a prediction model is
to maximize the ability to discriminate between at risk groups and to correctly classify true risk,
known as discrimination and calibration. Typically these indices are not evaluated in etiological
models. Thirdly, prediction models must be generalizable in other populations to which the
model can be applied. Typically etiological models fit the data used to generate the relative risk
1
estimate as tightly as possible and as a result may not be reproducible using data in other settings
or may not be applicable to describe risk in another population. These goals change the criteria
for model assessment and concomitantly the methodological framework being used.
In medicine, prediction models, in the form of risk algorithms, have been used as tools for
patient decision-making. A risk algorithm is a tool used to estimate the absolute risk of an
expressed as the probability of dying or developing a disease in a given time period (6). In
medicine, risk algorithms have contributed important advances in individual patient treatment
and disease prevention. One of the most utilized risk tools is the Framingham Heart Score (7).
This tool is used to calculate the probability that a patient will develop coronary heart disease in
5 or 10 years, and has been widely integrated into cardiovascular disease prevention and
management throughout the world (8-12). Risk algorithms are widely recommended by medical
societies for appropriate identification of patients that will benefit from specific interventions.
Typically, risk prediction tools are used clinically and applied at the individual level.
Several potential benefits may be realized by extending the application of these tools to the
population level. Similar to the individual level, at the population setting predictive risk tools
have the potential of providing insight into the future burden of a disease in an entire region or
nation and the influence of specific risk factors. These tools can support health care decision-
making, including the effective and efficient allocation and distribution of health care resources
and planning for effective disease prevention interventions. To date, prediction tools specifically
designed for use at the population level are neither created nor used for planning.
2
Epidemiology of Diabetes
Diabetes is a chronic endocrine disorder affecting the body‘s metabolism and resulting in
structural changes affecting the organs of the vascular system. Serious complications resulting
from diabetes include coronary heart disease, stroke, retinopathy, renal failure, peripheral artery
disease, and neuropathy (14). The two main forms of diabetes are type I diabetes and type 2
diabetes. Type 1 diabetes is a result of pancreatic islet beta-cell destruction usually due to an
autoimmune response which results in insulin deficiency requiring exogenous insulin to prevent
abnormal insulin secretion(15). In people with type 2 diabetes, blood sugar must be controlled
either through diet, with oral hypoglycemic drugs or in severe cases with exogenous insulin (16,
17). Type 2 diabetes accounts for over 90% of all diabetes cases worldwide (18) and is the focus
Diabetes and its complications is a leading cause of death and disability, reducing overall
life expectancy and healthy life expectancy. Global estimates place the number of people with
diabetes at approximately 200 million and increasing rapidly (19). Over 2 million Canadians
have diabetes and prevalence is increasing annually (20). In Canada‘s largest province, Ontario,
the 2005 prevalence estimate of diabetes had already exceeded the rate that was predicted by the
World Health Organization (WHO) for 2030 (21). There is a growing concern that if left
unchecked, these trends may slow or even reverse life expectancy gains in the US and other
developed countries (18). In Ontario alone, diabetes has been shown to reduce healthy life
expectancy by 2.7 and 2.9 years respectively (22). The economic and medical consequences of
3
complications arising from diabetes, including cardiovascular disease and kidney failure, are
significant. Diabetes and its associated complications pose a significant economic burden to the
health care system in Canada, which was estimated at 1.6 billion in 1998 (23).
Excess body weight or obesity is associated with insulin resistance (24) and is
overwhelmingly associated with incidence of type 2 diabetes (25-29). Furthermore, it has now
been demonstrated through randomized trials that diabetes can be prevented or delayed through
reduction in body weight (30-33) and consequently intervention strategies that target weight
reduction or prevention of weight gain are largely recognized as integral to primary prevention
of diabetes (34, 35). Currently, more than 1.1 billion adults are overweight worldwide including
312 million who are obese (36). In Canada, obesity doubled between 1985 and 1998 (37) and
currently approximately 6.8 million adults ages 20 to 64 are overweight (including 4.5 million
who are obese) (38). Therefore, obesity in Canada is a significant problem (39) that must be
In addition to obesity several lifestyle, environmental and demographic factors have been
associated with diabetes, the main ones being: ethnicity, physical activity, alcohol and tobacco
(40). There is growing evidence that certain ethnic groups are at increased risk for developing
type 2 diabetes. Globally, non-European populations have a higher proportional burden of type 2
diabetes compared to the other regions of the world(41). The highest diabetes rates in the world
are seen in aboriginal population, including those in Australia(42, 43), United States(44, 45), and
Canada(46, 47). In the United States, studies have shown that those of African and Hispanic
decent are at increased risk for developing diabetes compared with non-Hispanic white
Americans (48-50). Throughout the world, those of South Asian decent have been shown to
carry an increase burden of type 2 diabetes compared with both non-white and white ethnicities
4
(51-53). The variability in risk between ethnicities may be due to a contribution of genetic
factors, which can influence insulin metabolism and predisposition to weight gain(54, 55). On
the other hand, the clustering of risk factors for type 2 diabetes within both families and cultural
groups may partly explain the variability in risk (56-59). The debate on the relative contributions
of genetic susceptibility versus environmental and lifestyle factors is ongoing and will continue
Both intervention and observational studies have supported the reduced incidence of type
2 diabetes associated with increased physical activity (61-71). Physical activity can reduce
diabetes risk by eliciting physiological changes at the metabolic level related to insulin-
stimulating pathways (72) in addition to aiding in the maintenance of a healthy body weight.
Several observational studies have now confirmed that shown that moderate alcohol
consumption is associated with a reduced risk of developing diabetes (73-77) whereas smoking
especially given the current high prevalence of diabetes, the startling rise in the occurrence of
obesity, and the substantial associated costs and consequences of the disease for the health
system and individuals. Population estimates of future diabetes burden can be found in the
literature using a variety of different techniques. Previous studies that estimate future diabetes
burden have either extrapolated overall trends in diabetes prevalence or indirectly incorporated
information on the influence of risk factors with various assumptions (19, 41, 48, 81, 82).
Previous studies of diabetes lifetime risk and life expectancy are not predictive, rather they
5
describe diabetes from a life-course perspective using a period or stationary population approach
(22, 48). Clinical risk algorithms have been developed for use in a clinical primary care setting.
These types of algorithms are limited in their ability to be used at the population level because
they are developed for the individual (versus the population) and require clinical data which is
infrequently available at the population level such as fasting blood sugar (83-85), or require
Population planning tools for diabetes in the form of a population-based risk algorithm
has not yet been developed. Estimates of future incidence of diabetes using a risk algorithm
approach and based on current baseline risk factors in the population will alert policy makers,
planners, and physicians to the extent of the diabetes epidemic in communities. In addition, the
target and how extensive a strategy is needed to stabilize or reduce the number of new cases. In
order to maximize the benefits of a population-based risk prediction tool for diabetes, it
necessarily must be available to the widest audience and tightly integrated into existing risk
factor surveillance systems to maximize efficiency. Therefore, an effective tool is required that is
Thesis Objectives
The goal of this thesis is to demonstrate and inform the methods of population-based diabetes
risk prediction. This goal will be investigated in three broad areas: (i) development and
validation of a population-based risk tool; (ii) measurement; and, (iii) modeling of obesity trends.
These components will be investigated using different types of data and distinct analytic
approaches.
6
The first objective of this thesis is to develop and validate a risk tool to predict the probability of
developing diabetes in the subsequent 10 years (10-year risk) for Canadians using widely
available data. Thos model developed in this study is known as the Diabetes Population Risk
Tool (or DPoRT). This is the first time that a risk tool has been developed specifically for the
purpose of population health assessment and using regularly collected national data. The
conceptual framework of this model can be extended to future studies to develop tools for
disease planning at the population level in Canada and to enrich the epidemiologic field on
predictive modeling strategies. Epidemiologists are in need of innovative and accessible tools to
assess population risk thereby making these risk algorithms an important scientific advance. The
development and application of this novel population-based risk tool for diabetes raises
important questions concerning the measurement of risk factors in the population and
Indeed, there are few issues more important to epidemiology than measurement. Though
significant research is devoted to this area for relative risk estimation, there exists little
information on how measurement error and types of variables influence a prediction algorithm.
Measurement in the context of a prediction algorithm can differ from measurement in the context
properties on risk prediction may differ from the influence on relative risk estimation in
etiological models because they rely on different endpoints for their assessment. Secondly, the
data needed for population prediction must be generalizable to the population in addition to
being regularly collected and reported on and this has implications for the type of data that can
be used to make the predictions. Specifically the scale of this type of data may require reliance
7
on methods (such as self-reporting), which may be more prone to error. In addition, the level of
detail available at the population level may be more limited than clinical data used in
epidemiological studies. The first objective in the measurement section is to use simulation to
useful tool for epidemiologists which allow one to carry out investigations which otherwise may
not be possible due to lack of data. Specifically this section will quantify and describe the effect
measurement error (differential and non-differential) in self-reported height and weight on the
performance (discrimination and accuracy) and outcome (predicted risk) of a diabetes risk
algorithm. The second objective in this section is to investigate the influence of scope of
variables (or lack thereof) available at the population level. Detailed ethnic information is
collected but not publicly reported in Canada. Therefore in keeping with the goal of DPoRT to
use publicly available population health surveillance data, DPoRT uses the form of ethnicity
which is publicly reported as ―white/non-white‖. The purpose of this section is to assess the
impact of this data restriction by comparing DPoRT to an algorithm for diabetes that uses
detailed ethnicity and to report on its relative importance in terms of improvement to the
predictive accuracy of the model and policy implications. This work also provides insight into
the role of ethnicity on diabetes risk in the context of other risk factors.
The third segment of this thesis is focused on studying obesity trends over time using
longitudinal data and multilevel growth modeling. The objective of this section is to study how
lifestyle and socio-demographic factors are associated with weight gain, which is key for
preventing diabetes since obesity is its most influential risk factor. The results from such an
analysis are useful for the Diabetes Population Risk Tool for two reasons. First it can provide
further insight into the likely trajectories of obesity over time and this can be used to improve
8
DPoRT predictions or to allow for predictions farther into the future. Secondly, the results of the
model can be used to determine characteristics of the population that can be used to identify
those who are most likely at risk for weight gain (and thus are increased risk for developing
diabetes). In addition to elucidating the factors that modify weight gain, the multilevel growth
model can identify factors that influence the rate at which it occurs over time. This can be used
analytic methods, and measurement in the population setting. This thesis describes the
development and validation of DPoRT, which can be applied using commonly-collected national
survey data. Conclusions drawn from the measurement analysis will inform research on the
influence of measurement properties (error and type) on modeling and statistical prediction.
Furthermore, the use of novel modeling strategies to model change of BMI provides important
information on factors that influence the relationship between obesity and diabetes and
9
1.2 References
10
15. Alberti KGMM, Zimmet PZ. Definition, diagnosis and classification of diabetes
mellitus and its complications. Part 1: Definition, diagnosis and classification of diabetes
mellitus provisional report of a WHO consultation. Diabetic Medicine 1998;15:559-3.
16. Gannon MC, Nuttall FQ. Control of blood glucose in type 2 diabetes without
weight loss by modification of diet composition. Nutrition and Metabolsim 2006;3.
17. Inzicchi SE. Oral antiglycemic therapy for type 2 diabetes - scientific review.
JAMA: The Journal of the American Medical Association 2002;287:360-72.
18. Zimmet P, Alberti KGMM, Shaw J. Global and societal implications of the
diabetes epidemic. Nature 2001;414:782-7.
19. Wild S et al. Global prevalence of diabetes - Estimates for the year 2000 and
projections for 2030. Diabetes Care 2004;27:1047-53.
20. Health Canada. Diabetes in Canada - Second Edition. H49-121. 2002. Ottawa,
Health Canada. 1999.
21. Lipscombe LL, Hux JE. Trends in diabetes prevalence, incidence, and mortality
in Ontario, Canada 1995-2005: a population-based study. Lancet 2007;369:750-6.
22. Manuel D, Schultz S. Health-related quality of life and health-adjusted life
expectancy of people with diabetes in Ontario, Canada, 1996-1997. Diabetes Care
2004;27:407-14.
23. Health Canada. Economic Burden of Illness in Canada. 1998. Ottawa, Public
Health Association of Canada.
24. Kahn SE, Hull RL, Utzschneider KM. Mechanisms linking obesity to insulin
resistance and type 2 diabetes. Nature 2006;444:840-6.
25. Colditz G et al. Weight as a risk factor for clinical diabetes in women. American
Journal of Epidemiology 1990;132:501-13.
26. Colditz G et al. Weight gain as a risk factor for clinical diabetes mellitus in
women. Annals of Internal Medicine 1995;122:481-6.
27. Perry IJ et al. Prospective study of risk factors for development of non-insulin
dependent diabetes in middle aged British men. British Medical Journal 1995;310:555-9.
28. Vanderpump MPJ et al. The incidence of diabetes mellitus in an English
community: a 20-year follow-up of the Wickham Survey. Diabetic Medicine
1996;13:741-7.
29. Wilson P et al. Prediction of incident diabetes mellitus in middle-aged adults.
Archives of Internal Medicine 2007;167:1068-74.
30. Knowler WC et al. Reduction in the incidence of type 2 diabetes with lifestyle
intervention or metformin. New England Journal of Medicine 2002;346:393-403.
11
31. Li G et al. The long-term effect of lifestyle interventions to prevent diabetes in
China Da Qing Diabetes Prevention Study: a 20-year follow-up study. Lancet
2008;371:1783-9.
32. Lindstrom J et al. Sustained reduction in the incidence of type 2 diabetes by
lifestyle intervention: follow-up of the Finnish Diabetes Prevention Study. Lancet
2006;368:1673-9.
33. Tuomilehto J et al. Prevention of type 2 diabetes mellitus by changes in lifestyle
among subjects with impaired glucose tolerance. New England Journal of Medicine
2001;344:1343-50.
34. Avenell A et al. Systematic review of the long-term effects and economic
consequences of treatments for obesity and implications for health improvement. Health
Technology Assessment 2004;8:1-+.
35. Mayor S. International Diabetes Federation consensus on prevention of type 2
diabetes. International Journal of Clinical Practice 2007;61:1773-5.
36. Haslam DW, James WP. Obesity. Lancet 2005;366:1197-209.
37. Katzmarzyk PT. The Canadian obesity epidemic. Obesity Research 2002;10.
38. Statistics Canada. Canadian Community Health Survey, 2005. 2005.
39. Katzmarzyk PT. The Canadian obesity epidemic. CMAJ 2002;166:1039-40.
40. van Dam RM. The epidemiology of lifestyle and risk for type 2 diabetes.
European Journal of Epidemiology 2003;18:1115-25.
41. World Health Organization. Report of a WHO consultation on obesity, Obesity:
preventing and managing the global epidemic. 1998. Geneva, World Health
Organization.
42. Odea K. Westernization, Insulin Resistance and Diabetes in Australian
Aborigines. Medical Journal of Australia 1991;155:258-64.
43. Odea K et al. Obesity, Diabetes, and Hyperlipidemia in A Central Australian
Aboriginal Community with A Long History of Acculturation. Diabetes Care
1993;16:1004-10.
44. Mokdad AH et al. The continuing epidemics of obesity and Diabetes in the United
States. JAMA: The Journal of the American Medical Association 2001;286:1195-200.
45. Pavkov ME et al. Changing patterns of type 2 diabetes incidence among Pima
Indians. Diabetes Care 2007;30:1758-63.
46. Harris SB et al. The prevalence of NIDDM and associated risk factors in native
Canadians. Diabetes Care 1997;20:185-7.
12
47. Young TK et al. Type 2 diabetes mellitus in Canada's First Nations: status of an
epidemic in progress. Canadian Medical Association Journal 2000;163:561-6.
48. Narayan KMV et al. Lifetime risk for diabetes mellitus in the United States.
JAMA 2003;290:1884-90.
49. Winkleby MA et al. Ethnic and socioeconomic differences in cardiovascular
disease risk factors - Findings for women from the third national health and nutrition
examination survey, 1988-1994. Jama-Journal of the American Medical Association
1998;280:356-62.
50. Brancati FL et al. Diabetes mellitus, race, and socioeconomic status - A
population based study. Annals of Epidemiology 1996;6:67-73.
51. Abate N, Chandalia M. Ethnicity and type 2 diabetes - Focus on Asian Indians.
Journal of Diabetes and Its Complications 2001;15:320-7.
52. Ramachandran A et al. Risk of noninsulin dependent diabetes mellitus conferred
by obesity and central adiposity in different ethnic groups: A comparative analysis
between Asian Indians, Mexican Americans and Whites. Diabetes Research and Clinical
Practice 1997;36:121-5.
53. Ramachandran A et al. Rising prevalence of NIDDM in an urban population in
India. Diabetologia 1997;40:232-7.
54. Qi L, Hu FB, Hu G. Genes, environment, and interactions in prevention of type 2
diabetes: A focus on physical activity and lifestyle changes. Current Molecular Medicine
2008;8:519-32.
55. Qi L. Genetic effects, gene-lifestyle interactions, and type 2 diabetes. Central
European Journal of Medicine 2008;3:1-7.
56. Carmelli D, Cardon LR, Fabsitz R. Clustering of Hypertension, Diabetes, and
Obesity in Adult Male Twins - Same Genes Or Same Environments. American Journal of
Human Genetics 1994;55:566-73.
57. DeFronzo RA, Ferrannini E. Insulin resistance. A multifaceted syndrome
responsible for NIDDM, obesity, hypertension, dyslipidemia, and atherosclerotic
cardiovascular disease. Diabetes Care 1991;41:173-94.
58. Martin BC et al. Familial clustering of insulin sensitivity. Diabetes 1992;41:850-
4.
59. Meigs JB et al. Risk variable clustering in the insulin resistance syndrome - The
Framingham Offspring Study. Diabetes 1997;46:1594-600.
60. Zimmet P, Alberti KGMM, Shaw J. Global and societal implications of the
diabetes epidemic. Nature 2001;414:782-7.
13
61. Bassuk SS, Manson JE. Epidemiological evidence for the role of physical activity
in reducing risk of type 2 diabetes and cardiovascular disease. Journal of Applied
Physiology 2005;99:1193-204.
62. Gill JMR, Cooper AR. Physical Activity and Prevention of Type 2 Diabetes
Mellitus. Sports Medicine 2008;38:807-24.
63. Haapanen N et al. Association between leisure time physical activity and 10-year
body mass change among working-aged men and women. International Journal of
Obesity 1997;21:288-96.
64. Helmrich SP et al. Physical activity and reduced occurrence of non-insulin-
dependent diabetes mellitus. New England Journal of Medicine 1991;325:147-52.
65. Hu FB et al. Walking Compared With Vigorous Physical Activity and Risk of
Type 2 Diabetes in Women . JAMA 1999;282:1433-9.
66. Hu G et al. Physical activity, body risk of type 2 diabetes in patients with normal
or impaired glucose regulation. Archives of Internal Medicine 2004;164:892-6.
67. Meisinger C et al. Leisure time physical activity and the risk of type 2 diabetes in
men and women from the general population. Diabetologia 2005;48:27-34.
68. Norman A et al. Total physical activity in relation to age, body mass, health and
other factors in a cohort of Swedish men. International Journal of Obesity 2002;26:670-5.
69. Petersen L, Schnohr P, Sorensen TIA. Longitudinal study of the long-term
relation between physical activity and obesity in adults. International Journal of Obesity
2004;28:105-12.
70. Villegas R et al. Physical activity and the incidence of type 2 diabetes in the
Shanghai women's health study. International Journal of Epidemiology 2006;35:1553-62.
71. Manson JE et al. Physical-Activity and Incidence of Non-Insulin-Dependent
Diabetes-Mellitus in Women. Lancet 1991;338:774-8.
72. Perseghin G et al. Increased glucose transport-phosphorylation and muscle
glycogen synthesis after exercise training in insulin-resistant subjects. New England
Journal of Medicine 1996;335:1357-62.
73. Davies MJ et al. Effects of moderate alcohol intake on fasting insulin and glucose
concentrations and insulin sensitivity in postmenopausal women - A randomized
controlled trial. Jama-Journal of the American Medical Association 2002;287:2559-62.
74. Friedman GD, Klatsky AL. Is alcohol good for your health? New England Journal
of Medicine 1993;329:1882-3.
75. Howard AA, Arnsten JH, Gourevitch MN. Effect of alcohol consumption on
diabetes mellitus - A systematic review. Annals of Internal Medicine 2004;140:211-9.
14
76. Marques-Vidal P et al. Relationships between alcoholic beverages and
cardiovascular risk factor levels in middle-aged men, the PRIME study. Atherosclerosis
2001;157:431-40.
77. Wei M et al. Alcohol intake and incidence of type 2 diabetes in men. Diabetes
Care 2000;23:18-22.
78. Manson J.E. et al. A prospective study of cigarette smoking and the incidence of
diabetes mellitus among us male physicians . American Journal of Medicine
2000;109:538-42.
79. Nakanishi N et al. Cigarette smoking and risk for impaired fasting glucose and
type 2 diabetes in middle-aged Japanese men. Annals of Internal Medicine 2000;133:183-
91.
80. Wannamethee SG, Shaper AG, Perry IJ. Smoking as a modifiable risk factor for
type 2 diabetes in middle-aged men. Diabetes Care 2001;24:1590-5.
81. Boyle JP et al. Projection of diabetes burden through 2050 - Impact of changing
demography and disease prevalence in the US. Diabetes Care 2001;24:1936-40.
82. King H, Aubert RE, Herman WH. Global burden of diabetes, 1995-2025 -
Prevalence, numerical estimates, and projections. Diabetes Care 1998;21:1414-31.
83. Eddy DM, Schlessinger L. Validation of the archimedes diabetes model. Diabetes
Care 2003;26:3102-10.
84. Hanley AJG et al. Prediction of type 2 diabetes using simple measures of insulin
resistance - Combined results from the San Antonio Heart Study, the Mexico City
Diabetes Study, and the Insulin Resistance Atherosclerosis Study. Diabetes 2003;52:463-
9.
85. Ito C et al. Prediction of diabetes mellitus (NIDDM). Diabetes Research and
Clinical Practice 1996;34:S7-S11.
86. Herman WH et al. A new and simple questionnaire to identify people at increased
risk for undiagnosed diabetes. Diabetes Care 1995;18:382-7.
87. Lindstrom J, Tuomilehto J. The diabetes risk score: a practical tool to predict type
2 diabetes risk. Diabetes Care 2007;26:725-31.
15
2. A Population Based Prediction Algorithm for the Development of Physician Diagnosed
Diabetes Mellitus: Development and Validation of the Diabetes Population Risk Tool
(DPoRT)
2.1 Abstract
Background: Diabetes prevalence is increasing in most countries and its burden poses a serious
threat to public health. National estimates of the magnitude of the upcoming diabetes epidemic
are needed to understand the distribution of diabetes risk in the population and inform health
Objective: The objective of this study is to create and validate a population based risk prediction
tool for incident diabetes mellitus (DM) using widely available public data.
Methods: Using a cohort design that links baseline risk factors to a validated population-based
diabetes registry, a model (Diabetes Population Risk Tool or DPoRT) to predict 9-year risk for
diabetes using commonly-collected national survey data was developed and validated. The
development cohort was the Ontario‘s 1996/7 National Population Health Survey (NPHS) linked
to the validated Ontario Diabetes Database (ODD), a provincial component of the National
Diabetes Surveillance System (NDSS) in Canada. Variables were restricted to factors routinely
measured in the population. The probability of developing diabetes was modeled using sex-
specific Weibull survival functions for those > 20 years, without diabetes and not pregnant at
baseline (N = 19,861). The model was validated in two external validation cohorts: the 2000/1
Canadian Community Health Survey (CCHS) in Ontario (N=26,465) and 1996/7 NPHS in
Manitoba (N= 9,899), both linked to administrative data for NDSS-defined physician diagnosed
diabetes. Predictive accuracy was assessed by comparing observed physician diagnosed diabetes
rates with predicted risk estimates from DPoRT. Discrimination of the model was assessed
16
using a C statistic and calibration was assessed with the Hosmer-Lemeshow chi-square statistic
(χ2H-L).
Results. In the development cohort, 9-year age-standardized diabetes rates were 6.7% for males
and 5.6% for females. Predictive factors included in the final model were: BMI, age, ethnicity,
hypertension, immigrant status, smoking, education status and heart disease. The DPoRT-
predicted incidence closely agreed with the observed diabetes incidence rates in the validation
cohorts. Calibration measured using the Hosmer-Lemeshow was satisfied (χ2H-L <20) and overall
observed and predicted differed by less than 0.4%. Good discrimination was found in both
validation cohorts. In CCHS Ontario C= 0.78 95% CI (0.76, 0.80) for males and C= 0.77 95% CI
(0.75, 0.79) for females. In NPHS-Manitoba C = 0.81 95% CI (0.78, 0.83) for males and C= 0.81
Conclusions. The algorithm accurately predicted diabetes incidence in two validation cohorts
and demonstrated good discrimination. This algorithm can be used by health planners to estimate
diabetes incidence and quantify the impact of interventions using routinely collected survey data.
17
2.2 Introduction
In medicine, prediction tools are used to calculate risk, defined as the probability of developing a
disease or state in a given time period. Within the clinical setting, predictive studies such as the
Framingham Heart Score (1) - used to calculate the probability that a patient will develop
coronary heart disease - have contributed important advances in individual patient treatment and
disease prevention (2). Similarly, applying predictive risk tools to populations can provide
insight into the influence of risk factors, the future burden of disease in an entire region or
Global estimates place the number of people with diabetes at approximately 200 million,
and increasing rapidly (3). There is a growing concern that if left unchecked these trends may
slow or even reverse life expectancy gains in the US and other developed countries (4). Planning
for health care and public health resources can be informed by robust prediction tools. Estimates
of future diabetes incidence will alert policy makers, planners, and physicians to the extent and
urgency of the diabetes epidemic. In addition, a population prediction tool for diabetes can
identify the optimal target groups for new intervention strategies, and determine how extensive a
strategy must be to achieve the desired reduction in new cases. This insight can improve the
Clinical risk algorithms have been used at the population level but with considerable
challenges (5). In particular, clinical risk tools usually require clinical data which is infrequently
available at the population level. For diabetes in particular, several clinical risk prediction tools
exist, but they require clinical data that are collected infrequently or not at all at the population
level, such as fasting blood sugar (6-8), or require detailed information such as diabetes family
history (9, 10). In addition, some prediction tools apply only to specific sub-groups of the
18
population, such as to individual within specific age ranges or only to individuals that have other
co-morbid conditions, and thus are not suitable for national population estimates (11-13).
