DIRECT COST OF OSTEOARTHRITIS IN CANADA: AN APPLICATION OF
MICROSIMULATION MODELING WITH UNCERTAINTY ANALYSIS
by
Behnam Sharif
M.Sc., The University of Manitoba, 2007
A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF
THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
in
THE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES
(Population and Public Health)
THE UNIVERSITY OF BRITISH COLUMBIA
April 2014
©Behnam Sharif, 2014
Abstract
Introduction: While OA is a debilitating disease with an immense economic burden on the Canadian society, there is a lack of understanding about OA’s direct costs and its future trend in Canada.
Objectives: The overall goal of this thesis is to illustrate the application of population- based disease microsimulation (PDMS) modeling in estimating the economic burden of a disease by performing the direct cost analyses for osteoarthritis (OA) using Population
Health Microsimulation Model for OA (POHEM-OA). Specific objectives were: 1) To estimate the average direct costs of OA from 2003 to 2010 in Canada; 2) To estimate the future direct cost of OA from 2010 to 2031 in Canada; 3) to estimate the uncertainty around the prevalence and total cost of OA in future years.
Methods: I used administrative health data from the province of British Columbia (BC),
Canada, a survey of a random sample of BC residents diagnosed with OA (Ministry of
Health of BC data), Canadian Institute of Health Information (CIHI) cost data and literature estimates to perform a bottom-up cost of illness (COI) study for OA. I then implemented the results of the COI study into POHEM-OA and constructed cost profiles for each individual. Finally, I developed a framework and adapted an ANOVA-based approach for performing uncertainty analysis (UA) for OA outcomes.
Results: I showed that the average cost increased from $735 to $811 between 2003 and 2010 (in 2010 $CAD). From 2010 to 2031, while the prevalence of OA increases from 13.8% to 18.6%, the total direct cost of OA is projected to increase from $2.9 billion (95% uncertainty interval (UI): $2.4-$3.1 billion), to $7.6 billion ($6.2-$9.1 billion), an almost 2.6-fold increase (in 2010 $CAD). From the highest to the lowest, the cost ii
components that will constitute the total direct cost of OA in 2031 are hospitalization cost, outpatient services, drugs, and out-of-pocket cost categories.
Conclusions: By further developing a PDMS model of OA, I were able to project trends in the cost of OA and identify the key cost drivers, while predicting significant shifts in distribution of cost in the future.
iii
Preface
In this thesis, I used the Population Health Microsimulation model of
Osteoarthritis (POHEM-OA), a microsimulation model previously developed by
Kopec et al. (1) and Statistics Canada. I further developed this simulation model by incorporating cost modules as described in Chapter 3 and Chapter 4. After implementing the coding into POHEM-OA, it was then corrected and validated by Mr.
Bill Flanagan, chief of microsimulation modeling at Statistics Canada, Health Analysis
Division. A section of this thesis has been published as a multi-authored paper in a refereed journals. Details of the co-authors’ contributions are provided below.
The section in Chapter 3 on the Cost of illness (COI) study for per patient-year cost was performed as part of a prior study, the Canadian Osteoarthritis Simulation
Team (COAST) by Dr. Jacek Kopec, Dr. Aslam Anis, Dr. Nick Bansback, and Mr.
Mushfiqur Rahman. Drs. Kopec and Bansback designed this section of the study and
Mr. Rahman performed the statistical analysis, using provincial population-based administrative data of British Columbia (BC). I have performed additional analysis on the procedure costs for inpatient cost components, surgery costs, rehabilitation costs, alternative care costs and out-of-pocket-costs. I have designed these analyses to estimate the per patient-year cost of OA as a function of individual characteristics and implemented the results of the aforementioned COI study into POHEM-OA, and performed the writing and revisions of Chapter 3. Dr. Kopec, Dr. Anis, and Dr. Bansback helped me in design, analysis and revisions of this Chapter.
A version of Chapter 5 has been published in the journal Epidemiology Research
International (Behnam Sharif, Jacek A Kopec, Hubert Wong, Philippe Finès, Douglas G iv
Manuel, David L. Buckeridge, Uncertainty analysis in Population-based disease microsimulation models, Epidemiology Research International, July 2012). With the assistance of Drs. Kopec, Wang, and Fines, I conceptualized and designed the study. I conducted the analysis with support from the Canadian Arthritis Network grant. Dr.
Kopec contributed to the interpretation and discussion of the study. Dr. Wang and Dr.
Fines contributed to the statistical analysis. I prepared the manuscript and submitted it for publication. All co-authors contributed feedback on my drafts of the manuscript and I completed all revisions.
Ethics approval for the studies described in Chapter 3 and 4 of this thesis was obtained from the University of British Columbia Behavioral Research Ethics Board
(certificate number H06-04000).
v
Table of Contents
Abstract ...... ii
Preface ...... iv
Table of Contents …..……………………………………..………………………………….vi
List of Tables ...... ix
List of Figures ...... xi
Abbreviations ...... xiii
Acknowledgements ...... xvi
Dedication ...... xvii
Chapter 1: Introduction ...... 1
1.1 Thesis Organization ...... 1
1.2 Overview ...... 3
1.3 Rationale ...... 11
1.4 Study Objectives ...... 13
Chapter 2: Background ...... 14
2.1 Osteoarthritis ...... 14
2.2 Cost-of-Illness Studies ...... 24
2.3 Economic Burden of OA ...... 33
2.4 Discussion ...... 54
Chapter 3: Application of a Simulation-Based Cost-of-Illness Study to Estimate
Average Direct Cost of OA From 2003 to 2010 ...... 56
3.1 Introduction ...... 56
3.2 Methods ...... 58
vi
3.3 Analysis of Input Parameters for Cost Components ...... 65
3.4 The Population Health Microsimulation Model (POHEM) ...... 83
3.5 Description of the Cost Algorithm ...... 89
3.6 Scenario Analysis for Drivers of the OA Direct Cost Burden ...... 92
3.7 Results: Average Costs From 2003 to 2010 ...... 95
3.8 Discussion ...... 106
Chapter 4: Projecting the Total Direct Cost Burden of Osteoarthritis Patients in
Canada From 2010 to 2031 Using the POHEM-OA Model ...... 111
4.1 Introduction ...... 111
4.2 Methods ...... 112
4.3 Results ...... 127
4.4 Discussion ...... 139
Chapter 5: Uncertainty Analysis in Population-Based Disease Microsimulation
Models……………………………………………………………………………………………….143
5.1 Introduction ...... 143
5.2 Uncertainty Analysis in Simulation Models ...... 144
5.3 Overview of Uncertainty Analysis in PDMS Models ...... 148
5.4 Steps for Performing Uncertainty Analysis ...... 153
5.5 Results of Uncertainty Analysis for Prevalence of OA ...... 165
5.6 Results of Uncertainty Analysis for Total Direct Cost of OA ...... 176
5.7 Discussion ...... 185
Chapter 6: Discussion ...... 189
6.1 Key Findings ...... 190
6.2 Integration and Implications of the Research ...... 195
6.3 Strengths and Limitations of the Research ...... 205 vii
6.4 Recommendations and Future Research ...... 215
6.5 Summary of Thesis ...... 221
References ...... 223!
Appendix A Details of Methods for Estimating the Direct Cost of OA………..….244
Appendix B Details of Methods for Projecting the Direct Cost of OA …………...271
Appendix C Details of Equations and Statistical Methods in Chapter 5 ...…..….281
viii
List of Tables
Table 2-1. Summary of risk factors for OA incidence ...... 20!
Table 2-2. Cost-of-illness (COI) study perspectives ...... 29!
Table 2-3. National cost of OA studies performed before 2006 ...... 37!
Table 2-4. National cost of OA studies from 2006 to 2012 ...... 39!
Table 2-5. Summary of the literature review for direct cost of OA ...... 43!
Table 3-1. Direct cost categories of OA implemented in the POHEM-OA direct cost module ...... 60!
Table 3-2. Probability of discharge to rehab destinations after hip/knee TJR surgery ... 73!
Table 3-3. Out-of-pocket costs for rehabilitation cost category ...... 75!
Table 3-4. Probability of incurring a non-zero cost for the formal caregiver category .... 77!
Table 3-5. Probability of visit to physiotherapist and chiropractic within the past year .. 80!
Table 3-6. Probability of visit to alternative care professionals within the last year ...... 81!
Table 3-7. Characteristics of OA patients and of general population ...... 96!
Table 4-1. Annual inflation rates for each cost component ...... 121!
Table 4-2. Characteristics of OA patients and of general population ...... 127!
Table 5-1. Components of a simplified version of the POHEM-OA model used in UA example ...... 166!
Table 5-2. Proposed steps for UA in PDMS and corresponding steps for the POHEM-
OA model ...... 168!
Table 5-3. Point estimates and 95% CI for hazard ratios of OA diagnosis implemented in the POHEM-OA model ...... 171!
ix
Table 5-4. Correlation matrix for hazard ratios of osteoarthritis in the POHEM-OA model
...... 172!
Table 5-5. List of parameters, data sources and their distributions used for UA of direct cost in the POHEM-OA model ...... 178!
Table 5-6. Results for the initial step for UA of total cost in the POHEM-OA model ... 181!
x
List of Figures
Figure 2-1. Cost of illness (COI) study types, methodologies, and approaches ...... 30!
Figure 2-2. Cost components in COI studies of OA and their respective perspectives in
Canada ...... 41!
Figure 3-1. Rehabilitation and home-care cost after surgery ...... 72!
Figure 3-2 . POHEM-OA: an osteoarthritis model …………………………………………84
Figure 3-3. POHEM-OA diagnosis module ...... 87!
Figure 3-4. POHEM-OA progression module ...... 88!
Figure 3-5. Average cost of OA patients in 2010 $CAD and current dollars ...... 97!
Figure 3-6. Average cost of OA for females and males from 2003 to 2010 ...... 97!
Figure 3-7. Average cost for OA patient by age categories from 2003 to 2010 ...... 98!
Figure 3-8. Average cost of an OA patient by cost categories in 2003 ...... 100!
Figure 3-9. Average cost of an OA patient by cost categories in 2010 ...... 100!
Figure 3-10. Number of OA patients by OA state in 2003, 2006 and 2010 ...... 102!
Figure 3-11. Average cost of OA by OA state in 2003, 2006 and 2010...... 104!
Figure 3-12. Sensitivity analysis: effect of cost drivers on average cost change ...... 106!
Figure 4-1. Prevalence of OA for males and females from 2010 to 2031 ...... 128!
Figure 4-2. Total direct cost of OA for different inflation scenarios ...... 131!
Figure 4-3. Sex-specific healthcare system and out-of-pocket cost of OA and effect of inflation ...... 132!
Figure 4-4. Total direct cost of OA by age category for males and females ...... 134!
Figure 4-5. Share of cost categories of total direct cost from 2010 to 2031 ...... 137!
xi
Figure 4-6. Total direct cost of OA by cost categories in 2010 and 2031: out-of-pocket and healthcare system cost ...... 138!
Figure 5-1. Tornado diagram for prevalence of osteoarthritis among females in year
2021 ...... 170!
Figure 5-2. Results of uncertainty analysis: sex-specific prevalence of OA in Canada from 2001 to 2021 as predicted by the POHEM-OA model ...... 175!
Figure 5-3. Results of uncertainty analysis: total direct cost of OA in Canada from 2010 to 2031 as predicted by the POHEM-OA model ...... 184!
xii
Abbreviations
American College of Rheumatology (ACR)
Analysis of Variance (ANOVA)
Body Mass Index (BMI)
Canadian Community Health Survey (CCHS)
Canadian cost per weighted case (CPWC)
Canadian Dollar (CAD)
Canadian Institute of Health Information (CIHI)
Canadian Joint Replacement Registry (CJRR)
Cancer Intervention and Surveillance Modeling Network (CISNET)
Cardiovascular Diseases (CVD)
Case Mix Groups (CMG)
Canadian Osteoarthritis Simulation Team (COAST)
Cost-effectiveness Acceptability Curves (CEAC)
Chronic Obstructive Pulmonary Disease (COPD)
Consumer Price Index (CPI)
Cost of Illness (COI)
Cost-benefit Analysis (CBA)
Cost-effective Analysis (CEA)
Disability-adjusted Life-years (DALYs)
Discrete-event Simulation (DES)
Fractional Factorial Design (FFD)
Gastrointestinal (GI) xiii
Gross Domestic Product (GDP)
Hazard Ratio (HR)
Hospital Discharge Abstract Database (DAD)
Human Immunodeficiency Virus (HIV)
International Classification of Disease, 9th Revision (ICD-9)
Kellgren-Lawrence Grading Scale (KL)
Latin Hypercube Sampling (LHS)
Magnetic Resonance Imaging (MRI)
Microsimulation (MS)
Model Generated Language (MODGEN)
Monte Carlo (MC)
National Physician Database (NPDB)
National Population Health Survey (NPHS)
Nonsteroidal Anti-inflammatory Drugs (NSAIDs)
Orthopedic Surgeon (OS)
Osteoarthritis (OA)
Osteoarthritis Policy Model (OAPol)
Over-the-counter (OTC)
Population Data BC Database (PDBS)
Population Health Microsimulation model (POHEM)
Population Health Microsimulation model of Osteoarthritis (POHEM-OA)
Population-attributable Fraction (PAF)
Population-based Disease Microsimulation (PDMS) xiv
Probability Distribution Function (PDF)
Probability Sensitivity Analysis (PSA)
Public Health Agency of Canada (PHAC)
Purchasing Power Parity (PPP)
Quality of Life (QOL)
Quality-adjusted Life-years (QALYs)
Resource Intensity Weight (RIW)
Rheumatoid Arthritis (RA)
Social Time Preference (STP)
Socioeconomic Status (SES)
Total Joint Replacement (TJR)
Uncertainty Intervals (UI)
Western Ontario and McMaster Universities OA Index (WOMAC)
World Health Organization (WHO)
xv
Acknowledgements
I offer my enduring gratitude to the faculty and staff at University of British Columbia, school of population and public health (SPPH) and Arthritis Research Centre, who have inspired me to continue my work in this field. I owe particular thanks to Dr. Jacek Kopec who enlarged my vision of science and whose penetrating questions taught me to question more deeply. I thank Dr. Aslam Anis helped me understand how to overcome obstacles both in academic research and in life. Special thanks are owed to Dr. Hubert
Wong and Dr. Philippe Fines for their great help and advice.
xvi
Dedication
I dedicate this thesis to my wife, Soulmaz, and my Brothers, Behzad and Behrang, who have supported me throughout my PhD studies. I specially dedicate this to my parents,
Khojaste and Abdolreza, who have supported me all my life and taught me to work hard.
xvii
Chapter 1: Introduction
1.1 Thesis Organization
This thesis examines the subject of applications of microsimulation modeling to estimate the direct cost burden of Osteoarthritis (OA) in Canada and the uncertainty associated with the cost estimates. The overall goal of this thesis is to gain a better understanding of the different cost components of the total direct OA cost, its distribution across different sub-populations and the probable changes to this pattern in the next 20 years in Canada. In addition to identifying the major sub-populations who have borne this cost, the ultimate purpose of this program of research is to provide evidence of the amount and structure of the direct cost burden of OA to patients and the health care system, both in the past and in future decades. The results of this thesis can be used to develop interventions targeted at these sub-populations in order to reduce the OA cost burden in the future.
This thesis consists of seven chapters. Chapter 1 is introductory and gives an overview of the economic burden of OA, including a summary of its key cost components, and applications of simulation modeling in descriptive epidemiology of chronic diseases. In addition, in Chapter 1, I present the theoretical framework and rationale of the thesis. Chapter 2 presents a detailed literature review of the epidemiology of OA, together with its risk factors and burden on societies across the world. In the first part of Chapter 2, I provide a summary and literature review related to
Cost of Illness (COI) studies; I describe a summary of definitions, different types, and methods for traditional COI studies, as well as their shortcomings as performed throughout the last 50 years. I then perform a systematic review of COI studies on OA 1
performed between 2006 and 2012. In Chapter 3, I perform a bottom-up cost of illness (COI) study, a micro-costing method, to estimate the per-patient-year cost for osteoarthritis (OA) from 2003 to 2010 including: hospitalization, physicians, drugs, rehabilitation, formal caregivers, alternative care, and the side effects of drugs.
In addition, I use the distribution of the OA population in Canada, as estimated by the
Population Health Microsimulation Model of OA (POHEM-OA), a microsimulation model of osteoarthritis (OA) in Canada, to report and analyze the trend of the average cost of OA during the past decade, i.e., from 2003-2010. Further scenario analyses have been performed to calculate the effect of cost drivers on increases in the average cost of OA. I also investigate the effects of an increase in longevity, the number of hip/knee total joint replacement (TJR) surgeries, and the inflation rates between 2003 and 2010.
In Chapter 4, I use the results of the micro-costing approach in Chapter 3, together with the POHEM-OA results, to predict the total direct cost of OA between 2010 and 2031.
Then I predict the overall cost of OA in fixed 2010 $CAN dollars by including the inflation rate for each cost category (hospitalization, physician, drugs, rehabilitation, and out-of-pocket). I report the overall direct cost of OA by sex and age categories, and cost components over the next 20 years. In Chapter 5, I provide definitions for uncertainty analysis in the context of population-based disease microsimulation (PDMS) models and discuss the necessity of performing UA in these models. Further, by adapting common concepts from published UA guidelines in decision analytic and environmental modeling literature (2,3), I develop a comprehensive, step-by-step approach to UA in PDMS models, including sample size calculation to reduce computational time. As an illustration, I perform UA for prevalence and total direct cost 2
of OA as predicted by POHEM-OA. in Chapter 5. Performing UA in microsimulation
(MS) models, in its traditional way through the use of brute force Monte Carlo (MC) analysis, are often time consuming and even unfeasible (2). In Chapter 5, I adapt the
ANOVA-based approach previously developed by O’Hagan et al. (4) to PDMS models. I generalize their approach to the case of aggregate outcomes in PDMS models. Finally, I apply the proposed approach to estimate the prevalence and total direct cost of OA and their 95% associated uncertainty intervals (UI’s), as projected by the POHEM-OA, over the future years. Chapter 6, the concluding chapter, synthesizes the findings from each study and discusses the strengths, limitations, and potential implications of the collective work.
1.2 Overview
Osteoarthritis (OA) is a long lasting disease, caused by a progressive loss of joint cartilage, that ranks among the most prevalent diseases in the US and Canada: i.e.,
10%-15% of the adult population, based on self-reported data (5). OA is the leading cause of chronic pain and mobility limitation in women and older individuals (6). A patient with OA, even in its early stages, consumes health care resources at almost double the rate of the general population without OA (7). The burden for OA patients in
Canada, in terms of costs and quality of life years (QOL) lost, is significant compared to those for other chronic diseases and will become alarming in the decades to come (8,9).
This might partly be due to the considerable rise in obesity and population aging in western countries, two of the main risk factors for OA presence and severity (10,11).
According to the Arthritis Alliance of Canada report in 2011 (9), OA will be the fourth leading cause of disability-adjusted life-years lost by the year 2020 globally (12). 3
Estimates suggest that the public health consequences of the increasing prevalence of knee OA will have almost doubled by the year 2020, due in part to increases in obesity and age of the population (13). However, the main driver of changes in OA incidence is aging of the population, which in turn is due to changes in birth rates and changes in duration of life (13).
Although several studies have noted the rise in future burden of OA (10,13), only computer simulation models were able to accommodate the complex and interacting risk factors in order to project the burden of OA in the future (8,12).
1.2.1 Population-based Disease Simulation Models
With the advent of fast computing tools and availability of large population-based datasets in recent years, computer simulation models are finding new applications in epidemiology, health economics, and risk analysis of diseases in environmental studies
(14–17). Computer disease simulations are models that stochastically generate a sample path for the progression of a disease in an individual or group of individuals by modeling the risk factor of disease, its treatment resources and disease impact on the population (18). Applications of simulation modeling in descriptive epidemiology of chronic diseases have been rising in recent years: from simulation models of infectious diseases such as Human Immunodeficiency Virus (HIV) infections (19,20), tuberculosis
(21) and hepatitis C (22), to models of non-infectious diseases such as cancer (23–26), asthma (27), diabetes (28), cardiovascular diseases (CVD) (29), and other conditions
(30,31). Models have evolved from simple decision analytic trees to complex individual- level simulation models, or microsimulation models, utilizing Monte Carlo analysis and discrete-event simulation techniques (14). A well-designed and validated model can 4
effectively synthesize data from different sources; reveal important gaps in the knowledge; provide new insights into the dynamics of a chronic disease; estimate the costs and impact on quality of life; and suggest optimal approaches to prevention and treatment (23,32).
Individual-level computer simulation models are useful for forecasting trends in health outcomes, predicting the impact of health programs and policies, and comparing the cost-effectiveness of different treatments or prevention strategies for diseases that have intricate natural history in terms of multi-level and interactive risk factors (14,33).
Such models are particularly effective when the questions are complex and require integration of data from different sources such as those developed for projecting healthcare consequences of chronic diseases and aging (34).