There are several ways in which an algorithm intended for population health application
differs from an algorithm intended for clinical use. For a population algorithm the input
for health policy decisions makers, available to a wide audience, and regularly collected so that
estimates can be updated frequently. Often algorithms used in clinical settings maximized
discrimination at the expense of accuracy, meaning that the algorithms do well at rank ordering
subjects but not as well at accurately predicting actual risk (14) . In contrast, algorithms used in
populations may favor accuracy over discrimination, because population health decision makers
rely more heavily on estimates of absolute risk and numbers of disease cases versus rank-
ordering of individuals.
The objective of this study was to design and validate a diabetes risk prediction tool for
use in populations using publicly available data. The creation and application of a population-
based risk algorithm for diabetes was feasible because the risk factors for diabetes are well
known and readily measured through self-reported questionnaires in many country‘s population
health surveys. Studies have overwhelmingly shown that the most influential risk factor for the
development of diabetes is excess weight, or obesity (15, 16), which is collected nationally in
Canada. DPoRT generates diabetes risk using data that are routinely collected in Canada through
health surveys.
The objective of this study is to create a risk algorithm for diabetes incidence that can be
applied at the level of populations using widely available public data. The Diabetes Population
19
Risk Tool or DPoRT was created and validated by individually linking three different provincial
20
2.3 Methods
The cohort used to develop the risk algorithm was derived from 23,403 Ontario residents that
responded to the 1996/7 National Population Health Survey (NPHS-ON) conducted by Statistics
Canada that had an overall 83% response rate (17) and were linkable to health administrative
databases in Ontario. In the NPHS, households were selected though stratified, multilevel cluster
sampling of private residences using provinces and/or local planning regions as the primary
sampling unit. The survey was conducted through telephone and all responses were self-
reported. Persons under the age of 20 (n = 2, 407) and those who had self-reported diabetes were
excluded (n = 894). Those who were pregnant at the time of the survey were also excluded (n =
241), due to the fact that baseline Body Mass Index (BMI) could not be accurately ascertained,
leaving a total of 19,861 individuals (Figure 1). Sixty-six males were further excluded when
applying the algorithm due to missing baseline BMI resulting in 9,177 males and 10,618 females
in the final cohort. Respondents from the survey were individually linked to a chart-validated
registry of physician-diagnosed diabetes, the Ontario Diabetes Database, allowing each member
of the cohort to be followed to determine their diabetes status for the next 9 years (to 2005/6).
The Manitoba respondents of the 1996/7 NPHS (NPHS-MB) were used as one of two validation
cohorts. The same sampling strategy and methodology used to carry out the Ontario portion of
the NPHS was used in Manitoba. There were N = 10,118 persons linked to health care data in
Manitoba. The same exclusion criteria were applied to NPHS-MB cohort resulting in 9,899
individuals for the validation. The second validation data set was from the Ontario portion of the
21
2000/1 Canadian Community Health Survey (CCHS, Cycle 1.1, N = 37,473), a national
telephone survey administered by Statistics Canada. The target population of the CCHS
consisted of persons aged 12 and over resident in private dwellings in all provinces and
territories, excepting those living on Aboriginal reserves, on Canadian Forces Bases, or in some
remote places. The CCHS included the same self-reported health questions as the NPHS. Like
the NPHS, this survey uses a multistage stratified cluster design and provides cross-sectional
data representative of 98% of the Canadian population over the age of 12 years, and attained an
80% overall response rate (18, 19). After the exclusion criteria were applied there were 26,465
The NPHS-MB cohort had a 9-year follow-up (1996-2005) and the CCHS-ON had a 5-
year follow-up (2000-2005) therefore the predicted risks from the DPoRT algorithm were
Females Males
N=23,403
1138 age < 20 years 1269 age < 20 years
N = 20, 996
241 pregnant at interview
N = 20, 755
452 Prior DM 442 Prior DM
N = 19, 861
No DM (N = 18, 451) DM (N = 1410)
N = 16, 819
22
Identifying respondents who develop diabetes
Survey data from development and validation cohorts were linked to provincial administrative
health care databases that include all persons covered under the government funded universal
health insurance plan. The diabetes status of all respondents in Ontario was established by
linking persons to the Ontario Diabetes Database (ODD). The Ontario Diabetes Database (ODD)
contains all physician diagnosed diabetes patients in Ontario identified since 1991. The database
was created using hospital discharge abstracts and physician service claims. A patient is said to
have physician diagnosed diabetes if he or she meets at least one of the following two criteria:
Clinical Modification code 250 (ICD9-CM) before 2002 or ICD-10 code E10 – E14 after 2002,
or (b) a physician services claim with a diabetes diagnosis (code 250) followed within two years
Individuals entered the ODD as incident cases when they were defined as having diabetes
according to the criteria described above. A hospital record with a diagnosis of pregnancy care
or delivery close to a diabetic record (i.e. a gestational admission date between 90 days before
and 120 days after the diabetic record date), were considered to relate to a diagnosis of
gestational diabetes and therefore were excluded. The ODD has been validated against primary
care health records, and demonstrated excellent accuracy for determining incidence and
regarding the vital statistics and eligibility for health care coverage for linked respondents was
captured from the Registered Persons Data Base (RPDB). The algorithm used to create the ODD
23
diabetes status in that province. The ODD algorithm is applied nationally using provincial
administrative registries (known as the National Diabetes Surveillance System (NDSS)) and has
Variables
To ensure that DPoRT would be widely applicable across different populations, variables
considered for the algorithm had to fulfil the following criteria: (i) be based on well established
evidence, (ii) be captured in a consistent manner across surveys (i.e. using the same questions),
(iii) be unlikely to be subject to serious self-reporting error (such as alcohol and dietary habits),
and (iv) be easily captured using survey data. All potential variables were obtained from self-
report responses to NPHS and CCHS telephone surveys, including: age, height and weight,
heart disease), ethnicity, immigration status, smoking status, highest level of achieved education,
total household income, alcohol consumption and physical activity (based on metabolic
equivalents). Body Mass Index (BMI) in kg/ m2 was used as an indicator of obesity. Derived
BMI, calculated by dividing the weight in kilograms by height squared (in meters squared)
directly from the NPHS, is only calculated for respondents aged 30 to 64; therefore, BMI was
calculated using weight and height values for the individual for those who fell outside the age
range of 30 – 64 (23). In order to facilitate future use in a variety of settings, all variable in the
risk algorithm were kept in the form which is released in the public use data.
24
Statistical Analysis
The goal of the model was to create a risk algorithm that would accurately predict diabetes risk
with high discrimination and calibration. Discrimination is the ability to differentiate between
those who are at high risk and those who are at low risk – or in this case those who will and will
not develop diabetes given a fixed set of variables. Calibration is the existence of close
agreement between observed diabetes risk – calculated by comparing observed diabetes rates and
predicted probabilities from DPoRT. Detailed definitions of discrimination and calibration are in
appendix 6.1. For each cohort member, the probability of physician-diagnosed diabetes was
assessed from the interview date until censoring for death or end of follow-up. The final model
was fit using a Weibull accelerated failure time model which provides simple and robust survival
probabilities. Importantly this model includes time in its equation, allowing the user to predict
probability for a range of follow-up periods. Alternative parametric models (exponential and log-
logistic) and a semi-parametric model (Cox Proportional Hazard model) were also considered
but either did not fit the data appropriately or did not perform as well, particularly when applying
to external cohorts. Furthermore, the ability to asses risk for a variety of follow-up times is an
Diabetes risk functions were derived separately for men and women above age 20
without prior diabetes diagnoses. The following covariates were considered as candidates for the
algorithm: age, BMI and their interactions, presence of hypertension, presence of heart disease,
ethnicity, immigration status, smoking status, education, income, and physical activity (based on
metabolic equivalents or METS). First unadjusted relationships with diabetes were inspected and
variables with a clinically significant unadjusted hazard ratio were considered for the final
model. Risk factors were added to the model in a nested fashion based on clinical importance,
25
and the marginal statistical and predictive significance of each group was evaluated, controlling
for variables already in the model. For both males and females, the BMI-age interactions were fit
first, then hypertension and heart disease, then ethnicity and immigration status, then remaining
risk factors. Covariates with missing information were also assessed for their relationship with
the outcome as missingness is often associated with higher risk, in this case, missing BMI for
females. Polynomial functions and splines were fit to try and accurate capture the relationship
Each variable was centred on the population mean before inclusion in the model to allow
for easier calibration with other cohorts. This means that when the algorithm is applied, all
variables are centered to the mean variables for the cohort to which it is being applied allowing
levels of risk to be reflective of the average baseline risk in the cohort of interest. Overall risk
(predicted probability) of diabetes for each person was calculated by multiplying the individual‘s
risk factor values by the corresponding regression coefficients, and summing the products (24).
The form of the model was assessed using likelihood ratio tests to compare nested parametric
models (25). A plot of Log(-logS(t))) vs log(t) was inspected for linearity to assess consistency
of the survival times with the Weibull distribution. Cox-Snell residual plots were also
constructed to assess the adequacy of the Weibull distribution assumption. The Cox-proportional
hazards model was fit to compare the risks with the parametric model.
Discrimination was measured using a C statistic modified for survival data developed by
Pencina et al. (26), analogous to the area under the ROC curve (27). Calibration was assessed
using a modified version of the Hosmer-Lemeshow χ2 statistic developed by Nam (28, 29). This
statistic is computed by dividing the validation cohort into deciles of predicted risk and
comparing the observed versus predicted risk in each decile using a chi-square statistic (see
26
appendix 6.1). To mark sufficient calibration χ2 = 20 was used as a cutoff (p<0.01), consistent
with D‘Agostino‘s method in validating the Framingham algorithms (28). Discrimination and
calibration statistics were also computed using the algorithm coefficients generated from the
validation cohort itself, and labeled ―own cohort‖. This was done to assess if the algorithms
generated from the validation cohort produced significantly different measures of predictive
accuracy than DPoRT. In addition to the Hosmer-Lemeshow test for calibration, graphical
representations of predicted and observed rates were produced to assess accuracy in prediction
across quintiles of risk. In order to generate these plots, predicted risks are ordered from smallest
to largest. The predicted probabilities are separated into quantiles of risk (in this study deciles or
quintiles). The number of predicted diabetes cases in each decile or quintile is determined by
summing the predicted risk within each quantile and the observed diabetes cases are also
summed within quantile. To present these numbers as proportions, the number of predicted and
observed cases are respectively divided by the number of individuals within each decile or
quintile. The observed and predicted cases are then plotted against each other at each quantile of
risk.
Re-calibration was achieved by substituting the mean values from the validation cohort to
define all variables (which were mean-centered) with the Framingham risk function. Due to
applied to predicted rates outside of Ontario (to reflect case ascertainment differences between
provinces). This adjustment was done by estimating the average ratio between observed and
predicted risk across decile (after taking into account differences in risk level through mean-
centering). The aggregate risk estimate was divided by the average amount of over-prediction
averaged across all risk groups. The assumption for this adjustment was that the differences are a
27
function of case-ascertainment (due to billing practices) between provinces. This approach has
been used in previous research involving the calibration of risk algorithms, including by Brindle
All estimates (including betas and variance estimates) incorporated bootstrap replicate
survey weights to accurately reflect the demographics of the Canadian population and account
for the survey sampling design based on selection probabilities and post stratification
adjustments. Variance estimates and 95% confidence intervals were calculated using bootstrap
survey weights (31, 32). All statistics were computed using SAS statistical software (version 9.1
28
2.4 Results
In the development cohort, 718 males and 692 females developed physician diagnosed diabetes
within the 9-year follow-up period (crude 9-year incidence rate of 7.78% and 6.13%
respectively). The age standardized (standardized to the 1991 Canadian population) 9-year
incidence rates in the development cohorts were 6.67 % for males and 5.59 % for females. The
5-year age standardized incidence rates in the development cohort were 3.59 % for males and
2.81% for females. In the Manitoba validation cohort 272 males and 258 females developed
physician diagnosed diabetes within the 9-year follow-up period (crude 9-year incidence rate of
7.22% and 4.75% respectively). The age standardized 9-year incidence rates in the development
cohorts were 6.55 % for males and 4.27 % for females. In the Ontario-CCHS validation cohort
559 males and 558 females developed physician diagnosed diabetes within the 5-year follow-up
period (crude 5-year incidence rate of 4.60% and 3.69% respectively). The age standardized
5-year incidence rates in the Ontario-CCHS validation cohort was 3.95 % for males and 3.35 %
for females. All baseline population characteristics in the derivation cohort and two validation
cohorts are shown in table 1. Both the Manitoba and Ontario validation cohorts differed from
the derivation cohort. There were similar in age distribution, however both NPHS-MB and
CCHS-ON had a higher proportion of obese individuals. The CCHS-ON cohort had significantly
higher number non-white ethnicities compared to the derivation cohort, whereas the derivation
cohort and Manitoba validation cohort were similar. Compared to the derivation cohort the
CCHS-ON had a higher baseline prevalence of hypertension and heart disease but a lower
prevalence of smoking. The Manitoba validation cohort had higher levels of hypertension and
29
Table 1. Baseline characteristics of development and validation cohorts. Categorical variables
are represented as a proportion (%) and continuous variables are represented as a mean/median.
MALES FEMALES
Development Validation Cohorts Development Validation Cohorts
Cohort Cohort
Risk Factor Ontario Manitoba Ontario Women Manitoba Ontario
NPHS NPHS CCHS NPHS NPHS CCHS
(N = 9,177) (N=4,670) (12, 020) (N = 10,618) (N=5,229) (N=14,445)
Body Mass Index 26.10/ 26.86/ 26.12/ 24.47/ 25.43/ 24.98/
(Kg/m2) 25.70 26.31 25.62 23.50 24.59 24.03
Age (mean/median) 44/42 44/42 44/42 46/42 47/42 46/42
30
For both males and females, diabetes risk was highly positively related to BMI and age (tables
2a&b). BMI was considered in both its continuous and categorical form. It was found the
relationship between BMI and diabetes was non-linear and complicated power transformations
were needed to correctly model the association. Continuous BMI not only had a complicated
exponential form but its relationship with diabetes varied significantly by age and sex. BMI
categories had the advantage of not being constrained to a specific mathematical function and
having regression coefficients that were easier to interpret than those from a series of polynomial
BMI terms. Ultimately, the best goodness-of-fit and calibration was achieved by categorizing
BMI and including its interactions with age since the effect of BMI was highly dependent on
age. This categorization was least likely to over-fit when applied to other data while at the same
remaining discriminating.
Non-white ethnicity, hypertension, and less than post secondary education were also
important factors associated with an increased risk of diabetes in the multivariate model. For
males, smoking and heart disease were also important independent risk factors that were found to
improve model characteristics; for women, immigrant status was an additional important risk
factor found to improve the model. Some variables were excluded from both models because
they did not improve the model or worsened predictive accuracy, including: income, physical
activity, and alcohol consumption. The majority of discrimination (C=0.70) was achieved by
including only age, BMI and their interaction. The sex-specific estimates for DPoRT algorithm
are shown in table 3. Figure 2 demonstrates how the risk coefficients in DPoRT were used to
31
Table 2a. DPoRT multivariate-adjusted hazard ratios and 95% confidence intervals for 9-year
physician diagnosed diabetes for males
Adjusted for all variables in the model simultaneously; all standard errors are computed
using survey bootstrap weights
32
Table 2b. DPoRT multivariate-adjusted hazard ratios and 95% confidence intervals for 9-year
physician diagnosed diabetes for females
Adjusted for all variables in the model simultaneously; all standard errors are computed
using survey bootstrap weights
33
Table 3. DPoRT Functions for predicting 9-year physician diagnosed diabetes for a) males and b) females
(a) (b)
Risk Factor Men Risk Factor Women
Intercept 10.5971 Intercept 10.5474
Hypertension Hypertension
No 0.00 Yes 0.00
Yes -0.2624 No -0.2865
Non-white Ethnicity Non-white Ethnicity
No 0.00 Yes 0.00
Yes -0.6316 No -0.4309
Heart Disease Immigrant Status
No 0.00 Yes 0.00
Yes -0.5355 No -0.2930
Current Smoker Education
No 0.00 < Post-secondary 0.00
Yes -0.1765 Secondary 0.2042
Education Age-BMI category
< Post-secondary 0.00 BMI<23*Age<45 0.00
Secondary 0.2344 23≤BMI<25*Age<45 -0.5432
Age-BMI category 25≤BMI<30*Age<45 -0.8453
BMI<23*Age<45 0.00 30≤BMI<35*Age<45 -1.4104
23≤BMI<25*Age<45 -1.2378 BMI≥35*Age<45 -2.0483
25≤BMI<30*Age<45 -1.5490 BMI = missing*Age<45 -1.1328
30≤BMI<35*Age<45 -2.5437 BMI<23*45≤Age<65 0.0711
BMI≥35*Age<45 -3.4717 23≤BMI<25*45≤Age<65 -0.7011
BMI<23*Age≥45 -1.9794 25≤BMI<30*45≤Age<65 -1.4167
23≤BMI<25*Age≥45 -2.4426 30≤BMI<35*45≤Age<65 -2.2150
25≤BMI<30*Age≥45 -2.8488 BMI≥35*45≤Age<65 -2.2695
30≤BMI<35*Age≥45 -3.3179 BMI = missing*45≤Age<65 -1.7260
BMI≥35*Age≥45 -3.5857 BMI<23*Age≥65 -1.0823
Scale 0.8049 23≤BMI<25*Age≥65 -1.1419
25≤BMI<30*Age≥65 -1.5999
30≤BMI<35*Age≥65 -1.9254
BMI≥35*Age≥65 -2.1959
BMI = missing*Age≥65 -1.8284
Scale 0.7814
34
Figure 2. Example use of DPoRT to predict the 9 year risk of a diabetes mellitus for a specific
high-risk male
2
Profile: Male; 55 years old; BMI = 29 kg/m , hypertension, white, non- smoker; heart disease & hypertension,
graduated secondary school
**Note all variables are centered
μ = 8.8816
m = log(365.25*9) – μ / σ = 8.09781– 8.8816/ 0.8049
m = -0.97374
9-year predicted risk for developing diabetes is:
-0.97374
P = 1-exp(-e )
P = 0.314542896
or 31.45%
The algorithm was re-calibrated to the mean of the validation cohorts. For example, the
coefficient for the ‗non-white ethnicity‘ variable in males is -0.6316 (table 3a). If the variable
were not mean centered this coefficient would be multiplied by 1 or 0 given the status of
ethnicity for that individual; therefore for someone with non-white ethnicity it would be -0.6316
x 1. By mean centering the value of the individual is related to the baseline rate of non-white
ethnicity in the cohort to which the algorithm is being applied. If the cohort where the algorithm
is being applied has a 10.828% baseline rate of non-white ethnicity for males, then the resulting
Observed and predicted diabetes rates differed ≤ 0.4% in the two validation cohorts. In
the CCHS cohort, the mean-calibrated DPoRT 5-year predicted (and observed) DM incidence
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
rates were 4.2% (4.6%) for males and 3.4% (3.7%) for females. Observed and predicted risks
for the CCHS cohorts, by decile of risk, are shown in figure 2. In the NPHS-MB cohort, the
mean-calibrated DPoRT 9-year DM predicted (observed) incidence rates were 8.2% (7.1%) for
males and 6.6% (4.7%) for females. After adjusting for case ascertainment differences between
provinces, the predicted (observed) incidence rates were 7.0% (7.1%) for males and 5.1% (4.7%)
Figure 2. Overall calibrated predicted versus observed physician diagnosed diabetes rates for
males and females in the two validation datasets.
Observed Predicted
8.0%
7.0%
6.0%
Diabetes incidence rate (%)
5.0%
4.0%
3.0%
2.0%
1.0%
0.0%
Males Females Males Females
CCHS-ON validation cohort NPHS-MB validation cohort
36
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Due to the smaller sample size in the NPHS-MB cohort, the risks are reported by quintiles. R-
squared between observed diabetes incidence rates and predicted probabilities from the DPoRT
algorithm across deciles of risk in both validation data sets exceeded 98% for both males and
females after re-calibration. C-statistics when applying DPoRT to the validation cohorts were
high (0.77 – 0.80) and were not appreciably lower than those generated from the ―own cohort‖
models (Table 4). Hosmer-Lemeshow χ2 statistics to assess calibration are shown in table 4.
For men and women, sufficient calibration was demonstrated in the CCHS validation cohort after
mean adjustment (χ2 <20). For NPHS-MB, to achieve goodness of fit additional calibration was
37
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Figure 3. Calibrated predictions for incidence of physician diagnosed diabetes for males and
females in two validation datasets across deciles or quintiles of risk.
Males Females
A. DPoRT applied to Ontario CCHS †
N = 12,018 C = 0.77 95%CI (0.76, 0.79) N = 14,442 C = 0.76 95%CI (0.74, 0.77)
N = 4,670 C = 0.79 95%CI (0.77, 0.82) N = 5,229 C = 0.80 95%CI (0.77, 0.82)
38
Table 4. C statistics with 95% confidence intervals and calibration χ2 statistics for DPoRT and
cohorts‘ own functions.
Men Women
NPHS ON CCHS ON NPHS MB NPHS ON CCHS ON NPHS MB
C statistic 95% CI
DPoRT 0.77 0.77 0.79 0.78 0.76 0.80
(0.76, 0.79) (0.76, 0.79) (0.77, 0.82) (0.76, 0.79) (0.74, 0.77) (0.77, 0.82)
Own Function† 0.80 0.78 0.80 0.80 0.77 0.80
(0.78, 0.83) (0.76, 0.79) (0.77, 0.82) (0.78, 0.83) (0.75, 0.79) (0.77, 0.82)
Calibration χ2
Uncalibrated DPoRT 4.33 13.23 136.13 5.22 24.84 35.07
Mean Calibrated … 13.04 32.616 … 18.27 25.83
DPoRT‡
Mean and CA Calibrated … … 18.35 … … 17.61
DPoRT¥
Own function … 8.89 8.32 … 10.44 4.88
† Own function is the factors of the algorithm applied using coefficients derived from the validation cohort‘s own
data
‡Calibrated DPoRt is function adjusted using the validation cohort‘s own means for factors; does not include case
ascertainment adjustment
¥ Based on binary H-L statistic (approximation to survival based H-L)
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
2.5 Discussion
This study demonstrated that diabetes risk can be accurately predicted at the population level
using self-reported age, sex, BMI and other measures available in population health surveys. In
addition to displaying good discrimination, DPoRT-predicted rates closely agreed with observed
rates for both males and females in both external validation cohorts, and this agreement was
generally maintained across deciles and quintiles of risk. To my knowledge, DPoRT is the first
validated risk tool that is integrated into commonly-collected population health survey data.
DPoRT offers advantages over existing methods used to estimate future diabetes risk in
populations. Previous studies that estimate future diabetes burden have either extrapolated
risk factors with various assumptions (3, 33-36). Previous studies of diabetes lifetime risk and
life expectancy are not predictive, rather they describe diabetes from a life-course perspective
using a period or stationary population approach (36, 37). Other studies focus on overall diabetes
burden, a useful approach, but one which does not enable users to directly assess the impact of
risk factors, such as BMI, on future diabetes. Furthermore, these studies did not assess how
future diabetes can be prevented by targeting risk factors since they do not directly quantify the
Complex modeling and simulation studies differ from the approach used in this study in
that they use additional information on how populations and risk factors change over time (38,
39). Other simulation studies add more detailed clinical information such as fasting blood sugar
level or information on diabetes family history, data not available at the population level. A
strength of these simulation models is that they can combine together different data sources and
study findings (40). However, these models are complex and often represent clinical or
40
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
theoretical populations, making their estimates difficult to validate in external populations that
are meaningful for population health planning. DPoRT could be incorporated into simulation
models that consider future changes in population composition and risk factors.
The nature of diabetes risk allowed us to discriminate and explain risk using a limited number of
variables – most importantly BMI. Discrimination of DPoRT (C statistic 0.77 – 0.80 and wide
range of predictive risk across deciles) is as high as or higher than many clinical risk prediction
tools used in clinical practice. The most stringent test of model accuracy is the application of the
model to a different population (41, 42). This study demonstrates that DPoRT is discriminating
and accurate in two external populations that varied across geography and time. The algorithm
was further calibrated using population means, which may attenuate differences between
populations since risk estimates are relative to baseline risk in the population.
Given current data in most countries, DPoRT is a more balanced approach to estimating
diabetes risk than methods used in previous research. A number of important clinical values are
excluded from DPoRT, such as hip to waist ratio, waist circumference, fasting blood glucose,
and family history (10;41;42). Although these variables may be clinically important for
assessing diabetes risk, adding these, or other detailed anthropometric measures, is not feasible
because they are not routinely collected in most populations. These omitted variables are
unlikely to have a major impact on the performance characteristics of the model due to the
clustering of risk factors, particularly when dealing with abnormalities of the metabolic system
(43-47). Variables not included in DPoRT, such as family history of diabetes or poor diet, are
also associated with the clustering of metabolic risk factors that are included in the algorithm
such as hypertension and BMI. There is increasing awareness that simple clinical risk tools,
including those with self-report risk factor data, perform as well or better than complex models.
41
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
(13, 43, 45-52). Adding more predictors or clinical measures results in negligible improvements
to discrimination, and this is likely due to the high correlation of these risk factors with obesity
and other variables already included in the model. Previous study have shown that additional
predictive variables need to have a high independent risk (odds ratio greater than 6.9 or greater)
achieved (53). This phenomenon was corroborated in the creation of DPoRT, as maximal levels
Obesity is the most important factor in predicting diabetes risk. BMI is the most
commonly used marker of obesity; however measures of central obesity may capture the entire
risk domain in a more comprehensive manner and be more meaningful across all age groups
(54). This may be considered a limitation of DPoRT since measures of central obesity are not
routinely collected on the population level and thus are not included. If included, these estimates
may increase both discrimination and accuracy of the algorithm; however debate about the
clinical utility of these measures still continues. A recent meta-analysis has shown that there is
no evidence of difference in estimates associated with incident diabetes between BMI, waist
circumference and waist/hip ratios (55). Furthermore, algorithms to identify individuals for
weight loss in populations did not differ if using BMI or waist circumference (56).
variables that were: based on established evidence, remained stable over time, were unlikely to
be subject to serious measurement error (such as alcohol and dietary habits), and were easily
captured using survey data in different populations. For example, physical activity has been
shown to have a protective effect on diabetes incidence (57) but was removed from the final
algorithm due to the inability to capture this in a reliable and reproducible manner across studies,
42
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
and because of it marginal improvement in the discrimination of diabetes risk in our creation
Using self-report measures is a limitation which could affect predictive risk accuracy
because these measures may be more subject to reporting error and/or bias than measured
anthropometric measures. In general there is a high correlation between self-report and measured
height and weight; however, validation studies do show that weight tends to be slightly
underestimated and height trends to be slightly overestimated and as a result reported BMI is
generally lower than measured BMI (58-60). The effect of self-reported BMI may depend on the
population where the algorithm is being applied since these patterns have been shown to vary
across gender and socioeconomic status (61-63). The ability of DPoRT‘s predictive estimates to
agree with observed diabetes risk in different populations will be reduced if systematic errors
associated with responses vary across populations or time. In chapter 3.1, simulation modeling
is used to specifically explore the impact of potential reporting biases on risk prediction.
true diabetes status (diagnosed plus undiagnosed). The estimates from DPoRT may exclude
diabetics in the population who are not yet identified by a physician. This could reflect patients
with less severe disease and/or poorer access to medical care. Physician-diagnosed diabetes is
currently the most commonly measured definition of diabetes at the level of populations.
Although estimates put the true prevalence of diabetes higher, advocates of the physician-
diagnosed outcome argue that it is meaningful to people with recognized diabetes and to the
treatment of these diagnosed patients within the health care system. If diabetes testing/screening
increases over time, predicted estimates could be lower than the observed estimates (under the
43
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
assumption that this would lead to increased case detection). DPoRT has been found to be
accurate in different populations for different time periods; however, DPoRT could be re-
calibrated to predict total diabetes cases using revised information on screening/testing or using
estimates of the number of undiagnosed cases relative to diagnosed cases in the population.
Finally, the potential for inaccuracy increases the longer into the future the predictions are made
or when unforeseen changes occur; therefore, it is recommended that predictions from DPoRT
are updated frequently by using the most recent data, limiting predictive calculations to 9 years
or less, and validating the risk tool in the population where it is being applied. This tool was
developed in Canada and is likely most appropriate in the Canadian setting; however like other
risk tools, DPoRT may be transportable once validated and calibrated as needed. Careful
interpretation and adjustment for contextual variables, such as immigrant status, will need to be
Curbing the diabetes epidemic has been identified by several governments and health
policy makers as one of their top priorities for improving, and even maintaining, the health of
their nations. Population-based prediction models such as DPoRT can be used to improve health
planning, explore the impact of prevention strategies, and enhance understanding of the
distribution of disease populations. This study demonstrates that DPoRT accurately predicts
diabetes incidence and is effective at discriminating risk at the population level. This algorithm
can be used by health planners to estimate diabetes incidence, to stratify the population by risk
and quantify the impact of interventions using routinely collected survey data. Validation of
DPoRT will continue as more effective methods to quantify variation in case ascertainment in
different populations are developed. As the surveillance of risk factors and diabetes advances,
44
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
DPoRT can be adapted to become even more accurate, while maintaining its accessibility for
decision makers.
45
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
2.6 References
46
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
15. Ford ES, Giles WH, Dietz WH. Prevalence of the metabolic syndrome among US
adults - Findings from the Third National Health and Nutrition Examination Survey.
Jama-Journal of the American Medical Association 2002;287:356-9.
16. Hu FB et al. Diet, Lifestyle, and the Risk of Type 2 Diabetes Mellitus in Women.
New England Journal of Medicine 2001;345:790-7.
17. Statistics Canada. 1996-97 NPHS Public Use Microdata Documentation. 1999.
Ottawa.
18. Statistics Canada. Canadian Community Health Survey Methodological
Overview. Health Reports 2002;13:9-14.
19. Canadian Community Health Survey 2000–2001. Ottawa: 2003.
20. Hux JE, Ivis F. Diabetes in Ontario. Diabetes Care 2005;25:512-6.
21. Lipscombe LL, Hux JE. Trends in diabetes prevalence, incidence, and mortality
in Ontario, Canada 1995-2005: a population-based study. Lancet 2007;369:750-6.
22. Health Canada. Responding to the Challenge of Diabetes in Canada. 2003.
Ottawa, ON.
23. Statistics Canada. 1996-7 National Population Health Survey: Derived Variable
Specifications. 1999. Ottawa.
24. Odell PM, Anderson KM, Kannel WB. New Models for Predicting
Cardiovascular Events. Journal of Clinical Epidemiology 1994;47:583-92.
25. Farewell VT, Prentice RL. A study of distributional shape in life testing.
Technometrics 1977;19:69-75.
26. Pencina M, D'Agostino RB. Overall C as a measure of discrimination in survival
analysis: model specific population value and confidence interval estimation. Statistics in
Medicine 2004;23:2109-23.
27. Campbell G. General Methodology I: Advances in statistic methodology for the
evaluation of diagnostic and laboratory tests. Statistics in Medicine 2004;13:499-508.
28. D'Agostino RB et al. Validation of the Framingham Coronary Disease Prediction
Scores. JAMA 2001;286:180-7.
29. Nam B-H. Discrimination and Calibration in Survival Analysis: Extension of the
ROC Curve for Descrimination and Chi-square test for Calibration. 2000. Boston
University.
30. Brindle P et al. Predictive accuracy of the Framingham coronary risk score in
British men: prospective cohort study. BMJ 2003;327:1267.
47
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
31. Yeo D, Mantel H, Lui TP. Bootstrap variance estimation for the National
Population Health Survey. 778-783. 1999. Baltimore, American Statistical Association.
32. Kovacevic MS, Mach L, Roberts G. Bootstrap variance estimation for predicted
individual and population-average risks. Proceedings of the Survey Research Methods
Section. 2008. American Statistical Association.
33. Boyle JP et al. Projection of diabetes burden through 2050 - Impact of changing
demography and disease prevalence in the US. Diabetes Care 2001;24:1936-40.
34. King H, Aubert RE, Herman WH. Global burden of diabetes, 1995-2025 -
Prevalence, numerical estimates, and projections. Diabetes Care 1998;21:1414-31.
35. World Health Organization. Report of a WHO consultation on obesity, Obesity:
preventing and managing the global epidemic. 1998. Geneva, World Health
Organization.
36. Narayan KMV et al. Lifetime risk for diabetes mellitus in the United States.
JAMA 2003;290:1884-90.
37. Manuel D, Schultz S. Health-related quality of life and health-adjusted life
expectancy of people with diabetes in Ontario, Canada, 1996-1997. Diabetes Care
2004;27:407-14.
38. Forouhi NG et al. Diabetes prevalence in England, 2001 - estimates from an
epidemiological model. Diabetic Medicine 2006;23:189-97.
39. Mainous AG et al. Impact of the population at risk of diabetes on projections of
diabetes burden in the United States: an epidemic on the way. Diabetologia 2007;50:934-
40.
40. Ford ES et al. Explaining the decrease in US deaths from coronary disease, 1980-
2000. New England Journal of Medicine 2007;356:2388-98.
41. Altman DG, Royston P. What do we mean by validating a prognostic model?
Statistics in Medicine 2000;19:453-73.
42. Harrell FE, Lee KL, Mark DB. Multivariable prognostic models: Issues in
developing models, evaluating assumptions and adequacy, and measuring and reducing
errors. Statistics in Medicine 1996;15:361-87.
43. Carmelli D, Cardon LR, Fabsitz R. Clustering of Hypertension, Diabetes, and
Obesity in Adult Male Twins - Same Genes Or Same Environments. American Journal of
Human Genetics 1994;55:566-73.
44. DeFronzo RA, Ferrannini E. Insulin resistance. A multifaceted syndrome
responsible for NIDDM, obesity, hypertension, dyslipidemia, and atherosclerotic
cardiovascular disease. Diabetes Care 1991;41:173-94.
48
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
45. Lorenzo C et al. The metabolic syndrome as predictor of type 2 diabetes - The
San Antonio Heart Study. Diabetes Care 2003;26:3153-9.
46. Meigs JB et al. Risk variable clustering in the insulin resistance syndrome - The
Framingham Offspring Study. Diabetes 1997;46:1594-600.
47. Schmidt MI et al. Clustering of dyslipidemia, hyperuricemia, diabetes, and
hypertension and its association with fasting insulin and central and overall obesity in a
general population. Metabolism-Clinical and Experimental 1996;45:699-706.
48. Mainous AG et al. A Coronary heart disease risk score based on pateint-reported
information. The American Journal of Cardiology 2007;1236-41.
49. Ambler G, Brady AR, Royston P. Simplifying a prognostic model: a simulation
model based on clinical data. Statistics in Medicine 2002;21:3803-22.
50. Chambless LE et al. Prediction of ischemic Stroke Risk in the Atherosclerosis
Risk in Communities Study. American Journal of Epidemiology 2004;160:259-69.
51. Wilson PWF et al. Prediction of coronary heart disease using risk factor
categories. Circulation 1998;97:1837-47.
52. DeFronzo RA. Insulin Resistance, Hyperinsulinemia, and Coronary-Artery
Disease - A Complex Metabolic Web. Journal of Cardiovascular Pharmacology
1992;20:S1-S16.
53. Pepe MS et al. Limitations of the odds ratio in gauging the performance of a
diagnostic, prognostic, or screening marker. American Journal of Epidemiology
2004;159:882-90.
54. Flint E, Rimm E. Obesity and cardiovascular disease risk among the young and
old - is BMI the wrong benchmark? International Journal of Epidemiology 2006;35:187-
9.
55. Vazquez G et al. Comparison of body mass index, waist circumference, and
Epidemiology 2007;29:115-28.
56. Mason C, Katzmarzyk PT, Blair SN. Eligibility for obesity treatment and risk of
mortality in men. Obesity Research 2005;13:1803-9.
57. Bassuk SS, Manson JE. Epidemiological evidence for the role of physical activity
in reducing risk of type 2 diabetes and cardiovascular disease. Journal of Applied
Physiology 2005;99:1193-204.
58. Nawaz H et al. Self-reported weight and height: implications for obesity research.
Journal of Preventive Medicine 2001;20:294-8.
59. Rowland M. Self-reported height and weight. American Journal of Clinical
Nutrition 2007;52:1125-33.
49
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
60. Shields M, Gorber SC, Tremblay MS. Estimates of obesity based on self-report
versus direct measures. Health Reports 2008;19:1-16.
61. Bostrom G, Diderichsen F. Socioeconomic differentials in misclassification of
height, weight and body mass index based on questionnaire data. International Journal of
Epidemiology 1997;26:860-6.
62. Niedhammer I et al. Validity of self-reported weight and height in the French
GAZEL cohort. International Journal of Obesity 2000;24:1111-8.
63. Wardle K, Johnson F. Sex differences in the association of socioeconomic status
with obesity. Internation Journal of Obesity and Related Metabolic Disorders
2002;26:1144-9.
50
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
3.1.1 Abstract
Background: Self-reported height and weight can be subject to several types of measurement
error. The impact of this measurement error on the estimated risk, discrimination, and calibration
of a model that uses height and weight expressed as body mass index (BMI) to predict 10-year
Objective: To use simulation to quantify and describe the effect of random and systematic
measurement error in self-reported height and weight on the performance of a model for
predicting diabetes. The three performances measures used are predicted risk, accuracy, and
discrimination.
and predicted probabilities of developing diabetes and C statistic were measured under various
levels of random and systematic error. Simulations were done 500 times with sample sizes
values similar to those generated by the algorithm that was derived from actual population data.
Increasing levels of random error in height and weight reduced the calibration and discrimination
of the risk algorithm. Furthermore, random error biased the predicted risk upwards. Systematic
51
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
error biased predicted risk in the direction of the under- or over-estimation and reduced
Conclusions: This study demonstrates that random and systematic errors have the potential to
influence the performance of risk algorithms. Overall predicted risk should be carefully
considered in the context of potential measurement errors. Further research that quantifies the
amount and direction of error can improve the performance of prediction tools by allowing for
52
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
3.1.2 Introduction
Measurement error is said to occur when the measured value of a variable does not equal the
‗true‘ value. Measurement error is a major concern in epidemiologic research and has been
discussed extensively in the literature(1-5). In general there are two classes of measurement
error: (i) random error is error that on average is equal to zero (ii) systematic error is error that on
average is not equal to zero (6). Measurement error has been mainly examined with respect to its
effect on risk estimates, such as risk ratios or hazard ratios(1, 7, 8). This research has led to
improvements in the critical appraisal and interpretation of epidemiologic findings and has
provided guidance for measurement in future studies. While useful for understanding the effects
of error on etiological estimates of disease, the findings from these studies do not directly apply
to risk algorithms.
Risk prediction is a key aspect of clinical work and has been recently has been applied to
population health through the Diabetes Population Risk Tool (DPoRT) (Chapter 2). Accurate and
valid prediction of the probability of developing a disease given a set of baseline risk factors is a
valuable tool for clinical management, health policy and population health decision making. In
risk algorithms disease prediction is based on a set of baseline variables that may contain
measurement error that could affect the prediction, discrimination and accuracy of the risk tool.
health surveys to make predictions. National health surveys are conducted through telephone
interviews and are based on self-reported responses. These self-reported responses can result in
random error due to imperfect recall or misunderstanding of the question. They can also result in
systematic error as a result of psychosocial factors, such as social desirability. The influence of
53
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
error from self-report surveys on disease prediction has not been systemically studied. In
particular, the influence that measurement error has on predictive accuracy has never been
could be made to correct for these errors and thus improve the accuracy and validity of a risk
algorithm. Furthermore, other developers of population risk tools can use this information to
better weigh the pros and cons of using different types of data.
The objective of this study is to use simulation to understand the effect of measurement
error in self-reported risk factors on the performance of a simple risk algorithm to predict
diabetes. This study will focus on the measurement of Body mass index (BMI), a function of
height and weight, because it has the greatest influence on diabetes risk (9-13). In general, there
is a high correlation between self-report and measured height and weight (14); however, a recent
systematic review to summarize the empirical evidence regarding the concordance of objective
and subjective measures of assessing height and weight found that the general trend was an
underestimation of weight and an overestimation of height (15). Currently, there are no linked
data available that allow assessment of self-reported versus measured or ‗true‘ BMI
simultaneously with the diabetes outcome. Simulation allows one to examine the impact of
varying levels of measurement error on discrimination and accuracy. These results will inform
DPoRT and assist in understanding how its predictions and validation are affected by the use of
Research from studies that examine the impact of measurement error on relative risk
shows that, in general, random error results in an estimate that is closer to the null value; whereas
54
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
the nature of the bias (1, 4, 7). These findings were the basis for the following a priori
hypotheses:
(i) Random (random) measurement error will affect both discrimination and calibration of a
model. The presence of measurement error will increase the observed variance in BMI and thus
widen the distribution of BMI in both diabetics and non-diabetics. This will affect discrimination
by increasing the overlap between the BMI distribution of those that are likely to develop
diabetes and the distribution of those that are not. This increased overlap will make it more
difficult to resolve them into populations of diabetics and non-diabetics. Calibration will be
affected because the predicted risk estimate will be higher or lower than the true estimate, due to
the over- or under-estimation of weight and height, and will result in misclassification of
(ii) Systematic error will have minimal affects on discrimination but significant effects on
predicted risk; however, this should not influence the ability to rank order subjects, under the
assumption that the error is not differential between persons that develop or not develop the
disease. The same rationale outlined in (i) applies to calibration under this scenario in that the
over- or under-estimation of weight and height will reduce the concordance with the observed
estimate. However the effect will be more apparent here since the error is occurring all in one
direction.
55
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
(iii) Random error will not affect the overall predicted risk of the model and systematic error
will influence the predicted risk in the direction of the systematic error. Random error increases
the variability, which increases the range of predicted risk equally in both directions and thus,
will not skew the distribution in any particular direction. Therefore, the overall predicted risk
estimate should be similar to the predicted risk estimate from the model without measurement
error.
56
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
3.1.3 Methods
Simulating Data
developing diabetes were created to closely match the actual data used to create DPoRT. For
this simulation, the algorithm predicting the binary outcome of physician-diagnosed diabetes was
modeled using a logistic regression form equation and included BMI and its quadratic term as
where p represents the probability of developing BMI in the next 10 years and is calculated as:
To randomly generate a person with diabetes, first height (m) and weight (kg) are generated as
Gaussian random variables with means and variances obtained from DPoRT development
cohort. Using the specified values for the regression coefficients β0, β1 and β2 the logit was
calculated from which the probability P of the person having diabetes is calculated. Then a 0 / 1
Bernoulli random variable with probability P is generated with 1 or 0 meaning that the person
BMI is calculated as the ratio of height in meters (m) to weight in kilograms (kg)
squared. Both height and weight were varied individually from the true value assuming different
levels of error in the observed estimate. The size of the random measurement error is defined by
the intraclass correlation coefficient (ICC), the proportion of the overall observed variance
57
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
attributable to the ―true‖ variance between subjects. The systematic levels of measurement error
in height and weight were taken from a recent systematic review that summarized the empirical
evidence regarding the concordance of objective and subjective measures of height and weight
(16). These error values are consistent with a recent study examining the difference between
measured and self-reported height and weight in the CCHS (17). The observed mean and
standard deviation of BMI, taken from the National Population Health Survey (NPHS) (18),
along with the parameter estimates from the linkage with the Ontario Diabetes Database (ODD)
(19, 20) (See chapter 2) were used in our simulation. This simpler algorithm that included only
BMI and not the entire set of algorithm variables used in DPoRT was chosen because it has been
shown that it explains the majority of variation in diabetes incidence across individuals in a
population and produces C statistics as high as 0.73. For this study, the simulation data were
created such that the discrimination value was equal to that observed in the actual DPoRT
development cohort ~ 0.7. The simulated data were created using observed data assuming no
measurement error and then re-fit using various values of systematic and random error in height
and weight. .
58
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Random Error
where:
σ2error is the variance, which is attributed to measurement error and σ2observed ≥ σ2true
By these definitions, in the case where no error exists σ2observed = σ2true the intraclass correlation
coefficient (ICC), denoted as ρ, is an estimate of the fraction of the total measurement variance
σ 2true
(II) ρ =
σ 2true + σ 2error
From (I) we can express σ2true as σ2true = σ2observed - σ2error and substituting this expression back
σ 2observed − σ 2error
(III) ρ =
σ 2obs erved
59
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
If no error exists in the measurement ρ = 1.0 and σ2error = 0 and from (I) the σ2observed = σ2true. In
other words, all the observed variance is due to the true variance among individuals. An example
of the influence of ICC on the distribution of BMI can be seen in the appendix section 6.4.
Systematic Error
where:
ωi represents the average amount of over- or under- estimation from the true value
Traditionally the E(ωi) = 0 and thus the given measurement is, on average, equal to the reported
measurement; however, in this simulation ωi will represent various levels of over- and
underestimation.
60
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
In the simulation, the errors associated with self-reported height and weight will be varied
independently; however there is a strong correlation between height and weight that must be
considered. All starting correlation values were taken from the original data. The following
definition will be used to express the correlation between height and weight.
Let be:
A1 = σheight
B1 = σweight * rhw(observed)
B2 = (σ2weight − B12 )
B1 and B2 will be used to define the variance in weight, such that B1 defines the portion of the
The correlation of height and weight will vary according to ICC values such that
rhw for the true height and weight under measurement error was defined as:
rhw(observed)
rhwTRUE =
𝐼𝐶𝐶ℎ𝑒𝑖𝑔ℎ𝑡 × 𝐼𝐶𝐶𝑤𝑒𝑖𝑔ℎ𝑡
61
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Let be:
M1 = A1 x Z1
M2 = B1 x Z1 + B2 x Z2 or
The correlation is a function of the B2, which is common to both height and weight
The true height and weight values were generated using the same approach except rhw(true) will
62
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Simulated values of height and weight were generated by applying a random component to the
standard deviation of the variable. By applying the definitions of M1 and M2 above and using
𝐻𝑒𝑖𝑔 ℎ𝑡 𝑖 𝑊𝑒𝑖𝑔 ℎ𝑡 𝑖
𝐻𝑒𝑖𝑔ℎ𝑡 = and 𝑊𝑒𝑖𝑔ℎ𝑡 =
𝑛𝑖 𝑛𝑖
Height = 𝐻𝑒𝑖𝑔ℎ𝑡 + M1
Weight = 𝑊𝑒𝑖𝑔ℎ𝑡 + M2
𝑊𝑒𝑖𝑔 ℎ𝑡
BMI =
𝐻𝑒𝑖𝑔 ℎ𝑡 2
*note that when using ICC = 1.0 it follows that σ2observed = σ2true, which is the way the data are
currently used when applying DPoRT. This analysis quantified the influence of random error by
simulating scenarios where a proportion of the observed variance was attributed to random error.
These steps were replicated for the given sample (9,177 for males and 10,618 for
females) 500 times. Ten-year cumulative incidence was maintained between 7—10 %, in
accordance with Ontario male and female diabetes rates taken from the Ontario Diabetes
Database. The model was first fit assuming that there was no error in the observed height and
weight values to ensure that the expected percent rejection of the null hypothesis, which would
be equal to the type I error rate (or α) under the null distribution is equal to 5% and that the
63
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
average C statistics were consistent with those seen in the DPoRT development cohort. All
simulations were done using SAS statistical software (version 9.1 SAS Institute Inc, Cary, NC)
and random numbers were generated using the RAN family of functions (RANUNI and
RANNORM).
Calibration was measured using the Hosmer-Lemeshow goodness of fit statistic (χ2H-L) where
observed and expected values are compared across deciles of risk (23-25). This statistic is
computed by dividing the validation cohort into deciles of predicted risk. Consistent with
D‘Agostino‘s approach for evaluating observed and predicted values using risk algorithms the
value 20, the 99th percentile of a chi square with 8 degrees of freedom , was used as a cutoff to
mark sufficient calibration (26). Power to detect calibration was therefore defined, as the
proportion of simulations out of the 500 simulations that achieved the χ2H-L = 20 cutoff. This
proportion was calculated under the various error situations. Discrimination was measured using
a C-statistic, analogous to the area under the ROC curve (27). The C-statistic was calculated
using rank statistics and verified by the output from the LOGISTIC procedure. A study flow
diagram and an example of the SAS code for females can be seen in the Appendix section 6.4.
64
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Assumptions
(i) The random or systematic error does not differ between persons that develop and do not
(ii) Other sources of error including error in ascertainment of diabetes status, selection bias in the
65
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
3.1.4 Results
The fitted risk algorithms for males and females are shown in Table 1. These algorithms predict
the probability of developing diabetes in 10 years given the respondents‘ BMI, which is a
Table 1a. Starting values taken from 1996-7 National Population Health Survey used in
simulation.
Parameter Males Females
mean (standard deviation) (N = 9, 177) (N = 10, 618)
Height (m) 1.768 (0.075) 1.627(0.069)
Weight (kg) 81.624 (13.805) 64.761 (12.320)
BMI (m/kg2) 26.076 (3.995) 24.495 (4.586)
Correlation for rhw = 0.475 rhw = 0.311
height and weight (rhw)
10-year DM incidence 9.17% 7.35%
Table 1b. Properties of actual risk equation relating BMI to probabilities of developing diabetes
from NPHS cohort and values from the simulation model.
66
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
The results from the simulations are presented by error type (random and systematic error) and
by the influence on three areas: predicted risk; calibration and validation. Due to the amount of
data, most results are displayed in graphical form; however tables of the results can be found in
appendix 6.9.
Random Error
Random error in height and weight were examined for ICC = 1.0, 0.9, 0.8, 0.7, 0.6 and 0.5.
Under the presence of random error the overall probability of developing diabetes predicted from
the risk algorithm was higher than the estimate from the algorithm applied to the data without
random error. In other words, the presence of random error biased the overall predicted risk
estimate upwards. The differences between the predicted risk without error and with error were
relatively small, with the biggest differences being 0.99% higher than the truth (N = 90 more
cases) for males and 0.89% higher for females (N = 95 more cases) under extreme levels of
random error. Random error in weight had a bigger influence on the predicted risk than random
error in height. When the ICC for weight was held at 1.0 (i.e. no error in weight) but ICCs in
height were allowed to vary (from 1.0 to 0.5), the largest overestimate in diabetes risk was 0.30%
(28 more cases) in males and 0.16% (70 more cases) in females. However, when the ICC for
height was held at 1.0 but the ICCs for weight were allowed to vary, the largest overestimate was
0.70% (64 more cases) for females and 0.66% (39 more cases) for males (Figure 1). In the
presence of random error, the observed distribution of predicted risk is skewed right compared to
the true distribution. Figure 2 shows the observed distribution of predicted risk under the
assumption of no random error (thus reported values = true values) compared with the true
67
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
distribution of predicted risk if the observed data had random error (ICC values of 0.8 and 0.6).
The fact that the true distributions have a narrower range relative to the observed distributions
indicates that if the data to which the algorithm was applied contained some level of random
error then the true distribution of predicted probability of developing diabetes would be narrower
than what is estimated from the algorithm. Further, the effect on the distribution largely
influences the right side of the distribution since the left side is bounded by zero.
Impact on Calibration
Under the assumption of no random error in the observed data, the χ2H-L value was < 20
(calibration cutoff achieved) 97% of the time for males and females. At increasing levels of
random error the proportion of simulations where H-L χ2H-L was less than the cutoff 20 decreased
steadily. Overall, ICCs of approximately 0.8 or higher resulted in the algorithm achieving the
calibration criteria, that is a H-L chi square less than 20, at least 80% of the time. In both males
and females, errors in weight lead to larger decreases in calibration than errors in height, such
that even a perfect height measurement (ICC height = 1.0) would result no calibration if the ICC
for weight drops below 0.8. On the other hand, if ICC for weight was 1.0 even if the ICC for
height was 0.6, the algorithm could still achieve calibration almost 80% of the time (Figure 3).