Population-based disease microsimulation (PDMS) models are individual-level simulation models that are integrated with (validated) demographic models to simulate cohorts of a population as their baseline model by incorporating statistical models to reflect past and future birth, immigration and mortality rates (e.g., population of a country) (32,34,35); PDMS models then use population-based surveys to model health determinants associated with a disease and its burden across a population (32,34,35).
Canada has an impressive, but still underutilized, population health database, comprised of large surveys and administrative databases (36). Incorporating these population health databases into a microsimulation framework enables us not only to predict the future burden of a disease, but also to identify the sub-populations who are affected most by the disease and its consequences. As a result, PDMS models would
5
provide us with the unique opportunity to use the results of large population-based surveys for disease-related outcomes in chronic diseases such as OA (37,38).
Use of a PDMS model in chronic diseases is rising in recent years; PDMS models have been used to investigate screening scenario for diseases such as breast cancer (39), Chronic Obstructive Pulmonary Disease (COPD) (40), HIV (41), in addition to projecting the burden of CVD (42) and diabetes (43). PDMS models have been developed in several countries including United States, e.g., disease models for cancer from the Cancer Intervention and Surveillance Modelling Network (CISNET) (39), United
Kingdom, e.g., Unal’s coronary heart disease model (42), Netherlands, e.g.,
Hoogendoorn’s COPD model (40) and Canada, e.g., POHEM models for OA (1), breast cancer, lung cancer and colorectal cancer (44–46).
In contrast to several applications of PDMS models in chronic disease such as
HIV, diabetes, CVD and cancer (39-46), very few PDMS models have been developed for arthritis models and OA (1). The focus of this thesis will be on further developing and implementing direct cost modules inside POHEM-OA, a PDMS model that was originally developed for predicting the burden of OA in future years in Canada by Kopec et al. (1).
1.2.2 Theoretical Framework
OA is a long lasting disease that starts between 40-50 years of age but becomes symptomatic often among patients of ages greater than 60 years (47). Both the incidence and progression of OA are functions of the complex interaction between risk factors such as age, sex, and Body Mass Index (BMI). Therefore, traditional statistical models, such as regression models, would not be appropriate alternatives for modeling the future burden of OA. Additionally, to incorporate heterogeneity within a population, 6
competing risks of mortality, disease incidence, and progression stages, warrants the use of simulation modeling (33). More importantly, the progression of OA for each patient is dependent on the past health history, in terms of risk factors, and this necessitates an individual-level simulation or micro-simulation modeling―rather than an aggregate-based or group-based simulation (21)
In the past 20 years, several studies used aggregate-level simulation models, mostly Markov models, to investigate cost-effectiveness of different interventions for OA care including different pharmacotherapy interventions (48) or joint replacement surgeries (49). Most of the simulation models in the literature, however, are the cost- effectiveness studies of oral treatments for osteoarthritis including, for example, selective and non-selective NSAIDs, often with gastro-protective agents (48). Markov or semi-Markov models (47,48) were used by modeling states of patients that represent different stages of OA, such as early OA, first phase medications (e.g., acetaminophen), second stage medications (NSAID, opiods) or end-stage OA in which the patient requires joint replacement surgery. Most of these simulation models assessed adverse effects of the medications such as gastrointestinal complications with one year time horizon or cardiovascular events that are usually incorporated in a longer (usually lifetime) time horizons to capture the full impact of cardiovascular events (47).
Very few recent studies used individual-level simulation models of OA. These include a discrete event simulation model for performing cost-effectiveness of total joint replacement (TJR) for hip and knee surgery in Australia (50) and state-based, individual-level Markov model for Knee-OA in US (51). The latter model is a validated, state transition, computer simulation model called Osteoarthritis Policy Model (OAPol) 7
that has been used in several recent studies for estimating the current burden of total knee-replacement surgery (52) and cost-effectiveness of future disease modifying drugs for knee-OA (50). In another recent study, the OAPol model is used in combination with population-level data to predict the burden of TKR among elderly patients in the next 10 years (53).
Although more and more simulation models are being developed for decision analytic purposes in chronic diseases such as OA (48-52), very few predictive simulation models use population-level data to generate and project individual-level risk factors (1, 50). For example, in a recent study using the OAPol model (51), a hypothetical population is generated based on the estimated distribution of age, sex and other characteristics of individuals according to population-level data. This is in contrast to another type of PDMS model where the individuals’ characteristics are directly generated using the population-level data (or use of bootstrap methods). This, in fact, would aid in representing the actual individuals within the population with the inherent correlation of their characteristics in terms of, for example, age, sex or disease risk factors such as BMI, physical activity levels, history of joint injury and socioeconomic status (SES) in the case of OA. These types of PDMS models have higher predictive power as they are capable of representing the heterogeneity between different individuals across the population and not just modeling the average patient. For example, POHEM-OA is a PDMS model that uses actual population-level data from
Canadian Community Health Survey (CCHS) in 2001 (54) to generate the initial population and its characteristics.
8
1.2.3 Population Health Microsimulation Model (POHEM)
A validated computer simulation model such as POHEM, developed by Statistics
Canada (44), is an ideal tool to integrate clinical, epidemiological, and population data.
POHEM framework was developed by the Health Analysis and Modeling Group at
Statistics Canada in early 1990’s (44). POHEM is an example of a continuous time model based on Monte Carlo (MC) analysis. In a continuous time model, hazards compete to determine the time of the next event in a person’s life (55). Current POHEM models include simulation models of lung cancer, breast cancer, preventive Tamoxifen, and more recently, colorectal cancer (1, 45, 46). In this thesis, I use POHEM-OA, an OA model within the POHEM framework previously developed by Kopec et al. (1).
POHEM-OA generates the entire adult population of Canada (around 20 million individuals) and their health-related characteristics in 2001, according to the Canadian
Community Health Survey of 2001 (54). POHEM-OA is a continuous longitudinal microsimulation in which every individual at the baseline is aged year-by-year, while his/her characteristics and OA risk factors, such as BMI and socioeconomic variables
(income, education) are also changed according to statistical models implemented within POHEM. Not only are the OA risk factors modeled in time, but also different OA- related events, such as OA diagnosis, visit to an orthopedic surgeon, hip/knee replacement surgery (TJR), and death, are modeled as competing events. As a result, the OA-related life history for each individual from 2001 until their death or the end time horizon of the model is simulated for the entire adult population of Canada (1). In addition to the health-adjusted life expectancy of patients with OA, POHEM-OA was also used to project the incidence and prevalence of OA between 2001 and 2021 (1). 9
POHEM-OA has been used to project the prevalence and incidence of OA into future years; however, no-one has yet created a PDMS model that projects the economic burden of OA in Canada or other countries. In this thesis, I have used
POHEM-OA, combined with population-based data and cost estimates to project the total direct cost burden of OA among Canadians over the next 20 years.
1.2.4 Direct Cost of OA
Arthritis constitutes a significant proportion of the total costs due to musculoskeletal diseases (around 10%). At the same time, the direct and indirect costs for OA make up around 40-50% of all arthritis costs (47). Direct costs reflect the diversion of resources towards the diagnosis, treatment, and management of the illness, and away from other productive uses (56). Indirect costs on the other hand, are productivity losses or the illness-related morbidity and mortality that render human resources unavailable for productive uses (57). In simpler terms, direct costs are the costs incurred for paid services, e.g., wages, professional fees, costs for supplies, while indirect costs are defined as the imputed valuations for unpaid resource services (57).
The comprehensive list of OA-related direct cost categories, paid for by patients or the health care system, includes medical expenditures (inpatient, outpatient, drugs, and the extra costs due to the side effects of drugs) and non-medical ones (formal caregiver, travelling, and community services) (58). Major cost components for the indirect costs of
OA include productivity losses of employed patients because of short or long-term disability due to OA, i.e., absenteeism and presenteeism, informal caregiver costs, and early retirement due to OA (58). In this thesis, I only focus on the direct costs of OA.
10
The direct cost associated with hospitalization and total joint replacement (TJR) surgery, physicians’ visits, and prescription-based drug costs have been the three major cost categories presented in almost all COI studies of OA in Canada, the US, Australia, and European countries (58). Hospitalization has been reported to be the most expensive of all the categories in almost all of these studies (58). Recent studies shed light on the significance of out-of-pocket costs for OA patients (59,60). In a review of cost categories in musculoskeletal diseases such as OA, Mittendorf et al. (61) identify rehabilitation costs and out-of-pocket costs, as the major cost categories that are often not reported in COI studies.
1.3 Rationale
This thesis addresses several important gaps in the literature. First, until now, there have been no unique study that used patient-level sources of data for performing a COI study of OA at the national level in Canada that includes a comprehensive list of all direct cost components. As was mentioned in a recent COI evaluation study (62), most of the traditional estimates and projections for the total direct costs of OA in
Canada are based on a gross percentage of the total cost for musculoskeletal or arthritic diseases. These top-down extrapolations across time for different sub- populations are often biased, as they do not consider the disease dynamics and patients’ heterogeneity (62). In this study, I used patient-level data sources, included a comprehensive list of cost components, and with the use of a simulation model showed how the average cost of OA has changed for the Canadian population in previous years, from 2003 to 2010.
11
Second, in the case of COI studies that use patient-level data, up until now, none has produced a methodological framework to utilize different data sources and project the costs of OA into the future. In this thesis, I presented a framework to use microsimulation modeling as a unifying tool that can facilitate the use of administrative data sources for inpatient, drug, and physician cost components, in addition to patient questionnaires and surveys to arrive at valid alternative care and out-of-pocket costs. I showed how the historical trends for the inflation rates of each health sector cost can also be used to model their future rates, in addition to the use of regression modeling to project changes in resource utilization for the number of hip/knee surgeries.
Third, there is a gap in the literature concerning the main cost drivers for the recent increases in the average cost of OA. In this thesis, I presented how a microsimulation model can be used to identify the degree to which different cost drivers are responsible for increases in the average cost of OA in Canada from 2003 to 2010 including: changes in number of hip/knee replacement surgeries due to policy changes, an increase in the median age of the population, changes in the body mass index (BMI) of the population, price inflation, and increases in the expected lifespan of individuals due to a reduction in mortality rates.
Fourth, there is no clear understanding of how changes in the demographics of the Canadian population and the obesity epidemic will affect the cost burden for OA in future years and which sub-population will be affected the most. In this thesis, I showed how, with a microsimulation model, I can identify the cost associated with OA sub- populations by age category, sex, and disease states at different time points. As baby boomers in Canada and in other western countries are entering their last decades of life 12
in upcoming years, and the obesity epidemic continues, there is no clear understanding of how these will affect the quality of life and healthcare cost of various sub-populations at different ages, sexes, and disease states (1).
Fifth, there has been no previous study on how to perform uncertainty analysis
(UA) in PDMS models. Although UA is generally known to be a necessary post-model practice in microsimulation (MS) models, there are no guidelines as to how to perform such tasks. Furthermore, no previous methods has been developed to reduce the computational burden of UA in PDMS model for estimating the uncertainty associated with parameters of these models.
1.4 Study Objectives
The objectives of this thesis are as follows: 1) to estimate the average direct cost of OA between 2003 and 2010 using microsimulation modeling and identify the key cost drivers for the increase in the average cost within this period among Canadians; 2) to project the total cost of OA in Canada over the next 20 years, from 2010 to 2031, using microsimulation modeling and identify the distribution of this cost across sub- populations of different age categories, sexes, and disease statuses; 3) to provide a framework to perform UA in PDMS models; estimate the uncertainty around the mean of any type of outcomes in a reasonable computational time and estimate the uncertainty around the prevalence and total direct cost of OA over the future decades as predicted by POHEM-OA.
13
Chapter 2: Background
2.1 Osteoarthritis
Osteoarthritis (OA) is a chronic, debilitating disease that ranks among the most prevalent diseases in high-income countries (5,47). It is estimated that, in the US, 65-
75% of all patients with arthritis and rheumatism have OA (63). OA is caused by a failed repair of joint damage caused by an excessive mechanical load on joint tissues (64).
Excessive mechanical load may be due, in part, to an excessive body mass index (BMI) of an individual, exerting a load on weight bearing joints during daily activities (65).
Although primarily affecting the articular cartilage of specific joints, OA affects the entire joint, including subchondral bones, ligaments, and even muscles (64).
Early OA-related biomechanical changes starts in the knee, hip, and other joints of patients when they are in their 40s and early 50s (64), and then continue through a lengthy induction period. The hand, knee, hip, and spinal areas are the most frequently afflicted sites, whereas the elbow, wrist, neck and ankle areas are less affected (47).
However, the most severe pain, loss of function, and long-term disability are associated with the hip and knee areas (47). The disease burden of OA is significant in terms of loss of quality of life and the economic burden placed on patients, their families, communities, and society (66).
2.1.1 OA As a Major Musculoskeletal Disease
Among non-communicable chronic diseases, OA is the most significant contributor to disability in Canada (67). At the world stage, high-income countries suffer the most from chronic diseases in terms of per capita cost and loss of quality of life (12).
The World Health Organization (WHO) has identified six major types of chronic 14
diseases as they are responsible for most disability-adjusted-life-years (DALYs) lost: cardiovascular diseases, diabetes, cancer, chronic lung diseases, chronic neurological disorders, and musculoskeletal diseases (68).
Musculoskeletal conditions include joint diseases, such as arthritis - that includes osteoarthritis (OA), rheumatoid arthritis (RA), soft tissue rheumatism, and other arthritic diseases- back and neck pain, osteoporosis and fragility fractures, sports and workplace-related injuries, and trauma commonly related to road traffic accidents (69).
All of these conditions cause physical disability, pain, and loss of independence, both personal and economic, and thus place a heavy burden on both patients and family members. Taking into account both disability and premature mortality, musculoskeletal conditions have the fourth greatest impact on the health of the world population by contributing to 6.8% of the loss in DALYs of the world population- after cardiovascular diseases (CVD) (11.8%), all neoplasms (7.6%), mental and behavioral disorders (7.4%)
(12). Arthritic diseases constitute a significant proportion (>20%) of all musculoskeletal diseases in US (70) and around the world (12).
2.1.2 The Most Common Form of Arthritis
OA is the most common form of arthritis (5), afflicting more than 50% of all those with arthritis in Canada (71) and the US (70). There are over 100 different types of arthritis such as rheumatoid arthritis, lupus, juvenile arthritis, and psoriasis arthritis (70).
There are relatively few studies in the literature reporting on OA disability and its cost burden compared to those focusing on rheumatoid arthritis or general arthritis (arthritis of any type) (58).
It is estimated that, worldwide, 9.6% of men and 18.0% of women aged ≥ 60 15
years have symptomatic osteoarthritis (5). In North America, arthritis has been identified as the leading cause of activity limitation and work-related disability among the adult population (70). In a recent Public Heath Agency Report in Canada more than 2.5% of all adults reported a long-term disability associated with arthritis (71). In Canada, over half the population aged 45 to 75 years that was afflicted with arthritis reported activity limitation (71). In a recent report, Arthritis Society of Canada reported that the average impact of arthritis on the average Quality of Life (QOL) of patients was greater than those of CVD, cancer, diabetes, and chronic bronchitis (72). In comparing the global burden of disease studies from 2000 to 2010, Murray et al. (12) shows that the increase in the burden of arthritis, in terms of loss of disability-adjusted life years of the world population, is the highest among all other musculoskeletal diseases. As with almost all chronic diseases, arthritis creates a burden that increases with the age of the population
(5, 46). In fact, arthritis was the most frequent cause of disability among the elderly in
Canada according to Public Heath Agency Report in 2010 (71). In terms of prevalence among musculoskeletal diseases, OA ranks third worldwide, after back and neck pain, respectively, accounting for 10% of all life years lost due to disabilities caused by musculoskeletal conditions in 2010 (12), i.e., OA was associated with 0.7% of all life years lost due to disabilities caused by any chronic conditions.
2.1.3 Incidence and Prevalence of OA
Although prevalence estimates of chronic diseases such as osteoporosis, diabetes or CVD have been reported in national population-based surveys, OA prevalence estimates are not well established (8,73). This is due in part to the different criteria used to define and identify prevalent cases. In addition to the self-reporting, two 16
other clinical diagnostic criteria are used to measure the prevalence of OA in administrative-based and other clinical studies: radiographic and symptomatic criteria.
Radiographic OA is defined on the basis of the x-ray features of the afflicted area(s); on the other hand, symptomatic OA is defined according to the existence and frequency of the stiffness and pain of different joints.
There exists a large variation between different prevalence estimates for both site-specific and general OA (74) - the same is also true for incidence estimates of OA.
One reason for the variation between the prevalence estimates is the lack of definitive diagnostic criteria for OA, due to its lengthy asymptomatic latency period (46). For example, an early radiographic signs of knee-OA is often seen in people between 55-
65, while the majority of these patients have no symptoms such as pain or stiffness
(46). In fact, according to the radiographic definition of OA, such as KL grade1, imaging techniques can be used for case ascertainment-, but many radiographic cases may not experience pain or stiffness. Therefore, radiographic OA has low predictive power in terms of disease burden, which is directly related to pain and other disease symptoms
(46). The definition used for the case ascertainment in each study is in part dependent on the goals and context of the study. For example, the case definition criteria developed by the American College of Rheumatology (ACR) for classification and reporting of joint specific OA (knee, hip (75) and hand (76)) include a combination of symptoms, including pain and stiffness due to OA and KL grades for radiographic signs
(measured by joint space narrowing and osteophytes). The reliability and specificity of
1 Kellgren-Lawrence Grading Scale (KL) is a score used for severity status of OA based on x-ray features (46).
17
this combined set of criteria, which was designed to separate OA from other inflammatory forms of arthritis, have been tested in clinical populations (75).
The majority of large incidence and prevalence studies of OA have used a prospective or retrospective cohort design and clinical, survey or administrative data sources (74). These include, for instance, the Framingham OA Study (77), Johnston
County Osteoarthritis Project (78), NHANES I (79) and OA studies in British Columbia using administrative data (74). These studies use case definitions involving radiographic
OA (NHANES I (78)), symptomatic OA (80) (Johnston County (77)), or a combined measurement, which is a combination of disease indications and symptoms (23,74).
Based on ACR criteria, two recent cohort-based studies (77,78) have estimated that over one third of adults greater than 65 years of age in the US have clinical OA of the knee.
Administrative Data Sources in OA Disease Burden Studies
Although the combined measure developed for site-specific OA by ACR
(mentioned above) is generally accepted, researchers’ use of administrative data for prevalence and incidence studies of chronic diseases such as OA has been rising in recent years in Canada and other countries (74). The availability of population-based linked administrative data for health resource utilization has fueled the increase in both cost studies and epidemiologic studies (81,82). According to consensus statements for the methodology, recent guidelines were developed for the use of administrative health data in rheumatic disease research and surveillance (83).
Using 10-years of data (1991-2001) from billing records that defined OA based on ICD-9 diagnostic codes (all joints combined), Kopec et al. (74) estimated an overall 18
population prevalence of 11% for doctor-diagnosed OA in 2001, and reported that by age 70-74 years, about one-third of men and 40% of women had OA in British Columbia
(BC), Canada. In both the incidence and prevalence estimates from the above study,
Kopec et al. (74) use the following criteria for extracting an all site incidence number for
OA from a population-based administrative database (i.e., Population Data BC (PDBC
(84)): “OA was defined as at least two visits to a health professional within 2 years or one hospitalization with the International Classification of Diseases, Ninth Revision
(ICD-9) code 715 “. This same definition was used by Kopec et al. (1) to estimate some parameters in the POHEM-OA model, a population-based simulation model of OA in
Canada. As I used POHEM-OA in all models of this thesis, I used the same definition for the prevalence and incidence of OA (Chapter 3-5).
2.1.4 Risk Factors for OA Incidence and Progression
Risk factors for OA incidence can be categorized into two main categories: biomechanical and systemic. OA’s major biomechanical risk factors include weight, muscle strength, joint injury, and job type, all of which affect the cumulative axial force on the joint (5). The cumulative axial force on the joint is the final link through which all biomechanical risk factors manifest themselves. Axial force on the joint affects the start of the degradation of cartilage that leads to an OA incident (5). Table 2.1 shows the risk factors for OA.
In addition to biomechanical risk factors, systemic risk factors in OA play a significant role in the incidence and progression of OA. These are non-modifiable risk factors, including genetic predisposition, mal-alignment in joints, gender and age. Mal- alignment in joints affects the force exerted on joints. Female gender has shown to be a 19
significant risk factor in both OA incidence and its progression (5). While the exact details of how OA initiates are as yet unknown, the major direct risk factors at this stage are shown to be genetic predisposition and joint injury (73).
The effect of age on OA incidence is very complex. While OA used to be considered a disease due to aging, there is accumulating evidence that OA initiates at younger ages (47). In the Framingham study, 14% of the population 26 years of age or older had radiographic OA and 5% had symptomatic OA (76). This increased to an average of 23% radiographic OA and 12% symptomatic OA among those older than 45 years of age (76). Additionally, recent studies using MRI-based diagnostic criteria revealed that changes in the cartilage often occur at younger age and in people without any x-ray signs or symptoms (64).