Impact on Discrimination
Discrimination was decreased in the presence of random error. Under the most severe
measurement error the C-statistic was reduced from a C-statistic of 0.69 (with no error) to 0.55 in
males and from 0.72 (with no error) to 0.63 in females. If the ICCs for height and weight were
higher than 0.8, then the differences in the C-statistic compared to the estimate that had no
68
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
random error were less than 0.02. As with calibration, error in weight had a bigger impact on the
69
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Figure 1. Difference in overall risk (observed – true) under random error for males and females
(averaged over 500 simulations).
70
Figure 2.Observed (top graphs) and true (middle and bottom graphs) distribution of the probability of developing diabetes under
increasing levels of random error (ICC = 0.8 and 0.6) compared with observed distribution.
Observed Distribution
ICC = 1.0
Mean: 7.20%
22.5
20.0
17.5
15.0
10.0
7.5
5.0
0
0.005 0.025 0.045 0.065 0.085 0.105 0.125 0.145 0.165 0.185 0.205 0.225 0.245 0.265 0.285 0.305 0.325 0.345
prob
20.0
Mean: 7.00%
17.5
12.5
Percent
10.0
7.5
2.5
0
0.005 0.025 0.045 0.065 0.085 0.105 0.125 0.145 0.165 0.185 0.205 0.225 0.245 0.265 0.285 0.305 0.325 0.345
Mean: 6.80%
prob
True distribution when ICC
ICC ==0.6
0.6
25
20
10
prob
Figure 3a. Ability to detect calibration in 500 replications under various levels of random error in
height and weight among males (top) and females (bottom).
Figure 4. Average C-statistic for 500 replications under various levels of random error in height and weight.
Systematic error
The values for systematic measurement error in height and weight used in this study were taken
from a recent systematic review that summarized the empirical evidence regarding the
concordance of objective and subjective measures of height and weight (16). The under-
reporting of weight was varied from 0 kg to 3.0 kg below true value (varied in increments of 0.5
kg). The over-reporting of height was varied from 0 cm to 3.0 cm above the true value in
increments of 0.5 cm. The average level of systematic error found in the systematic review was
height and weight taken from the systematic review (+2.5 cm and -1.7 kg) would result in 0.86%
reduction (91 fewer cases) in the overall predicted probability of developing diabetes for females
and a 0.91% reduction (84 fewer cases) for males (Figure 5). The presence of random error in
conjunction with systematic error slightly reduced the amount of underestimation (Figure 6).
Impact on Calibration
resulted in the algorithm achieving the calibration at least 80% of the time. None of the 500
simulations achieved calibration under the maximum biases in reported height and weight
(Figure 7). It must be noted that there is no evidence that these extreme biases are likely in self-
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
reported height and weight; rather, they were investigated to illustrate of the range of results of
under- and over-reporting. The presence of random error in conjunction with systematic error
Impact on Discrimination
Under- or over-reporting of weight and height did not have a significant effect on the
discrimination of the model (Figure 8). C-statistics were reduced very slightly in the most
from the true estimate was never more than 0.01 for both males and females. When both random
error and systematic error were imposed, the C-statistic was reduced; however this was due to
75
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Figure 5. Difference in overall risk (observed – true) under systematic error for males and
females (averaged over 500 simulations).
76
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Figure 6. Difference in overall risk (observed – true) under systematic error and random error in
height and weight for males and females (averaged over 500 simulations).
77
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Figure 7. Ability to detect calibration in 500 replications under various levels of systematic error
in males and females.
Percent that achieved calibration in H-L test among males in the presence of bias - 500 Replications
Males Females
100%
90%
80%
70%
60%
% H-L <20
50%
40%
30%
20%
10%
0%
None "-0.5 kg" "-1.0 kg" "-1.5 kg" "-2.0 kg" "-2.5 kg" "-3.0 kg" "+0.5 "+1.0 "+1.5 "+2.0 "+2.5 "+3.0
cm" cm" cm" cm" cm" cm"
Bias
78
Figure 8. Average C-statistic for 500 replications under various levels of systematic error.
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Below is a summary of the a priori hypotheses and findings of the study in response to those
hypotheses:
(i) Random measurement error will affect both discrimination and calibration of the risk
algorithm.
This was confirmed in this study. Increasing levels of random error (generally ICCs below 0.8)
reduced the discrimination and calibration compared to a model without random error.
(ii) Systematic error will have minimal affects on discrimination but large effects on calibration
This was confirmed in this study. Systematic error resulted in reductions in calibration;
(iii) Random error will not affect the overall predicted risk of the model and systematic error
will influence the predicted risk in the direction of the under- or over-estimation of height and
weight.
This hypothesis was only partly confirmed in this study. It was confirmed that systematic error
influenced the predicted risk in the direction of the systematic error. However, contrary to this
hypothesis, random error also biased the overall estimate of predicted risk and the direction of
80
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
3.1.5 Discussion
This study systematically examined the impact of measurement error in the context of a
prediction algorithm. Simulation studies are flexible and permit a range of assumptions about the
magnitude and direction of errors. This simulation study reveals several interesting aspects of the
influence of measurement error (systematic and random) on the performance of a risk algorithm.
Two out of the three a priori hypotheses were realized in this study.
From equation (I) σ2observed = σ2true + σ2error , it can be seen that that the observed variance is
greater than the true variance in the presence of measurement error. In the context of Canada‘s
population health surveys, the observed variance of self-reported BMI would be greater than the
true variance because self-reported BMI may contain some random error. This study confirmed
that measurement error increases the observed variance of BMI, thus, making the observed BMI
distribution from the public use data wider than the true distribution. This affects both diabetics
and non-diabetics resulting in greater overlap between the BMI distributions making assigning
risk according to different levels of BMI more difficult to achieve. Even though random error in
height and weight should on average correctly estimate the true BMI in the population (since it
does not skew the mean in a particular direction), it can still influence the performance of a
prediction model due to decreased precision, which leads to greater dispersion in the BMI
distribution.
This study confirmed that systematic error in height and weight will result in bias in the
predicted risk estimates. This affects calibration, which is not surprising since the concordance
of observed and predicted events would be influenced by the under- or over-reporting of the
level BMI. In other words, persons that are over- or under-reporting their weight will then be
81
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
over- or under-estimated by the risk model and thus result in disagreement with observed
estimates. As hypothesized, systematic error did not influence the ability to rank order subjects.
In other words, the ability to discriminate between who will and will not develop diabetes was
not affected by systematic error, when variance due to random error is held constant. This was
reflected by the stability of the C-statistic under varying degrees of systematic error. The way
that the systematic error was examined in this study was such that the distribution of BMI was
shifted to the left (as a result of underestimating weight or overestimating height or both)
compared to the true distribution. This is an overall effect and the decreased precision or
increased variability as seen with random error is not observed. Even though the distribution is
shifted to the left – those with higher BMI still have a higher probability of developing diabetes
compared to those with lower BMI, even though the absolute levels of risk will be under-
estimated in both groups. This is a classic example of how discrimination and calibration are
often discordant. Due to the nature of probability, it is possible for a prediction algorithm to
exhibit perfect discrimination – i.e. it can perfectly resolve a population into those who will and
will not experience the event - and at the same time have deficient accuracy (meaning that the
predicted probability of that person experiencing the event does not agree with what will be
actually observed) (28). This study did not impose systematic error with respect to disease
status, but it could be hypothesized that if the systematic error were differential between
diabetics and non-diabetics that this could indeed affect discrimination. This is a topic of future
research.
The finding that random error resulted in the overall predicted risk estimated to be biased
upwards was contradictory to the hypothesis that only systematic error will bias the risk estimate.
As mentioned, random error increases the variability of a measurement and increases the range
82
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
of predicted risk. From other simulations not reported here it was shown that that if the true
prediction was greater than 0.5 then the estimated probability based on variables measured with
error would be smaller than the observed probability. That is, in both cases the prediction using
variables with measurement error pushes the estimated probabilities closer to 0.5
Not surprisingly, the error in predicted risk resulting from under-reporting weight or
over-reporting height is in the anticipated direction i.e. if weight is under-reported the observed
risk will be underestimated and not surprisingly, based on the above discussion, the addition of
random error to this type of systematic error slightly reduced the amount of underestimation
because the random and systematic errors work in opposite directions. In another situation,
random error could potentially augment the error in predicted risk. Such would be the case if
There are several results from this study that have implications for DPoRT. DPoRT
relies on self-report survey data, which is likely to suffer from some form of random error. This
study confirms that random error, which makes up to 20% of the total observed variance (ICC of
0.8 or higher), is unlikely to affect the performance or validation of the model. Research shows
that the random error in height and weight reporting is unlikely to exceed that amount (14).
However, the level of random error in the self-reporting of height and weight in the national
health surveys need to be confirmed to ensure that it does not make up more than 20% of the
total observed variance. Interestingly, the effects on the estimate of predicted diabetes risk in the
population were relatively minor, even in situations of high under- or over- reporting of weight
and height. This is likely because BMI has such a strong relationship with the outcome of
diabetes such that increased risk is apparent even with significant underestimation. In other
words, the true distributions of BMI in diabetics and non-diabetics are so distinct that even in the
83
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
presence of underreporting these populations has dissimilar risk for developing diabetes. Had
this misclassification affected a variable which did not have such a strong relationship with the
outcome, the effect on predicted risk may have been more severe. Furthermore, in this study
systematic error in self-reported height and weight was taken as an overall effect in the
population. If self-reporting error were significantly more likely to occur in those who were
more likely to develop diabetes then the impact of this bias could be augmented. This is another
This study examined a range of error values found from validation studies looking at self-
report and measured height and weight compared with measured height and weight. A recent
study by Shields et al. (17) examined agreement between self-report and measured BMI in a sub-
sample of the CCHS population. Overall systematic error in females was +0.5 cm for height and
-2.5 kg for weight which corresponded to an average underestimate of BMI of 1.2 kg/m2. In
males the bias was +1.0 cm for height and -1.8 kg for weight which corresponds to an average
underestimate of BMI of 0.9 kg/m2. According to these values, DPoRT predicted risks may be
underestimated by ~1% for both sexes; however, this underestimation is in the context of no
systematic error, which as discussed above may minimize underestimation. The relatively small
amount of underestimation in predicted risk that occurs due to systematic error in self-reported
height and weight is almost exactly proportional to the magnitude of error i.e. BMI is
underestimated by ~1 (0.9 & 1.2 for males and females) and overall predicted risk is
If DPoRT were applied in a setting where reporting error is thought to be higher than in
the NPHS (where the tool was developed), this increased error must be taken into account when
predicting risk and validating the tool. If the amount of error can be quantified, this study
84
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
outlines a way to correct predicted risk estimates using the ICC and equation (IV) σ2error =
- ρ). Since this study shows that systematic error has the ability to influence the
discrimination and accuracy of a risk algorithm performance, research into understanding and
This study focused on the overall trend of self-reporting error seen in several validation
studies, that is an underestimation of weight and an overestimation of height (15); however these
patterns may also vary across subpopulations such as gender and socioeconomic status.
Generally women tend to underestimate weight more so than men and men tend to overestimate
height more so than women(14, 29). Socioeconomic status has been shown to modify these
associations such that those of lower socioeconomic status may actually overestimate their
weight and/or underestimate their height (30, 31). These subgroups may also have differential
diabetes risk and the effect to which this error influences population risk prediction is a topic of
future research.
There are several limitations to consider in the context of this study. Conclusions drawn
from this simulation study will relate only to the scenarios simulated and may not apply to all
risk algorithm situations. Simulation programs that reflect the specific study conditions to which
interpreting the findings of this study is that models examined in this exercise are simpler than
complicated multivariate risk algorithms encountered in practice. This simpler model allows us
to focus on the height and weight error, which is the greatest potential source of error in DPoRT.
It should be noted that one of the assumptions of this study is that the only sources of error are
that in self-reported height and weight. Other sources of error including, error in diabetes status
and selection bias in the survey or in sampling, are assumed to be absent. We cannot confirm
85
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
that the results of this study would be the same in the presence of the above mentioned sources of
error.
This study has provided novel information about the influence of measurement error in a
algorithm performance, efforts can be made to correct for these errors and thus improve the
accuracy and validity of a risk algorithm. Future research will include investigation into
systematic error with respect to disease status or other characteristics of the population. Further,
efforts must be made to understand the nature of error in self-reporting measurements. Ongoing
work to improve the quality of measurements used in risk algorithms will improve model
performance. Researchers developing and validating risk tools must be aware of the presence of
measurement error and its impact on the performance of their risk tools.
86
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
3.16 References
1. Flegal KM, Keyl PM, Nieto FJ. The effects of exposure misclassification on
estimates of relative risk. Epidemiology 1986;123:736-51.
8. Weinstock MA, Colditz GA, Willet WC. Recall (report) bias and reliability in the
retrospective assessment of melanoma risk. American Journal of Epidemiology
1991;133:240-5.
9. Colditz G et al. Weight as a risk factor for clinical diabetes in women. American
Journal of Epidemiology 1990;132:501-13.
10. Colditz G et al. Weight gain as a risk factor for clinical diabetes mellitus in
women. Annals of Internal Medicine 1995;122:481-6.
11. Perry IJ et al. Prospective study of risk factors for development of non-insulin
dependent diabetes in middle aged British men. British Medical Journal 1995;310:555-9.
14. Nawaz H et al. Self-reported weight and height: implications for obesity research.
Journal of Preventive Medicine 2001;20:294-8.
15. Gorber SC et al. A comparison of direct vs. self-report measures for assesing
height, weight, and body mass index: a systematic review. Obesity Reviews 2007;8:307-
26.
87
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
16. Gorber SC et al. The feasibility of establishing correction factors to adjust self-
reported estimates of obesity. Health Reports 2009;19.
17. Shields M, Gorber SC, Tremblay MS. Estimates of obesity based on self-report
versus direct measures. Health Reports 2008;19:1-16.
18. Statistics Canada. 1996-97 NPHS Public Use Microdata Documentation. 1999.
Ottawa.
20. Lipscombe LL, Hux JE. Trends in diabetes prevalence, incidence, and mortality
in Ontario, Canada 1995-2005: a population-based study. Lancet 2007;369:750-6.
21. Deyo RA, Diehr P, Patrick DL. Reproducibility and responsiveness of health
status measures: Statistics and strategies for evaluation. Controlled Clinical Trials
2008;12:142S-58S.
22. Fleiss J. Statistical Methods for Rates and Proportions. New Jersey: John Wiley &
Sons, 2003.
23. Hosmer DW, Lemenshow S. Applied Logistic Regression. New York: Wiley,
1989.
24. Hosmer DW et al. A comparison of goodness-of-ft tests for the logistic regression
model. Statistics in Medicine 1997;16:965-80.
25. Hosmer DW, Lid Hjort N. Goodness-of-fit processes for logistic regression:
simulation results. Statistics in Medicine 2002;21:2723-38.
28. Diamond GA. What price perfection? Calibration and discrimination of clinical
prediction models. Journal of Clinical Epidemiology 1992;45:85-9.
29. Niedhammer I et al. Validity of self-reported weight and height in the French
GAZEL cohort. International Journal of Obesity 2000;24:1111-8.
88
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
89
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
3.2.1 Abstract
Background: Certain ethnic groups have been shown to be at increased risk for type 2 diabetes.
The current form of the Diabetes Population Risk Tool (DPoRT) includes a non-specific
Having to include detailed ethnic information would limit the applicability of the risk tool since
this information is not routinely collected at the population level. Given the importance of
ethnicity in influencing diabetes risk and its significance in Canada‘s multi-ethnic population, it
is prudent to determine whether detailed ethnic information would significantly improve the
Objective: To apply and compare the Diabetes Population Risk Tool (DPoRT) with a modified
version that includes detailed ethnic information in Canada‘s largest and most ethnically diverse
province, Ontario.
Methods: Two diabetes prediction models were built using the same principles and data as
DPoRT. The 2 models created in this study were: (i) a model that contained predictors specific to
the following ethnic groups: White, Black, Asian, South Asian, and First Nation and (ii) a
reference model which did not include a term for ethnicity. In addition to discrimination and
calibration measures of model performance, 10-year diabetes incidence rate and predicted
number of new diabetes cases in Ontario using the different algorithms were compared. The
algorithms were developed using the 1996-7 National Population Health Survey in Ontario (N =
19,861) and validated in the 2000/1 Canadian Community Health Survey in Ontario
(N=26,465).
90
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Results: All non-white ethnicities were associated with higher risk for developing diabetes than
white ethnicity. Among non-white ethnicities South Asian ethnicity had the highest hazard ratio
for diabetes in both males and females. Discrimination and calibration were similar across all
algorithms (0.75 – 0.77). Sufficient calibration (χ2H-L < 20) was maintained in the development
and validation cohort for all models except the detailed ethnicity models for males ( χ2H-L =
33.9). For both males and females, applying DPoRT resulted in the lowest overall ratio between
observed and predicted diabetes risk compared to the other two algorithms. DPoRT appears to
identify more cases at high risk than the other two algorithms in males, whereas in females both
DPoRT and the full ethnicity model identified more high risk cases compared to the algorithm
without ethnic information. Overall across decile of risk the DPoRT and full ethnicity algorithms
were very similar in terms of predictive accuracy and estimated risk in the population.
ethnicity may be important, when predicting new cases of diabetes at the population level and
accounting for other risk factors, detailed ethnic information did not improve the discrimination
and accuracy of the model or identify significantly more diabetes cases in the population.
91
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
3.2.2 Introduction
Planning for health care and public health resources needed to address the significant burden of
diabetes (1) is as an important aspect of population health management, which can be informed
by robust prediction tools, such as the Diabetes Population Risk Tool (DPoRT)( Chapter 2). This
tool can aid policy makers, planners, and physicians by providing reliable estimates of the
can be improved by knowing which groups to target and how extensive a strategy is needed to
Risk prediction tools for estimating disease risk are common in clinical settings and are
used for clinical decision-making (2). One of the limitations of clinical risk prediction tools for
population prediction is the reliance on physical measurements or special risk questions, such as
fasting blood sugar (3-5) or diabetes family history (6, 7) in the case of diabetes. At the
population level in Canada, these measurements are not easily, accurately, or systematically
captured. Currently, data on the prevalence of risk factors in the population are only collected
through national population health surveys using self-reported measures. One of the key
attributes of DPoRT is its accessibility to a broad audience. This is achieved by using data from
surveys that are publicly available. Detailed ethnic information, though collected, is not publicly
reported. Ethnic information from the surveys are reported publicly as ―white/non-white,‖ and
thus this form for ethnicity was used in DPoRT in order to ensure that the tool can be applied to
public data.
There is growing evidence that certain ethnic groups are at increased risk for developing
type 2 diabetes. Globally, non-European populations have a higher proportional burden of type 2
diabetes compared to the other regions of the world(8). The highest diabetes rates in the world
92
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
are seen in aboriginal population, including those in Australia(9, 10), United States(11, 12), and
Canada(13, 14). In the United States, studies have shown that those of African and Hispanic
decent are at increased risk for developing diabetes compared with non-Hispanic white
Americans (15-17). Throughout the world, those of South Asian decent are another ethnic group
which has been shown to carry an increase burden of type 2 diabetes compared with both non-
white and white ethnicities (18-20). Data from Ontario demonstrate that overall immigrant and
ethnic minority populations suffer from a higher burden of diabetes and its complications (21).
The importance of ethnicity when considering those at high risk for developing diabetes in the
clinical setting has been emphasized through diabetes guidelines that recommend people of
Aboriginal, Hispanic, south Asian, Asian, or African descent should be targeted for screening
Immigrants account for 18-20% of Canada‘s population (24), and this percentage is expected to
increase over time. Estimates of immigrant populations are as high as 50% for major urban
Though clinically and epidemiologically important risks associated with ethnicity are
apparent, it is not clear how ignoring ethnic-specific predictors will affect a population-based
prediction tool for diabetes. Currently, given that DPoRT performed well in 2 external validation
cohorts, it is assumed that diabetes risk is sufficiently estimated using the current form of the
model. However, given the significance of ethnicity in Canada and its important influence on
diabetes risk, it may be possible that failing to apply ethnic-specific predictors will reduce the
ability to identify high risk groups. In order to have confidence in applying this tool in Canada, it
needs to be determined if the inclusion of detailed ethnic predictors will significantly change the
performance of DPoRT.
93
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
The purpose of this study was to assess the impact of including detailed ethnic information in a
prediction algorithm for diabetes. Specifically this study described the relative benefits to
predictive accuracy and model outputs that are gained with the addition of ethnic specific
predictors to the model. In addition to informing the application of DPoRT, this work also
provides insight into the independent role of ethnicity on diabetes risk once additional risk
factors are considered. No study has previously taken such an approach (i.e. prediction) to
94
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
3.2.3 Methods
Development Cohort
The study cohort was derived from 23,403 people from Ontario that responded to the 1996/7
National Population Health Survey (NPHS-ON) conducted by Statistics Canada. In the NPHS,
households were selected though a stratified, multilevel cluster sampling of private residences
using provinces and/or local planning regions as the primary sampling unit. The survey,
conducted via telephone, had an overall 83% response rate and all responses were self-reported
(15). Persons under the age of 20 (n = 2, 407) and those who had previously diagnosed diabetes
or self-reported diabetes were excluded (n = 894). Those who were pregnant at the time of the
survey were also excluded (n = 241), due to the fact that baseline Body Mass Index (BMI) could
not be accurately ascertained, leaving a total of 19,861 individuals (Figure 1). Sixty-six males
were further excluded due to missing baseline BMI resulting in 9,177 males and 10,618 females
Validation Cohorts
The DPoRT algorithm was validated in two external cohorts. Further details on the DPoRT
validation are provided in Chapter 2. One validation cohort was used in this study to compare the
performance of the 3 risk functions in an external cohort. The validation study used in this cohort
was derived from the Ontario portion of the 2000/1 Canadian Community Health Survey (CCHS,
Cycle 1.1, N = 37,473), a national telephone survey administered by Statistics Canada. The
target population of the CCHS consisted of persons aged 12 and over resident in private
dwellings in all provinces and territories, excepting those living on Aboriginal reserves, on
95
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Canadian Forces Bases, or in some remote places. The CCHS included the same self-reported
health questions as the NPHS. Like the NPHS, this survey uses a multistage stratified cluster
design and provides cross-sectional data representative of 98% of the Canadian population over
the age of 12 years, and attained an 80% overall response rate (25, 26). After the exclusion
criteria were applied there were 26,465 individuals in the validation cohort.
Survey data from development and validation cohorts were linked to provincial administrative
health care databases that include all persons covered under the government funded universal
health insurance plan. The diabetes status of all respondents in Ontario was established by
linking persons to the Ontario Diabetes Database (ODD). The Ontario Diabetes Database (ODD)
contains all physician diagnosed diabetes patients in Ontario identified since 1991. The database
is created using hospital discharge abstracts and physician service claims. A patient is said to
have physician diagnosed diabetes if he or she meets at least one of the following two criteria:
Clinical Modification code 250 (ICD9-CM) before 2002 or ICD-10 code E10 – E14 after 2002,
or (b) a physician services claim with a diabetes diagnosis (code 250) followed within two years
Individuals entered the ODD as incident cases when they were defined as having diabetes
according to the criteria described above. A hospital record with a diagnosis of pregnancy care
or delivery close to a diabetic record (i.e. a gestational admission date between 90 days before
and 120 days after the diabetic record date), were considered to represent gestational diabetes
and so were excluded. The ODD has been validated against primary care health records as an
96
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
specificity of 97%) (27, 28). Information regarding the vital statistics and eligibility for health
care coverage for linked respondents was captured from the Registered Persons Data Base
(RPDB). The algorithm used to create the ODD was applied to Manitoba‘s administrative health
care data to ascertain physician-diagnosed diabetes status in that province. The ODD algorithm is
applied nationally using provincial administrative registries (known as the National Diabetes
Surveillance System (NDSS)) and has been used and validated in several Canadian provinces
(29).
Variable Definitions
Variables used in this study were obtained from responses in the NPHS and CCHS, including:
age, Body Mass Index (BMI), presence of chronic conditions diagnosed by a health professional
(including hypertension and heart disease), ethnicity, immigration status, smoking status, highest
level of achieved education. Body Mass Index (BMI) in kg/ m2 was used as an indicator of
obesity. Derived BMI, calculated by dividing the weight in kilograms by height squared in
meters-squared directly from the NPHS, is only calculated for respondents aged 30 to 64;
therefore, BMI was calculated using weight and height according to derived variable
specification for those who fell outside the age range of 30 – 64 (30). Ethnicity was ascertained
by the question, 'To which ethnic or cultural groups do your ancestors belong?' Classification of
ethnic groups were: White, Black, Asian, South Asian, and First Nation according to Statistics
Canada‘s definition (31). Statistics Canada releases public-use micro data files of the national
health surveys; however, certain variables are suppressed or modified in these files to protect
privacy. In the public-use file ethnicity is only categorized as white or non-white, derived from
97
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
the response to the ethnicity question (20). The shared population health survey files, which are
available at the provincial level, contain more detailed information (including detailed ethnicity).
Access to the shared data files is highly restricted, which is why DPoRT was developed using
variables from the public use file rather than the shared file. In this study the shared file was used
in order to allow for both forms of the variable (white/non-white vs 5 ethnic groups) to be
compared.