Table 2-1. Summary of risk factors for OA incidence
Biomechanical risk factors Systemic risk factors
Diet and physical activity Genetic predisposition, mal-alignment
Weight, muscle strength, joint injury, in joint, gender and age and job type
Risk factors for the progression of OA are different from those for incidence (75).
For example, while a patient’s high levels BMI are reported to be a major risk factor for the incidence of symptomatic and radiographic knee-OA, their effect on OA progression to more severe stages, is shown to be rather mild (73). Progression of OA has been measured through either radiographic, symptomatic, e.g., through patient’s
20
questionnaire such as The Western Ontario and McMaster Universities Osteoarthritis
Index (WOMAC)2, or a combined measure (5). Both radiographic and/or symptomatic criteria are related to the loss of articular cartilage of the joint (5, 73). Data from studies using X-rays, i.e., radiographic criteria, can be used for information on the progression of OA according to joint space narrowing and KL grade (64). However, recent studies have used MRI to provide detailed data on the percentage of cartilage loss for different joint compartments (85).
Once OA progresses to the severe stages, joint replacement surgery is the ultimate treatment. Apart from replacement surgery, other types of interventions, including medication, physiotherapy, and alternative therapies are used mainly for pain relief and improvement in function of OA patients (86). Since, up to now, there is no cure for OA, further studies into the modifiable risk factors of OA, including obesity and physical activity are warranted (87). Such studies can help design policies that will prevent or modify OA’s progression to the more severe stages, and reduce the significant burden of OA on society in general (87).
2.1.5 OA Disease Burden
Symptomatic OA places a significant burden on patients, their families and society in general, in terms of physical and psychosocial disability. Due to its high prevalence and long duration, the OA disease burden is among the highest, in terms of both societal cost and patients’ loss of quality of life, when accumulated over all patients, compared to other chronic diseases. OA ranked 11th in 2010, in terms of its
2 WOMAC score: a questionnaire to measure the severity of a general (all-site) OA that has three measures for pain, function and stiffness. It is between 0-100 and higher scores represent worse symptoms (5). 21
effect on disability-adjusted-life-years (DALYs), which reflect the combined impact on mortality and quality of life, and became the fastest increasing major health condition in the world as reported in the recent global burden disease study by Murray et al. (12).
This relates not only to the aging of the population, but also to increased obesity (12), and other risk factors of OA as discussed in the previous section.
Although relatively low in the younger population, disability due to OA is the highest among the elderly population in the US and Canada, as compared to other causes of disability (7, 63). In the Framingham Study, disability due to OA was equal to that of cardiovascular disease and highest among all other chronic conditions among the elderly, including diabetes, other musculoskeletal diseases and other conditions3
(76). Population-based studies have shown that a high level of comorbid conditions exist among those diagnosed with OA, such as depression, cardiovascular disease, diabetes, and respiratory diseases (88). Although OA rarely causes death, the medications used for relieving its symptoms, such as NSAIDS and coxibs, might cause gastrointestinal (GI) problems and increase the risk for CVD and stoke (5).
2.1.6 Projecting the OA Burden with the Use of Modeling
Since the majority of disabilities from OA occur in older adults, the number affected in the future will increase substantially as the baby boom generation ages (34).
In addition, age combined with the obesity epidemic will place enormous demands on the healthcare system and society (89). As described above, the major risk factors for the incidence and progression of OA in the knee, hip, and hand are BMI and age, with
3 Disability was defined in that study as needing help walking or climbing stairs (76).
22
different sex-specific rates for each OA site. As baby boomers enter the last stages of life and the obesity epidemic is increasingly affecting the health of the populations of high-income countries, the prevalence of OA and its disease burden will be rising in the future years in these countries (1).
Various studies have used historical data and modeling to predict the musculoskeletal and arthritis burden in future years (10, 34), while recent studies have focused more on disease specific burdens such as OA-related burden (1, 64,68). For example, Badley et al. (10) used regression modeling and projected that the number of people with arthritis and arthritis-attributable disability in Canada will double by 2031. In another study conducted in the US, the authors used regression modeling and estimated that the number of people with arthritis will increase by 57% by 2020, while activity limitations associated with arthritis will increase by 66% (13). A review of existing data as part of the Bone and Joint Monitor Project, in collaboration with the
WHO’s Global Burden of Disease 2000 project (68), estimated that OA would be the fourth leading cause of disability-adjusted life-years lost by the year 2020. Most of the above studies only include changes in the age-structure of the population and do not consider the longitudinal changes in average of BMI, which has been increasing in high- income countries. Using only historical trends and extrapolating the future changes by using linear regression models, Hunter et al. (64) estimate that the prevalence of knee
OA will double by the year 2020, due in part to increases in obesity and longevity.
Although hospital discharge data was used in this study, their estimates did not capture the interaction between age and BMI in future years in the sense that the effect of BMI varies by age of the patients. 23
Simulation modeling allows for the use of inter-related regression models for projecting the burden of chronic diseases, such as OA, with time-dependent, interacting risk factors, such as age and BMI. Kopec et al. (1) used POHEM to predict the OA burden in future years (between 2001 and 2021) by including population-based data as their initial population, while considering age and BMI as major risk factors for the disease. Their results show that, from 2010 to 2031, the prevalence of OA in females would increase from 16 to 21%, while the male prevalence would increase from 11%-
16% (1). This model uses a validated demographics module from Statistics Canada that includes projected birth, immigration, and mortality rates in addition to models estimated from National Population Health Survey (NPHS) for projecting the BMI of the population
(1).
Simulation modeling have been using in recent decades to project the disease burden of several chronic diseases (12). In the Global Burden of Disease report, simulation modeling was used to calculate the effect of each chronic disease on the average DALYs. Complex types of imputation approaches and data mining approaches have been combined to conduct simulation modeling in the Global Burden of Disease report (12). Simulation modeling has been used to estimate and project the burden of knee and hip surgery in the US population (52, 53).
2.2 Cost-of-Illness Studies
Calculating the economic burden associated with (1) a disease, (2) a group of diseases (such as arthritis), (3) a medical procedure or intervention and (4) the overall healthcare system in a country, has been the goal of different cost studies over the past
24
50 years (90). Since the first studies on health-related costs in the 1960s, the methodology has evolved significantly, and the quality of the studies has improved.
In general terms, the total economic cost of health care is represented by the value of all the resource services, including labour, materials, and capital equipment.
However, the values of these resources should be measured in terms of their highest valued alternative use (90). According to this definition, healthcare costs due to illness have been defined as “reductions in consumption possibilities that are brought about by the occurrence of illness within the population“ (57). These reductions in consumption possibilities result in tangible and intangible costs. Tangible costs are those that can be expressed in monetary terms and are included in cost studies. The first type of tangible costs comprises the direct costs that reflect the “diversion of resources towards the diagnosis, treatment, and management of the illness, and away from other productive uses” (57). The second type of tangible cost is the indirect costs (i.e., productivity losses) that reflect how “illness-related morbidity and mortality render human resources unavailable for productive uses “(57).
On the other hand, intangible costs are considered illness burdens rather than losses in scarce resources and are best expressed in non-monetary measures, such as
DALYs or quality-adjusted life-years (QALYs) foregone (90). In recent studies, only tangible costs are defined as losses in economic resources and included in cost studies.
The valuation of so-called intangible costs in monetary terms is controversial, according to several studies (90). In fact, some of these studies consider the practice inappropriate because, according to the definition of healthcare costs I mentioned above, intangible costs do not cause reductions in the availability of tangible resources, 25
i.e., they do not incur direct or indirect costs (90). Therefore, I will not include intangible costs in the cost of disease study in this thesis.
As mentioned above, economic tangible costs that are included in healthcare cost studies are categorized into two distinct types: direct and indirect costs. In simple terms, direct costs are the money costs incurred for paid services, e.g., wages, professional fees for a formal caregivers, costs for supplies (57), while an indirect cost is defined as an imputed valuation for unpaid resource services, e.g., productivity loss such as patient’s foregone income while being absent from the job (absenteeism) (57). I will discuss different cost components of direct and indirect cost for OA in Section 2.3.4.
Direct costs are categorized further into medical and nonmedical costs. Direct medical costs include hospital inpatient, physician and specialist’s fees, emergency department outpatient, nursing home care, hospital care, rehabilitation care, and other health professionals’ care costs, diagnostic tests, prescription drugs and drug sundries, and medical supplies (90). Nonmedical direct costs include transportation costs to health care providers; relocation expenses; and costs of making changes to one’s diet, house, car, or related items. However, some nonmedical direct costs are generally not included in cost-of-illness studies, such as research, training, and capital costs (e.g., construction). It can be difficult to attribute these costs to a particular disease. On the other hand, indirect costs are categorized into three major types: mortality costs
(premature mortality), morbidity costs due to absenteeism and presenteeism, and informal caregiver care costs, in terms of (consumption) opportunities lost and the cost of hiring outside care, e.g., unpaid caregiver time and caregiver productivity loss (90).
26
According to a recent review of economic studies of healthcare (57), healthcare cost studies can be categorized into the following four types: 1) cost evaluation studies, including cost-effectiveness studies, health technology assessment studies, and cost- benefit analyses; 2) cost impact analysis; 3) cost surveillance studies; and 4) cost-of- illness studies (COI). Each one of these study types utilizes a costing method within a specific context of health, such as a total cost approach for a specific disease, or top- down approaches for evaluating a procedure cost (57). In recent years, a significant proportion of cost studies in healthcare lay within the field of chronic diseases and included cost evaluation studies for specific interventions and COI studies, while fewer were cost impact analyses (62). My focus here will be only on the COI approaches.
2.2.1 Cost-of-Illness Studies: Definition and Methods
Cost-of-illness (COI) studies are “descriptive analyses assessing the economic burden of health problems on the population” (57). COI studies have been invaluable tools in public policy in the last 40 years (57). The significance of COI studies can be seen in their frequent use by policy makers, and in studies that attempt to highlight the importance of a particular disease. For instance, a COI study of diabetes illustrating the magnitude of its burden, performed by the American Diabetes Association in 2002, was cited in several other informative public policy studies (90). Another advantage of COI studies on a practical level is that cost figures from these studies could be used in cost- effectiveness and cost-benefit analyses. For example, a cost-effectiveness study of treating Alzheimer’s disease with Donepezil (91) uses the cost figures from an earlier
Alzheimer’s disease cost-of-illness study by Rice et al. (92).
COI studies have several limitations in their applicability. For example, while 27
these types of studies can demonstrate which diseases may require an increased allocation of prevention or treatment resources, they are limited in determining how the resources are to be allocated (62). More importantly, COI studies employ varying methods, which results in high variation in their outcomes and can limit the comparability of their findings (62). In addition to differences in COI methodology, these studies can vary by perspective, sources of data, inclusion of indirect costs, and the time-period of costs (62, 90). One of the major reasons for variation among different
COI studies is due to the perspective of the study.
Perspective of the COI Study
The perspective of a COI study is defined according to those who are affected when choices are made about the allocation of resources, such as by governments, private funders, the patients, and/or society, or those on whose behalf the decisions are made (93). In other words, the perspective of the COI study is defined according to the policies relevant in the study (i.e., context dependent). For example, regarding the costs or losses associated with a disease, the main focus, from a third party payer’s perspective such as an insurer, is on the impact of absenteeism and lost productivity as shown for the costs of Alzheimer’s disease to US businesses (91). Table 2.2 summarizes the different study perspectives found in the current COI literature where X refers to those cost categories that are included in the COI study with the relevant perspective (90, 93).
28
Table 2-2. Cost-of-illness (COI) study perspectives
Perspective Medical Morbidity Mortality Transportation/ costs costs costs non-medical costs Societal X* X X X Healthcare X system Third party X X payer partially partially Industries X X X /Businesses partially presenteeism and absenteeism only Patients/ X X X X Families out-of- lost wages lost out-of-pocket cost pocket and caregiver wages cost only cost * X refers to those cost that are included for each perspective
Although a high variation exists among the outcomes of COI studies, cost-of- illness studies represent an important analytic tool in public health policy. Here, I describe a categorization for COI types, methodologies, and approaches.
COI Study Methods and Approaches
COI studies can be evaluated according to both the type of data they utilize and the methods they use according to the design of the study (COI methodology). COI types include top-down and bottom-up approaches, while COI methodologies are a collection of two major methods: 1) incremental cost vs. total cost method; and 2) incidence-based vs. prevalence-based cost method (62). Figure 2.1 depicts this categorization of the COI studies. These dichotomies are independent. For example, a top-down study is described as incidence-based and incremental. However, according
29
to Akobundu et al. (57), the COI approaches were only used in total or incremental cost studies.
Figure 2-1. Cost-of-illness (COI) study types, methodologies, and approaches*
* Every COI study can be categorized into a type, methodology, and approach, as shown in each level of this figure, respectively. COI type reflects the type of data used in the study, while the COI methodology category reflects the methods researchers use to design the study. COI approach reflects the technical and statistical approaches implemented in total cost or incremental cost studies only (57). These dichotomies are independent. For example, a bottom-up study that utilizes individual-level cost data could use an incremental and incident-based cost methodology with a regression-based approach. COI studies are evaluated according to their type, methodology and approach.
30
The top-down COI study type also referred to as the epidemiological or attributable risk approach (90), measures the proportion of a disease that is due to exposure to the disease or risk factor. This approach utilizes the published expenditure reports, by cost component, and then uses aggregated data, along with a population- attributable fraction (PAF) to calculate the attributable costs. Top-down models have been criticized for their limited ability to reflect variations across sub-populations and the possibility of a resulting bias in estimates of the disease cost (62). For instance, biased estimates can be produced in top-down COI type estimates when adjusted relative risks of a disease are used to calculate overall population PAFs (90). More importantly, top- down COI-type studies should use a selection of co-morbid diseases that are attributable to the primary condition (57). Furthermore, top-down COI-type studies do not account for the variation of resource use across different patients (57).
On the other hand, bottom-up, micro-costing, or person-based approaches assign resource use and productivity loss to individuals with the health condition of interest, either from detailed actual data from real cases covering the timeframe of the study, or using evidence from other sources to construct hypothetical cases, i.e., use of modeling (90). Then, total cost can be calculated by the resulting mean per-person cost and appropriate population prevalence or incidence data. While critics have pointed out several advantages to using bottom-up COI types rather than top-down, the main challenge is the lack of availability of data on the disease status, in addition to a lack of availability of data regarding the actual resource used by each individual patient over the time period of the study (62).
31
Incidence-based and prevalence-base cost studies are two methodologies used in COI studies (90). Incidence-based cost studies estimate the lifetime costs of a disease by measuring the cost from the time of the onset of the disease to death, for all cases within the period of the study. Incidence-based cost estimates include the discounted, lifetime direct and indirect costs for the incidence cohort using a longitudinal data source (90). Prevalence-based cost studies estimate disease costs within a specific time period (which is often one year) by measuring the direct and indirect costs associated with all cases using a cross-sectional data source (90).
As shown in Figure 2.1, total and incremental cost methods are other methodologies used in COI studies, in addition to prevalence and incidence-based methodologies (62). Total disease costs provide estimates for the potential savings from eradicating the illness –even if eliminating the disease is unrealistic (90). However, this method does not calculate the costs causally related to the disease as it sums all the cost for those with a specific diagnosis, i.e., sum-diagnosis approach, or those with specific medical treatments, i.e., sum-all-medial approach. For example, the overall estimate from a total cost method with sum-all-medical costs approach may include additional cost due to comorbidities associated with a disease and therefore, does not reflect the cost causally associated with the disease in question. On the other hand, incremental cost studies “estimate the increase in costs that is solely attributable to the presence of the disease” (90). In other words, incremental cost studies utilize different approaches to deduct the unrelated costs and therefore, they are more precise in estimating the costs that are causally associated with the disease in question.
Depending on the total or incremental cost methodology, different approaches 32
have been identified in the literature for performing a COI study (57). These are the technical and statistical approaches that have been used in each study to analyze the data. As shown in Figure 2.1, two different approaches have been identified in the COI literature for each of the above-mentioned methods (57): 1) regression-based and matched cohort studies for incremental cost method, and 2) the sum-diagnosis and sum-all-medical cost for total cost method. The sum-diagnosis approach aggregates all the costs of individuals who have same type of diagnosis, while the sum-all-medical cost approach sums over all treatment costs for patients who have the same type of treatment (57). As a result, a COI study can be evaluated according to its type, methodology, and approach. For instance, Anis et al.’s (94) study that estimated the overall cost of obesity-related diseases in Canada is a top-down, prevalence-based study that calculated the total cost using a sum-diagnosis approach.
In a review of all the COI studies done between 1996 and 2005, of the 387 articles identified, 57% used the total cost method with a sum-diagnosis approach; 18% used the incremental cost method with a matched cohort approach; 16% used the total cost method with a sum-of-all –medical costs approach; and only 9% used the incremental cost method with a regression-based approach (57).
2.3 Economic Burden of OA
While there are some COI studies of arthritis and OA in Canada, earlier COI studies focused on all musculoskeletal diseases (69). Although, since 2004, COI studies in OA and arthritis have been moving toward more comprehensive cost listings, no standardized cost listings for OA in Canada have been published yet (61). On the other hand, Mittendorf et al. (61) have developed guidelines on how to implement 33
standardized cost categories for musculoskeletal diseases. Here, I review the musculoskeletal and arthritis cost studies and, in the next section, I perform a systematic review of COI studies in OA.
Musculoskeletal diseases comprise a significant part of all non-communicable chronic diseases in terms of cost and disease burden in both high-income and low- income countries (69). Using a top-down, prevalence-based, total-cost approach based on expenditure data from 2002, Health Canada estimated the annual cost of all musculoskeletal diseases to be around $ 22.5 billion $CAD4, including both direct and indirect costs (medical care and lost wages only) (95). In 2002, the cost for musculoskeletal diseases was the second highest cost burden among all chronic diseases, after cardiovascular diseases (95). A 2007 US study estimated the cost of arthritis to be over $80.8 billion as compared to $64.8 billion in 1997 including both direct and indirect cost (earning losses) (96).
Arthritis constituted almost 33% of the overall cost burden of musculoskeletal conditions in Canada in 2000 (95). As mentioned before, OA is the most prevalent disease among all types of arthritis (around 50-60% of all arthritis cases), and its cost has been shown to be around 35-40% of the overall arthritis cost (85). Also, in a 1997 study using expenditure reports, OA was reported to be the most common disorder among all musculoskeletal diseases in five industrialized countries (97). Knee and hip
OA are the main contributors to these costs – and, while hand OA is higher in frequency, its cost is not as high as that of hip and knee OA mostly due to TJR surgery
4 All costs in this chapter are in 2010 Canadian dollars ( 2010 $CAD) unless mentioned otherwise. The costs were translated into 2010 $CAD using overall consumer price index (CPI) according to Statistics Canada rates (details are discussed in Chapter 3). 34
costs (98).
2.3.1 Systematic Review of Cost-of-Illness Studies in OA
For this study, I developed a search strategy to identify the literature related to the cost of osteoarthritis, with a focus on those studies aimed at measuring the national total or incremental cost of OA including direct and indirect costs. The review was performed in MEDLINE (99). I performed an additional search in four databases: 1) the
Canadian Research Index (100); 2) The Database of Abstracts of Reviews of Effects; 3) the National Health Services Economic Evaluations Database (NHS EED); and 4) the
Health Technology Assessment database (HAT). The last three databases, which include reviews as well as cost-effective analysis (CEA) and COI studies from all over the world, were all provided in the Centre for Review and Dissemination, Canada (99). I selected publications from all of the above databases between May 2006 and February
2013. I did not search any publications before April 2006, as Xie et al. (101) performed a systematic review using the same strategy as ours for 1966-April 2006. Table 2.3 depicts the results performed by Xie et al. (101) for which I have associated the type, method and approach of the COI study.
To identify the most relevant articles on COI studies performed after 2006, I used the two overlapping strategies described by Petitti et al. (102). First, I searched for
“osteoarthritis“ with “direct cost”, “economic burden”, “expenditures”, and” cost-of- illness” in abstracts of all publications within the texts. Next, I searched “osteoarthritis“ with “costs” in the titles of publications. I then included the resulting references from both strategies in one directory and reviewed the titles of identified references to exclude any duplicates. Overall, 117 articles were selected. As these were not all COI 35
studies, I modified the list in two steps: first, I removed the CEA and cost-benefit analysis (CBA) studies or clinical trials and observational studies that examined the costs as part of their study. As a result, 34 studies remained, from which I excluded 19 studies related to specific interventions. Finally, I also excluded four general reviews of the cost of OA in recent years (58, 101,103,104). As a result, for this review I selected
11 OA cost-of-illness studies investigating the cost burden of OA, including either direct or indirect costs or both. Seven studies estimated the overall national cost burden of OA in industrialized countries, including Canada (94, 105), the US (60,106), European countries (107), (108), and Singapore (109) as shown in Table 2.4, while four studies calculated the burden of OA among special sub-populations or provided site-specific burden of OA (110–113). I will review the cost of OA among sub-populations in Section
2.3.4.
2.3.2 Literature Synthesis
The number of COI studies conducted in the past six years that reported on the national burden of OA including either direct, indirect, or both types of costs, are shown in Table 2.4. As it can be compared to the results of the systematic review for cost burden of OA done by Xie et al. (101) for 1966 to 2007, shown in Table 2.3, an almost equal number of COI studies of OA were performed in the last 6 years, as opposed to those since the beginning of COI studies publication.
36
Table 2-3. National cost of OA studies performed before 2006
Type Methods Data Approaches
Canada: Maetzel et al. 2004 (114) Bottom-up Incremental Survey Regression-based Gupta et al. 2005 (115) Bottom-up Total cost Survey Sum- all medical
US: Lanes et al. 1997 (119) Bottom-up Total cost Database Sum-diagnosis Maclean et al.1998 (120) Top-down Total cost Database Sum-all medical
Europe: Levy et al.1993 (116) Top-down Total cost Database Sum-all medical Le Pen et al. 2005 (117) Top-down Total cost Database Sum-all medical Leardini et al. 2004 (118) Bottom-up Total cost Survey Sum-all medical
Asian countries: Woo et al., 2003 (121) Bottom-up Total cost Survey Sum-all medical
*Eight out of ten studies reported in Xie et al. (101) are shown in Table 2.3; two studies prior to
1993 were not presented here.
Out of overall 15 selected studies shown in Tables 2.3 and 2.4, the majority of
COI studies in OA are from Europe (5 studies), US (4 studies), Canada (4 studies), and
Asia (2 studies from Singapore). Compared to studies prior to 2006 (Table 2.3), most of the studies performed after 2006 (Table 2.4) were bottom-up COI types rather than top- down (85% in Table 2.4 vs. 63% in Table 2.3), and used more national databases rather
37
than primary survey data (72% in Table 2.4 vs. 50% in Table 2.3). Studies in recent years used large size (insurer) claim data in the US studies (60,106) and population- based administrative data in the Canadian ones (94,105) rather than small sample sized studies, such as Gupta et al. (115) and Maetzel et al. (114). As mentioned in the last section, this pattern has been seen in COI studies of other diseases, in recent years, as well (57). According to internal and external validity criteria, the use of large sized studies increases the quality of COI studies (62).
Almost 70% of all (5 out of 7) studies after 2006 used regression and matched cohort (incremental cost) approaches, compared to 12% (1/8) of studies prior to 2006, as shown in Tables 2.3, 2.4. This fits with the claim by Akobundu et al. (57) that recent
COI studies are using new techniques to estimate incremental costs, whether by using matched cohort data, by performing regression-based analysis to control for relevant comorbidities, or by using both techniques (i.e., gold standard). As mentioned as a limitation of COI studies (62), types of the cost components used were inconsistent across the COI studies of OA. In the next section, I present the list of all components that were common among the COI studies of OA.
38
Table 2-4. National cost of OA studies from 2006 to 2012
Type Method Data Approach
Canada: Tarride et al. 2012 Bottom- Incremental Database Regression (105)* up (linked and CCHS) matched cohort Anis et al. 2010 (94) ** Top- Total cost Database Sum -all down US: Kotlarz et al. 2009 Bottom- Incremental (Claim) Regression- (60) up database based
Bottom- Incremental (Claim) Matched Le TK et al. 2012 up database cohort (106)
Europe: Loza et al. 2009 Bottom- Incremental Survey Matched (107) up cohort
Rabenda et al. 2006 Bottom- Incremental Survey Matched (108) up cohort
Asian countries: Xie et al. 2008 Bottom- Total cost Database Sum- (109) up diagnosis
* This study only estimated the cost of OA in one province of Canada (Ontario); ** This study estimated the cost of OA based on the Economic Burden of Illness in Canada (EBIC) 1998 study
(95).
39
2.3.3 Cost Components in COI Studies of OA
The comprehensive list of OA-related direct cost categories, paid for by patients or the health system, includes medical expenditures (inpatient, outpatient, drugs, and those due to the side effects of drugs) and non-medical ones (formal caregiver, travelling cost to a physician’s office and other disease-related destination, and other out-of-pocket cost of patients and community services costs such as home remodeling, or purchase of medical aids) (61,114,115). Figure 2.2 depicts the overall cost components for both the direct and indirect costs of OA according to different study perspectives. Cost components that were common across all 7 studies of Table 2.4 are listed in Figure 2.2. Those cost components that are shown in parenthesis in Figure 2.2 have only been used in two studies or less, while others used in more than two studies.
For example, the cost associated with caregivers’ productivity loss was included in calculating the indirect cost of OA only in one of the studies (i.e., Gupta et al (115)) that had the societal perspective.
40
Figure 2-2. Cost components in COI studies of OA and their respective perspectives in Canada
Society*
Patient * • Productivity!loss!at!work!including!absenteeism,!presenteeism,! • Over *All cost components are either due to OA or OA-related comorbidities including side effect of medications from the societal perspective; ** components in a parenthesis has only used in 1 or 2 studies while others used in more than 2 studies; *** comprehensive list of OA drugs are shown in Table A10, Appendix A. Acetaminophen and some of the NSAIDs are over-the-counter and are out- of-pocket cost. While the main focus of this study is on the direct cost burden of OA, I have nonetheless reviewed studies on both the direct and indirect costs of OA. Here, I discuss the overall economic burden of OA based on both direct and indirect costs, by summarizing all the publications in Table 2.4. 41 2.3.4 Direct Cost of OA The total cost for OA has been reported to be around two times higher than the direct cost, i.e., the ratio of direct to indirect cost is about 1:1 (85). On the other hand, in rheumatoid arthritis (RA), this ratio has been reported to be between 1:2 and 1:4 (85). Table 2.5 shows the summary of the results of the literature review for all studies after 2000 for average cost and total cost of OA, in addition to the cost components included in each study. Below I compare different results separately for average costs, and total costs of OA according to the countries of the study including Canada, US, European countries and Asian countries. 42 Table 2-5. Summary of the literature review for direct cost of OA (all costs are in 2010 $CAD) Author' Country' Year'of' Cost'studied' Direct' Average'direct' Total' study'1' cost'components'' cost' direct' cost' Tarride!et!al.2!(105)!!! Canada! 2010! Direct!! Physician,!hospitalization!! $1,200! N/A! ! Anis!et!al.3!(94)!! Canada! 2006! Direct!and!indirect! Physician,!prescription! N/A! $1.2! ! drugs,! billion! hospitalization,! Maetzel!et!al.!(114)! Canada! 2000! Direct!and!indirect! Physician,!drugs,! $3,797! N/A! hospitalization,!outIofI pocket! Gupta!et!al.!(115)! Canada! 2002! Direct!and!indirect! Physician,!drugs,! $2,123! N/A! hospitalization! Kotlarz!et!al.!(60)! US! 2005! Direct! OutPhysician,!IofIpocketdrugs,!! $7,846!women! $234.28! hospitalization,!! $5,974!men!! billion!! outIofIpocket!! ! Le!TK!et!al.!(106)! US! 2010! Direct! Physician,!prescription! $13,381!! N/A! drugs,!hospitalization! White!et!al.4!(113)! US! 2010! !Direct!and!indirect! Physician,!drugs,!outIofI $14,530!! N/A! pocket! ! ! ! ! ! $2,323!mild!! N/A! Dibonaventura!et!al.!4! US! 2010! Direct! Physician,! $8,390!moderate!! (112)!!!!!!!!!!! drugs,!hospitalization,! $17,379!severe!OA! ! outIofIpocket! ! ! ! ! 43 Author! Country' Year'of' Cost'studied! Direct' Average'direct' Total' study'1! cost'components'! cost! direct' cost! Le!Pen!et!al.!(117)!! France! 2003! Direct! Physician,!prescription! $379!! $1.89! ! drugs,!hospitalization! billion! Leardini!et!al.!(118)! Italy! 2001! Direct!and!indirect! Physician,!drugs,! $1,178! N/A! hospitalization,! outIofIpocket! Loza!et!al.!(107)! Spain!! 2003! Direct!and!indirect! Physician,!drugs,! $1,552! $5.28! hospitalization,! billion! outIofIpocket! Rabenda!et!al.!(108)! Belgium! 2003! Direct!and!! Physician,!drugs,! $662! N/A! indirect! hospitalization! Woo!et!al.!(121)! Hong! 2001! Direct!and!indirect! Physician,!drugs,! $7,873! $388! Kong! hospitalization,! million5! outIofIpocket! Table 2.5 notes: 1 Only studies performed after 2000 are included here; annual average and total cost of OA costs are deflated to 2010 $CAD using consumer price index in Canada (134) and purchasing power parity of different countries (149); 2; Year of study refers to the year of data for each study; 3 Anis et al. (94) estimated the cost of OA based on the Economic Cost of Illness in Canada (EBIC) 1998 study (95); 4 White et al. (113) and Dibonaventura et al. (112) performed their study only among employed OA population; 5 This is for both direct and indirect costs. ! 44 Average Cost To estimate the average cost of OA in Canada, Maetzel et al. (114) performed a bottom-up study in 2004 and estimated the incremental costs of three diseases including OA, high blood pressure, and rheumatoid arthritis (RA). In this study, 140 OA patients were surveyed and the annual average cost of OA was estimated to be $5487 per OA patient, using a Canadian cost template from 2000 (2010 $CAD), where the direct cost portion was $3,797 (114). As shown in Figure 2.2, they included all the components of indirect and direct costs in their study, except for the loss of leisure time of both the patient and the caregiver. On the other hand, Gupta et al. (115) obtained $2123 for the average direct cost and $11,993 for the indirect costs of OA in 2005 (all in 2010 $CAD). Gupta et al.’s study (115) was the only study that included both patients’ and caregivers’ productivity losses. Tarride et al. (105) was a most recent COI study for OA that was performed for only one province of Canada (Ontario) in 2012. They used a bottom-up, incremental methodology and performed a regression-based matched cohort study from a linked CCHS data in 2010 with provincial administrative data to estimate the average cost of OA. They have reported that the incremental average direct cost of OA in the province of Ontario in 2010 was around $1,200 (95 %CI: $1077, $1323) for an average OA patient that did not include out-of-pocket costs (105). The average cost of OA in the US was reported to be much higher than in Canada. In a recent US study, Le TK et al. (106) that used a large population-based insurer database, the average incremental direct cost for an OA patient was reported to be $13,381. In a 2008 study by White et al. (113), the medical and drug direct costs were $10,828 and $3,702, respectively, while work productivity losses, including those 45 related to absenteeism and short/long-term disability due to OA were $5,795 (all in 2010 $CAD dollars). European estimates for the average cost of OA were reported be lower than Canadian estimates. For example, Loza et al. (107) reported $1552 for the direct cost of OA in Spain, while in Italy, Leardini et al. (118) estimated those costs to be $1178. In a 2003 French study by Le Pen et al. (117), the average direct cost of OA was estimated to be $379 including physicians, drugs and hospitalization cost using a population- based administrative database (all in 2010 $CAD). Using survey data for about 1000 patients across Spain, Loza et al. (107) reported that the direct costs represented 86% of the total cost. Total Cost In one of the first reviews of OA cost studies in 1997, March et al. (97) looked at the global cost of osteoarthritis and found the cost of osteoarthritis in the USA, Canada, the UK, France, and Australia to account for between1 and 2.5% of the gross domestic product (GDP) for those countries (11). Using clinical data, Yelin et al. (13) estimated that the total cost of OA was around 2.65 times the total cost of RA in the US (in 1999 $US). Within the arthritis literature, the majority of COI studies focused on either the overall cost of arthritis or the RA cost burden, even in recent decades (after 2000) (103). In the US, among the very few recent studies that include both out-of-pocket and insurer costs, Kotlarz et al. (60) estimated that the annual aggregate OA-attributable medical expenditures, using 1996–2005 Medical Expenditure Panel Survey data on US adults with health insurance, added up to a total annual cost of $234.28 billion, including both out-of-pocket and insurer’s cost of OA (in 2010 $CAD) (60). As shown in Table 2.4, they 46 used a sample of patients with OA and performed a regression-based analysis, while controlling for the medical conditions, socioeconomic status, and region of the patients for whom the cost incurred. The results of this study have been criticized to be overestimating the overall cost burden of OA (122). Most of the COI studies on OA in Canada were conducted prior to 2006. According to the systematic review, only two COI studies for OA, Anis et al. (94) and Tarride et al. (105), were performed in Canada since the 2006 review by Xie et al. (101), and only one estimated the total cost of OA (94) (Table 2.5). Using aggregate data on cost expenditures, Anis et al. (94) employed a top-down method to arrive at their estimate of around $1.2 billion in 2010 $CAD for the total direct cost of OA (without out- of-pocket costs); however, they calculated the cost related to obesity, for which OA is only included as one component. In older studies, the total direct cost for arthritis in Canada in 1994 was reported to be between $2.1 and $2.6 billion dollars per year, in 2010 $CAD (123). Most recent national cost burden reports in Canada used data from studies prior to 2006 (9, 71, 95) and therefore, I did not include them in the systematic review (Table 2.5). For example, the 2010 Public Health Agency of Canada (PHAC) (9) used estimates from the Economic Burden of Illness in Canada performed in 1998 (95). They also only reported the overall cost of arthritis and not the specific cost due to OA. According to PHAC report in 2010 (9), the overall cost to patients due to arthritis was $7.7 billion in 2008 $CAD, with only $2.2 billion of that due to direct costs. In Europe, the total cost of OA was much lower than that in Canada and the US. While the total direct cost of OA in Spain was reported to be $5.28 billion in 2010 $CAD (107), the COART study (117) performed in France estimated the direct cost of OA at 47 1.89 billion in (2010 $CAD), which is 1.7% of the overall expenses of the French healthcare system. Comparing their results to another French study, Le Pen et al. (117) showed that the overall direct medical cost due to OA increased by 156% in a 10-year period. As shown in Table 2.6, all of the survey-based studies included the out-of- pocket cost, while Le Pen et al. (117) that used a population-based administrative database, only reported cost components paid by the healthcare system. Cost Components Several studies have reported hospitalization costs as the most expensive category among all other cost categories of the total direct medical cost of OA (58). While Lanes et al. (119) reported that nearly half (46%) of the total direct medical cost was due to hospitalization and surgery costs for US patients with OA using one year administrative data in a managed care setting, Leardini et al. (118) reported this proportion to be 42% among a cohort of 254 Italian OA patients followed for one year. In a recent study of Le TK et al. (106) using matched cohort of around 200,00 patients and non-OA individuals, average incremental inpatient costs was $6,007, an almost 45% of the average cost of OA which was $13,381 (in 2010 $CAD). In most studies, the physician’s visit and drug costs were reported as the next most important cost categories after hospitalization (106-108, 118, 119). At the same time, some studies report smaller portion for hospitalization costs compared to physician costs. For example, Tarride et al. (2010) used a matched cohort study with around 1,400 patients and reported physician cost ($604) and hospitalization cost ($548) to be the most significant cost components, respectively- out of the average cost of $1200 (105). In another study, using a claim database of privately insured OA 48 patients in the US, White et al. (113) reported that 38% of the average cost direct cost of OA was due to physician’s visits, 33% due to inpatient while 29% was due to drugs. The difference can be, in part, due to the composition of patients included in these studies; in both of the above studies (105, 113), very few patients during or after the TJR surgery were included in the OA cohort (i.e., selection bias). Most of the studies in this review that used patient surveys reported the out-of- pocket costs associated with OA (114, 115, 118). Among these studies, Leardini et al. (118) shows that nearly 40% of the total direct cost of OA is due to non-medical costs, including transportation, cost of auxiliary device (aids), and paid temporary caregivers. Using a survey of OA patients in Canada, Maetzel et al. (114) found that the out-of- pocket costs including community services (meals on wheels, transportation), in addition to cost medical aids for OA patients were around 14% of the average direct cost ($266 out of $1976 in 1998 $US), while this proportion was 9% for patients with RA and 1% for patients with high blood pressure. On the other hand, very few studies utilizing database such as population-based administrative data or insurer claim registries reported the out-of-pocket costs (60,105). In a recent US study, Kotlarz et al. (60) reported that around 20% of the total cost of OA in the US was due to patients’ out- of-pocket costs, using a national health database in 2007. Knee and Hip Total Joint Replacement (TJR) Cost OA is the most frequent cause of hip and knee replacements since 83% and 95% of the total number of primary hip and knee total joint replacements (TJRs) in Canada in 2011 was due to OA as reported by Canadian Joint Replacement Registry (CJRR) (124). Knee and hip TJRs have shown to be cost effective or even cost-saving in 49 several healthcare settings including those in US, Australia, and Canada (49, 50,125). Several cost-utility analyses in Canada rank this procedure among the top cost-effective interventions with a cost per QALY of $3268 in 2010 $CAD, averaged over the lifetime of the prosthesis (125). According to CJRR data from 1981/1982 to 2009/2010, the average length of stay (LOS) for a primary hip and knee TJR decreased from an average of 25 days to an average of 6 days (126). Almost the same rate of decrease was observed for hip and knee revision surgeries (126). As a result, hospitals were able to increase the volume of surgeries they performed, which, in turn, resulted in the reduction in cost of surgeries; for example, from 1988 to 1996, the cost for a hip procedure decreased from $10,000 to $8,500 in Canada (103). At the same time, however, the costs of hip and knee prostheses increased from $800 to $1384 (all in 2003 $CAD) (103). I will discuss the details of calculating the cost of hip and knee TJRs and their projections in time in Chapters 3 and 4. Out-of-Pocket Cost As discussed in previous sections, out-of-pocket cost of OA is one of the main cost components of the direct cost of OA that has often been neglected in COI studies (61). The out-of-pocket costs for OA patients include: 1) the cost of purchasing equipment and medical aids; 2) renovations to the home, including the purchase of tub benches or bars and/or raised toilet seats or the movement of furniture to reduce the risks of patient injury; 3) over-the-counter medications; 4) paid caregivers; 5) alternative care professional visits that are not covered by patients’ insurance; and 6) community services such as meals on wheels and transportation fees (Figure 2.2) (61,114, 115). As mentioned earlier, around 20% of the total cost of OA in the US (60) and around 50 14% of the average cost of OA in Canada (114) was reported to be due to patients’ out- of-pocket cost. According to Kotlarz et al. (60), 28% of the average cost of OA for women and 17% for men were due to out-of-pocket expenditures. In an older study, not included in the review, for out-of-pocket cost for OA patients in Australia, Lapsley et al. (127) found that women with OA incur higher out-of-pocket costs than men with the same level of OA severity. The range for annual out-of-pocket costs in their study was found to be from 0-$120 in1994 Australian dollars (127). Cost of OA Among Specific Sub-Populations Several recent studies have reported the OA costs among the sub-populations suffering from this disease. These include studies comparing cost of OA among women versus men (60), patients waiting for TJRs (128,129) versus those after TJRs, as well as employed patients versus non-employed ones (110-114). In terms of total costs, women with OA cost much more than men, in part due to higher number of women with OA. Kotlarz et al. (60) reported that women cost almost $75.8 billion higher than that of men in 2007 in US including out-of-pocket and insurer-paid cost (in 2010 $CAD). The severity of OA is one of the major factors that affect the patients cost profile. Only two studies reported the cost by OA stages. Dibonaventura et al.’s (112) reported the cost according to a self-reported classification for OA severity for mild, moderate or severe OA (112). In this study, the authors studied the cost of OA among only employed members of the US population, using bottom-up, incremental cost methodology and the matched cohort approach. After consulting a national health and wellness survey with 4,876 self-reported OA cases and 34,896 non-OA comparator subjects, they estimated that the total annual incremental cost per worker was $2,223 for mild self-reported OA, 51 $8,390 for moderate OA, and $17,379 for severe OA. In another study, Gupta et al. (115), the total costs attributable to OA in the hip and knee were three-times higher in patients with high versus low disability ([WOMAC] score >55 vs. <15). In addition to the high cost burden both for patients and the healthcare system, waiting for a hip/knee TJR can cause a considerable reduction in one’s quality of life (59, 85). Therefore, many studies have estimated OA costs for patients waiting for joint replacement surgery or those with post-surgery status (128,129). Hawker et al. (128) reported the mean annual cost for patients after the TJR surgery to be $287 less than their matched controls with OA prior to surgery - the average cost reported included out- of-pocket costs for patients. 