Statistical Analysis
Creation of DPoRT
The goal in the creation of DPoRT was to create a risk algorithm that would accurately predict
diabetes risk with high discrimination and calibration using risk factors that are measured
reliably from health survey data. A detailed description of the development of DPoRT can be
found in Chapter 2. Briefly, for each cohort member, the probability of physician-diagnosed
diabetes was assessed from the interview date until censoring for death or end of follow-up using
a Weibull accelerated failure time model Diabetes risk functions were derived separately for
men and non-pregnant women above the age of 20 without a prior diabetes diagnosis. Each
variable was centred on the population mean before inclusion in the model to allow for easier
calibration with other cohorts. This means that when the algorithm is applied, all variables are
centered to the mean variables for the cohort to which it is being applied allowing levels of risk
to be reflective of the average baseline risk in that cohort. Overall risk (predicted probability) of
diabetes for each person was calculated by multiplying the individual‘s risk factor values by the
corresponding regression coefficients, and summing the products (32). The form of the model
was assessed using likelihood ratio tests to compare nested parametric models (33). A plot of
98
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Log(-logS(t))) vs log(t) was inspected for linearity to assess consistency of the survival times
with the Weibull distribution. Cox-Snell residual plots were also constructed to assess the
All estimates (including betas and variance estimates) incorporated bootstrap replicate
survey weights to accurately reflect the demographics of the Canadian population and account
for the survey sampling design based on selection probabilities and post stratification
adjustments. Variance estimates and 95% confidence intervals were calculated using bootstrap
survey weights(34, 35). All statistics were computed using SAS statistical software (version 9.1
Two additional models were created in the NPHS-ON development cohort as described above
except the models were modified to either include ethnic-specific predictors or remove any
(ii) DPoRT
(iii) DPoRT plus detailed ethnic information – called ―Full ethnicity model‖
In DPoRT (model (ii)) ethnicity is grouped as in the public use file as white/non-white.
In model (iii) ethnicity is broken up into the categories consistent with the diabetes screening
guidelines (36): White, Black, Asian, South Asian, and First Nation.
99
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
The performance of the model was measured by the discrimination and calibration values in the
external validation cohort. Discrimination was measured using a C statistic modified for survival
data developed by Pencina et al. (37), analogous to the area under the ROC curve (38) .
developed by Nam (39, 40). This statistic is computed by dividing the validation cohort into
deciles of predicted risk and compared the observed versus predicted risk in each decile using a
chi-square statistic (see appendix 6.1). To mark sufficient calibration χ2 = 20 was used as a
cutoff (p<0.01), consistent with D‘Agostino‘s validation of the Framingham algorithms (39).
Observed versus predicted cases of diabetes were also compared across ethnicities to examine
The policy implications of the 3 models were assessed by applying the model to the
2000/1 data to predict 10-year diabetes incidence rates and cases and then to compare these
values between the algorithms. The proportion of the population who were identified as high risk
were also reported and compared across algorithms. In this study the probability of developing
diabetes was stratified into the following categories: <2%, 2-5%, 5-10%, 11-20%, ≥20% where
In order to describe the impact of disagreement between observed and predicted diabetes
risk as a function of the proportion of the population where that disagreement exists, an index
called the Population Disagreement Index (PDI) was developed. PDI was summarized across
100
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
𝑛
𝑂𝑖
PDI = × 𝑃𝑝𝑖
𝑖=0 𝑃𝑖
𝑂𝑖
Where 𝑃𝑖 = Ratio between observed: predicted in subgroup i
The unweighted ratio between observed and predicted were calculated to demonstrate the
influence of the distribution of the subgroup (i.e. 𝑃𝑝𝑖 ) This was calculated by sampling taking
𝑛
𝑂𝑖
the overall ratio, i.e. .
𝑖=0 𝑃𝑖
101
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
3.2.4 Results
The observed 10-year diabetes risk (cumulative incidence rate) in the Ontario development
cohort was 8.8% for males and 7.3% for non-pregnant females aged 20 years and older at
baseline. In addition to non-white ethnicity the other attributes in the model which were
previously validated were: BMI, age (and its interactions), hypertension, smoking, heart disease
and immigrant status. The DPoRT development section shows the multivariate-adjusted hazard
ratios for the risk factors in the DPoRT algorithm for males and females.
Ethnicity
Adjusted hazard ratios for the ethnic categories in DPoRT and the full ethnicity algorithm are
shown in Figure 1. Non-white ethnicity has a hazard ratio of 2.14 95% CI (1.74, 2.63) in males
and 1.71 (1.35, 2.16) in females, adjusted for all other variables in the risk algorithm. In the full
ethnicity model, hazard ratios for specific ethnic groups ranged from 1.11 to 3.02 compared to
white ethnicity. All non-white ethnic groups were at higher diabetes risk than white ethnicity.
South Asian ethnicity had the highest hazard ratio for diabetes in both males and females.
102
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Figure 1. Adjusted Hazard Ratios (white ethnicity as reference) and 95% CIs for Ethnic
Variables in DPoRT.
103
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Model Performance
All 3 models showed good discrimination in the development and validation cohorts (C-statistic
ranging from 0.75 – 0.77). The algorithm with white/non-white ethnicity predictors had slightly
higher discrimination versus a model with no ethnicity (i.e. DPoRT versus no ethnicity
algorithm) (table 2). The full ethnicity algorithm achieved the same discrimination as DPoRT for
males and females. Sufficient calibration ( χ2H-L < 20) was maintained in the development and
validation cohort for all models except the detailed ethnicity model for males (χ2H-L = 33.9). All
models had a similar ratio of observed to predicted risk across decile of risk (figure 1a&b). All 3
models under-predicted risk in south Asian males and the largest under-prediction was in the
model without any ethnic information (figure 2a&b). Of the three risk algorithms DPoRT had
the lowest overall average ratio between the observed and predicted (Figure 3). Weighting by
population proportion significantly reduces the overall disagreement in the population due to the
Table 1. C statistics with 95% confidence intervals and calibration χ2 statistics for 3 algorithms.
Males Females
Cohort Development Validation Development Validation
C-Statistic
Model
0.75 0.76 0.77 0.76
No ethnicity (0.74,0.77) (0.75,0.78) (0.75,0.78) (0.74, 078)
0.76 0.77 0.77 0.76
DPoRT (0.74,0.77) (0.75,0.78) (0.75,0.79) (0.74,0.78)
0.76 0.77 0.77 0.76
Full ethnicity (0.74,0.77) (0.76,0.79) (0.76,0.79) (0.74,0.78)
H-L χ2
Model
No ethnicity 3.22 11.38 9.01 6.05
DPoRT 1.54 15.36 7.02 8.00
Full ethnicity 6.27 33.91 7.18 9.11
104
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Figure 1a. Observed versus predicted in the validation cohort by decile of risk using the
algorithm without ethnicity, with ‗white/non-white‘ ethnicity, and with detailed ethnic predictors
for males.
Males
70000
No Ethnicity
60000
50000
40000
Observed
30000
Predicted
20000
10000
0
1 2 3 4 5 6 7 8 9 10
70000
DPoRT
60000
50000
40000
Observed
30000
Predicted
20000
10000
0
1 2 3 4 5 6 7 8 9 10
70000
Full Ethnicity
60000
50000
40000
30000 Observed
20000 Predicted
10000
0
1 2 3 4 105
5 6 7 8 9 10
Deciles of Risk
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Figure 1b. Observed versus predicted in the validation cohort by decile of predicted risk using
the algorithm without ethnicity, with ‗white/non-white‘ ethnicity, and with detailed ethnic
predictors for females.
Females
70000
No Ethnicity
60000
50000
40000
Observed
30000
Predicted
20000
10000
0
1 2 3 4 5 6 7 8 9 10
70000
DPoRT
60000
50000
40000
Observed
30000
Predicted
20000
10000
0
1 2 3 4 5 6 7 8 9 10
70000
Full Ethnicity Females
60000
50000
40000
Observed
30000
Predicted
20000
10000
0 106
1 2 3 4 5 6 7 8 9 10
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Figure 2a. Observed versus predicted 5-year risk and number of cases by ethnicity for males.
Males
15.00 160
No Ethnicity
Thousands
80
5.00
40
0.00 0
White Black Asian First South Other
Nation Asian
15.00 160
DPoRT
Thousands
120
10.00
80
5.00
40
0.00 0
White Black Asian First South Other
Nation Asian
15.00 160
Full Ethnicity
Thousands
5-year DM rate (%)
120
10.00
80
Predicted number of DM cases
5.00
40
0.00 0
White Black Asian First South Other
Nation Asian
Figure 2b. Observed versus predicted 5-year risk and number of cases by ethnicity for females.
Females
15.00 160
No Ethnicity
Thousands
80
5.00
40
0.00 0
White Black Asian First South Other
Nation Asian
Thousands
120
10.00
80
5.00
40
0.00 0
White Black Asian First South Other
Nation Asian
15.00 160
Full Ethnicity
Thousands
120
10.00
80
5.00
40
0.00 0
White Black Asian First South Other
Nation Asian
108
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Figure 3. Unweighted average ratio between observed and predicted cases and weighted average
ratio between observed and predicted (i.e. PDI).
Males
2.0
1.89
Average ratio observed:predicted
1.8
1.6
1.4
1.32
1.2
1.17
1.09 1.08
1.03
1.0
A lgorithm DPoRT Full ethnicity No ethnicity DPoRT Full ethnicity No ethnicity
Unweighted Weighted
Females
1.3
1.27
Average ratio observed:predicted
1.2
1.1
1.09
1.08
1.07
1.06
1.0
0.93
0.9
A lgorithm DPoRT Full ethnicity No ethnicity DPoRT Full ethnicity No ethnicity
Unweighted Weighted
109
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
In males the overall 10- year predicted risk ranged from 9.85% for the no ethnicity
model, 10.11% in DPoRT and 10.06% in the full ethnicity model. In females average 10 year
predicted risk ranged from 7.83% for the no ethnicity model, 7.95% in DPoRT and 7.97% in the
full ethnicity model. There were 9,660 more predicted cases in males and 5, 013 predicted cases
in females in DPoRT than with the model without ethnicity. In males, 1,409 less cases were
predicted in the full model compared to DPoRT and in females 934 more cases were predicted in
Ethnic specific risk diverged more in males and females when including ethnic specific
terms (Figure 4). The distribution of risk in the population for males and females can be seen in
Figure 5. Overall the distribution of risk (i.e. the proportion of the population belonging to each
risk category) is similar across the algorithms for females and males. Overall, DPoRT appears to
identify more cases at high risk than the other two algorithms in males, whereas in females both
DPoRT and the full ethnicity identify more high risk cases and are not substantially different.
Across decile of risk the number of diabetes cases predicted using DPoRT and full ethnicity
algorithms were very similar in both males and females (Figure 6).
Table 2. 10-year risk and predicted new diabetes cases in Ontario from baseline in the 2001
CCHS.
Males Females
DIABETES Med Rate, DIABETES Med Rate,
rate (%) % # new cases rate (%) % # new cases
No ethnicity 9.85 5.78 386,964 7.83 4.18 321,724
DPoRT 10.11 6.50 396,624 7.95 4.74 326,737
Full
ethnicity 10.06 6.74 395,215 7.97 4.69 327,671
110
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Figure 4. Average 10-year diabetes risk for males and females by ethnicity using 3 algorithms
Males
io
n ian
Nat As
ian k t er ut
h te
ac rs th hi
As Bl Fi O So W
DPoRT Full Ethnicity
20
15
10-year risk (%)
10
5
No Ethnicity
20
15
10
5
n k n er n te
ia ac t io th ia hi
As Bl a O As W
tN ut
h
rs
Fi So
Ethnicity
Panel variable: Algorithm
Females
io
n ian
at As
ian k tN er ut
h
hi
te
ac rs th
As Bl Fi O So W
DPoRT Full Ethnicity
20
15
10-year risk (%)
10
5
No Ethnicity
20
15
10
5
n k n er n te
ia ac io th ia hi
As Bl Nat O As W
t h
rs ut
Fi So
Ethnicity
Panel variable: Algorithm
111
Doctor of Philosophy Epidemiology (PhD) Dissertation Laura C.A. Rosella
Figure 5a. Distribution of diabetes risk (left) and number of new cases
by risk group in the population (below) using 3 algorithms for males.
112
Doctor of Philosophy Epidemiology (PhD) Dissertation Laura C.A. Rosella
Figure 5b. Distribution of diabetes risk (left) and number of new cases
by risk group in the population (below) using 3 algorithms for females.
113
Doctor of Philosophy Epidemiology (PhD) Dissertation Laura C.A. Rosella
Figure 6. Predicted diabetes cases in 10 years by decile of predicted risk for males and females using three algorithms
140000 140000
Males Females
120000 120000
No Ethnicity No Ethnicity
DPoRT DPoRT
Full Full
100000 100000
80000 80000
60000 60000
40000 40000
20000 20000
0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Decile of Risk Decile of Risk
114
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
3.2.5 Discussion
The aim of this study was to assess the impact of including detailed ethnic predictors in a
population-based risk tool for diabetes in Canada. In addition to identifying relative hazards of
developing diabetes by ethnicity, this study provides estimates of the predicted number of cases
in a provincial population by ethnicity for the next 10-year period. Using a population-based
cohort, this study confirmed that those of non-Caucasian descent are at increased risk for
developing diabetes and consistent with previous research, hazard ratios were highest among
South Asians (18). In terms of overall model performance, no additional predictive value was
detected when adding detailed ethnic predictors. At the population level, distribution of risk was
similar across of different risk levels in the population, particularly between DPoRT and the full
ethnicity model. This study suggests that using DPoRT in its current form is sufficient for
accurately predicting diabetes cases in ethnically diverse population similar to Ontario. The
funding that the algorithm to predict diabetes that uses detailed ethnicity did not significantly
differ from one that uses a broad categorization of ethnicity can be driven by two mechanisms
There are several reasons why a clinically important risk factor may not improve the
outcome, it may not provide incremental improvements in test characteristics which are relevant
for prediction (i.e. discrimination and calibration) in the context of existing predictors. This
phenomenon has also been shown for other clinical predictors and outcomes such as C-reactive
protein for cardiac risk prediction (41). In fact, research has shown that although a battery of
novel risk factors have been developed for the prediction of major coronary heart disease (CHD)
events, these novel factors have been generally unimpressive in their ability to improve CHD
115
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
prediction (42). Furthermore, it has been shown that for a variable to make significant
improvements in discrimination (i.e. improvement in AUC from 0.8 to 0.9) its multivariable
odds ratio must be 6.9 or greater (43) suggesting that in order for detailed ethnicity to improve
the algorithm beyond its current discrimination, the adjusted hazard ratio must be very large in
magnitude. Interestingly, the model without ethnicity was not detectably worse in terms of model
performance such that discrimination and calibration were only marginally decreased compared
to DPoRT. This is likely due to the fact that many of the reasons that ethnicity plays a role in
diabetes risk are related to other factors captured in the model. In particular, socioeconomic
status, obesity (particularly younger onset of obesity), and other lifestyle factors have been
shown to be related to both ethnicity and diabetes risk (18). Most importantly, immigration
status, which is already captured in the model, may explain a significant amount of variability in
diabetes incidence that is associated with ethnicity. The diminishing return on model
performance when adding statistically significant predictors to the model was also noted in the
model building process of DPoRT and was one of the reasons that DPoRT maintained good
attributable risk (PAR). PAR describes the impact of a risk factor on population risk as a
function of the prevalence of the exposure and the relative risk of disease (44). In this study the
estimate of relative risk in PAR is translated to the disagreement between observed and predicted
and the prevalence of the exposure is translated to the prevalence of the population where the
disagreement exists. Therefore, PDI described the impact of disagreement between observed and
predicted risk for a risk tool as a function of the proportion of the population where that
disagreement exists. PDI exemplifies how population risk is driven by where the cases lie in the
116
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
population. This means that a large relative discrepancy between observed and predicted that is
concentrated in a subgroup that covers a small proportion of the population will have less impact
on the overall population estimate of diabetes compared with disagreement of the same
magnitude that affects a larger proportion of the population. This finding emphasizes an
but these differences may not be as critical if applied for population estimates. This also
identifies a potential difference in the way that algorithms must be independently validated,
depending on whether they are intended for use on the individual or in small subpopulations. Of
course, in the same way PAR is affected by the prevalence of the risk factor in the population,
the influence of disagreement within ethnic groups is affected by the ethnic composition of the
population. Ethnic composition can change over time and the impact of this on the validity of the
algorithm should be regularly assessed. The use of prediction tools at the individual level or in
small subpopulations must be independently validated in specific subpopulations and used with
The purpose of this study was to examine the impact of ethnicity on population risk
prediction and not to validate it for use within specific ethnic groups. Nonetheless, looking at
performance within ethnic groups provides important information about diabetes risk by
ethnicity. DPoRT performed well in all ethnic groups (especially in females), with the exception
of South Asian males where even with the inclusion of full ethnic information the algorithm
resulted in an under-prediction of diabetes risk. This can indicate that there is an aspect of risk in
this population which is not captured by either the variables in DPoRT or detailed ethnicity.
This result is consistent with emerging evidence about the nature of metabolic risk in South
117
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Asian males. This population has significantly more insulin resistance than Caucasian
populations even in the absence of excessive obesity (18). It has been proposed that the excessive
insulin resistance in Asian Indians could be explained by an abdominal fat distribution which
may be genetically determined (45). A recent study looking at detailed radiographic and
anthropometric measures in Asian and Caucasian men showed that for a given BMI or waist
circumference, South Asian men had approximately 6% higher total body fat than Caucasian
men. Other studies have shown that adjustments for BMI or waist circumferences to define
obesity do not entirely account for possible differences in inherent insulin resistance in the South
Asian population (46). Several physiological mechanisms for this occurrence have been
proposed including that South Asian men have a defect in adipose tissue metabolism, which
tissue metabolism are concomitant with insulin resistance (47). These studies indicate that there
may be an important aspect of diabetes risk which is not captured by simply including ethnicity
and BMI along with the other predictors of the model. The type of detailed physiological
information which may be needed is not captured at the population level nor is it feasible to
include in a tool such as DPoRT. Regardless, these differences did not affect the performance of
Another difficulty in estimating the ethnicity-diabetes risk among males is the possibility of
confounding by physical activity. Immigrant men are more likely to engage in jobs that require
physical activity on a daily basis (48) which has been shown to reduce the risk of developing
diabetes (49). This may explain why the full ethnicity algorithm actually performs worse than
DPoRT for some ethnic groups. Inclusion of full ethnic information may result in over-fitting of
the model. This phenomenon was also seen during DPoRT creation.
118
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Previous studies indicate that weight cut-offs may differ in their associated risk for diabetes
within ethnicity groups and that different cut-offs should be used to identify those at high risk
(50, 51). Our study suggests that as long as additional risk factor differences among ethnic
groups are captured in the prediction algorithm, the difference may not actually be as substantial
as previous noted. This difference may be due to the fact that previous studies did not fully
account for possible confounders including age and additional metabolic disorders (50). This
study examined the interaction between age-specific BMI and ethnicity and found no significant
differences.
There are several limitations to consider when interpreting the results of this study. Firstly,
the minimal difference detected between DPoRT and the full ethnicity algorithm may not be
found in other populations with different ethnic compositions. Secondly, using self-report
measures from the health survey is a limitation which could affect predictive risk accuracy since
these measures may be more subject to reporting error and bias than clinical measures. For self-
reported height and weight, in general there is a high agreement; however, validation studies
show that weight tends to be slightly underestimated and that height may be slightly
overestimated and as a result reported BMI is generally lower than measured BMI (52-54),
The possibility of misclassification is also possible with the use of self-reported ethnicity,
even though it is the most common measure of acquiring ethnic information in epidemiological
federal agencies, such as the US and Canadian Census and the National Center for Health
Statistics (56). This is due to the fact the self-identification with ethnicity is most important for
119
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
may be more problematic when examining the genetic associations with disease (57). Another
generalizability of this research to all ethnicities in Canada. The cohorts covered in the NPHS
and CCHS exclude those living on Aboriginal reserves. Therefore, estimates for First Nations
people apply only to those living off-reserve and are not intended to represent First Nations on-
reserve. Previous studies show that First Nations are at greater risk for diabetes than other
members of the Canadian population including off-reserve First Nations counterparts (13, 14, 58-
60). This is an important component of the population to consider for diabetes prevention and a
population risk algorithm developed specifically for on-reserve populations would be beneficial
Overall there are several key messages to be taken from the results of this study. Firstly,
this analysis provides adjusted hazard ratios and risk estimates to quantify the impact of ethnicity
on diabetes risk using a prospective population-based cohort study in Ontario. This is the first
study that reports 10-year risk and number of cases of diabetes from a prediction model
according to ethnicity. These estimates provide key information for predicting diabetes risk at the
Secondly, this study shows that DPoRT in its current form is as effective or in some cases better
than the algorithm with full ethnic information for predicting diabetes risk at the population
level. Furthermore, it also appears to work well within ethnic groups, in particular for women.
Though overall model performance was good, analysis by ethnicity shows that further research is
required to improve model fit in South Asian males. This study, consistent with other prediction
tools, has reaffirmed that using a measure that has a statistically significant association with a
120
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
3.2.6 References
1. Wild S et al. Global prevalence of diabetes - Estimates for the year 2000 and
projections for 2030. Diabetes Care 2004;27:1047-53.
4. Hanley AJG et al. Prediction of type 2 diabetes using simple measures of insulin
resistance - Combined results from the San Antonio Heart Study, the Mexico City
Diabetes Study, and the Insulin Resistance Atherosclerosis Study. Diabetes 2003;52:463-
9.
7. Lindstrom J, Tuomilehto J. The diabetes risk score: a practical tool to predict type
2 diabetes risk. Diabetes Care 2007;26:725-31.
11. Mokdad AH et al. The continuing epidemics of obesity and Diabetes in the United
States. JAMA: The Journal of the American Medical Association 2001;286:1195-200.
12. Pavkov ME et al. Changing patterns of type 2 diabetes incidence among Pima
Indians. Diabetes Care 2007;30:1758-63.
13. Harris SB et al. The prevalence of NIDDM and associated risk factors in native
Canadians. Diabetes Care 1997;20:185-7.
121
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
14. Young TK et al. Type 2 diabetes mellitus in Canada's First Nations: status of an
epidemic in progress. Canadian Medical Association Journal 2000;163:561-6.
15. Narayan KMV et al. Lifetime risk for diabetes mellitus in the United States.
JAMA 2003;290:1884-90.
18. Abate N, Chandalia M. Ethnicity and type 2 diabetes - Focus on Asian Indians.
Journal of Diabetes and Its Complications 2001;15:320-7.
21. Manuel D, Schultz S. Diabetes in Ontario: An ICES Practice Atlas. Hux, J. E.,
Booth, G. L., Slaughter, P. M., and Laupacis, A. 4.77-4.94. 2003. Toronto, Institute for
Clinical and Evaluative Sciences.
22. Calonge N et al. Screening for type 2 diabetes mellitus in adults: U.S. Preventive
Services Task Force recommendation statement. Annals of Internal Medicine
2008;148:846-U63.
23. Norris SL et al. Screening adults for type 2 diabetes: A review of the evidence for
the U.S. Preventive Services Task Force. Annals of Internal Medicine 2008;148:855-
U82.
24. Newbold KB, Danforth J. Health status and Canada's immigrant population.
Social Science & Medicine 2003;57:1981-95.
122
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
28. Lipscombe LL, Hux JE. Trends in diabetes prevalence, incidence, and mortality
in Ontario, Canada 1995-2005: a population-based study. Lancet 2007;369:750-6.
30. Statistics Canada. 1996-7 National Population Health Survey: Derived Variable
Specifications. 1999. Ottawa.
31. Statistics Canada. Population and Dwelling Counts, for Census Divisions, Census
Subdivisions (Municipalities) and Designated Places, 2001 and 1996. 2001.
32. Odell PM, Anderson KM, Kannel WB. New Models for Predicting
Cardiovascular Events. Journal of Clinical Epidemiology 1994;47:583-92.
33. Farewell VT, Prentice RL. A study of distributional shape in life testing.
Technometrics 1977;19:69-75.
34. Yeo D, Mantel H, Lui TP. Bootstrap variance estimation for the National
Population Health Survey. 778-783. 1999. Baltimore, American Statistical Association.
35. Kovacevic MS, Mach L, Roberts G. Bootstrap variance estimation for predicted
individual and population-average risks. Proceedings of the Survey Research Methods
Section. 2008. American Statistical Association.
36. Berg AO et al. Screening for type 2 diabetes mellitus in adults: Recommendations
and rationale. Annals of Internal Medicine 2003;138:212-4.
40. Nam B-H. Discrimination and Calibration in Survival Analysis: Extension of the
ROC Curve for Descrimination and Chi-square test for Calibration. 2000. Boston
University.
42. Wilson PWF et al. C-reactive protein and risk of cardiovascular disease in men
and women from the Framingham Heart Study. Archives of Internal Medicine
2005;165:2473-8.
123
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
43. Pepe MS et al. Limitations of the odds ratio in gauging the performance of a
diagnostic, prognostic, or screening marker. American Journal of Epidemiology
2004;159:882-90.
45. Banerji MA et al. Body composition, visceral fat, leptin, and insulin resistance in
Asian Indian men. Journal of Clinical Endocrinology and Metabolism 1999;84:137-44.
46. Chandalia M et al. Relationship between generalized and upper body obesity to
insulin resistance in Asian Indian men. Journal of Clinical Endocrinology and
Metabolism 1999;84:2329-35.
47. Abate N et al. Adipose tissue metabolites and insulin resistance in nondiabetic
Asian Indian men. Journal of Clinical Endocrinology and Metabolism 2004;89:2750-5.
48. Norman A et al. Total physical activity in relation to age, body mass, health and
other factors in a cohort of Swedish men. International Journal of Obesity 2002;26:670-5.
50. Barba C et al. Appropriate body-mass index for Asian populations and its
implications for policy and intervention strategies. Lancet 2004;363:157-63.
51. Diaz VA et al. How does ethnicity affect the association between obesity and
diabetes? Diabetic Medicine 2007;24:1199-204.
52. Nawaz H et al. Self-reported weight and height: implications for obesity research.
Journal of Preventive Medicine 2001;20:294-8.
54. Shields M, Gorber SC, Tremblay MS. Estimates of obesity based on self-report
versus direct measures. Health Reports 2008;19:1-16.
55. Comstock RD, Castillo EM, Lindsay SP. Four-year review of the use of race and
ethnicity in epidemiologic and public health research. American Journal of Epidemiology
2004;159:611-9.
57. Burchard EG et al. The importance of race and ethnic background in biomedical
research and clinical practice. New England Journal of Medicine 2003;348:1170-5.
124
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
59. Horn OK et al. Incidence and prevalence of type 2 diabetes in the first nation
community of Kahnawa : ke, Quebec, Canada, 1986-2003. Canadian Journal of Public
Health-Revue Canadienne de Sante Publique 2007;98:438-43.