2.3.5 Indirect Cost of OA As defined by Phillips et al. (130), indirect costs, from the societal point of view, are those that “relate to ‘losses’ to society incurred as a result of the impact of the disease, [including those for] injury, disability, premature death and the (side effects of) treatments [that] prevent people from engaging in ‘normal’ daily activities, such as work, domestic responsibilities, and social and leisure engagements” (130). In fact, most indirect costs are productivity costs that are incurred outside the healthcare sector and relate to losses in production. Some of the indirect cost categories of OA include OA-related productivity losses due to missed work days (absenteeism), at-work productivity losses (presenteeism), informal caregiver’s productivity losses, i.e., care provider either giving up work or sacrificing leisure time to provide care to the patient that would otherwise have been provided by formal care agencies, work transition productivity losses, i.e., work 52 transition due to illness includes a change in one‘s type of job, a change in the number of hours on the job, early retirement due to illness, and premature mortality. While there seems to be some debate in healthcare economic literature over whether or not to include the leisure time and productivity loss of informal caregiver’s as part of the indirect cost (115), the majority of studies only include absenteeism, presenteeism, early retirement and premature mortality; recently, however, the informal caregiver cost category has been discussed as one of the major cost components of indirect cost among musculoskeletal diseases and OA (61,115,131). Informal caregivers’ losses in productivity constitutes a major portion of the indirect cost of OA, i.e., approximately 40% of the indirect cost of OA excluding work transitions and early retirement according to Gupta et al. (115). Early retirement also constitute a significant part of indirect cost of OA. In a retrospective cohort study by Gignac et al. (132), it was found that approximately 36.9% of the working population had left the labor force as a result of having OA during the 4-year study period. Most COI studies report the average (i.e., per capita) indirect cost of OA to be less than or very close to its direct cost (105,106,107,114). Leardini et al. (118) reported an average indirect cost of $1560 per patient per year, while the average direct cost was $1178. In a study in Spain, Loza et al. (107) calculated the average indirect cost to be $251 by including only absenteeism and informal caregiver costs, which were much lower than the average direct medical cost of $1552 (all in 2010 $CAD). In a Canadian study by Maetzel et al. (114), the average indirect cost of OA was $1690, almost a third of the average direct cost of $3797 (all in 2010 $CAD). In their study, indirect costs were associated with productivity losses of patients, including absenteeism and 53 presenteeism, in addition to the informal caregiver costs (114). They used a patient’s questionnaire for utilization data, e.g., number of hours lost, which they then transferred into monetary terms, using standardized unit costs. Only one study in Canada (115) included leisure time productivity loss of patients and their informal caregiver and reported a much higher average for indirect cost of OA compared to other studies. Gupta et al. (115) reported $11,993 for the average indirect cost including absenteeism, presenteeism, informal caregivers as well as caregivers’ productivity and leisure time losses. The average direct cost of OA was estimated to be $2123 (all in 2010 $CAD) (115). In their study, they have used a cross-sectional sample of around 800 OA patients from Ontario, Canada and estimated the cost according to the questionnaire by performing an incremental COI methodology using a regression- based approach to control for SES, age and other characteristics of patients (115). The total indirect cost of OA, including absenteeism and presenteeism, were reported to be between $2-8 billion in 1997 $US by March et al. (97). In another study, Kotlarz et al. (133) estimated only the absenteeism cost of OA (US) to be $13 billion. According to Loza et al. (107), in Spain, the total indirect cost of OA including the cost associated with informal caregivers was $785 million (all in 2010 $CAD). 2.4 Discussion In this Chapter, I have provided a background on OA risk factors, its prevalence and incidence rates, in addition to different types and methods of COI studies. Next, I performed a systematic review of previous COI studies reporting the national burden of OA including either direct, indirect, or both types of costs. I first categorized these studies according to the COI types and methods for studies prior to 2006 and those 54 after 2006 (Tables 2.4 and 2.5). According to these results, recent studies performed after 2006 are utilizing more bottom-up types and incremental methodology with large population-based databases rather than top-down types and total cost methodology with survey-based design. I have provided the summary for the outcomes of the COI studies performed after 2000 that reported the direct cost of OA in Table 2.5. According to the review, average direct cost of OA across different countries range from $379 in France (117) to $14,530 in US (113) (all in 2010 $CAD). The average direct cost of OA was higher in US, Hong Kong and Canada compared to European countries. The direct cost of OA within each country had less variation. In Canada, direct cost of OA was reported to be between $1,200 (105) and $3,797 (114) (all in 2010 $CAD). The variation of COI studies within each country’s healthcare system was due to different COI types and methodologies, in addition to different cost components included in the study. Highest share of the average and total cost of OA was reported to be due to hospitalization, physician and drugs cost in most of the studies. None of the traditional COI studies were able to integrate different types of data sources (62). As result, none of the studies that used large administrative databases reported out-of-pocket costs. At the same time, COI studies using survey data were not representative of all sub-populations with OA. Very few studies reported the average and total cost by OA states and recognized the heterogeneity among the OA patients (112, 115). 55 Chapter 3: Application of a Simulation-Based Cost-of-Illness Study to Estimate Average Direct Cost of OA From 2003 to 2010 3.1 Introduction The average direct cost of OA reported in COI studies is subject to substantial variation even within one country (101). For example, in three recent Canadian studies, the average direct cost of OA was reported to be $1,200 (105), $2,123 (114) and $3,797 (115), all in 2010 $CAD. The variation in direct costs of OA across countries is much higher, i.e., from $379 in France (117) to $2,123 in Canada (114), $7,873 in Hong Kong (121) and $14,530 in US (113) (all in 2010 $CAD). At the same time, while most COI studies report the average indirect cost of OA to be less than or very close to the direct cost (105,106,107,114), a few studies report higher indirect costs (115,118). For example, Gupta et al. (115) reported an individual’s indirect cost to be $11,993 and direct cost to be $2,123 in 2010 $CAD. The variation within traditional COI studies is in part due to different data sources, COI methodologies, and heterogeneity within the OA population (57, 62). In this Chapter, I aim to integrate the results of a bottom-up COI study into a microsimulation model of OA to be able to integrate different data sources and account for the heterogeneity across the OA population. Although it has been noted that the total cost burden of OA patients is significant compare to that of other chronic diseases (95), it is not clear how much of it is due to OA’s high prevalence and how much due to its average cost (59). In fact, while the number of people with OA has increased significantly over the past years and affected the total cost of OA, the rate of utilization of healthcare resources, their prices and other 56 cost drivers have caused the average cost to increase as well. On the other hand, little information is available about the trend for average cost of OA in the past decade in Canada, its distribution across different cost components and among different sub- populations. That’s why it is important to analyze the average cost of OA and its drivers separate than the total cost. The goal of this Chapter is to estimate the average direct cost of OA in the past decade from 2003 to 2010. The projection of cost components into the future and estimating the total cost of OA in Canada from 2010-2031 will be discussed in Chapter 4. PDMS models are novel tools that have tremendously assisted public health researchers by integrating different methods and data sources in addition to provide unique test bed to perform what-if scenarios (32). Although several COI studies have reported the total and average direct costs of OA at one point in time (94,105-120), none has reported and analyzed the historical trend for the direct cost of OA, including both health care system expenditures and out-of-pocket costs of OA. PDMS models are unique tools that can be used to address this gap in knowledge with regard to COI studies. As a special type of Individual-level simulation models, PDMS models would allow us to impute the missing cost data in the past in the same way as in a retrospective cohort study (14). By integrating cost and utilization estimates for different cost components from various data sources within an individual-level simulation model, cost associated with each individual is calculated while individuals inside the model age and disease-related events occur. Additionally, PDMS models would allow us to perform scenario analyses to investigate the effects of different variables on the population-level 57 outcomes (32). Details of advantages in use of PDMS models for chronic diseases such as OA have been discussed in Chapter 1. In this Chapter, I perform an incidence-based bottom-up COI study integrated with a microsimulation model to analyze recent trends in the average cost of OA in the past decade, between 2003 and 2010. In the bottom-up COI study, I used a British Columbia (BC) population-based database (84), other data sources, and literature estimates to calculate the per-patient cost in an average year for OA patients. Finally, I used the POHEM-OA model to simulate the life and health histories of the entire adult Canadian population including OA patients, from 2003 to 2010. In conjunction with the result of the COI study, I calculated the average per-patient direct cost associated with having OA over the period of 2003 to 2010. In addition, I performed a what-if scenario to report the effect of major cost drivers including increase in life expectancy, price inflation, and increase in number of hip and knee TJR surgeries on the increase of the average cost of OA during the study period. 3.2 Methods The direct cost of OA in Canada contains both publicly funded cost categories (physician, outpatient, and inpatient costs) in addition to patient expenditures (out-of- pocket costs) together with other categories that are partially funded by the healthcare system, including home care and rehabilitation, alternative care, and prescription drugs (114, 115,128). In this study, I assumed all alternative care cost categories including physiotherapists’ and other alternative care professionals’ OA-related costs are part of the out-of-pocket cost. Rehabilitation costs were divided into two parts for the out-of- pocket and healthcare system cost. 58 The OA cost categories can further be divided into cost components: (1) physician visits and other outpatient procedures, (2) hip and knee TJR hospitalization, other in-patient procedures, (3) rehabilitation costs (alternative care such as physiotherapy both home care and at the hospital) (3) drugs (prescription-based and over-the-counter) and (4) other out-of-pocket costs of OA (formal caregiver, community costs including meals on wheels, transportation, home remodeling and medical aids [114, 115]). List of all the cost categories and their associated data sources/references are shown in Table 3.1. As shown In Table 3.1, the input parameters associated with each cost component in this study were estimated using different approaches; the average per patient-year cost for hospital procedures other than hip and knee TJR surgeries, physician’s visit, outpatient services, and drug costs (both prescription and over-the- counter) were estimated according to a state-based approach using the bottom-up micro-costing approach in the prior research study (the COAST study). On the other hand, according to an event-based approach, unit costs for each event were estimated for hip and knee TJR surgeries, formal caregiver, alternative care (physiotherapist, chiropractic and other alternative care professionals), rehabilitation, and the side effects of OA drugs. All costs in this Chapter are in 2010 Canadian dollars (2010 $CAD) unless mentioned otherwise. The costs were translated into 2010 $CAD using overall consumer price index according to Statistics Canada rates (134). In the next section, I describe the COI study used to estimate the input parameters associated with each cost component that were implemented in the direct cost module of the POHEM-OA. 59 Table 3-1. Direct cost categories of OA implemented in the POHEM-OA direct cost module Cost category Cost component Parameter Data sources (year) 1. Hospitalization and Hip/knee TJR surgery Hazard rates of surgery by age PDBC1 (1987-2003) inpatient costs and sex Average unit cost of surgery CIHI-HP2 (2010) Number of surgeries per year and CJRR3 (1996-2010) future trend in surgery rates Other in-patient procedures Average patient-year unit cost PDBC (2003) for OA (non-hip/knee TJR Average procedure costs CIHI-HP (2010) Surgery) Number of procedures per year St. Paul’s hospital4 (2006) 2. Physician and Physician visits and medical Average patient-year unit costs PDBC (2003) outpatient care costs tests (X-ray, MRI, blood work and other outpatient care) 3. Drugs Prescription drugs Average patient-year unit costs PDBC (2003) Over-the-counter drugs Average patient-year unit costs NPHS5 (2006) 4. Rehabilitation and Healthcare system paid Probability for home/hospital Oldmeadow et al. home care costs (after rehabilitation and home care discharge strategy (2000) [135] hip/knee surgery)*** costs Unit costs of each discharge Coyte et al. (2001) strategy [136] Out-of-pocket-cost during Average patient-year unit costs8 March et al. (2002) rehab [137] Mitchell et al. (2005) [138] 5. Formal caregiver Formal caregiver, Probability and unit costs Gupta et al. (2005) and other costs (before transportation and community [115] hip/knee surgery) costs (medical aids and home remolding) 6. Alternative care Physical therapy, Odds ratio for MOH-OA6 (2007) costs Chiropractic’s, acupuncture visits and other complementary care Unit costs per MSP7 (2003) visit 7. Side effect of drugs Cardiovascular disease Utilization of drug use PDBC (2003) (CVD’s) Stroke Probability of side effects for each Sayre et al. (2010) Gastrointestinal (GI) type of drug [139] Dyspepsia Life time cost for CVD and stroke8 Birnbaum (2003) [140],Taylor et al. ((1999) [141] Incidence-based cost for GI and Rahme et al. (2001) dyspepsia8 [142], Moayyedi et al. (2002) [143] 60 Table 3.1 notes: 1 PDBC: Population data BC from 1987-2003 (84); 2 Canadian institute of health information hospitalizations cost data in 2010 (144); 3 CJRR: Canadian Joint Replacement Registry data for number of surgeries from 1996-2010 (124); 4 St. Paul’s hospital unit cost model in 2006 (145); 5 NPHS: National Population Health Survey 2006 (146); 6 MOH-OA: British Columbia Ministry of Health’s OA survey (147); 7 MSP fees in 2003: Medical Services Plan fees (148). 8The purchasing power parity (PPP) was used to transfer the cost from other countries to $CAD (149). 3.2.1 Data Sources The main data source for estimating the input parameters of this study was the Population Data BC (PDBC) (84), a pan-provincial population health data service that provides access to the health care services’ databases for all BC residents registered in the province’s publicly funded universal insurance program. The following components of health resource utilization were retrieved from the PDBC database: 1) The Discharge Abstracts Database (DAD) for hospital separations (150); 2) Medical Services Plan (MSP), which includes records of physician visits at any facility (office, hospital, home, nursing home, and continued care facility but not emergency departments visit), in addition to laboratory tests, ambulatory care, and outpatient services (84); 3) the provincial PharmaNet system that contains the quantity and day’s supply for every prescription drug dispensed to members of the provincial health insurance plan (84). All hospital admissions, as well as office visits covered by DAD and MSP for the fiscal years 1986 to 2003 were obtained for this study5. 5 Fiscal year is from April 1 of a given year to March 31 of the following year - referred to as ‘year’ in this thesis. 61 As shown in Table 3.1, the British Columbia Ministry of Health’s OA (MOH-OA) survey (147) was also used in this study to estimate the input parameters of alternative care categories, including physiotherapy, chiropractic, massage therapy, and others (Section 3.3.5). Trends for the number of knee and hip TJRs and their cost estimates were derived from the Canadian Institute of Health Information (CIHI) CJRR data in 2011 (124) and hospitalization cost data from the CIHI hospitalization cost database (144), respectively. In addition, I used St. Paul’s Hospital Cost Model in 2006 (145), a BC hospital, and CIHI hospitalization cost database (144) to calculate hospitalization cost for procedures other than hip and knee TJR. Price indexes for the average hospitalization cost reported by Cost per Weighted Case (CPWC) across Canada for the years 2003-2010, along with those for drugs, physicians and out-of-pocket were taken from the CIHI Price Index Report for Hospitals (151), Drugs (152), Physicians and Orthopedic surgeons (153,154) and overall economy Consumer Price Index (CPI) report from Statistics Canada in 2012 (134). I discuss the inflation rates in Section 3.5.2. 3.2.2 Patient Characteristics and OA States The severity of OA varies according to different stages of the disease; patients at early stages of OA are often prescribed medications to alleviate their symptoms in addition to exercise or other types of behavioral interventions. On the other hand, those at later stages of OA who experience severe pain are often referred for hospital procedures such as total joint replacement surgery. In the POHEM-OA cost model, different variables representing patient characteristics and OA states have been used to capture the heterogeneity among OA populations and account for the variation in their resource utilization. These include: age categories, sex, OA state, and time in each 62 state. OA state is defined by the OA-related events with relation to the patient’s flow in the healthcare system: State 1) OA diagnosis; State 2) orthopedic surgeon (OS) visit; State 3) hip and knee TJR surgery; State 4) hip/knee TJR revision surgery. From the perspective of health services used by OA patients, these variables are the most important cost drivers for direct cost of OA (cost drivers are factors that can change the cost of an activity) (61). Use of services might change according to the time spent in each OA state. To reflect these in the model, the time periods were defined according to preliminary descriptive analyses of data in the COAST study (unpublished data). As a result, the time in each OA state is categorized as 0-2 years, 2-5 years and >5 years. The above-mentioned patient characteristics are the key drivers of the direct cost of OA that were implemented in the simulation model as described in Section 3.4. The study population for performing the COI study was generated from BC population data and stratified according to the patient characteristics and OA states as discussed below. 3.2.3 Study Population The study population was generated from PDBC (84) from 1986/7 to 2003 for the COI study to estimate the input parameters associated with OA cost categories paid by the healthcare system including physician, prescription-based drugs, and hospitalization procedures other than hip\knee TJR surgery. As mentioned in Section 3.2, although prescription-based drugs are not paid fully by the healthcare system I assumed all were part of the healthcare system paid cost categories of OA. For the drug cost calculation, a stratified random sample of 100,000 individuals was selected from the PharmaNet database in the prior COAST study. The sample was stratified according to OA state (pre OA, OA diagnosis, 1st orthopedic surgeon visit, 1st 63 hip/knee TJR surgery, 1st hip/knee TJR revision surgery), sex and age (in 10 year age bands starting at 0 and ending at 90+), with a greater proportion in the 50 to 90 age groups. All the revision cases (total of 5,146 cases) in the cohort were selected from 1986/87 to Dec 31st, 2003 who were alive on Dec 31st 2003. Further, 20,000 primary surgery cases were selected from the data between 1986/87 and Dec 31st, 2003, who were alive on Dec 31st 2003, and who were not selected as revision cases. From the cases from 1996/97 to Dec 31st, 2003, who were alive on Dec 31st 2003, 20,000 patients were selected who had made OS visits and who were not selected as primary and revision cases, for the sample (first OS visit with a 5-year run-in period). Another 27,427 OA cases from the period of 1996/97 to Dec 31st, 2003, who were still alive on Dec 31st 2003, and who were not selected previously, were selected from the data. A sample of 27,427 non-OA controls who were alive on Dec 31st 2003 was also selected and used for the calculation of the stratum-specific differences in drug costs between persons with and without OA. Since appropriate weights were assigned to each of the selected individuals, the sum of the weights equaled the BC population. Details of the weights for the sampling scheme in the COAST study are shown in Table A3, Appendix A. I used the study population as discussed above to calculate the per patient-year cost according to patients characteristics by dividing the total cost for all patients within each strata (defined by their characteristics) by the total number of years that patients spent in that strata. Tables A7-A10 in Appendix A presents the tables associated with the final parameters calculated for the per patient-year cost of hospitalization costs, 64 drugs costs and physicians cost according to patients characteristics. Below I discuss the details of calculating the input parameters for each cost component. 3.3 Analysis of Input Parameters for Cost Components Based on standard cost categories of OA-related direct costs (61), six main cost categories are included in the analysis of input parameters in this study: hospitalization, physician’s and other outpatient services, drugs, rehabilitation and home care, alternative care, out-of-pocket and formal caregiver costs (Table 3.1). 3.3.1 Hospitalization Hospitalization cost categories associated with OA include those for hip/knee total joint replacement (TJR) surgery, and other hospital procedures such as shoulder TJR, arthroplasty other than TJR, and other OA-related short stay procedures. In the following sections I discuss cost estimation for each of these categories. Hip/knee Total Joint Replacement Surgeries The cost for a hip/knee TJR was implemented in the model as an event-based cost (Section 3.2). To this end, I needed to estimate the costs of both a knee/hip TJR and the number of TJR events in the simulation model (i.e., time from diagnosis to surgery is implemented in POHEM-OA, based on longitudinal data from PDBC (1987- 2003) (84). Number of hip/knee TJRs (calibration): Between 2002/2003 and 2009/2010, the number of primary TJR surgeries in Canada increased by 75% (126). In the model, I used the number of hip/knee surgeries for each age/sex group as reported in the CJRR for the fiscal years 1994/95 through 2009/10 (126). CJRR captures approximately 35% of the hip and 28% of the total knee replacements performed in Canada (126). 