60. Kaler SN et al. High rates of the metabolic syndrome in a first nations community
in western Canada: Prevalence and determinants in adults and children. International
Journal of Circumpolar Health 2006;65:389-402.
61. Pencina MJ et al. Evaluating the added predictive ability of a new marker: From
area under the ROC curve to reclassification and beyond. Statistics in Medicine
2008;27:157-72.
125
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
4. Determining predictors of body mass index and change in body mass index from 1994 –
2004 using a multilevel growth model
4.1 Abstract
Abstract
Background: Increases in obesity and weight are major contributors to chronic disease around
the world, particularly for type II diabetes. Understanding determinants and trajectories of
weight change is an important aspect of public health prevention of chronic disease through
obesity reduction. Few studies have examined the change in Body Mass Index (BMI) over time
and its association with lifestyle and demographic factors using longitudinal population-based
data.
Objective: To understand the predictors of BMI and BMI trajectories in the Canadian population
Methods: This study uses a population-based sample of 14, 123 adults in cycles 1 – 6 of the
longitudinal National Population Health Survey. BMI (at baseline and over time) was modeled
separately for men and women using multilevel growth models with random and fixed effects.
Demographic and lifestyles variables were investigated for their association with these BMI
Results: The multilevel analysis showed that age and initial BMI were associated with higher
BMI and increased physical activity, immigrant status, and smoking were negatively associated
with BMI. Those who were older and had higher BMI had significantly lower rates of BMI
increase over time. Female immigrants were less likely to increase BMI over time compared to
non-immigrants. Adjusting for all factors in the model, an important interaction between income
126
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
and sex was found in which, compared to those with higher income, low-income males had
Conclusions: Lifestyle and demographic factors are associated with BMI and BMI change
over time. Longitudinal data and appropriate analytic techniques, such as multilevel growth
127
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
4.2 Introduction
There are few risk factors so influential to the development of a disease as increased Body Mass
Index (BMI) on the development of type 2 diabetes (1-5). The past and projected increases in
diabetes incidence throughout the world have been attributed mainly to the increasing incidence
of obesity in the population (6). From a primary prevention perspective, understanding factors
which influence excess weight is integral in implementing and planning effective diabetes
prevention strategies.
There have been several cross-sectional studies that have examined correlates of BMI;
however, longitudinal studies of weight change are less common. Previous analysis of change in
self-reported weight in Canada has been analyzed using ordinary linear regression (OLS) (7). In
this study, Orpana et al. found, on average, a trend of weight gain in the Canadian population
from 1996/7 to 2004/5 and recommend further research to identify the correlates and causes of
this trend. However, as stated by the authors of this study, OLS is not as efficient as other
statistical methods because it does not exploit all the information present in longitudinal data.
Accordingly, this study uses longitudinal data and multilevel growth modeling techniques to
expand on previous research to achieve a better understanding of the determinants of weight and
Hierarchical or multilevel growth models can be used to model individual change over
time in an intuitive and flexible way. These analytic methods extend the concept of multilevel
modeling, taking into account the hierarchical structure of the data, such that individuals (level-
1) are nested into groups (level-2). Multilevel growth models treat level-1 variables as within-
person differences over time and level-2 variables as between-person differences independent of
time (8). In addition to understanding what factors influence BMI at each time point, multilevel
128
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
growth models allows one to study the factors that influence BMI change. Compared to other
models, this technique exploits all of the available data by using variability that exists within-
and between individuals to inform estimates of associations. In doing so, this approach uses
longitudinal data to model the predictors of weight and predictors of weight change distinctly,
Using Canadian population-based longitudinal data, this study employs multilevel growth
129
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
4.3 Methods
Study Population
This analysis was done using the longitudinal National Population Health Survey
Canada (9). The NPHS longitudinal sample contains 17,276 persons, from all ages, sampled in
1994/1995. These same persons were interviewed at regular cycles (i.e., every two years) and
will continue to be interviewed up to a total period of 18 years, i.e. 10 cycles. This study uses the
first 6 cycles of the survey, up to 2004/5 and covering a 10-year period. Households were
selected though a stratified, multilevel cluster sampling of private residences using local
planning regions as the primary sampling unit, excluding residents of Indian reserves, long-term
care institutions, prisons, remote areas, and Foreign Service personnel. The survey design was a
two-stage probability sample. The overall response rates were cycle 1, 83.6%; cycle 2, 92.8%;
cycle 3, 88.2%; cycle 4, 84.8%; cycle 5, 80.6%; and cycle 6, 77.4%. The cumulative attrition rate
(i.e. i.e. those who did not complete the questionnaire in all 6 cycles) were: cycle 2 9.3%, cycle 3
15.4%, cycle 4 21.3%, cycle 5 27.3 % and cycle 6. 32.7%. The most significant causes of
attrition were inability to trace and refusal. Average non-response for questionnaire items was <
0.1%. Nevertheless, the methods used in this study do not require respondents to have complete
data for all waves of the survey. Population weights to reflect the population characteristics were
computed based on selection probabilities and post stratification adjustments. Subjects with
missing BMI at the first wave of data collection or those under the age of 18 were excluded from
this analysis.
130
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Variables
Self-reported measures of exposures (highest level of education achieved, income, rural status,
chronic conditions, ethnicity, physical activity, and smoking) were obtained from the NPHS, at
every cycle. The outcome, BMI, was based on self-reported height and weight at each cycle.
BMI is the most commonly measured metric of relative weight and is calculated as weight [in
kilograms (kg)] divided by height [in meters (m)] squared (kg/m2). Baseline BMI (BMI at the
first wave of data collection) was included in the model in order to assess its effect on BMI
change. Categories of physical activity were based upon a Physical Activity Index (PAI). The
PAI is based on an individual's leisure time metabolic energy expenditure (EE). EE is calculated
using the frequency and duration of several physical activities, as well as their metabolic (MET)
value. The MET is the energy cost of the activity expressed as kilocalories (kcal) expended per
kg of body weight per hour of activity, doing a physical activity and the number of times and
time spent on each activity. A PAI < 1.5 kcal/kg/day was considered inactive, moderate was
defined by a 1.5 ≤PAI < 3.0 kcal/kg/day, and active was defined as ≥3.0 kcal/kg/day (10).
Income was assessed using income adequacy which is calculated as the dollar distance between
the individual‘s gross household income in the past 12 months and the low-income cut-off
Statistical Analysis
Descriptive statistics on baseline characteristics and average BMI at each survey were reported
as means (for continuous variables) and percentages (for categorical variables). As part of the
descriptive and exploratory analysis, BMI change (expressed as a slope) and BMI at the first
wave of data collection (expressed as intercept) were calculated using OLS regression. For each
131
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
individual a regression line was fit using BMI values at each time point. Then the slopes and
intercepts were averaged across individuals. Unadjusted BMI and rate of BMI change were
compared between sociodemographic and lifestyle factors and tested using a t-test or one-way
ANOVA depending on if the variable was categorical or continuous. All descriptive statistics
were calculated using survey weights to accurately reflect the demographics of the Canadian
population and account for the survey sampling design. Significance and variance estimates
The multilevel growth model was fit on the longitudinal data using PROC MIXED in
SAS (8) including both fixed and random effects. A random effect describes an estimate in the
model that is generated from a subset of all possible subjects in the population; therefore when
estimating a random effect, the variance for that parameter is also estimated. The multilevel
growth model allows the user to examine both within-individual factors and systematic
differences in growth trajectories that occur across groups. The model solves the within- and
between-individual variation in two stages and can include random effects of variation in slopes
and intercepts between individuals (12, 13). This allows the variances of these parameters to be
The level-1 model describes within-person variation for BMI. This model can be written as:
Where
π1i is the slope or rate of change for the individual i‘s BMI (i.e. average rate of change
across cycles)
132
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
εij is the deviation from the individuals i‘s trajectory from linearity on occasion j (wave).
BMIij is the subject i‘s BMI value at time j (j goes from 0 – 5). Level-1 equations predict the
smallest data-unit, which in this study are BMI-time values. For example, if 10 subjects are
measured 6 times there are 60 level-1 units. Variables that change over time (time-varying
covariates) may also enter into level-1 equations (Figure 1). In this study, the functional form of
the time variable (WAVE) was tested in both its linear and quadratic form to determine which
The level-2 model represents the between-person differences in the change trajectories
(intercept and slope of BMI over time) and time-invariant characteristics of the individual. The
level-1 model only allows people to vary in the values of their individual growth parameters but
the level 2 model ascribes differences in individuals‘ slopes or intercepts such that individuals
who share common predictor values should, on average, vary only according to their individual
change trajectories. The level-2 or between-person model represents the association between a
predictor (shown here as one dichotomous predictor, Xi = 1 or 0) with each subject‘s estimated
initial BMI (i.e. intercept) and rate of change over time (i.e. slope):
π1i = γ10 + γ11Xi + δ1i (Xi is the predictor of rate of change of BMI)
ζ0i 0 𝜎2 2
𝜎01
where ~𝑁 , 20
ζ1i 0 𝜎10 𝜎12
The γs represent the structural or fixed effects of the model. The fixed effects capture systematic
inter-individual differences such as age, ethnicity, or immigrant status. In the above model:
133
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
γ00 and γ10 are the level-2 intercepts which represent the mean initial BMI (intercept) and mean
Variance components
δ0i and δ1i are level-2 error variables (residuals) which represent the stochastic variation in the
individual growth models allowing each individual‘s growth parameters (slopes and intercepts)
to differ from the population average (See figure 2). In other words, δ0i and δ1i represent the
between-individual stochastic variation for the level-1 predictors (slope and intercept).
Specifically:
δ0i = Difference between population average initial BMI and individual i‘s level of BMI
δ1i = Difference between population average rate of change of BMI and individual i‘s rate of
change of BMI.
δ0i and δ1i represent the unexplained variation in initital BMI and BMI rate of change. Their
variances, defined by the matrix above (σ02 , σ12 & σ012) represent the population variation in the
individual intercept and slope around the mean initial BMI and the mean rate of change as
134
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
σ012 = population covariance between baseline BMI and BMI rate of change
Interpreting these values (σ02 , σ12 & σ012) allow us to comment on how much heterogeneity
exists in the change parameters after accounting for the variables in the model.
Moderately Inactive
Active
active
Time-varying
BMI
Time-invariant
1 2 3
Wave
135
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Figure 2. Ten randomly selected growth trajectories from individuals in the cohort to
represent variation in growth trajectories between individuals in the cohort.
136
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
In growth models, level-1 units typically represent time and are nested within individuals.
Therefore, using the earlier example of 10 subjects measured 6 times, we would have 10 level-2
units.
Therefore, our full joint multilevel growth model can be written as:
BMIij = γ00 + γ01Xi + γ10* (WAVEij) + γ11Xi* (WAVEij) + δ1i δ0i + εij
Multilevel growth modeling makes use of intra-individual variation and inter-individual variation
(thus accounting for clustering within subjects) in an efficient manner since coefficients are
iteratively estimated at both levels. Although attrition can be a significant problem with
longitudinal data, multilevel models are able to accommodate missing or unbalanced data. The
underlying assumption in multilevel growth models is that each individual‘s observed data are a
random sample from their underlying growth trajectory allowing them to be estimated even
without complete data (14). Also, the model is able to handle between-individual variation in
the timing and frequency of measurements because the timing of each measurement occasion is
treated as an explanatory variable in the model; therefore, data from individuals‘ different
measurement patterns (e.g. some of whom may have been measured only once and others at
several irregularly spaced intervals) can be analyzed simultaneously in the same model. Each
BMI-trajectory is a combination of the observed trajectory for the individual based on the
measured waves, and the model-based trajectory based on the predictor variables.
137
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
The following covariance structures were examined and compared to determine the best
fit with the data: Unstructured (UN), Compound Symmetry (CS), Autoregressive(1) (AR1), and
Heterogeneous AR(1) (ARH). The final covariance structure was chosen based on the most
favorable information criteria [Akaike‘s Information Criteria (AIC) and Bayesian Information
Criteria (BIC)], log-likelihood statistics, and number of iterations. Standard diagnostics with
level-1 and level-2 residuals were performed to check model assumptions of normality and
homoschedasicity.
multilevel cluster sampling of private residences using local planning regions as the primary
sampling unit. In order to account for the way that the respondents were sampled normalized
weights that account for selection probabilities were used. Normalized weights represent the
section) divided by the total sample size; accordingly, the sample size is not inflated but the
differential probabilities associated with the survey design are accounted for. All statistical
analyses were carried out using SAS version 9.2 (SAS Institute, Inc, Cary, NC).
138
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
4.4 Results
There were 17, 276 respondents in the first wave of the longitudinal NPHS. Those under the age
of 18 (n = 3,159) and without BMI information in wave 1 or 2 (n = 354) were excluded from the
cohort. The final sample size was 14, 123, or, separated by sex, 7,496 women and 6,627 men
(Figure 3).
Starting Cohort
N = 17,276
354 excluded
3,159 excluded <
no BMI in wave 1
18 yrs
or 2
N=13,763
FINAL COHORT
N = 7,496 females
N = 6,627 males
Baseline characteristics of the cohort are shown in table 1. Males had an average baseline BMI
of 25.9 kg/m2 and women at 24.7 kg/m2. Male BMI increased on average by 4.4% and female
BMI increased on average by 5.3% over the 10-year period (Figure 4). OLS regression analysis
revealed that baseline BMI and average change in BMI were associated with a number of
139
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
population characteristics (Table 2 & 3). Higher levels of physical activity and smoking status
were associated with lower baseline BMI; however, these factors also tended to be positively
associated with higher rates of BMI increase over time (not statistically significant). Low
income status was negatively associated with baseline BMI in men but the opposite was true for
women. In a cross-sectional analysis at the amid-point of the study (2003) the association
between income quintile and percentage of individuals that reached the obese cutoff (BMI ≥ 30
kg/m2) were not strongly associated in men; however the likelihood of being obese was strongly
140
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Figure 4. Mean unadjusted Body Mass Index (BMI) (+/- standard error) across time
[National Population Health Survey (NPHS) 2-year cycles), by sex.
28
Males
27.1
26.9
27
Females
26.6
26.3
BMI (kg/m2)
26.0 26.0
26 25.9
25.8
25.5
25.2
24.9
25 24.7
24
1 2 3 4 5 6
Cycle
141
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Table 2. Average BMI at the first wave of data collection (i.e. baseline BMI) (+/- standard
error) (i.e. intercepts from OLS) and tests for significant differences between cohort
characteristics by sex.
MALES FEMALES
N = 6,627 N = 7,496
Mean (se)* P-value Mean (se) P-value
or difference from or difference from
reference (se) reference (se)
Overall 25.88 (0.07) 24.70 (0.07)
Age
18-30 24.48 (0.12) (ref) 23.18 (0.14) (ref)
31-60 +1.86 (0.14) <0.0001 +1.88 (0.18) <0.0001
≥ 61 + 1.54 (0.20) <0.0001 +2.25 (0.20) <0.0001
Low Income
No 25.93(0.07) (ref) 24.56 (0.07) (ref)
Yes - 0.59 (0.19) 0.0026 +0.57 (0.19) 0.0027
Immigrant
No 25.92 (0.08) (ref) 24.75 (0.07) (ref)
Yes - 0.48 (0.20) 0.0155 -0.22 (0.19) 0.2842
Physical Activity
Low 25.97 (0.09) (ref) 24.97 (0.09) (ref)
Medium -0.27 (0.17) 0.1139 -0.48 (0.18) 0.0058
High -0.38 (0.16) 0.0152 -1.16 (0.19) <0.0001
Education
< Secondary School 26.12 (0.13) (ref) 25.63 (0.15) (ref)
Secondary School Grad -0.15 (0.23) 0.6139 -0.94 (0.25) 0.002
Post-secondary School Grad -0.49 (0.17) 0.0047 -1.35 (0.18) <0.0001
Marital Status
Married 26.44 (0.08) (ref) 25.17 (0.10) (ref)
Common-law -1.22 (0.22) <0.0001 -0.16 (0.31) <0.0001
Single -2.00 (0.21) <0.0001 -2.05 (0.18) <0.0001
Separated or Divorced -0.60 (0.21) 0.0039 -0.06 (0.24) <0.0001
Widowed -0.65 (0.78) 0.4072 +0.15 (0.22) 0.7904
Current Smoker
No 26.03 (0.08) (ref) 24.92 (0.09) (ref)
Yes -0.66 (0.14) <0.0001 -0.76 (0.15) <0.0001
Any chronic condition
Yes 25.60 (0.11) (ref) 23.94 (0.12) (ref)
No +0.42 (0.13) 0015 +1.28 (0.15) <0.0001
se: standard error - all standard errors are computed using survey bootstrap weights
142
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Table 2. Average change in BMI (+/- standard error) (i.e. slopes from OLS) for significant
differences between cohort characteristics by sex.
MALES FEMALES
Mean (se) P-value Mean (se) P-value
or difference from or difference from
reference (se) reference (se)
Overall 0.206 (0.01) 0.221 (0.01)
Age
18-30 0.365 (0.02) (ref) 0.327 (0.02) (ref)
31-60 -0.186 (0.03) <0.0001 -0.108 (0.02) <0.0001
≥ 61 - 0.414 (0.04) <0.0001 -0.420(0.03) <0.0001
Quartile of baseline BMI
Q1 0.339 (0.03) (ref) 0.265 (0.01) (ref)
Q2 -0.109 (0.03) 0.0005 -0.055 (0.02) 0.0150
Q3 -0.186 (0.03) <0.0001 -0.131 (0.02) <0.0001
Q4 -0.272 (0.03) <0.0001 -0.225 (0.03) <0.0001
Low Income
No 0.207 (0.01) (ref) 0.193 (0.01) (ref)
Yes -0.007 (0.03) 0.8375 -0.050 (0.0.03) 0.0463
Immigrant
No 0.191 (0.03) (ref) 0.187 (0.07) (ref)
Yes - 0.058 (0.20) 0.0750 -0.057 (0.19) 0.0198
Physical Activity
Low 0.162 (0.013) (ref) 0.1588 (0.02) (ref)
Medium +0.047 (0.02) 0.044 + 0.038 (0.02) 0.0913
High +0.038 (0.03) 0.145 + 0.155 (0.03) 0.0091
Education
< Secondary School 0.138 (0.02) (ref) 0.063 (0.02) (ref)
Secondary School Grad +0.008 (0.03) 0.7963 + 0.140 (0.03) <0.0001
Post-secondary School Grad +0.070 (0.02) 0.0040 +0.155 (0.02) <0.0001
Marital Status
Married 0.130 (0.01) (ref) 0.152 (0.03) (ref)
Common-law +0.101 (0.03) <0.0001 +0.145 (0.04) 0.0007
Single +0.035 (0.04) 0.0017 +0.153 (0.04) <0.0001
Separated or Divorced +0.130 (0.04) <0.0001 +0.09 (0.03) 0.0232
Widowed -0.017 (0.03) 0.4115 -0.207 (0.03) <0.0001
Current Smoker
No 0.161 (0.02) (ref) 0.151 (0.01) (ref)
Yes +0.059 (0.02) 0.0029 +0.091 (0.02) <0.0001
Any chronic condition
Yes 0.222 (0.01) (ref) 0.2120 (0.01) (ref)
No -0.080 (0.02) 0.0015 -0.062 (0.02) 0.0007
se: standard error - all standard errors are computed using survey bootstrap weights
143
Figure 6. Conceptual diagram of the mulilevel growth model showing variables in the model at each level.
24
22 21.95
21.01
20.7 p < 0.1109
20 19.98 19.76
19.15
18.86
18
17.75
% Obese
16.9
16
14
Males
13.18
Females p < 0.0001
12
10
Lowest Lower-Middle Middle Upper Middle Highest
Income Adequacy Quintile
Figure 6. Conceptual diagram of the mulilevel growth model.
Age (γ01)
σ02 Baseline BMI (γ02)
Low income(γ03)
Immigrant (γ04) σε2
Ethnicity (γ05)
Intercept
(π0i)
BMI (y)
Wave
(π1i)
Time-varying covariates
Intercept (γ10) (π2i - π8i)
Age (γ11)
Baseline BMI (γ12) Physical Activity
Smoking Status
σ12 Marital Status
The conceptual diagram of the multilevel growth model can be seen in figure 6 and the results
from the model are shown in table 4. All four covariance structures (CS,ARH, AR, and UN)
produced similar results. However, AIC, BIC and log-likelihood statistics showed ARH(1)and
UN produced the best fit. Since iteration time was quicker for UN, this covariance structure was
chosen for subsequent analyses. Quadratic forms of time were considered but found not to be
statistically significant and not to improve the model properties; therefore, the linear form of
time was used in the model. All variables were assessed for their variability over time. Physical
activity, marital status, and smoking status, varied across waves and thus were kept in their time-
varying form. Income adequacy was kept in its time-invariant form due to its minimal variation
The multilevel growth model shown in table 4 reveals several important predictors for BMI
and its trajectory. The full specification of the multilevel model for males is shown below:
146
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Combined Model:
γ09(Ethnicity – other/mixed)
Reference groups:
147
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Table 4 shows the multilevel results from 3 models. The first model (model 1) described the
model without covariates or the effects of time. This model is also known as the unconditional
growth model since the effects of initial BMI and rate of change are is estimated independent of
any terms in the model. The second model (model 2) shows only age and initial BMI and was
used for comparison when building the multivariate model. The third model (model 3) is the
fully adjusted model. Each multilevel model has two intercepts, one for baseline BMI at wave 1
(γ00) and another for subsequent change in BMI over time 1 (γ10). Average BMI and BMI change
in the unconditional growth model was 26.03 km/m2 and 0.2031 kg/m2 for males 24.92 kg/m2
and 0.2245 kg/m2 for females. When each term was added to the second model, results were
compared to the unconditional and conditional growth model and improvement tested using the
Controlling for all variables in the model, males have an average increase in BMI of 0.1655
(± 0.0157) kg/m2 and females have an average increase of 0.2216 (± 0.0110) kg/m2. In model 3
in men and women, increasing age and baseline BMI were associated with an increase in BMI;
however those that were older and/or had a higher BMI at baseline increased at a reduced rate.
Prototypical trajectories, based on the multivariate model in females, demonstrating the effects of
In the final model, marital status, physical activity, smoking status, and chronic conditions
(for females) were modeled in their time-varying form due to their fluctuations over the 10 year
predictor. Controlling for the effects of time and other variables in the model, increasing levels
of physical activity were associated with lower levels of BMI and the effect was stronger in
148
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
females than males. Smoking was also associated with lower levels of BMI and being married
was associated increased BMI. Having a chronic condition was not associated with BMI or BMI
rate of change in males but was positively associated with average BMI in females. Income,
immigrant status and ethnicity were modeled in their time invariant form. In the multivariate
models, immigrant status was associated with lower BMI; however the effect was stronger in
females. In addition, South Asian ethnicity was associated with increased BMI compared to
white ethnicity females only. Adjusted for all other factors, in men low income was associated
with lower BMI compared to those in the medium or high income category whereas in women,
low income status was associated with increased BMI. When the models were analyzed with
only the subset of individuals with complete, balanced data, the findings did not differ in
direction or statistical significance; however, the magnitude of the associations was slightly
larger.
The random variance components for the intercept and slope of the multivariate models
are also shown in table 4a&b. The random effects for both initial BMI status (σ02) and rate of
change (σ12) were statistically significant (P<0.0001) in all models, indicating significant
variation in baseline levels of BMI and rate of change within the population after controlling for
all variables in the final model. Furthermore, the covariance between baseline BMI and BMI rate
of change and (σ012) was also significant (P<0.0001) meaning that variance in BMI trajectories
depends on initial BMI. The majority of the variation in initial BMI was explained by the fixed
effects added to the model (σ02 decreased from 13.625 to 0.0781 in males and 22.839 to 0.1143
in females). Age and baseline BMI were the main fixed effects that reduced the variance in
initial BMI.
149
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
finding in multilevel growth models (14). The variation in initial status was almost double the
magnitude in females versus males indicating that BMI is much more variable in Canadian
150
Table 4a. Results of multilevel growth model in males. All variables in the model were centered to their overall mean.
Goodness of fit
minus2logL 154417 139376 112726
p-value versus model 1 15041 41691
(df =4; p<0.0001) (df =19; p<0.0001)
versus model 2 26650
(df =15; p<0.0001)
152
Doctor of Philosophy Epidemiology (PhD) Dissertation Laura C.A. Rosella
Table 4b. Results of hierarchical linear model fitting in females. All variables in the model were centered to their overall mean.
Common law
π5i γ50 Marital Status -
Single -0.1821 (0.038) 0.6996
π6i γ60 Marital Status –
Separated or divorced -0.2045 ( 0.0425) <.0001
π7i γ70 Marital Status -
Widowed -0.0808 ( 0.0545) <.0001
π8i γ80 Current Smoker
-0.3129 (0.03219) <.0001
Goodness of fit
minus2logL 188503 171270 143960
p-value versus model 1 17233 44913
(df =4; p<0.0001) (df = 20; p <0.0001)
versus model 2 27310
(df = 16; p <0.0001)
AIC (smaller is
better)
AIC versus model 1
Improvement 188515 171290 143973
versus model 2 yes yes
BIC BIC (smaller is better)
Improvement yes
154
Doctor of Philosophy Epidemiology (PhD) Dissertation Laura C.A. Rosella
155
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Figure 7. Prototypical plot of BMI change over time for quartiles of baseline age in females
at population average levels for all other characteristics of the cohort.
156
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
4.5 Discussion
are related to BMI and BMI change over time. Several demographic and lifestyle factors
significantly influence baseline BMI (intercept) and rate of change (slope) over 10 years.
Though factors which influence BMI and BMI change over time were generally similar between
men and women there was an important difference in the influence of income, namely that low
income status was protective in men and a risk factor in women. Furthermore, all associations
There are several strengths to this study. This sample is population-based with high
participation (response rates ≥ 77%) and representative of the Canadian population (15).
Furthermore, the longitudinal design allowed this study to overcome limitations associated with
cross-sectional designs or single follow-up cohort designs which are unable to consider
trajectories of growth or predictor variables that vary over time. The multilevel growth model
used in this study assessed patterns of growth trajectories, taking into account both within- and
between-individual variation. Other approaches to study change over time, including OLS
regression, are unable to model these within-individual variations since the data are collapsed to
the level-2(16). An additional strength was the use of time-dependent covariates, which has not
been considered in previous studies. This is particularly important to consider with lifestyle
variables that are known to change over time, such as physical activity or smoking status. A
variable that fluctuates over time that is modeled in its time invariant form can lead to the
157
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Previous analysis of change in self-reported weight in Canada used OLS (7). Orpana et
al. did not examine the effects of the factors investigated in this study; however they did examine
how change in BMI was influenced by sex and age group. Consistent with our study, they found
that weight gain is higher among young and middle aged-adults compared with those greater
than 65, and that females were more likely to increase over time compared to males.
Evidence from cross-sectional studies have generally agreed with the finding that
physical activity is inversely associated with BMI (18-21); however, longitudinal studies have
been less consistent (22-24) though it is unclear if methodological differences, the definition or
previous study has considered the time-dependent nature of leisure activity, as was done in this
study, and this may have strengthened the ability to find a significant inverse association. The
association between low BMI and smoking is well documented (25-27) and confirmed in this
study, even when considering the time-varying nature of smoking status, which has not been
done in previous studies. Previous studies have also shown, in general, that immigrant status is
associated with decreased BMI (28-31); and though acculturation is thought to lessen this effect,
and even reverse it after a substantial amount of time, that phenomenon was not observed in the
time period of this study. This study revealed that BMI tends to increase with time for younger
people, and even more so for people who start with normal BMI values (< 25 kg.m2) as in
previous studies (7, 16, 32). This emphasizes the need to consider population demographics
particularly important in young adults. It is possible that targeting this group will have the most
impact for preventing obesity in the population, since they are at highest risk for weight gain.
158
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
An interesting finding from this study was the opposite association between low income
status and obesity by sex. In their landmark analysis, Sobal and Stunkard (1989) pointed out that
in developed nations SES influences on weight affect men and women differently and is perhaps
related to attitudes and cultural influences (33). It has been shown consistently that adult women
tend to have a strong inverse relationship between SES indicators and obesity; that is, the higher
the SES the less likely the women were to have a higher BMI and/or gain weight (33-36).
Furthermore the prevalence of obesity and being overweight is paradoxically higher in women
with food insecurity(37-41). There are a variety of explanations for these results. Unhealthy
patterns in nutrition (high-density/low nutrient diets) occur in women to a greater extent than
men. Lower income women are less likely to report reducing calories and snacks and limiting
meal proportions to healthy size (42). Emotional eating, described as the practice of consuming
large quantities of food (usually unhealthy food) in response to feelings instead of hunger, is
more likely to affect women versus men (43). Low SES women may be more likely to
experience emotional eating due to feelings of lack of control over their own lives, depression,
emotional stress and poor self-esteem associated with low social class. Women who lack self-
esteem, have high psychological demand and low controls in their occupation are less likely to
exercise and/or have healthy dietary habits (44). A clear demonstration of the SES effects in
women is the contradictory trend of increased obesity among food-insecure women. This is
known as the ‗hunger-obesity paradox‘ in which under-nutrition and obesity co-exist (40). In
periods of food insecurity, women often change their eating patterns to ensure sufficient food for
their children and/or husbands (43). The psychosocial and emotional changes that occur in times
of food restriction result in an increased likelihood to binge eat when food is available and eat
high-calorie foods (one of the key reasons why food restriction is an ineffective method of
159
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
dieting). This phenomenon occurs in a dose-response manner such that the more severe the
food insecurity the more likely the women is to have a binge pattern of eating (39). Men are less
food insecurity. Furthermore, men are not as susceptible to fluctuations in food consumption
during times of food insecurity compared to women (41). An additional difference is the
potential protective effect of low-income labor. Low SES jobs for men tend to be more
physically active than high SES jobs (44). This increased exposure to physical activity burns
calories and is potentially protective of BMI. Our analysis revealed that relationship status had
an impact on future weight, such that those that were married had higher levels of BMI
compared to those who were in other types of relationship or not in a relationship. Previous
research using an individual-level fixed effects model on a national longitudinal survey from the
U.S. also found a significant relationship between marriage and weight gain and proposed
explanations related to shared meals, decreased individual physical activities, and social
obligations (45).
There are several limitations to consider in this study. First, all factors are self-reported
including height and weight, thus possibility subject to measurement error. Non-differential
error would lead to an underestimation of the true associations(46, 47). In general, there is a high
correlation between self-report and measured height and weight (48), however, some studies
have shown that under or over-estimation of height can be related to factors such as sex and
socioeconomic status (49-51). If present, these self-reporting biases may have affected the
direction of association in our study. Recent research into methods to correct for self-reporting
error in these health surveys have recently been introduced. Incorporating these correction
factors is a subject of future research (52). Second, due to the complex nature of the modeling
160
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
process, the models were kept relatively simple to ensure convergence; therefore, additional
variables related to labor, social support, and specific co-morbid conditions were not considered.
Third, additional parameters important to weight gain were not included in the survey‘s data file.
Specifically, there was no detailed information on reproductive factors or dietary habits, which
are a potentially important aspect of weight and weight gain. With respect to the SES
relationship in women, a previous study demonstrated that a large proportion of the gradient was
explained by reproductive factors (44). Not controlling for reproductive factors could cause
significant confounding in women since women of lower SES are more likely to have children
(and more of them), putting them at increased risk for obesity. Consequently, inability to control
for reproductive factors may overestimate SES effects in women. Finally, attrition occurs across
waves and, though this model can handle missing data, the decrease in participants in later waves
may reduce statistical power to detect relationships or higher powers in the functional form for
change.
disease burden, particularly for diabetes (53). This study describes the lifestyle and demographic
factors associated with BMI and BMI change over time. The study‘s population-based
longitudinal design and multilevel growth model analyses offered a unique opportunity to
investigate new aspects of these relationships. This study shows that weight and weight gain are
influenced by age, sex and baseline BMI. Of the lifestyle variables, physical activity was
protective for BMI, controlling for the effects of time and accounting for the time-dependent
nature of this variable. Furthermore, this study shows differential effects of lifestyle and socio-
demographic factors on BMI by sex. This study shows that young and low-income women are at
higher risk for increased BMI compared to older women and compared to men of the same ages.
161
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Given that young women are particularly at risk of for developing diabetes in the next 10 years
(Rosella et al, under review) particular attention should be paid to this sub-group. Though not
specifically carried out in this study, this model can be explored for its potential as a tool to
predict future BMI in the population, which may be used when generating future estimates of
162
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
4.6 References
1. Colditz G et al. Weight as a risk factor for clinical diabetes in women. American
Journal of Epidemiology 1990;132:501-13.
2. Colditz G et al. Weight gain as a risk factor for clinical diabetes mellitus in
women. Annals of Internal Medicine 1995;122:481-6.
6. Mokdad AH et al. The continuing epidemics of obesity and Diabetes in the United
States. JAMA: The Journal of the American Medical Association 2001;286:1195-200.
7. Orpana HM, Tremblay MS, Fines P. Trends in weight change amoung Canadian
adults. Health Reports 2007;18:9-16.
8. Singer JD. Using SAS PROC MIXED to fit multilevel models, hierarchical
models, and individual growth models. Journal of Educational and Behavioral Statistics
1998;23:323-55.
10. Statistics Canada. 1996-7 National Population Health Survey: Derived Variable
Specifications. 1999. Ottawa.
11. Yeo D, Mantel H, Lui TP. Bootstrap variance estimation for the National
Population Health Survey. 778-783. 1999. Baltimore, American Statistical Association.
12. Demidenko E, Stukel TA. Efficient estimation of general linear mixed effects
models. Journal of Statistical Planning and Inference 2002;104:197-219.
13. Stukel TA, Demidenko E. Two-stage method of estimation for general linear
growth curve models. Biometrics 1997;53:720-8.
14. Singer JD, Willett JB. Applied Longitudinal Data Anlaysis:Modeling Change and
Event Occurrence. Oxford: Oxford University Press, 2003.
163
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
15. Statistics Canada. NPHS Public Use Microdata Documentation. 1999. Ottawa,
Canada, Statistics Canada.
16. Heo M et al. Hierarchical linea models for the development of growth curves: an
example with body mass index in overweight/obese adults. Statistics in Medicine
2003;22:1911-42.
17. Streiner DL, Norman GR. Measuring Change. Health Measurement Scales. New
York: Oxford University Press, 2003:194-212.
18. Haapanen N et al. Association between leisure time physical activity and 10-year
body mass change among working-aged men and women. International Journal of
Obesity 1997;21:288-96.
19. Hu FB et al. Walking Compared With Vigorous Physical Activity and Risk of
Type 2 Diabetes in Women . JAMA 1999;282:1433-9.
25. Fogelholm M et al. Predictors of weight change in middle-aged and old men.
Obesity Research 2000;8:367-73.
27. Molarius A et al. Smoking and relative body weight: An international perspective
from the WHO MONICA project. Journal of Epidemiology and Community Health
1997;51:252-60.
164
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
31. Wandell PE et al. Country of birth and body mass index: A national study of
2,000 immigrants in Sweden. European Journal of Epidemiology 2004;19:1005-10.
32. Bjorvell H, Rossner S. A ten year follow-up of weight change in severely obese
subjecrs treated in a behavioural modification program. International Journal of Obesity
1990;14:88.
33. Sobal J, Stunkard AJ. Socioeconomic status and obesity: a review of the
literature. American Psychological Association 1989;105:260-75.
34. Ball K, Crawford D. Socioeconomic status and weight change in adults: a review.
Social Science in Medicine 2005;60:1987-2010.
37. Adams E et al. Food insecurity is associated with increased risk of obesity in
California women. Journal of Nutrition 2003;133:1070-4.
38. Basiotis P, Lino M. Food insufficiency and prevalence of overweight among adult
women. Family Economics and Nutrition Review 2003;15:55-7.
40. Olson C. Nutrition and health outcomes associated with food insecurity and
hunger. Journal of Nutrition 1999;129:521S-4S.
165
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
43. Parker S, Keim K. Epic perspectives of body weight in overweight and obese
women with limited income. Journal of Nutrition Education and Behavior 2004;36:282-9.
45. Averett SL, Sikora A, Argys LM. For better or worse: Relationship status and
body mass index. Economics and Human Biology 2008;6:330-49.
48. Nawaz H et al. Self-reported weight and height: implications for obesity research.
Journal of Preventive Medicine 2001;20:294-8.
50. Niedhammer I et al. Validity of self-reported weight and height in the French
GAZEL cohort. International Journal of Obesity 2000;24:1111-8.
52. Gorber SC et al. The feasibility of establishing correction factors to adjust self-
reported estimates of obesity. Health Reports 2009;19.
166
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
5. Thesis Conclusion
Each of the studies presented in this thesis are distinct in their data and methodological approach
but are linked in their common goal to inform population-based risk prediction for diabetes.
Taken together, these studies can inform public health aspects of diabetes & obesity and
epidemiological methods.
In section 2, a novel risk tool - The Diabetes Population Risk Tool (DPoRT) - used to
estimate the incidence of type 2 diabetes that can be applied at the population level using
publicly available data was created. Four important goals were achieved with this work. First an
algorithm to predict the incidence of diabetes with good discrimination and accuracy was
developed. Secondly, an important policy advantage was achieved by building the tool so that it
can be applied to the current risk factor surveillance data (routinely collected survey data) that is
publicly available in Canada. This allows DPoRT to be used by a wide audience of health
planners to accurately estimate diabetes incidence and quantify the impact of interventions.
Thirdly, the vigor of the validation of DPoRT demonstrates a framework, which should be
applied to the validation of other population-based risk algorithms. Finally, the novel application
of a risk algorithm at the population level reveals and important way to understand distribution
of diabetes risk in populations. The subsequent sections of this thesis were built from specific
aspects related to the performance of the risk tool (as affected by measurement) and BMI change
One of the key aspects of any epidemiological study is the accurate measurement of
exposures and outcomes. To achieve the goals of DPoRT a balance between data availability and
data detail was sought. This choice led to several questions regarding the consequence of public
survey data. In Chapter 3.1, the impact of measurement error (systematic and random) in self-
167
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
reported height and weight on the performance of a diabetes risk prediction tool was studied.
This study demonstrated that systematic and random error decreased the calibration of a
prediction model and only random error reduced the model‘s discrimination. One of the
particularly interesting findings of this study was the finding that of non-differential error can
introduce systematic bias on predicted risk estimates. This study confirms that level of error in
self-reported height and weight in Canada‘s health surveys unlikely to affect the performance or
validation of DPoRT. Importantly, this study provides a framework to quantify the influence of
measurement error on risk prediction using simulation. This work reaffirms that researchers
developing and validating risk tools must be aware of the presence of measurement and its
impact on the performance of their risk tools. Further efforts must be made to understand the
nature of error in self-reporting measurements and ongoing work to improve the quality of
measurements used in risk algorithms will improve model performance. Understanding the
consequence of measurement error on risk prediction is not only important for population risk
tools but can provide further insight for understanding the influence of measurement properties
which can be used to provide evidence for making decisions about data utilization for different
applications in epidemiology.
The results from the investigation into the impact of ethnicity on diabetes risk (3.2) have
several important implications for application of DPoRT in addition to providing insight into the
independent role of ethnicity in the development of diabetes. From this study we find that
although from the individual risk perspective, focusing on different ethnicity may be important,
when predicting new cases of diabetes at the population level and accounting for other risk
factors, detailed ethnic information did not significantly improve DPoRT population estimates of
new diabetes cases. This indicates that for health planning purposes including detailed ethnic
168
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
information may not improve population estimates of future diabetes risk, which has significant
implications on how the tool can be used. It also demonstrates the diminishing return that is
often seen in prediction tools, such that maximum levels of discrimination and goodness of fit
can often be achieved without detailed information on risk factors. While the clinical appeal of a
using detailed ethnicity for estimating diabetes risk is significant, the statistical reality is that
classifying ethnicity in broader categories which are available to the public work as well as a
model which include detailed ethnic information in the Canadian population. Conclusions drawn
from both studies within the measurement section can be used to inform research on the
influence of measurement properties (error and type) on modeling and statistical prediction.
In building the risk tool for diabetes it was demonstrated that that BMI (a relative
measure of weight for height) overwhelmingly influences the predictions for developing diabetes
in the future. For that reason, clarifying determinants of weight and weight change is essential
when developing strategies to prevent or reduce the future diabetes burden. In monitoring trends
over time researchers are often faced with the dilemma of separating trends between individuals
and trends within individuals. Multilevel growth models allow us to model both these aspects
which strengthen the ability to model trends that vary between and within individuals. In
Chapter 4 predictors of weight and weight change were modeled in a longitudinal sample of
Canadians. Specifically importance of age on baseline obesity levels and rate of change are
quantified. The fact that younger individuals are at greatest risk of increasing obesity reinforce
the fact that obesity prevention is most important in younger adults. DPoRT also reinforces that
those with high BMI at lower ages will have a higher probability of developing diabetes than
those that have higher weight at older ages. The fact that factors that influence BMI substantially
depend on gender (particularly income) reveal that interventions for reducing weight may need
169
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
to be considered differentially between genders. The multilevel growth model provides a more
efficient way for predicting weight change over time and can help inform DPoRT and improve
predictive estimates. In addition, weight interventions modeled using DPoRT can be better
informed by the findings of this model by targeting interventions for those at highest risk for
weight gain. Finally, Chapter 4 demonstrates an important use for multilevel growth models in
epidemiology to understand trends of risk factors or diseases that change over time.
There are several follow-up studies which will be conducted based on the findings of this
thesis. Validation of DPoRT will continue as more effective methods to quantify variation in
case ascertainment in different populations are developed. As the surveillance of risk factors
and diabetes improves, DPoRT can be adapted to become even more accurate, while maintaining
its accessibility for decision makers. Future research on measurement error will include
investigation into differential error with respect to disease status or other characteristics of the
population. Though not specifically carried out in this thesis, the multilevel growth model can be
used to predict future BMI in the population, which may be used when generating future
Taken together, this thesis represents a body of literature focused on diabetes risk
prediction at the level of populations. Policy makers and health planners can apply these findings
to estimate and plan for the upcoming diabetes epidemics in a country, such as Canada. The
target and how extensive a strategy is needed to stabilize or reduce the number of new cases.
and public health. The methods and data used for DPoRT development and validation can
potentially be applied for other diseases in Canada and represent a new way to understand
170
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
population health. The results and methods from investigating measurement and obesity are
171
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
6. Appendix
172
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Discrimination: Discrimination is the ability to differentiate between those who are high risk
and those who are low risk – or in this case those who will and will not develop diabetes given a
The ROC curve is a good way to measure discrimination. An ROC curve repeats all possible
pairings of subjects in the sample who exhibit the outcome and do not exhibit the outcome and
calculates the proportion of correct predictions- essentially being and index of resolution of the
model. This proportion under the receiving operator curve is equal to the C statistic which can
be used to assess the degree of discrimination - 1.0 being perfect discrimination and 0.5 being
no discrimination (1-3). A perfect prediction model would perfectly resolve the population into
those who get diabetes and those who do not. Accuracy is unaffected by discrimination,
Calibration (This concept is also knows as accuracy and/or reliability in the literature):
Calibration is achieved in a prediction model if it is able to predict future risk with accuracy i.e.
if the predicted probabilities closely agree with the observed outcomes. A model that is not
reliable will have significant over- or under- estimation of risk in the overall population and/or
within certain subgroups. A model with good accuracy model will maintain reliability across
various risk groups and other important subpopulations. Accuracy is not an issue if the purpose
173
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Calibration in the Framingham prediction models has been assessed statistically by a statistic
developed by D‘Agostino (4). It is calculated by dividing the cohort into deciles of predicted
risk and comparing observed versus predicted risk resulting in a modified version of Hosmer-
Lemenshow χ2. Other measures of assessing accuracy include graphical methods and
The Hosmer-Lemeshow statistic is then compared to a chi-square distribution with (g-n) degrees
of freedom. As with the other GOF tests, evidence of lack of fit is demonstrates as the chi-
174
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Reference List
(1) Harrell FE, Lee KL, Mark DB. Multivariable prognostic models: Issues in developing
models, evaluating assumptions and adequacy, and measuring and reducing errors.
Statistics in Medicine 1996; 15:361-387.
(2) Harrell FE. Regression Modeling Strategies With Applications to Linear Models, Logistic
Regression, and Survival Analysis. New York: Springer, 2001.
(4) D'Agostino RB, Grundy S, Sullivan LM, Wilson P. Validation of the Framingham
Coronary Disease Prediction Scores. JAMA 2001; 286(2):180-187.
Acronyms
175
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
6.2 Ethics
For the DPoRT work approval to link survey data (NPHS 1996/7, CCHS 2000/1) to
hospital administrative data held at ICES was received from Sunnybrook Health Sciences
Centre Research Ethics Board in January 2005. The study design and use of hospital
administrative data meets the requirements of the ICES research agreement with the
Ontario Ministry of Health and Long Term Care. Ethics approval was sought from the
University of Toronto Research Ethics Board and Sunnybrook and Women‘s College
Health Sciences Centre Research Ethics Board and granted in January 2005. Use of the
Manitoba data was approved by the Health Information Privacy Committee and University
Longitudinal Health Survey at the Research Data Centre and the University of Toronto was
176
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
177
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
SAS code
***** INPUT PARAMETERS *****;
%LET TOTREP = 500; %LET N = 10618;
%LET BETA1 = 0.4565 ; %LET SEED1 = 55 ;
%LET MEANHEIGHT = 1.627; %LET MEANWEIGHT = 64.761;
%LET STDHEIGHT_OBS = 0.069 ;%LET STDWEIGHT_OBS = 12.320 ;
%LET SEEDU = 8971 ;
%LET BETA2 = - 0.00509; %LET PREV = 0.0736;
%let intercept = -10.8967;
%LET SEEDE = 4567; %LET STDBMIe = 5.00 ;
%LET SEED2 = 123 ; %LET SEED3 = 15 ;
%LET SEED4 = 2345 ; %LET SEED5 = 4513;
%LET SEED6 = 9876; %LET CORRHW = 0.311;
%LET ICCHEIGHT = 0.9; %LET ICCWEIGHT = 0.9;
%let b_wt = 0;%let b_ht = 0;
data time;
format start time.;
start = time();
output;
run;
DATA TEMP ;
A1 = &STDHEIGHT_OBS ;
B1 = &STDWEIGHT_OBS * &CORRHW ;
B2 = SQRT(&STDWEIGHT_OBS**2 - B1**2) ;
VAR_OBS_HT = &STDHEIGHT_OBS*&STDHEIGHT_OBS ;
if &ICCwEIGHT = 1.0 then varwt_e = 0;
else if &ICCwEIGHT ne 1.0 then VARWT_E =
&STDWEIGHT_OBS*&STDWEIGHT_OBS*(1-&ICCWEIGHT);
VARWT_TRUE = &STDWEIGHT_OBS*&STDWEIGHT_OBS - VARWT_E;
VAR_OBS_WT = &STDWEIGHT_OBS*&STDWEIGHT_OBS ;
178
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
RUN;
Y1 = A1 * Z1 ;
Y2 = B1 * Z1 + B2 * Z2 ;
Y3 = A2 * Z1 ;
Y4 = B3 * Z1 + B4 * Z2 ;
HEIGHT_OBS = &MEANHEIGHT + Y1 ;
WEIGHT_OBS = &MEANWEIGHT + Y2 ;
BMI_OBS = WEIGHT_OBS/(HEIGHT_OBS*HEIGHT_OBS);
BMI_OBS_SQ = BMI_OBS*BMI_OBS;
HEIGHT_TRUE = meanheight_true + Y3 ;
WEIGHT_TRUE = meanweight_true + Y4 ;
BMI_TRUE = WEIGHT_TRUE/(HEIGHT_TRUE*HEIGHT_TRUE);
BMI_TRUE_SQ = BMI_TRUE*BMI_TRUE;
output;
end ;
end ;RUN;
179
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
DIABETES = 0 ;
IF U LE PROB THEN DIABETES = 1 ;
NEWDIABETES = 1 - DIABETES;
diabetes_obs = 0;
IF U LE PROB_obs THEN DIABETES_obs = 1;
NEWDIABETES_OBS = 1 - DIABETES_OBS;
RUN;
180
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
Sensitivity = CValue1 + 0;
run;
/* C stat generation*/
181
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
ranks deciles;
by rep;
run;
data hlgof;
set hldiab;
N=_freq_;
chi = (N*(diabsum - trueprobsum)**2)/(trueprobsum*(N-trueprobsum));
run;
182
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
data time;
set time;
format stop time.;
stop = time();
TIMETORUN = stop - start;
format timetorun mmss.
output;
run;
proc print data=time;
title 'TIME TAKEN TO RUN THE PROGRAM';
RUN;
183
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
The distribution of BMI in our survey data (for females) can be represented below:
10
6
Percent
0
6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43
BMI_OBS
In the case where ICC = 1.0 (no error in self-reported weight) the true BMI distribution will look
identical:
184
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
ICC = 1.0
10
6
Percent
0
6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43
BMI_TRUE
However if that observed BMI has 20% of its variability due to random error (eg. Non-
differential reporting error) then the true distribution of BMI will be:
ICC = 0.8
10
6
Percent
0
6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43
BMI_TRUE
185
Doctor of Philosophy Epidemiology (PhD) Dissertation
Laura C.A. Rosella
And in the case of even more error it would look like as so:
ICC = 0.6
12
10
8
Percent
0
6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43
BMI_TRUE
Therefore, an incrase in reporting error results in a wider variance and distribution of BMI
186
Doctor of Philosophy Epidemiology (PhD) Dissertation Laura C.A.