65 Additionally, 82% and 95% of the TJRs for hips and knees, respectively, were due to OA (132). The number of hip/knee TJR surgeries in POHEM was previously implemented using Kaplan-Meyer curves estimated from 12 years of longitudinal data between1991- 2003 from PDBC, as a function of age, sex, and the current OA status of a patient and time in each state (1). However, the projected number of TJR surgeries from the model for the years 2005-2012 was lower than the actual number of surgeries in Canada during that period. The faster-than-projected growth in the number of TJRs may have been due to external factors not included in the POHEM-OA, such as government policies aimed at reducing surgical wait times in Canada (52). I used the number of surgeries reported in the CJRR for the years 2003 to 2010 to calibrate the model by adjusting the multiplier for the hazard ratio of hip/knee TJR surgery events until the results from the model agreed with the CJRR data (124). Cost of a hip/knee TJR: In addition to the number of TJR surgeries, I also estimated the age-specific unit cost of a TJR hip and knee surgery (separately for primary TJR and revision). The overall cost of a TJR equals the weighted average of the three types of TJR performed on OA patients, i.e., unilateral/bilateral, hip/knee, and with/without infection. Below I discuss how each procedure cost was calculated according to different components of the TJR surgery. Further details regarding cost calculation for each procedure from 2003-2012 are provided in Appendix A3. The cost of a hip/knee TJR has two major components: 1) hospital stay and procedure cost and 2) orthopedic surgeon cost. The first component includes inpatient, outpatient and operating room nursing, as well as procedure costs during and post- 66 operation (medical imaging, surgery, and implant costs and staff cost during surgery). For each of the three aforementioned TJR procedure and age categories, the average cost of a procedure was calculated according to length of hospital stay, multiplied by the per diem cost rates, which reflect the per-day resource intensity for a procedure. The final results need to be multiplied by the cost per weighted case (CPWC) for each hospital or healthcare setting. To reflect the overall Canadian rates, I used the average Canadian CPWS from the CIHI hospitalization cost database for each year from 2003- 2012 (144). The cost of an orthopedic surgeon visit was not reported in the CIHI database and therefore, I used the orthopedic surgeon fee based on data from the MSP in 2003 (148). The results of expected length of stay multiplied by per diem rate for resource intensity and further multiplied by average CPWC would give us the average cost of each procedure. However, to calculate the overall weighted average with respect to the volume of each procedure, I also included the DAD volume for hip/knee TJRs by age and procedures (144). Details of the DAD volumes and calculations are provided in Table A4, Appendix A. There was a change in the cost of TJR surgery over time, due to both changes in the average length of stay (which has decreased within the last decade) and price inflation for CPWC reported in the CIHI Hospital Price Index Report (151). Therefore, at each year from 2003 to 2012, I used different length of stay according to those reported in CJRR (124). Additionally, to implement the changes in the hospitalization price over the study period, I used the average Canadian CPWC for each year from 2003-2010 as reported by the hospital price index from the CIHI report (151). Finally, to take into 67 account the inflation rate for the cost for an orthopedic surgeon, I used the average inflation rate for orthopedic surgeons during 2003 to 2010 from CIHI Physicians Fee report, which was on average 2.1% (154). I did not include the cost of drugs during hospitalization, rehabilitation cost or home care cost as part of the TJR costs. These costs are calculated separately and discussed in Sections 3.3.3 for drug costs and 3.3.4 for rehabilitation costs. The resultant primary and revision knee/hip TJR surgery costs for each age category between 2003 and 2010 are shown in Table A5, Appendix A. OA-related Hospital Procedures (Other Than Hip/Knee TJR) To estimate the average per patient-year cost of non-hip/knee TJRs, I accessed DAD linked to the study population in PDBS as discussed in Section 3.2.3. Up to 25 diagnosis codes and a maximum of 12 procedure codes, along with the type of procedure (principal, primary, secondary or emergency), were found for each admission date on hospital discharge summaries. The list of all non-hip/knee TJR procedures is provided in Table A6, Appendix A. The hospitalization cost associated with non- hip/knee TJRs has two components: the hospital stay cost and the procedure cost. For the hospital stay cost, I first calculated the aggregate number of days of hospital stay for all the non-hip/knee TJRs for all OA patients in the study population (calculated from DAD) and then categorized them by patients’ characteristics as defined in Section 3.2.2. Then the number of hospital stays was multiplied by the unit cost of each day of the patient’s stay at the hospital, which I garnered from St. Paul’s Hospital’s Cost Model (145). For the procedure cost, the average procedure cost by age and sex was calculated from the CIHI hospitalization cost database (144), by matching the non- hip/knee TJRs of International Classification of Disease, 9th Revision (ICD-9) and Case 68 Mix Groups (CMG+), i.e., lists of inpatient procedures provided by the CIHI hospitalization cost database (144) to calculate the hospitalization costs as shown in Table A4, Appendix A. Finally, by adding the procedure and hospitalization stay costs for the OA sample and dividing that by the total number of person-years, I calculated the average person-year cost of hospitalization for non-hip/knee TJRs based on patient characteristics. The resultant input parameters representing the per patient-year costs for all OA-related hospitalization procedures (other than hip and knee TJR surgery) can be seen in Table A7, Appendix A. 3.3.2 Physician Visits and Other Outpatient Services The MSP database contains information regarding the date and type of a physician visit or ambulatory care service, the physician’s specialty code, the patient’s birth and (if applicable) death dates, gender, postal code, area-based socioeconomic status6, and MSP registration start and exit dates for approximately 4 million residents of BC (84). Patients were identified using the ICD-9 code of 715.x. Physician visit were classified according to specialty into 3 types: 1) general practitioner, 2) orthopedic surgeon (OS) and 3) other specialist. It should be noted that the ICD code identifies the diagnosis for billing purposes and that only one code, for one condition, is allowed in MSP for a given visit. In addition, the MSP database includes information on ambulatory services such as x-rays, MRI, blood work, other diagnostic tests and other types of services. All physician visits for OA and OA-related inpatient procedures, as well as those related to hip/knee TJR surgeries, were also identified. Physicians’ fees for the 6 Socioeconomic status is not part of MSP. It is added to the database as a derived variable based on postal code. 69 visits and the ambulatory procedures fees were obtained from MSP data for the calendar year 2003, and then I used the physician wage inflation rate reported in (153), for each year from 2003 to 2012 (orthopedic surgeon fees were calculated differently according to data from (154) as previously discussed in Section 3.3.1). Final results for physician cost category according to patients’ characteristics and all types of physician services are listed in Table A11 in Appendix A. 3.3.3 Drugs To calculate cost parameters related to the prescription drugs, the costs of prescription drugs were included in the prior COAST study during the fiscal year of 2003/04 for all OA patients in the study population linked to BC’s provincial PharmaNet database as described in Section 3.2.3. All OA-relate drugs were categorized into four types: acetaminophen, NSAIDs, coxibs and opiods. There was total of around 30 different sub-categories for all of the drug types as listed in Table A10 in Appendix A. First, I calculated the average drug cost plus the fee claim for each drug type. Next, for each drug type, the total costs were calculated over all OA patients according to patients’ characteristics (i.e., age, sex, OA state and time in each state as discussed in Section 3.2.2). The resulting total cost was then divided by the total person-years in each stratum of patients’ characteristics to calculate the average per person-year cost. As the same types of drugs were also used for diseases other than OA, I calculated the costs for the same drug categories among the non-OA individuals from the random study sample (Section 3.2.2) and deducted it from the OA-patients’ costs within the strata defined by patients’ characteristics. The difference was interpreted as the cost of prescription drugs utilized by OA patients to reduce OA-related symptoms. The final 70 results for per patient-year costs of prescription drugs are shown in Table A8 in Appendix A. Over-the-Counter Drugs For the over-the-counter drug costs, I used the results of the study by Sayre et al. (139). This study included the drug categories that are available over-the-counter in Canada (acetaminophen, some NSAIDS and some opiods). It should be noted that while these drugs can be also prescribed, I excluded the prescription drugs in the over- the couther category. For over-the-counter drugs, costs were estimated using a weighted drug model from the National Population Health Survey (NPHS) as described in Sayre et al. (139). Data from the 2006 cycle of the NPHS (146) were used to calculate the proportion of the drug costs accounted for by over-the-counter drugs. The final results for per patient-year costs of over-the-counter drugs are shown in Table A9 in Appendix A. 3.3.4 Rehabilitation and Home Care I have also implemented the rehabilitation cost category after hip/knee TJR surgery in the model. In recent decades, the decrease in post-surgical LOS has resulted in an increase in rehabilitation procedures after discharge, at the patient’s home or a rehab facility (135). At the same time, recent studies reveal the benefits of early rehabilitation and patient mobilization (135,136). Discharge planning for a hip/knee TJR patient consists of post-acute care rehabilitation performed either in an inpatient post-acute care facility or at the patient’s home. To model rehab procedures, I assumed the following four possible targets for OA patients after hip/knee TJR that match the current clinical pathways at Canadian 71 hospitals (135): (1) a rehab facility stay (between 5-20 days) followed by home care paid by the healthcare system, (2) a rehab facility stay followed by self-care paid by patients (out-of-pocket), (3) discharged home with paid home care (4) discharged home without home care, i.e., self-care after the acute care (135,136). Patients can move to different types of rehab care centers, where they incur different costs. Figure 3.1 shows the discharge pathways for rehabilitation and homecare algorithm implemented in the POHEM-OA after the TJR surgery. Figure 3-1. Rehabilitation and home-care cost after surgery * Probability of each discharge destination is shown in Table 3.2; **cost of each discharge destination is consisted of out-of-pocket cost component (OOP) and healthcare system paid component (Care). I estimated the overall probability of a patient moving along each pathway after surgery, based on the average rate of discharge from the CIHI hospitalization cost data 72 (1994-2000) reported by Coyte et al. (136). The probabilities of discharge were then calculated for different age and sex categories based on odds ratios reported by Oldmeadow et al. (135) that were estimated using a scoring system for predicting the discharge destinations of patients after replacement surgery (Table 3.2). Details of the cost and resulting parameters implemented in POHEM-OA for rehabilitation costs are shown in Appendix A4. Table 3-2. Probability of discharge to rehab destinations after hip/knee TJR surgery* Homecare! Home Rehab Rehab Self-care** care!Home-care facility!Self-care facility!Home- care Male < 50 0.343 0.509 0.109 0.038 50 - 64] 0.324 0.481 0.145 0.051 [65 -74] 0.311 0.461 0.169 0.059 ≥75 0.298 0.442 0.193 0.067 Female < 50 0.275 0.409 0.234 0.082 [50,64] 0.259 0.384 0.264 0.092 [65,74] 0.249 0.370 0.282 0.099 ≥75 0.239 0.356 0.300 0.105 * Probabilities were calculated from literature estimates (Coyte et al. (136)) and (Oldmeadow et al. (135)); **for rehabilitation pathways, from a!b: ‘a’ is the destination within 1 month after the surgery, and ‘b’ is the second destination, within the next 3-12 months. Three main cost categories for rehab care include a) rehab facility costs, b) home care costs paid by the healthcare system and c) out-of-pocket costs paid by patients for rehabilitation and home care services (128). These were calculated using estimates 73 from the literature (128). Table 3.2 shows the out-of-pocket costs during rehab that include formal caregiver costs (out-of-pocket portion), transportation, equipment, medical aids, and community services, all borne by patients due to OA (128). As can be seen in Table 3.3, those who are discharged to rehab have higher out-of-pocket costs for every year after the first year ($160), as compared to those who discharged to home care ($80). The total out-of-pocket rehab cost then consists of these costs, in addition to the cost of physiotherapy after surgery. Details of these calculations are provided in Appendix A4. 74 Table 3-3. Out-of-pocket costs for rehabilitation cost category Discharged Type of Time period after TJR surgery destination surgery after surgery^ Every 2-3 4-6 7- 8 9-12 First month year months months months months after ** Home care Knee $192.7* $264.8 $54.45 $41.20 $29.43 $80 surgery Hip $129.5 $213.37 $35.32 $20.60 $23.54 $80 surgery Average $161.1 $239.1 $44.88 $30.90 $26.49 $80 Rehab Knee $338.45 $235.45 $88.29 $47.09 $7.36 $160 facility*** surgery Hip $117.72 $173.6 $79.46 $47.09 $75.05 $160 surgery $228.09 $204.5 $83.88 $47.09 $41.20 $160 Average ^ “Surgery” here refers to hip or knee TJR surgery; *All costs are in 2010 CAN$; Costs are taken from Hawker et al. (128) and translated into 2010 Canadian Dollars using CPI for the years between 2003 and 2010 (134); ** Every year after the first year (annual cost); *** All of the rehab discharge destination pathways can be divided into two major groups: those discharged to the rehab facility (hospital) after surgery and those discharged to home. 3.3.5 Out-of-Pocket Costs Patients’ out-of-pocket costs due to OA have been shown to be one of the major costs borne by patients that have increased during the past few decades (115). In this study, I have included four main categories of out-of-pocket costs: 1) formal caregiver, transportation and community costs before surgery; 2) out-of-pocket costs after surgery 75 (included in rehabilitation costs, discussed in Section 3.3.4); 3) alternative care cost categories; (4) over-the-counter drugs (Section 3.3.3). Formal Caregiver, Transportation and Community Costs Although very few studies have assessed formal caregiver cost, transportation, and community costs (including medical equipment, home remodeling, meals on wheels), these cost categories were shown to make up a significant portion of the total direct cost (115). I separated these cost categories into before and after a hip/knee TJR surgery. I included the out-of-pocket cost of a post-surgical period in the rehabilitation cost (Section 3.3.4). For OA patients’ out-of-pocket costs due to formal caregiver and community services, I used available literature estimates (115). To implement these costs into the simulation model, I took a two-step probabilistic approach: in the first step, based on the age and sex of the person being simulated, POHEM-OA simulates whether the individual incurs a non-zero cost or not. Table 3.4 shows the probabilities for a non-zero cost. In the second step, if a non-zero cost is incurred, then a mean cost is assigned to the patient. For the base case (without uncertainty analysis), I assigned a fixed cost for the services of a formal caregiver as reported by Gupta et al. (115) - results are shown in Appendix A4. To reflect he inflation in the out-of-pocket costs, I used the CPI for the overall economy as reported by Statistics Canada (134). The CPI’s for years 2003 to 2010 that was implemented in POHEM-OA are presented in Table B1, Appendix B. 76 Table 3-4. Probability of incurring a non-zero cost for the formal caregiver category* Age category Male (95%CI) Female (95%CI) <60 0.12 (0.04,0.19) 0.25 (0.16,0.33) 60-69 0.11 (0.06,0.18) 0.23 (0.16,0.29) 70-79 0.14(0.09,0.20) 0.26(0.19,0.33) 80-89 0.19(0.12,0.25) 0.36(0.30,0.41) ≥90 0.52(0.40,0.63) 0.72(0.61,0.79) * Formal care category is part of the out-of-pocket cost category and includes paid help, transportation, equipment and medical aids, and community services, all borne by patients due to OA (before the hip/knee TJR surgery). I used the odds ratios reported in Gupta et al. (115) and calculated the probabilities based on the baseline rate in their study population. I adjusted their baseline to the Canadian population used in the POHEM-OA in 2001 for sex and age categories (details in Appendix A4). Alternative Care Cost Alternative care cost categories include professional care, such as by a physiotherapist, a chiropractor, and/or other complimentary care categories, such as massage therapy, acupuncture, and others. Physiotherapy and other alternative care visits after the hip and knee TJR surgery were included in the rehabilitation cost. I used MOH-OA data (147) to estimate the out-of-pocket costs associated with alternative care visits before the hip and knee TJR surgery. Alternative care costs occurred after the hip\knee TJR surgery was calculated using literature estimates. Here, I first discuss the alternative car cost before the hip\knee TJR surgery. 77 As I will describe in Section 3.4, in the simulation model, I differentiate between visit to an orthopedic surgeon and hip\knee TJR surgery to reflect the waiting time for the TJR surgery. Additionally, according to the MOH-OA data, there was a significant difference between rates of alternative care professional visits for patients who have seen by the orthopedic surgeon and those who have not. Therefore, I performed a separate analysis for probability of visits by alternative care professionals for OA patients before the OS visits and those after the OS visit. I used MOH-OA data (147) to perform the analysis for two types of alternative care categories: 1) physiotherapists and chiropractors for which data was available for previous visits and the number of visits within the last year; and 2) complementary care categories including massage therapy, acupuncture, and others, for which data was available for the probability of (at least) one visit and the number of visits within the past 6 months. Based on the MOH dataset, I reported 7 different complementary care categories for the second type of alternative care professionals that include: massage therapy, acupuncture, naturopathy, homeopathy, traditional Chinese medicine, herbal remedies and Ayurvedic medicine. Among the complementary care categories, massage therapy had the highest frequency of at least one visit in the past 6 months with 11% of all OA patients (out of 1,124 OA patients) and Ayurvedic medicine had the lowest frequency of visits with 2%. Frequencies of visits to all alternative care professionals are listed in Table A12, Appendix A. The analysis for calculating the per-patient cost for the alternative care categories was performed in two steps. I first performed a logistic regression to estimate the probability of a non-zero cost in each of the above two categories for OA patients of 78 different age and sex groups, before and after visit to an orthopedic surgeon. Next, I estimated the mean number of visits for each age/sex group. The results of the logistic regression models and the odds ratios of the three age groups for males and females are reported in Table A13, Appendix A. In the next step, I calculated the baseline rate for the reference category in the MOH-OA data and, using the odds ratios and baseline probability, I calculated the probability of visiting (at least) 1 alternative care professional for each age and sex group, before and after visit to an orthopedic surgeon. Tables 3.5 and 3.6 show the probabilities for each group of alternative care professionals, respectively. Next, the weighted average fee for visiting each of the above type of alternative care facilities were calculated using the BC MSP billing data for 2003 which was $67 in 2003 $CAD for the first group (physiotherapists and chiropractors) and $61 for the second group (complementary care categories) (148). For the fee inflation rates from 2003 to 2012, I used the average rate from the physicians’ CPI that was 3.3% during the study period (153). I implemented these probabilities, number of visits, and unit costs based on the fees for each visit in POHEM-OA. The probabilities were translated into annual hazards ratios in order to model each visit as an event. This was performed taking into account the time period for the results of complementary care categories (6 months). Further details regarding the calculations for probabilities and results for mean numbers of visits that were implemented in POHEM-OA are provided in Appendix A4. 79 Table 3-5. Probability of visit to physiotherapist and chiropractic within the past year* n=1124 Before orthopedic surgeon visit After orthopedic surgeon visit Male (95%CI) Female (95%CI) Male (95%CI) Female (95%CI) <60 0.13 (0.06,0.25) 0.27(0.10,0.40) 0.21(0.09,0.34) 0.43(0.21,0.68) 60-69 0.12 (0.04,0.27) 0.25(0.12,0.38) 0.19(0.08,0.29) 0.40(0.22,0.59) 70-79 0.25 (0.15,0.36) 0.47(0.31,0.57) 0.40(0.28,0.58) 0.75(0.49,0.91) ≥80 0.32 (0.22,0.42) 0.56(0.33,0.72) 0.45(0.33,0.60) 0.86(0.52,0.90) * Estimates are calculated from regression models in Appendix A4 and using MOH-OA data (147). All the alternative car professional visits in this section are prior to hip/knee TJR replacement surgery. Those after TJR surgery were included in the rehabilitation cost category. For post-surgery alternative care cost calculations, I used the reported results in studies of March et al. (137) and Mitchell et al. (138) for the physiotherapy visit cost as the MOH-OA dataset lacked both discharged destination data and sufficient number of post-surgery subjects (147). The resulting costs were included in the rehabilitation cost category. The cost of physiotherapy, after the hip/knee TJR surgery, depends on the discharge destination, including rehab facility or home care as shown in Figure 3.1. For the post-surgery costs, I did not have literature data on alternative cost categories other 80 than physiotherapy and, therefore, only physiotherapy costs were included. As mentioned in Section 3.2, while all the alternative costs were included as out-of-pocket costs, the physiotherapy costs after hip\knee TJR surgery were included in the out-of- pocket costs of the rehabilitation category. The results for parameters representing the probability of non-zero visits and mean of visits are presented in Table A14, Appendix A. Table 3-6. Probability of visit to alternative care professionals within the last year n=1124 Before orthopedic surgeon visit After orthopedic surgeon visit Male (95%CI) Female (95%CI) Male (95%CI) Female (95%CI) <60 0.06 (0.02,0.11) 0.27(0.13,0.39) 0.13(0.04,0.32) 0.46(0.30,0.53) 60-69 0.21(0.12,0.35) 0.25(0.12,0.38) 0.38(0.22,0.49) 0.43(0.31,0.55) 70-79 0.2 (0.12,0.33) 0.47(0.31,0.58) 0.37(0.24,0.48) 0.67(0.44,0.77) ≥80 0.16(0.08,0.32) 0.54(0.32,0.72) 0.32(0.23,0.48) 0.75(0.54,0.87) * Probabilities are only for complementary care categories that includes Massage therapy, Acupuncture, Naturopathy, Homeopathy, Traditional Chinese Medicine, Herbal remedies, Ayurvedic medicine in order of highest frequency- all are prior to hip/knee replacement surgery. Results are from MOH-OA data (147). Details are discussed in Appendix A4. 81 3.3.6 Side Effects of OA Drugs Four types of side effects from drugs have been implemented in this study: gastrointestinal (GI), cardiovascular disease (CVDs), stroke, and dyspepsia. The rate of drug use has been modeled based on the Health Utility Index (HUI) and patients characteristics using data from NPHS data in 2002 (146). The drug module in the POHEM-OA model was previously developed by Sayre et al. (139) and discussed in Appendix A1. The probability of occurrence of each side effect has been modeled for each drug type including GI complications from NSAIDS and coxibs, dyspepsia from opiods, CVD and stroke from coxibs and NSAIDs. I assumed no side effects for acetaminophen (139), in accordance with the POHEM-OA model developed by Kopec et al. (1). The hazard rates of the side effects for each drug were calculated and implemented in POHEM-OA according to differences in rates of side effects between persons on each drug compared to those on placebo per 1000 person-years for each drug type as shown in Table A2, Appendix A1. In this study, I calculated the lifetime and per-case cost of side effects from the literature. Details of these cost calculations are shown in Appendix A4. To arrive at the lifetime direct cost for Cardiovascular Disease (CVD) across different age and sex categories, I used a Canadian study (140), in addition to other studies (156, 157,159). For the lifetime cost associated with stroke, I used a US study (141) in addition to a recent study (158) for reporting average direct lifetime costs for different types of stroke as side effects of NSAIDs or coxibs (141). To calculate the average GI-related costs resulting from the use of NSAIDs among OA patients, I used 82 the results of Rahme et al. (142) study that used population-based data in Quebec, Canada, and computed the GI cost, including the healthcare system direct costs (which also include the costs of the medications used to treat the GI complications). For dyspepsia, I calculated per person (per event) costs from the results of Moayyedi et al.’s study (143). Results for stroke and CVD life-time cost estimates are provided in Table A15, A16 in Appendix A. Finally, I used the overall economy CPI (134) to translate the lifetime and per case cost of each side effect to 2003 $CAD, in addition to their inflation rates from 2003 to 2010. The Purchasing Power Parity (PPP) rates were used to transfer other countries’ monetary values to $CAD (149). 