Rosella
187
Doctor of Philosophy Epidemiology (PhD) Dissertation Laura C.A. Rosella
Predicted Probability
Height Weight Mean Mean P- % H-L <
ICC ICC Mean Median Min, Max rHW H-L value 20 rpo O-t Sensitivity Specificity
1.0 1.0 9.31% 8.17% (0.16%, 38.11%) 0.475 9.95 0.3689 97.0% 1.0000 0% 100.00% 100.00%
1.0 0.9 9.18% 8.17% (0.22%, 36.68%) 0.501 11.50 0.2919 92.6% 0.9999 0.13% 99.14% 99.76%
1.0 0.8 9.05% 8.17% (0.32%, 34.92%) 0.531 16.33 0.1397 71.6% 0.9997 0.27% 98.12% 99.52%
1.0 0.7 8.91% 8.17% (0.45%, 32.75%) 0.568 25.21 0.0323 30.8% 0.9991 0.40% 96.94% 99.26%
1.0 0.6 8.76% 8.17% (0.67%, 30.12%) 0.613 40.44 0.0023 3.4% 0.9982 0.55% 95.49% 98.97%
1.0 0.5 8.62% 8.17% (0.10%, 26.85%) 0.672 65.12 0.0000 0.0% 0.9965 0.70% 93.70% 98.65%
0.9 1.0 9.26% 8.17% (0.16%, 37.33%) 0.501 10.16 0.3595 96.6% 0.9980 0.05% 98.78% 99.82%
0.9 0.9 9.13% 8.17% (0.23%, 35.77%) 0.528 12.56 0.2519 90.6% 0.9977 0.18% 98.49% 99.64%
0.9 0.8 8.99% 8.17% (0.32%, 33.87%) 0.560 18.52 0.0990 62.8% 0.9971 0.32% 97.62% 99.41%
0.9 0.7 8.85% 8.17% (0.47%, 31.56%) 0.598 28.95 0.0155 18.2% 0.9961 0.46% 96.46% 99.15%
0.9 0.6 8.71% 8.17% (0.69%, 28.75%) 0.646 46.44 0.0006 1.2% 0.9944 0.60% 94.97% 98.86%
0.9 0.5 8.56% 8.18% (1.11%, 25.28%) 0.708 74.17 0.0000 0.0% 0.9915 0.75% 93.11% 98.53%
0.8 1.0 9.21% 8.18% (0.17%, 36.50%) 0.531 10.79 0.3269 95.0% 0.9906 0.10% 97.41% 99.62%
0.8 0.9 9.07% 8.18% (0.23%, 34.80%) 0.560 14.43 0.1875 81.8% 0.9893 0.24% 97.28% 99.46%
0.8 0.8 8.94% 8.18% (0.33%, 32.76%) 0.594 21.79 0.0544 47.0% 0.9873 0.37% 96.57% 99.25%
0.8 0.7 8.79% 8.18% (0.48%, 30.30%) 0.635 34.26 0.0053 9.0% 0.9842 0.52% 95.49% 99.00%
0.8 0.6 8.65% 8.18% (0.72%, 27.32%) 0.685 54.58 0.0001 0.2% 0.9792 0.66% 94.04% 98.71%
0.8 0.5 8.50% 8.18% (1.11%, 23.65%) 0.751 86.65 0.0000 0.0% 0.9703 0.81% 92.10% 98.38%
0.7 1.0 9.15% 8.19% (0.17%, 35.59%) 0.568 12.28 0.2636 90.2% 0.9751 0.16% 95.75% 99.39%
0.7 0.9 9.02% 8.19% (0.24%, 33.76%) 0.598 17.56 0.1137 66.6% 0.9713 0.29% 95.60% 99.24%
0.7 0.8 8.89% 8.19% (0.34%, 31.58%) 0.623 26.98 0.0204 23.6% 0.9660 0.29% 95.03% 99.04%
0.7 0.7 8.74% 8.19% (0.50 %, 28.97% 0.678 42.31 0.0011 2.2% 0.9579 0.57% 94.04% 98.80%
0.7 0.6 8.59% 8.19% (0.75%, 25.82%) 0.733 66.78 0.0000 0.0% 0.9446 0.59% 92.56% 98.52%
0.7 0.5 8.44% 8.19% (1.12%, 21.95%) 0.803 106.77 0.0000 0.0% 0.9187 0.74% 90.48% 98.17%
0.6 1.0 9.01% 8.19% (0.17%, 34.74%) 0.613 15.42 0.1615 78.0% 0.9462 0.30% 93.65% 99.13%
0.6 0.9 8.97% 8.20% (0.24%, 32.68%) 0.646 23.06 0.0423 41.0% 0.9377 0.34% 93.50% 98.98%
188
Doctor of Philosophy Epidemiology (PhD) Dissertation Laura C.A. Rosella
0.6 0.8 8.83% 8.20% (0.35%, 30.36%) 0.685 35.55 0.0037 7.8% 0.9254 0.48% 92.85% 98.78%
0.6 0.7 8.68% 8.20% (0.51%, 27.60%) 0.733 55.64 0.0001 0.2% 0.9066 0.63% 91.93% 98.54%
0.6 0.6 8.53% 8.20% (0.77%, 24.26%) 0.791 87.75 0.0000 0.0% 0.8740 0.78% 90.36% 98.25%
0.6 0.5 8.38% 8.21% (1.24%, 20.15%) 0.867 142.23 0.0000 0.0% 0.8035 0.93% 87.90% 97.88%
0.5 1 9.05% 8.20% (0.17%, 33.63%) 0.672 22.22 0.0518 46.2% 0.8942 0.26% 91.04% 98.81%
0.5 0.9 8.91% 8.20% (0.25%, 31.55%) 0.708 33.670 0.0052 10.4% 0.8761 0.40% 90.77% 98.65%
0.5 0.8 8.77% 8.21% (0.35%, 29.09%) 0.751 51.852 0.0001 0.2% 0.8496 0.54% 90.13% 98.45%
0.5 0.7 8.63% 8.21% (0.52%, 26.14%) 0.803 80.593 0.0000 0.0% 0.8070 0.68% 88.92% 98.20%
0.5 0.6 8.48% 8.21% (0.80%, 22.57%) 0.867 128.457 0.0000 0.0% 0.7273 0.83% 86.96% 97.88%
0.5 0.5 8.33% 8.21% (0.13%, 18.07%) 0.950 225.916 0.0000 0.0% 0.5130 0.99% 83.39% 97.42%
189
Doctor of Philosophy Epidemiology (PhD) Dissertation Laura C.A. Rosella
Predicted Probability
Height Weight Mean Mean P- % H-L <
ICC ICC Mean Median Min, Max rHW H-L value 20 rpo O-t Sensitivity Specificity
1.0 1.0 7.20% 5.82% (0.04%, 33.49%) 0.311 10.45 0.3356 3.2% 96.8% 1.0000 0 100%
1.0 0.9 7.07% 5.82% (0.06%, 33.02%) 0.328 11.69 0.2861 7.2% 92.8% 0.9991 0.13% 99.30%
1.0 0.8 6.95% 5.82% (0.83%, 32.38%) 0.348 15.62 0.1561 25.20% 74.8% 0.9996 0.25% 98.46%
1.0 0.7 6.82% 5.82% (0.12%, 31.22%) 0.372 23.00 0.0456 56.6% 43.4% 0.9989 0.38% 97.50%
1.0 0.6 6.68% 5.82% (0.19%, 27.77%) 0.402 34.71 0.0043 92.60% 7.4% 0.9977 0.52% 96.33%
1.0 0.5 6.54% 5.82% (0.29%, 27.83%) 0.440 52.83 0.0001 99.8% 0.2% 0.9955 0.66% 94.89%
0.9 1.0 7.16% 5.82% (0.04%, 33.29%) 0.328 10.50 0.3361 3.8% 96.2% 0.9990 0.04% 99.05%
0.9 0.9 7.04% 5.82% (0.06%, 32.70%) 0.346 12.32 0.2617 9.80% 90.2% 0.9991 0.16% 98.96%
0.9 0.8 6.91% 5.82% (0.08%, 31.84%) 0.367 16.94 0.1273 30.40% 69.6% 0.9991 0.29% 98.29%
0.9 0.7 6.77% 5.82% (0.13%, 30.64%) 0.382 25.05 0.0291 67.00% 33.0% 0.9989 0.43% 97.36%
0.9 0.6 6.63% 5.82% (0.19%, 29.04%) 0.423 37.72 0.0021 97.0% 3.0% 0.9983 0.57% 96.22%
0.9 0.5 6.49% 5.82% (0.30%, 26.94%) 0.464 57.42 0.0000 100.0% 0.0% 0.9972 0.71% 94.75%
0.8 1.0 7.12% 5.82% (0.04%, 33.02%) 0.348 10.83 0.3198 4.8% 95.2% 0.9953 0.08% 97.75%
0.8 0.9 7.00% 5.83% (0.06%, 32.32%) 0.367 13.37 0.2219 13.0% 87.0% 0.9955 0.21% 98.09%
0.8 0.8 6.87% 5.83% (0.08%, 31.33%) 0.389 18.62 0.0941 35.4% 64.6% 0.9955 0.33% 97.68%
0.8 0.7 6.73% 5.83% (0.13%, 29.89%) 0.416 27.72 0.0172 75.4% 24.6% 0.9955 0.47% 96.88%
0.8 0.6 6.59% 5.83% (0.20%, 28.24%) 0.449 41.81 0.0009 98.2% 1.8% 0.9950 0.61% 95.80%
0.8 0.5 6.45% 5.83% (0.31%, 25.99%) 0.492 62.90 0.0000 100.0% 0.0% 0.9942 0.75% 94.38%
0.7 1.0 7.08% 5.83% (0.04%, 32.68%) 0.372 11.48 0.2929 6.80% 93.2% 0.9879 0.12% 96.61%
0.7 0.9 6.95% 5.83% (0.06%, 31.85%) 0.392 19.40 0.1726 19.4% 80.6% 0.9877 0.25% 96.87%
0.7 0.8 6.82% 5.83% (0.09%, 30.73%) 0.416 21.21 0.0563 47.8% 52.2% 0.9873 0.38% 96.65%
0.7 0.7 6.69% 5.83% (0.13%, 29.25%) 0.444 31.47 0.0081 87.2% 12.8% 0.9867 0.51% 96.02%
0.7 0.6 6.55% 5.83% (0.20%, 27.36%) 0.480 47.02 0.0002 99.8% 0.2% 0.9856 0.65% 95.03%
0.7 0.5 6.40% 5.83% (0.32%, 24.95%) 0.526 70.37 0.0000 100.00% 0.0% 0.9839 0.80% 93.65%
0.6 1.0 7.04% 5.83% (0.04%, 32.29%) 0.402 12.80 0.2360 11.00% 89.0% 0.9748 0.16% 95.03%
0.6 0.9 6.91% 5.84% (0.06%, 31.34%) 0.423 17.36 0.1116 29.40% 70.6% 0.9736 0.29% 95.31%
190
Doctor of Philosophy Epidemiology (PhD) Dissertation Laura C.A. Rosella
0.6 0.8 6.78% 5.84% (0.09%, 30.09%) 0.449 24.86 0.0292 68.60% 31.4% 0.9720 0.42% 95.20%
0.6 0.7 6.64% 5.84% (0.13%, 28.48%) 0.480 36.84 0.0026 95.20% 4.8% 0.9698 0.56% 94.72%
0.6 0.6 6.50% 5.84% (0.21%, 26.44%) 0.518 54.59 0.0000 100.00% 0.0% 0.9666 0.70% 93.84%
0.6 0.5 6.36% 5.84% (0.33%, 23.88%) 0.568 80.58 0.0000 100.00% 0.0% 0.9616 0.84% 92.51%
0.5 1 7.00% 5.84% (0.04%, 31.84%) 0.440 15.28 0.1582 20.20% 79.8% 0.9525 0.00% 93.08%
0.5 0.9 6.87% 5.84% (0.06%, 30.77%) 0.464 21.25 0.0547 50.00% 50.0% 0.9494 0.33% 93.36%
0.5 0.8 6.74% 5.84% (0.09%, 29.38%) 0.492 30.81 0.0084 85.2% 14.8% 0.9452 0.46% 93.31%
0.5 0.7 6.60% 5.84% (0.14%, 27.64%) 0.526 45.26 0.0003 98.80% 1.2% 0.9395 0.60% 92.87%
0.5 0.6 6.46% 5.84% (0.21%, 25.46%) 0.568 66.06 0.0000 100.00% 0.0% 0.9313 0.74% 92.07%
0.5 0.5 6.31% 5.84% (0.34%, 22.76%) 0.622 96.18 0.0000 100.00% 0.0% 0.9186 0.89% 90.80%
191
Doctor of Philosophy Epidemiology (PhD) Dissertation Laura C.A.
Rosella
MALES FEMALES
C-statistic C-statistic
Height Weight
ICC ICC Mean StDev Min Max Mean StDev Min Max
1.0 1.0 0.6858 0.0082 0.6628 0.7134 0.7179 0.0091 0.6933 0.7424
1.0 0.9 0.6760 0.0091 0.6512 0.7031 0.7097 0.0094 0.6811 0.7358
1.0 0.8 0.6649 0.0094 0.6397 0.6943 0.7002 0.0096 0.6710 0.7277
1.0 0.7 0.6525 0.0097 0.6243 0.6806 0.6894 0.0010 0.6605 0.7202
1.0 0.6 0.6377 0.0099 0.6090 0.6654 0.6770 0.0100 0.6470 0.7035
1.0 0.5 0.6199 0.0102 0.5928 0.6519 0.6625 0.0103 0.6342 0.6920
0.9 1.0 0.6827 0.0090 0.6597 0.7093 0.7158 0.0092 0.6899 0.7401
0.9 0.9 0.6727 0.0092 0.6462 0.6991 0.7074 0.0009 0.6803 0.7316
0.9 0.8 0.6613 0.0095 0.6345 0.6903 0.6978 0.0096 0.6668 0.7271
0.9 0.7 0.6482 0.0097 0.6198 0.6758 0.6868 0.0099 0.6579 0.7166
0.9 0.6 0.6327 0.0100 0.6058 0.6607 0.6743 0.0100 0.6469 0.7033
0.9 0.5 0.6131 0.0102 0.5862 0.6467 0.6592 0.0104 0.6296 0.6869
0.8 1.0 0.6787 0.0091 0.6558 0.7033 0.7131 0.0093 0.6873 0.7381
0.8 0.9 0.6683 0.0094 0.6439 0.6949 0.7044 0.0095 0.6757 0.7315
0.8 0.8 0.6565 0.0096 0.6291 0.6853 0.6949 0.0097 0.6643 0.7246
0.8 0.7 0.6428 0.0098 0.6164 0.6720 0.6836 0.0098 0.6566 0.7139
0.8 0.6 0.6265 0.0101 0.5978 0.6575 0.6706 0.0101 0.6424 0.6989
0.8 0.5 0.6067 0.0101 0.5791 0.6407 0.6553 0.0105 0.6250 0.6841
0.7 1.0 0.6735 0.0092 0.6507 0.7009 0.7097 0.0093 0.6828 0.7353
0.7 0.9 0.6625 0.0095 0.6364 0.6885 0.7009 0.0095 0.6725 0.7291
0.7 0.8 0.6499 0.0096 0.6251 0.6778 0.6909 0.0097 0.6586 0.7220
0.7 0.7 0.6354 0.0098 0.6102 0.6666 0.6795 0.0099 0.6493 0.7090
0.7 0.6 0.6181 0.0101 0.5888 0.6499 0.6661 0.0101 0.6394 0.6980
0.7 0.5 0.5960 0.0102 0.5700 0.6266 0.6502 0.0106 0.6215 0.6795
0.6 1.0 0.6659 0.0094 0.6407 0.6930 0.7053 0.0094 0.6799 0.7339
0.6 0.9 0.6542 0.0095 0.6293 0.6806 0.6960 0.0095 0.6667 0.7242
0.6 0.8 0.6409 0.0098 0.6149 0.6683 0.6858 0.0010 0.6556 0.7157
0.6 0.7 0.6250 0.0099 0.5959 0.6552 0.6741 0.0101 0.6451 0.7017
0.6 0.6 0.6054 0.0102 0.5772 0.6347 0.6601 0.0102 0.6336 0.6913
0.6 0.5 0.5796 0.0104 0.5520 0.6145 0.6435 0.0106 0.6104 0.6719
0.5 1 0.6550 0.0095 0.6279 0.6810 0.6993 0.0095 0.6720 0.7286
0.5 0.9 0.6423 0.0099 0.6148 0.6708 0.6898 0.0097 0.6625 0.7202
0.5 0.8 0.6272 0.0098 0.6014 0.6535 0.6790 0.0099 0.6516 0.7068
0.5 0.7 0.6089 0.0103 0.5824 0.6387 0.6664 0.0102 0.6392 0.6947
0.5 0.6 0.5852 0.0103 0.5591 0.6188 0.6518 0.0103 0.6238 0.6804
0.5 0.5 0.5492 0.0108 0.5193 0.5841 0.6344 0.0107 0.5987 0.6616
192
Doctor of Philosophy Epidemiology (PhD) Dissertation Laura C.A. Rosella
Height ICC Weight ICC Bias Height Bias Weight Mean o-t Median Min, Max
1.0 1.0 0 0 9.31% 0.00% 8.17% (0.16%, 38.11%)
1.0 1.0 0 Underestimate by 0.5 kg 9.54% -0.23% 8.40% (0.17%, 38.34%)
1.0 1.0 0 Underestimate by 1.0 kg 9.77% -0.46% 8.64% (0.18%, 38.57%)
1.0 1.0 0 Underestimate by 1.5 kg 10.00% -0.69% 8.88% (0.19%, 38.78%)
1.0 1.0 0 Underestimate by 2.0 kg 10.25% -0.94% 9.12% (0.20%, 38.99%)
1.0 1.0 0 Underestimate by 2.5 kg 10.49% -1.18% 9.37% (0.21%, 39.19%)
1.0 1.0 0 Underestimate by 3.0 kg 10.74% -1.43% 9.63% (0.23%, 39.39%)
1.0 1.0 Overestimate by 0.5 cm 0 9.53% -0.22% 8.38% (0.17%, 38.43%)
1.0 1.0 Overestimate by 1.0 cm 0 9.76% -0.45% 8.60% (0.17%, 38.74%)
1.0 1.0 Overestimate by 1.5 cm 0 10.00% -0.69% 8.82% (0.17%, 39.04%)
1.0 1.0 Overestimate by 2.0 cm 0 10.23% -0.92% 9.06% (0.18%, 39.33%)
1.0 1.0 Overestimate by 2.5 cm 0 10.47% -1.16% 9.30% (0.18%, 39.60%)
1.0 1.0 Overestimate by 3.0 cm 0 10.72% -1.41% 9.54% (0.18%, 39.87%)
1.0 1.0 Overestimate by 2.5cm Underestimate by 1.7 kg 10.22% -0.91% 9.09% (0.20%, 39.01%)
1.0 1.0 Overestimate by 2.5 cm Underestimate by 3.2 kg 12.14% -2.83% 11.04% (0.26%, 40.70%)
1.0 1.0 Overestimate by 5.0 cm Underestimate by 1.7 kg 12.70% -3.39% 11.58% (0.25%, 41.28%)
1.0 1.0 Overestimate by 5.0 cm Underestimate by 3.2 kg 13.56% -4.25% 12.50% (0.29%, 41.62%)
0.9 0.9 Overestimate by 2.5cm Underestimate by 1.7 kg 10.04% -0.73% 9.10% (0.28%, 36.85%)
0.9 0.9 Overestimate by 2.5 cm Underestimate by 3.2 kg 11.97% -2.66% 11.04% (0.37%, 39.02%)
0.9 0.9 Overestimate by 5.0 cm Underestimate by 1.7 kg 12.52% -3.21% 11.59% (0.35%, 39.83%)
0.9 0.9 Overestimate by 5.0 cm Underestimate by 3.2 kg 13.39% -4.08% 0.24% (0.41%, 40.36%)
0.8 0.8 Overestimate by 2.5cm Underestimate by 1.7 kg 9.85% -0.54% 9.11% (0.40%, 34.00%)
0.8 0.8 Overestimate by 2.5 cm Underestimate by 3.2 kg 11.77% -2.46% 11.05% (0.53%, 36.56%)
0.8 0.8 Overestimate by 5.0 cm Underestimate by 1.7 kg 12.33% -3.02% 11.60% (0.51%, 37.54%)
0.8 0.8 Overestimate by 5.0 cm Underestimate by 3.2 kg 13.21% -3.90% 12.52% (0.60%, 38.27%)
193
Doctor of Philosophy Epidemiology (PhD) Dissertation Laura C.A. Rosella
Height ICC Weight ICC Bias Height Bias Weight Mean o-t Median Min, Max
1.0 1.0 0 0 7.20% 0.00% 5.82% (0.04%, 33.49%)
1.0 1.0 0 Underestimate by 0.5 kg 7.42% -0.22% 6.03% (0.05%, 33.57%)
1.0 1.0 0 Underestimate by 1.0 kg 7.64% -0.44% 6.26% (0.05%, 33.64%)
1.0 1.0 0 Underestimate by 1.5 kg 7.87% -0.67% 6.49% (0.05%, 33.70%)
1.0 1.0 0 Underestimate by 2.0 kg 8.10% -0.90% 6.72% (0.05%, 33.76%)
1.0 1.0 0 Underestimate by 2.5 kg 8.33% -1.13% 6.96% (0.06%, 33.80%)
1.0 1.0 Overestimate by 3.0 cm Underestimate by 3.0 kg 8.57% -1.37% 7.20% (0.06%, 33.84%)
1.0 1.0 Overestimate by 2.5cm 0 7.39% -0.19% 5.99% (0.04%, 33.59%)
1.0 1.0 Overestimate by 2.5 cm 0 7.58% -0.38% 6.17% (0.04%, 33.68%)
1.0 1.0 Overestimate by 5.0 cm 0 7.77% -0.57% 6.35% (0.04%, 33.75%)
1.0 1.0 Overestimate by 5.0 cm 0 7.97% -0.77% 6.54% (0.04%, 33.81%)
1.0 1.0 Overestimate by 2.5cm 0 8.18% -0.98% 6.74% (0.04%, 33.86%)
1.0 1.0 Overestimate by 2.5 cm 0 8.39% -1.19% 6.94% (0.04%, 33.90%)
1.0 1.0 Overestimate by 5.0 cm Underestimate by 1.7 kg 8.06% -0.86% 6.67% (0.05%, 33.76%)
1.0 1.0 Overestimate by 5.0 cm Underestimate by 3.2 kg 9.79% -2.59% 8.42% (0.07%, 34.00%)
1.0 1.0 Overestimate by 2.5cm Underestimate by 1.7 kg 10.18% -2.98% 8.77% (0.07%, 34.04%)
1.0 1.0 Overestimate by 2.5 cm Underestimate by 3.2 kg 11.02% -3.82% 9.69% (0.08%, 34.04%)
0.9 0.9 Overestimate by 5.0 cm Underestimate by 1.7 kg 7.89% -0.69% 6.68% (0.07%, 33.22%)
0.9 0.9 Overestimate by 5.0 cm Underestimate by 3.2 kg 9.63% -2.43% 8.43% (0.10%, 33.86%)
0.9 0.9 Overestimate by 3.0 cm Underestimate by 1.7 kg 10.02% -2.82% 8.78% (0.09%, 33.96%)
0.9 0.9 Overestimate by 2.5cm Underestimate by 3.2 kg 10.87% -3.67% 9.70% (0.11%, 34.00%)
0.8 0.8 Overestimate by 2.5cm Underestimate by 1.7 kg 7.77% -0.57% 6.68% (0.11%, 32.12%)
0.8 0.8 Overestimate by 2.5 cm Underestimate by 3.2 kg 9.47% -2.27% 8.43% (0.15%, 33.37%)
0.8 0.8 Overestimate by 5.0 cm Underestimate by 1.7 kg 9.85% -2.65% 8.79% (0.14%, 33.66%)
0.8 0.8 Overestimate by 5.0 cm Underestimate by 3.2 kg 10.71% -3.51% 9.70% (0.17%, 33.83%)
194
Doctor of Philosophy Epidemiology (PhD) Dissertation Laura C.A. Rosella
195
Doctor of Philosophy Epidemiology (PhD) Dissertation Laura C.A. Rosella
196
Doctor of Philosophy Epidemiology (PhD) Dissertation Laura C.A. Rosella
C-statistic C-statistic
Height Weight
ICC ICC Bias Height Bias Weight Mean StDev Min Max Mean StDev Min Max
1.0 1.0 0 0 0.6858 0.0082 0.6628 0.7134 0.7179 0.0091 0.6933 0.7424
1.0 1.0 0 Underestimate by 0.5 kg 0.6847 0.0088 0.6623 0.7107 0.7126 0.0090 0.6931 0.7468
1.0 1.0 0 Underestimate by 1.0 kg 0.6834 0.0087 0.6607 0.7098 0.7145 0.0089 0.6918 0.7383
1.0 1.0 0 Underestimate by 1.5 kg 0.6821 0.0086 0.6592 0.7081 0.7127 0.0089 0.6898 0.7369
1.0 1.0 0 Underestimate by 2.0 kg 0.6810 0.0085 0.6556 0.7083 0.7110 0.0088 0.6882 0.7338
1.0 1.0 0 Underestimate by 2.5 kg 0.6798 0.0084 0.6560 0.7072 0.7093 0.0087 0.6880 0.7336
1.0 1.0 Overestimate by 2.5cm Underestimate by 3.0 kg 0.6786 0.0084 0.6531 0.7066 0.7078 0.0085 0.6846 0.7316
1.0 1.0 Overestimate by 2.5 cm 0 0.6856 0.0088 0.6639 0.7118 0.7174 0.0090 0.6940 0.7428
1.0 1.0 Overestimate by 5.0 cm 0 0.6852 0.0087 0.6626 0.7116 0.7169 0.0089 0.6944 0.7422
1.0 1.0 Overestimate by 5.0 cm 0 0.6848 0.0086 0.6616 0.7120 0.7163 0.0088 0.6928 0.7397
1.0 1.0 Overestimate by 2.5cm 0 0.6844 0.0086 0.6590 0.7126 0.7157 0.0089 0.6926 0.7408
1.0 1.0 Overestimate by 2.5 cm 0 0.6841 0.0084 0.6599 0.7111 0.7152 0.0087 0.6929 0.7383
1.0 1.0 Overestimate by 5.0 cm 0 0.6838 0.0083 0.6601 0.7095 0.7147 0.0086 0.6918 0.7391
1.0 1.0 Overestimate by 5.0 cm Underestimate by 1.7 kg 0.6815 0.0085 0.6561 0.7086 0.7118 0.0087 0.6888 0.7353
1.0 1.0 Overestimate by 3.0 cm Underestimate by 3.2 kg 0.6762 0.0079 0.6519 0.7025 0.7037 0.0080 0.6818 0.7263
1.0 1.0 Overestimate by 2.5cm Underestimate by 1.7 kg 0.6776 0.0078 0.6551 0.7039 0.7055 0.0078 0.6845 0.7264
1.0 1.0 Overestimate by 2.5cm Underestimate by 3.2 kg 0.6737 0.0076 0.6520 0.6988 0.7000 0.0075 0.6774 0.7204
0.9 0.9 Overestimate by 2.5 cm Underestimate by 1.7 kg 0.6687 0.0087 0.6428 0.6925 0.7014 0.0089 0.6778 0.7259
0.9 0.9 Overestimate by 5.0 cm Underestimate by 3.2 kg 0.0066 0.0082 0.6379 0.6926 0.6938 0.0081 0.6715 0.7178
0.9 0.9 Overestimate by 2.5cm Underestimate by 1.7 kg 0.0067 0.0079 0.6413 0.6906 0.6958 0.0079 0.6737 0.7173
0.9 0.9 Overestimate by 2.5 cm Underestimate by 3.2 kg 0.0066 0.0000 0.6368 0.6888 0.6903 0.0077 0.6656 0.7104
0.8 0.8 Overestimate by 5.0 cm Underestimate by 1.7 kg 0.6528 0.0090 0.6293 0.6791 0.6894 0.0092 0.6621 0.7163
0.8 0.8 Overestimate by 5.0 cm Underestimate by 3.2 kg 0.6486 0.0084 0.6218 0.6742 0.6821 0.0082 0.6590 0.7036
0.8 0.8 Overestimate by 2.5cm Underestimate by 1.7 kg 0.6500 0.0083 0.6244 0.6796 0.6840 0.0080 0.6622 0.7052
0.8 0.8 Overestimate by 2.5 cm Underestimate by 3.2 kg 0.6467 0.0080 0.6214 0.6771 0.6789 0.0077 0.6547 0.7005
197