3.4 The Population Health Microsimulation Model (POHEM) The POHEM model is a continuous-time, dynamic microsimulation model that was designed in 1990 to represent the lifecycle dynamics of the Canadian population (44). The model is built upon the evolution and interactions of sets of disease-related risk factors and health determinants, such as age, sex, comorbidity, mortality rates due to disease, etc. Several different diseases have been modeled using the POHEM models including heart disease, diabetes, osteoarthritis, and different types of cancer (lung, breast, and colorectal cancer). The cancer models include individual-level data on health care costs and utilization, together with a measure of health-related quality of life (45, 46). In a recent study, Kopec et al. (1) developed POHEM-OA, a new development of the model to quantify the incidence of osteoarthritis and its effect on health-related quality of life of Canadians. I further develop POHEM-OA to include cost modules in addition to performing uncertainty analysis (UA). Figures 3.2 shows the diagram of the 83 POHEM-OA model. Here, I discuss details of the POHEM-OA, as developed by Kopec et al. (1). Figure 3-2 . POHEM-OA: an osteoarthritis model* age (time) Death self- reported Obesity-OA NPHS 1994- BMI and association 2004 OA in 2000 BCLHD AGE and 2002 Relative risk of Individual’s developing OA individual’s incidence age, sex, according to change in rates by age survival data for Sex region, BMI category weight over and sex each transition education, time income, and past BMI BMI time to Region BMI category OA surgery (hip Surgery incidence or knee) Income higher weight, lower HUI for lower HUI use of drugs changes Individual’s those with OA (NPHS model) time to surgery initial value (NPHS) of BMI Education use of drugs PHARMANET lowers PAIN HRQOL CCHS 2001 Individual’s drugs may have initial value side-effects of HUI SIDE-EFFECTS surgery may have (Stroke, Death, ...) side-effects HRQOL declines with age (modeled from NPHS) surgery improves HRQOL * This figure is from Statistics Canada (unpublished); Rectangle in the left represent an individual with his/her characteristics defined for age, sex, region, etc. Cylinders inside the figure are data sets used for estimation of parameters; first in top left: NPHS 1994-2004: National Population Health Survey, a longitudinal household survey by statistics Canada with a sample of size of over 17,000 persons that started in 1994, 5 cycles (1996-2006) have used in BMI trajectory model and two cycles (2000-2002) used in estimation of relative risks (hazard ratios) of OA; ** Rectangle in the bottom left corner: CCHS-2001 (Canadian Community Health Survey (54)) used for generating the starting population for the microsimulation with characteristics listed on the left column of the figure; ** Rectangle in the top right corner: BCLHD is the same as PDBC (Population Data BC (84)): Administrative billing data for the province of BC including data for > 98% residents of BC from 1994-2003, used for estimating incidence of OA. PharmaNet is linked to PDBC and has drugs-related data for all BC population. 84 3.4.1 Population Health Model of Osteoarthritis (POHEM-OA) As discussed in Chapter 1, in modeling OA’s cost burden, the complexity of progression and interaction of the main risk factors in addition to the heterogeneity of OA patients warranted the use of microsimulation models (22). Therefore, in this study, I used the POHEM-OA model, an individual-level and continuous-time simulation model. In fact, POHEM-OA samples the time-to-events from the appropriate hazard distributions of OA and health–related events such as birth, death and hip\knee TJR surgery (1). Figures 3.3 and 3.4 show the diagram of the POHEM-OA model that consists of two major components to model OA diagnosis and its progression. POHEM-OA Diagnosis Module The POHEM-OA model uses the Canadian Community Health Survey (CCHS) in 2001 to sample individuals for its initial population. Sample size of CCHS in 2001 was about 130,000 (54), where each individual inside the CCHS sample represents a group of individuals corresponding to adult population of Canada (~ 20 million) 7. The main risk factors for the incidence of OA in POHEM-OA are age, sex and body mass index (BMI). The model simulates an individual’s life history in terms of OA-related events, through inter-related stochastic processes; at each calendar year, the events are assigned to individuals, based on their characteristics (e.g., OA diagnosis is assigned based on age, sex and BMI). BMI is modeled as an auto-regression model that includes age, sex, income, education, region, and previous BMI. At the same time, based on the current 7 Survey weights were provided because the sample is a stratified cluster sample in which some population groups are over or under sampled and the weights are needed to make the sample representative of the Canadian population (54) 85 value of BMI, a survival regression model was used to estimate the time of the OA diagnosis. Appendix A1 provides details of the statistical models inside the POHEM-OA model. Due to the continuous-time nature of POHEM-OA, the time of the next event in the queue is assigned to all possible events in the near future, based on the competing hazards approach. For example, the hazards of both an OA diagnosis and death for a given individual are randomly sampled from a distribution of hazard functions that represent a continuous period of time and are functions of the individual’s characteristics; the event that comes first, is assigned to that individual. After all events are assigned to each individual, the cost module is updated as discussed in Section 3.2. Appendix A1 provides a step-wise algorithm of how POHEM-OA assigns an event to an individual. 86 Figure 3-3. POHEM-OA diagnosis module Age (time) Death Self-reported NPHS BMI and OA Obesity-OA AGE 1994-2004 in 2000 and Association 2002 Relative risk of Individual's developing OA according age, sex, region, Individual's SEX to BMI category education, change in weight income, and over time past BMI OA REGION BMI BMI category incidence INCOME Individual's Incidence initial value rates by age of BMI and sex EDUCATION CCHS Pop-Data 2001 BC * POHEM-OA: Population Health Microsimulation model for osteoarthritis (OA) (1); ** Rectangles inside the figure are data sets used for estimation of parameters; first in top left: NPHS 1994-2004: National Population Health Survey, a longitudinal household survey by statistics Canada with a sample of size of over 17,000 persons that started in 1994 (146), 5 cycles (1996-2006) have used in BMI trajectory model and two cycles (2000-2002) used in estimation of relative risks (hazard ratios) of OA; ** Rectangle in the bottom left corner: CCHS-2001 (Canadian Community Health Survey (54)) used for generating the starting population for the microsimulation with characteristics listed on the left column of the figure; ** Rectangle in the bottom right corner: PDBC (Population Data BC (84)): Administrative billing data for the province of BC including data for > 98% residents of BC from 1994-2003, used for estimating incidence of OA. POHEM-OA Progression Module The progression of OA in POHEM-OA is modeled in terms of patient flow in the healthcare system, as shown in Figure 3.4. After an OA diagnosis, the first OA 87 progression state modeled was a visit to the orthopedic surgeon (OS), referred to by a general practitioner (GP). There may have already been several physician visits (GP, or specialist) by the time of the first OS visit, while most cases of OA never visit an OS. An OA patient may visit an OS several times (up to 10 times in the cost module of the POHEM-OA). If the orthopedic surgeon considers it necessary, patients may be hospitalized for a short stay of 1 to 2 days and undergo an inpatient procedure (Section 3.3.1). Figure 3-4. POHEM-OA progression module * It is possible to go through more than three OS visits prior to surgery. In the case of severe OA of the hip or knee, the OS may book a hip or knee TJR surgery for the patient. Due to the different patterns of resource utilization among OA patients who undergo hip/knee TJR, this procedure is modeled as an event for which the time is dependent on sex, age and time from the OS visit. A patient may undergo, at 88 most, four hip/knee primary surgeries. If necessary, a patient could undergo revision surgery of the same joint that comprises the last phase of OA progression in the model. The time to the next progression event is estimated through a survival model based on a Weibull model using PDBC (84)- this model is further described in Table A1 in Appendix A1. 3.5 Description of the Cost Algorithm The direct cost module of the POHEM-OA model consists of the input parameters needed to determine the average costs of the resources utilized by OA patients, and a cost algorithm to dynamically assign these costs to the individuals involved, as the simulation model evolves. The input parameters were estimated according to a bottom-up COI approach described above in Section 3.4. All OA-related resource usages are divided into direct cost categories as shown in Table 3.1. 3.5.1 Direct Cost Algorithm In the direct cost module of POHEM-OA, each individual is assigned a cost profile that keeps track of the values of that individual’s characteristics as they change throughout the simulation. The cost algorithm is coded so that, whenever the cost factor of an individual changes, the cost profile of a patient gets updated. For example, when patients change their age group or enter a new OA progression state, e.g., visit an orthopedic surgeon (OS) or have a primary surgery, the cost profile updates as the utilization rate is now changed according to the estimated parameters for these categories. The cost profile keeps track of all cost categories described in Table 1 and gets updated for each cost category at a specific time, based on a change in the patient’s characteristics (defined in Section 3.2.2). 89 The cost categories in the cost profile of each patient are implemented by two methods: state-based and event-based methods. The costs for inpatient (non hip/knee TJR) surgeries, physician’s visit, outpatient services, and drug costs (both prescription and over-the-counter) are updated based on the state-based cost approach: whenever the characteristics of a patient, including age and OA status, change, the cost profile of the patient gets updated based on the input parameters associated with these cost categories. The input parameters of the state-based approach include the average per patient-year cost, based on sex, age group, and OA status, which are estimated using the administrative PDBS (84). As discussed in Section 3.2.2, individuals’ characteristics were defined according to cost drivers including age, sex, OA status, and time in each state. As a result, the cost profile of each patient will be calculated by multiplying the per patient-year cost by the time the patient has spent in each OA state. On the other hand, I implemented other cost categories, including the costs for hip and knee TJR surgeries, formal caregiver, alternative care (physiotherapist, chiropractic and others), rehabilitation, and the side effects of OA drugs using an event- based approach. All of the parameters related to these cost categories were estimated from (cross sectional) OA surveys or extracted from the literature. The latter categories were implemented based on per patient cost, i.e., events related to each of these categories are simulated in the POHEM-OA and, once the event has occurred, the average per patient cost is assigned to the patient whose costs will be calculated based on the parameters inside the model. The probability of events in the event-based cost categories include: 1) hip/knee TJR surgery, 2) occurrence of a non-zero formal caregiver cost, 3) visit to a physiotherapist and other alternative care physicians and (4) 90 discharge to a rehab facility or home care. The probabilities were estimated as a function of the OA patient’s characteristics (i.e., age, sex, and OA status). The POHEM- OA model translates the probabilities into hazards, calculates the time to each event, and probabilistically assigns them to each individual (i.e., a discrete event simulation framework). Details of the POHEM-OA framework are discussed in Appendix A1. 3.5.2 Cost Outcomes In the POHEM-OA model (and other simulation models), parameters are fixed and can only be changed in the sensitivity or what-if analysis. For example, a relative risk or cost input parameters estimated from different data sources in this study are parameters. On the other hand, variables take new values in a simulation run, and are updated. For example, age, OA state and cost profile of OA patients are variables. After all the input parameters are implemented in POHEM-OA, the cost algorithm assigns a cost profile to each patient. As the POHEM simulates the life of each individual by stochastically assigning possible events, the cost module can be triggered according to OA-related events and the cost profile is updated. As a result, a patient’s cost profile is updated by assigning new values from the input parameters for each cost category. The total OA-related cost for each year would be the sum of the costs incurred by all OA patients in that year. The total direct cost would be the weighted sum of the time spent in each OA state multiplied by the average person-year cost for that state plus the event-based per-patient costs multiplied by the number of events. The average direct cost for each year would be the total cost divided by the number of OA patients during that year. The final reported average cost would be from the societal perspective and can be interpreted as the average direct cost for each year. As defined 91 in Chapter 2, societal perspective includes both the costs paid by patients and those paid by the healthcare system. Inflation from 2003 to 2012 The rate of inflation has been used to backdate all the unit costs to 2003 $CAD at baseline. For instance, for the out-of-pocket unit costs, I used studies from 2001, 2002 and 2005 (127, 137,138), but deflated them back to 2003 using the overall economy CPI for each year (134). For the costs reported in US dollars and Australian dollars, I used the PPP rates for exchanging US and Australian dollars into Canadian dollars (149). For the healthcare system paid cost categories, including (part of) drugs, physician and hospitalization, I used administrative data from the PDBC database, which were again deflated back to 2003 using the overall economy CPI for each year (134). However, to calculate how unit costs changed from 2003 to 2010, I used price indexes as reported separately for each category based on CIHI reports on physician price indexes (153), drug price indexes in 2003-2008 (152), and CPWAC rates based on hospital price index reports from 2003-2010 (151). I assumed the same trend for 2005-2010 and 2008-2010, for hospitalizations, drugs and physician visits. The final results were shown in both current prices and 2010 $CAD using 3.5% as the annual discount rate according to Boardman et al. (155). 3.6 Scenario Analysis for Drivers of the OA Direct Cost Burden I have performed three separate analyses for investigating the effects of three important cost drivers on the change in average cost of OA between 2003 and 2010: 1) increase in mean age of the population (due to reduced mortality rates), 2) increase in number of hip/knee TJR surgeries, and 3) price inflation. The remaining cost drivers 92 include change in the prevalence of OA ( due to change in the mean BMI, change in birth and migration rates), and change in the individuals’ behavior toward increase utilization of drugs and physician visits that was not included in the model. Scenario 1. Effect of an Increase in Life Expectancy of the Population Change in the mean age of the population in POHEM is partly associated with change in mortality rates, in addition to change in birth and migration rates. In this scenario, I investigated the effect of increase in the mean age only through the decrease in time-dependent mortality rates, which is the reason for increase in life expectancy of the population. The mortality rate in the POHEM-OA model is modeled as a function of year, age cohort, sex, age categories, and a 17-item set of disease conditions (causes of death). Based on a demographic module developed by Statistics Canada, I applied the mortality parameters to the POHEM-OA model by age cohorts from the year 1881 up to 2051 (44)- details of the POHEM model is discussed in Appendix A1. In the first scenario analysis, to examine the effect of stabilizing the longevity of the Canadian population, I changed mortality rates for the individuals entering the model either by immigrating to Canada or by birth from 2004-2012 to be the same as those in 2003 (for all individuals greater than 20 years old). In other words, I assumed no increase in the mean age of the population due to mortality rate- although the mean age could still increase due to change in birth and migration rates. Scenario 2. Effect of an Increase in the Number of Hip/Knee TJR Surgeries As described in Section 3.3.1, the hazard rates for hip/knee TJR surgeries were modeled using the longitudinal data for 1986-2003 from PDBC. However, I calibrated 93 the number of TJR surgeries in the POHEM-OA model for years after 2003, to match the numbers reported by the CJRR from 2003 to 2010 (124). The actual number of hip/knee TJR surgeries has increased by about 100% in the past 10 years according to the 2012 CJRR report (126). Although some part of this increase was observed in the baseline results (without calibration), which was due to the increase in number of people with OA, the additional number of hip and knee TJR surgeries that were added after the calibration was due to the change in the policy and increased hospital capacity over the past decade in Canada (52). To calculate this additional policy-related hip and knee TJR surgery over the past decade, in this scenario, I first run the POHEM-OA model without the calibration, and then performed the simulation with the calibrated rates. Scenario 3. Effect of the Health-Care-Sector Price Inflation As discussed for the base-case scenario in Section 3.5.2, I used the appropriate health-care-sector inflation rates for each cost category to inflate the costs, starting from 2003 to 2012. The inflation rate I used for each cost category was taken from different CIHI price index reports for physicians (153), orthopedic surgeon (154), drugs (152), and hospitalization (151) between 2003 and 2010. In this scenario, I investigated the change in the average cost of OA with and without the specific health-care-sector inflation rates. I have executed the POHEM-OA model by performing the above-mentioned scenarios independently for the three major cost drivers. The goal was to investigate how much of the increase in the average cost of OA from 2003 to 2010 was due to change in each cost driver. I first calculated the differences in the resulting average cost of OA from 2003 to 2010 by comparing the results of each scenario with the baseline 94 results. Then, the three differences were compared to show the effect of each cost driver on the increase in the average cost of OA from 2003 to 2010 as described in the results section. 3.7 Results: Average Costs from 2003 to 2010 Between 2003 and 2010, the OA population size grew from 2.9 million to 3.6 million, an almost 24% increase, while the adult population of Canada had a 12% increase in size (Table 3.7). As shown in Table 3.7, during the study period the average age and the mean BMI of the OA population increased up to 67.4 years of age and the mean BMI increased to 27.2 (kg/m2). Both were higher than that of the general adult population of Canada (those 20 years of age and older). The results for average cost of all OA patients in Canada between 2003 and 2010 are shown in Figure 3.5. Using 3.5% annual discount rate (155), average cost increased from $735 to $811 between 2003 and 2010 in 2010 $CAD. I also presented the results according to every year’s current dollar value in Figure 3.5, where the average direct cost of an OA patient rose from $578 in 2003 to $811 in 2010 in current $CAD, an almost 40% increase- the current dollar values do not include any discount rates and represent each year results according to that year $CAD. 95 Table 3-7. Characteristics of OA patients and of general population 2003 2006 2010 Average age (OA population) 66.5 66.9 67.4 Average age (General population) 46.2 46 .9 47.8 Average BMI* (OA population) 26.8 26.8 27.2 Average BMI (General population) 25.6 25.8 26.1 OA population size** 2,947,145 3,248,026 3,653,787 General population size old** 23,343,419 24,658,256 26,260,862 * Body mass index (BMI) unit: (kg/m2); ** both OA and general population are 20 years of age and older As shown in Figure 3.6, the average cost for a female and male OA patient was $597 and $543, respectively, in 2003, and increased to $959 and $757 in 2010 (in current dollars). In fact, average OA cost for a female OA patient are 22% and 20% higher on average than those for their male counterparts in 2003 and 2010, respectively (Figure 3.6). 96 Figure 3-5. Average cost of OA patients in 2010 $CAD and current dollars * Costs are all in 2010 $CAD and were discounted using a 3.5% discount (155). Figure 3-6. Average cost of OA for females and males from 2003 to 2010 $900.00% $850.00% $800.00% $750.00% $700.00% $650.00% $600.00% $550.00% $500.00% 2003% 2004% 2005% 2006% 2007% 2008% 2009% 2010% Average%cost%(all:current$)% Average%cost%(female)% Average%cost%(male)% * Costs are reported in current dollars (inflation is included in all of the figures in this Chapter) 97 With respect to costs for different age categories, it can be observed from Figure 3.7 that the average cost of OA patients in the age group 70-80 years old was the highest both in 2003 and in 2010 ($608 and $822 in current dollars). Additionally, the average cost for OA patients in the age groups 50-60, 60-70 and 70-80 years old was higher than that for the other three age groups (i.e., less than 50 years old, 80-90, and higher than 90 years old) throughout the study period. However, the average cost of the age group less than 50 years old and 50-60 increased with the highest rate than the other age groups, they had an almost 42% and an 39 % increase, respectively. Figure 3-7. Average cost for OA patient by age categories from 2003 to 2010* $900.0$ $850.0$ $800.0$ $750.0$ $700.0$ $650.0$ $600.0$ $550.0$ $500.0$ $450.0$ $400.0$ $350.0$ $300.0$ $250.0$ $200.0$ $150.0$ $100.0$ $50.0$ $0.0$ 2003$ 2010$ [min,50[$ [50,60[$ [60,70[$ [70,80[$ [80,90[$ [90,max]$ *Costs are all reported in current $CAD 98 Average Cost by Cost Categories The out-of-pocket share of the average cost for OA patients in 2003 was $192 (29%), while the health care system’s share was $386 (71%). During the study period, the share for the health care system’s costs increased to $624 (77%) in 2010. The main reason for this increase is due to the increase in the share of hospitalization cost. As shown in Figure 3.8 and 3.9, the share of hip/knee surgery and rehabilitation costs increased (21% to 32% and 10% to 11%), while the share of other cost categories (such as, for example outpatient services) decreased during the study period. The distribution of average costs, with respect to cost categories, is very similar in males and females (figure not shown) with two major exceptions. First, the alternative care and formal caregiver share of the average cost was higher for a female patient (5% and 7%, respectively), than a male patient (4% and 6%) in both 2003 and 2010. Second, the cost of the hip/knee TJR surgery and its share of the total OA cost were higher for males than females. Between 2003 and 2010, the average cost of TJR costs grew from $132 to $251 for males and from $125 to $234 for females. 99 Figure 3-8. Average cost of an OA patient by cost categories in 2003 *Rehabilitation cost represents the sum of out-of-pocket and heath care system costs for the rehab/homecare cost categories. Figure 3-9. Average cost of an OA patient by cost categories in 2010 *Rehabilitation cost represents the sum of out-of-pocket and heath care system costs for the rehab/homecare cost categories. 100 Average Cost by OA States The number of OA patients and their associated average cost varied significantly across different OA states (OA states were defined in Section 3.2.2). Figures 3.10 and 3.11 compare the number of patients and their associated average cost at different OA states. The highest average cost for the years 2003 and 2010 was incurred by patients in state 4 (during and after revision surgery) and state 3 (during and after primary surgery), respectively as shown in Figure 3.11. The average cost of those in state 3 was 7.1 and 6.8 times higher than that for state 1 patient in 2003 and 2010, respectively. However, the majority of OA patients were in state 1, after OA diagnosis and before orthopedic surgeon visit, in 2003 and 2010 compared to sate 3 and 4, i.e., 82% of the OA population in 2003 and 81% in 2010 were in state 1. As a result, total cost of OA in 2003 for patients in state1 was $856 million in2003 and increased to almost $1.4 billion in 2010, while total cost of OA patients in state 3 was $598 million $CAD in 2003 and increased to $983 million in 2010. In other words, while the average cost of patients who had a knee and hip TJR primary or revision surgery was almost 7 times higher than those OA patients who have not visited the orthopedic surgeon (and do not have sever OA), the total cost associated with patients in state 1 was higher throughout the study period. As shown in Figure 3.11, the highest rate of increase in average cost occurred among patients in state 2, those who have visited the OS, but waiting for TJR surgery, with almost 40% increase in the average cost from 2003 to 2010. These OA patients are in waiting line for hip an knee TJR surgery, have severe OA and are often experiencing very high level of pain and loss of function. Majority of the average cost of 101 OA patients in state 2 are due to out-of-pocket cost including formal caregiver and alternative care (54%), hospitalization costs other than hip and knee TJR (42%), physician cost (3%) and drugs (3%) in 2010. Figure 3-10. Number of OA patients by OA state in 2003, 2006 and 2010 4,000,000$ 3,500,000$ 3,000,000$ 2,500,000$ 2,000,000$ 1,500,000$ 1,000,000$ 500,000$ 0$ 2003$ 2006$ 2010$ Revision$surgery$ 15,515$ 21,097$ 27,274$ Primary$surgery$ 232,457$ 276,963$ 308,708$ OS$visit$ 266,565$ 281,781$ 340,505$ OA$diagnosis$ 2,432,609$ 2,668,184$ 2,977,301$ * OA diagnosis (state 1) The first stage shows the average cost of those who have been diagnosed with OA before a visit to the orthopedic surgeon (OS); *OS visit (state 2): this shows the average cost of those who have seen the orthopedic surgeon (OS), but are before the Knee/hip primary surgery. Surgery (state 3): During and after Primary surgery ; Revision (state 4): During and after revision surgery. 102 The average cost for a patient in state 1 increased by 32% from 2003 to 2010, the second highest rate of increase among other states (all in current $CAD). For patients in state 1, physician cost (43%), drugs (26%) and out-of-pocket (26%) constituted the highest share of the average cost among other cost components in 2010, respectively. Out-of-pocket cost included alternative care (13%) and formal caregiver cost (13%). Average cost of OA patients in state 3 and state 4 increased at the lowest rate compared to other states, i.e., 24% and 27% increase from 2003 to 2010 (Figure 3.11). This, in part, represents the decrease in the LOS of surgery during the study period (Section 3.3). As a result, patients were released earlier from the hospital, which caused the rise in the cost of rehabilitation after the surgery. For example, the share of rehabilitation cost in state 3 increased from 18% to 23% from 2003 to 2010. The share of different cost components in states 3 and 4 in 2010 were as follows: cost of TJR surgery (65% for state 3, 68% for state 4), rehabilitation (23% for state 3, 19% for state 4), drugs (8%, 10%), physician (3%, 2%) and other non-TJR hospitalization procedures (1%, 1%). 103 Figure 3-11. Average cost of OA by OA state in 2003, 2006 and 2010. $10,000$ $9,000$ $8,000$ $7,000$ $6,000$ $5,000$ $4,000$ $3,000$ $2,000$ $1,000$ $0$ 2003$ 2006$ 2010$ Revision$surgery$ $2,671.9$ $3,077.3$ $3,405.0$ Primary$surgery$ $2,576.0$ $2,729.3$ $3,185.7$ OS$visit$ $1,011.2$ $1,245.3$ $1,422.1$ OA$diagnosis$ $352.8$ $411.5$ $468.2$ * OA diagnosis (state 1): the first state shows the average cost of those who have been diagnosed with OA before a visit to the Orthopedic Surgeon (OS); *OS visit (state 2): this shows the average cost of those who have seen the orthopedic surgeon (OS), but are before the knee/hip primary surgery; Primary surgery (state 3): During and after Primary surgery; Revision surgery (state 4): During and after revision surgery. Cost Drivers of OA As shown in Figure 3.12, surgery numbers were the strongest cost driver of the average cost and accounted for 33% of the difference between the average costs in 2003 and 2010. As discussed earlier in this section, in the baseline scenario, the 104 average cost of OA increased from $578 in 2003 to $811 in 2010, an almost 40% increase (in current $CAD). After performing the scenario related to the increase in the number of surgeries (Scenario 2), for which the number of surgeries were not calibrated and therefore the extra increase in the number of surgeries were not included, the average cost of OA increased from $578 in 2003 to $734 in 2010. I then calculated the difference between the resulting average cost, which showed to be 67% of the difference of the average cost in 2003 and 2010 in the baseline scenarios. As a result, 33% (100-67=33) of the difference between average cost of OA between 2003 and 2010 was due to the increase in the number of surgeries. The same calculation was performed for other scenarios for which I investigated the price inflation and change in life expectancy of the population. Price inflation was responsible for 30% of the overall increase in average cost of OA and change in life expectancy of the population was responsible for 10% of this increase (all in current $CAD). The remaining 27% was due to other cost drivers, such as a change in mean age (due to birth and migration rates) and/or mean BMI of the population. 105 120%# Figure 3-12. Sensitivity analysis: effect of cost drivers on average cost change in 2010 100%# increase#in#longevity# 80%# Increase#in#number#of# hip/knee#surgery# (policy#change)# 60%# InflaBon# 40%# 20%# Other#(change#in# mean#age,# BMI,demographics)# 0%# *Increase in longevity is the increase in the life expectancy due to the reduction in mortality rates during the study period 9form 2003 to 2010) as modeled in the POHEM-OA model. 3.8 Discussion I performed a simulation-based cost-of-illness (COI) using the POHEM-OA model, a microsimulation model previously developed by Kopec et al. (1), to estimate the average direct cost associated with OA in Canada from 2003 to 2010. For this purpose, I first preformed a traditional bottom-up (micro-costing) COI study to estimate input parameters for per patient-year and per-patient cost associated with each cost component including drugs, physician visits, hospitalization and out-of-pocket cost. 106 Next, I implemented the estimated input parameters into the simulation model to estimate the average cost of OA from 2003-2010. This is the first study to provide the average cost of OA from the societal perspective from 2003-2010, in addition to their distribution across different cost components and sub-populations. In fact, I were able to impute the missing cost data for average cost of OA from 2003-2010 in Canada by integrating different data sources using a population-based microsimulation model. According to the results of this study, between 2003 and 2010, the average cost of OA has increased from $577 to $811 in current dollars, an almost 40% increase in the average cost. When discounting was taken into account, this increase was from $735 to $811, about 10%, in 2010 $CAD. The largest cost category was hospitalization cost due to hip/knee TJR surgery that constituted around 21% of the overall average cost in 2003 and increased to 35% in 2010. Physician costs and drug costs were the second and third largest categories (24% and 18% in 2003, respectively). The average cost for females was higher than for males. Additionally, the difference between the average cost for females and males increased from 10% in 2003 to almost 13% in 2010. Using the POHEM-OA model, I investigated the drivers of the increase in average cost during the study period. Increase in the number of surgeries, price inflation and increase in life expectancy of the population due to lower mortality rates (longevity) was the most significant drivers during the study period, respectively. The results of this study are similar to those reported in the CIHI report of the cost drivers for the overall health expenditure in Canada (161). In their report, increase in population size, price inflation, increase in the mean age of the population and other cost drivers were 107 introduced as the most important cost drivers for the health expenditure in Canada from 2001-2009, respectively (161). The average costs for OA patients reported in COI studies using population- based data were very close to those estimated in this study. For instance, population- based studies of Le Pen et al. (117) estimated $390 for average cost of OA in France in 2003 using administrative data and Tarride et al (105) estimated $1200 in Canada in 2010 using linked CCHS data- my estimate was $811 in 2010 (all in 2010 $CAD). As mentioned in Akobundu et al. (57), survey-based COI studies tend to overestimate the cost of diseases since they do not account for different disease states due to selection bias. For example, Canadian survey-based studies of Maetzel et al. (114) in 2000 estimated $3,797, and Gupta et al. (115) in 2004 estimated $2,123 for the average cost of OA. Although they estimated higher average costs of OA in Canada than my estimates, very few OA patients were included in their sample, i.e., 410 OA patients and around 1,000 in Maetzel et al. (114) and Gupta et al. (115), respectively. As shown in Table 2.5, US studies reported more than 10 times higher average cost of OA than my estimates, e.g., $13,381 (106) and $14,530 (113). According to a 2007 US study of Kotlarz et al. (60), average cost of OA was $7,846 for women and $5,974 for men, which corresponds to my results that women have higher average costs than men (all in 2010 $CAD). As shown in Table 3.7 of this Chapter, in 2003, 2,748,505 people in Canada had OA, which, based on the average per patient-year cost of $577 (CAN$ 2003), adds up to a total cost of $1.58 billion in 2003 $CAD. The total direct cost of OA was reported to be $1.15 billion in 2006 (2006 $CAD) according to the Anis et al. (94) study that did not 108 include the out-of-pocket costs. This estimate was close to the direct cost reported by my study in 2006, which was $1.5 billion in 2006 $CAD considering only the health care system costs. In addition, in the health expenditure report of Anis et al. (94), the largest part of the direct costs was those due to hospitalization (~ 25% in 2003) - a result very similar to ours (22%). As discussed in this Chapter, the hip/knee TJR surgery was the most significant cost component during the study period from 2010 to 2031. I have used a linear trend model to estimate the future number of hip/knee TJR surgeries as discussed in the methods section of this Chapter. Although the number of surgeries has increased significantly within the past decade (around 100% from 1999 to 2009), the rate of increase was smaller in the previous decades. While I have used a linear trend to model the number of surgeries for each age and sex category, it is possible that the rate of increase would slow down or plateau, mostly because of the limited capacity of the hospitals to perform TJR surgeries. The results provided in this Chapter are conditional on the fact that the current policies that determine of rate of increase for TJR surgeries will remain fixed over the next 20 years from 2010 to 2031. In my future studies, I plan to use other models (e.g., asymptotic growth curves) to investigate possible scenarios in terms of different rate of increase for the number of surgeries. In this study, I reported the out-of-pocket costs for OA patients, while top-down studies using the expenditure data such as (71, 94) do not include the out-of-pocket costs due to OA. According to the results shown in Figure 3.9 in this Chapter, 34% of the average cost of OA in 2010 was due to out-of-pocket cost of OA, from which 11% was due to formal caregivers’ cost, 7% was due to over-the-counter drugs, 9% was due 109 to alternative care cost before the hip/knee TJR surgery and 7% was due to the out-of- pocket portion of the rehabilitation cost after the surgery. Detailed discussion for disadvantages and strengths of this study is provided in Chapter 6. 110 Chapter 4: Projecting the Total Direct Cost Burden of Osteoarthritis Patients in Canada From 2010 to 2031 Using the POHEM-OA Model 4.1 Introduction Based on an economic cost analysis of musculoskeletal disorders in five industrialized countries (Australia, Canada, France, United Kingdom, and United States), the total economic costs of musculoskeletal disorders has been projected to increase from 1%-2.5% of the gross domestic product (GDP) of western countries in 1997 to 2%-4% in 2020 (97). The costs of OA have been reported to be around 10-15% of the overall cost of all musculoskeletal disorders (13). Most of the existing estimates and projections for the total direct costs of OA in Canada are based on a gross percentage of the total cost for musculoskeletal or arthritic diseases (71). For example, in a 2010 PHAC report, the cost for all arthritis diseases were estimated using the 2000 Canadian health expenditure data (71), but assuming that 35%-40% of the total direct arthritis cost is associated with OA (85), the total direct cost of OA was $ 0.9-1.1 billion dollars in 2010 (2010 $CAD). In another study, by using a top-down, prevalence- based COI approach, Anis et al. (2010), estimated the total direct cost of OA to be $1.29 billion dollars (2010 $CAD) based on a National Health Expenditure 2006 report (94).8 Many top-down and prevalence-based COI studies have been criticized because of inconsistencies in their estimates and their unreliability with regard to health care system decision-making (57). Direct cost estimates of OA in the US have been reported 8 We converted all costs reported here into 2010 $CAD according to overall economy CPI (134). 111 to have a 10-fold variation across different studies (60). More importantly, these estimates are at the macro level and their extrapolation across different sub-populations and time are often biased, as they do not consider the disease dynamics and patient heterogeneity (114). The complexity of OA progression and the statistical interaction (effect modification) between its main risk factors such as age, sex and BMI, along with the heterogeneity of OA patients, warrant the use of microsimulation model (14). I have discussed the advantages of the use of microsimulation in Chapter 1. In this study, I propose to use a simulation model in conjunction with a COI study for predicting the future costs of OA in order to enhance the usability and consistency of the resulting cost estimates. I further aim to use the POHEM-OA model, integrated with a micro-costing COI approach, to estimate future trends in the direct OA-related costs for: 1) physician visits and other outpatient procedures, 2) hospitalization and rehabilitation, 3) medications, and 4) non-medical and out-of-pocket costs of OA for the period of 2010-2031. 4.2 Methods The outcomes of this study are OA-related direct total costs from the societal perspective, as defined in Chapter 2. As mentioned in Chapter 3.4, OA-related costs incurred each year for a patient would be the sum of all the state-based and event- based costs for that patient (details discussed in Chapter 3, Section 3.4). The state- based costs are those for which I estimated per patient-year cost for input parameters and therefore, the total direct cost associated with them are represented by the area under the cost profile curve for each patient. These include costs associated with physician visits, drugs and hospitalization costs other than hip and knee TJR surgeries. 112 On the other hand, for event-based costs including hip and knee TJR surgery, physiotherapy visits, side effect of drugs, rehabilitation and formal caregiver costs, the input parameters reflect per-patient cost associated with each event. As a result, the total direct cost would be the summation of all individual costs for each year for both state-based and event-based costs. As discussed in Chapter 3, in the simulation model (POHEM-OA) cost module, I define a cost profile at the individual level that represents the direct costs of OA borne by each patient (Section 3.4). The details for estimating input parameters for each cost component and the way they were implemented into the POHEM-OA simulation model are discussed in Chapter 3 (Section 3.3). Table 3-1 depicts all cost categories associated with direct cost of OA and data sources used for their calculations. In this Chapter, I discuss how parameters related to projection of each cost component into future years were estimated from data sources and literature estimates. These parameters were then implemented into the POHEM-OA model to project the total cost of OA from 2010-2031. In this Chapter, I also discuss how each cost component has been projected into future years using the POHEM-OA model. 4.2.1 Projecting Health-Care-Sector Prices and Service Utilization Rates I computed the input parameters in the model using the PDBC database for 2003 (84), Ministry of Health OA survey data from 2007 (147), and literature estimates from different years between 2001 and 2010 (as discussed in Section 3.3). However, since the time period for this study is from 2010 to 2031, I needed to bring all the cost parameters to 2010 $CAD at baseline. In addition, the unit prices within each category are projected to increase between 2010 and 2031 due to health-care-sector inflation. As 113 reported in Cost Drivers of Health Expenditure in Canada (161), the health-care-sector inflation in preceding years was higher than the overall economy’s inflation; however, the inflation is different across the cost categories. Therefore, I used different price indexes from 2003 to 2008 for each category, including hospitalization (151), physicians (153), drugs (152), and the overall economy CPI (134) for out-of-pocket costs (formal caregiver, medical aids, transportation and other costs). I used historical data from national reports for each price index. On top of the health-care-sector inflation, the utilization rates for hospitalization, drugs, and physicians have been increasing during the past decade due to government policies, patient demand, and technology improvements (161). In this study, I only implemented the increase in hospitalization rates for the hip/knee TJR surgeries. Other cost drivers, such as changes in the population distribution, age and BMI or increases in expected life of the population, were already implemented in the POHEM-OA model (Appendix A1), while I did not include potential per-capita increase in drugs and physician utilization in this study. In the following section, I discuss how I have implemented the increases in health services utilization and healthcare prices. Hospitalization Costs The hospitalization costs for OA are associated with several types of hospital procedures performed on joints (knees, hips, hands, shoulders, ankles, feet, and other joints). Hip/knee TJR surgeries stand out as the most costly and resource intensive of all OA-related procedures, with long lengths of stay after surgery (an average of 5-6 days in 2010 (126)). Other OA-related hospital procedures such as shoulder TJR are also included in the analysis and a separate analysis was performed for those. 114 For projecting the cost of hip/knee TJR over the next 20 years, I considered both the future trend for the number of TJR events as well as the future trend in unit costs for knee/hip TJRs. The final costs for hip/knee TJRs were calculated and extrapolated for the next 20 years, separately for primary and revision surgeries and for different age categories. Projecting the number of hip/knee TJRs: As mentioned in Chapter 3, the number of hip/knee surgeries was calculated from the data provided in a report by the Canadian Joint Replacement Registry (CJRR) for the period 1994/95-2009/10 (124). The CJRR captures approximately 42% of the hip replacements surgeries and 38% of the knee replacement surgeries performed in Canada between 2009 and 2010 (124). Based on CJRR data, the number of primary TJR surgeries in Canada increased by 100% during the 10 years from 1999/2000 to 2009/2010 (124). To include these changes, I extrapolated the number of primary surgeries for each sex, and year, into the future using the previous 15 years of data with a linear model (Figure B2, Appendix B). In addition, I calculated the age/sex distribution of the number of surgeries for that period, based on data from the CJRR (124). Table B3 in Appendix B shows the overall number of surgeries from 2010 to 2031. In the next step, I calibrated the current POHEM-OA results for the number of surgeries. To this end, I compared the extrapolated number of surgeries with the results of the POHEM-OA model and then calculated the multipliers for each age and sex category, as defined by dividing the extrapolated results by the POHEM-OA results. I then implemented the calibrated results into the POHEM-OA model using the reciprocal 115 of the multipliers for TJR hazard ratios that were systematically changed until the extrapolated results for each age and sex category were reached. Projecting the expected length of stay after hip/knee TJR surgery: The length of acute stay (LOS) after a TJR surgery is one of the major components of the cost of knee/hip TJR surgeries. According to the data from CJRR (124), the LOS after TJR surgery has shortened throughout the past decade. Between 2006 and 2007, knee replacement recipients spent a median of four to five days in hospital, compared to eight days in 1996–1997, a decrease of 50% for males and 37.5% for females (126). The length of stay for hip TJR was equal to those having knee TJR from 2006 onward (126). To extrapolate the LOS into future years, this analysis was performed separately for primary and revision replacement surgery and for each age categories (18-60, 60-80 and 80+). Although the observed LOS differed slightly for females and males, I assumed the same expected length of stay (ELOS) for both sexes in the model. I also averaged the hip and knee ELOS as I only have on type of surgery in the POHEM-OA model. To estimate the ELOS of primary and revision TJRs for the next 20 years, I used a negative exponential model that fit the first 10 years of the observed LOS well ( R2=0.97). This model is shown in Figure B3, Appendix B that represents the ELOS from 2010 to 2031 for primary surgery and averaged for all ages. Projecting the unit cost of hip/knee TJR: I used the future trend for ELOS after a TJR surgery to calculate the cost of TJR over the next 20 years. As discussed in Section 3.2.3, the cost of a hip/knee TJR has three major components: 1) hospital stay cost that includes nursing wages and other costs which was calculated based on length of hospital stay and per diem cost rates taken from the CIHI hospitalization cost 116 database (144)- per diem cost rate is defined as a total Resource Intensity Weight (RIW) divided by LOS for each procedure (144); 2) procedure cost and cost of implant, which includes medical imaging, surgery, and the cost of materials and staff during surgery; 3) orthopedic surgeon cost. The procedure cost was calculated using functional area percentages from the CIHI database for the different types of TJRs. The cost of the implant is the cost of the materials used in the artificial joint that is replacing the actual joint of the patient during surgery. The cost of TJR surgeries changed over time due to changes in length of stay and increases in the costs of hospitalization and implants due to inflation (52). I first calculated the cost of surgery at baseline in 2010 from CIHI hospitalization cost database (144), without considering the inflation. The weighted average for a TJR surgery at each year was calculated across different types of surgeries, including unilateral/bilateral, hip/knee, and with/without infection, separately, for primary and revision surgeries. I used the volume of each TJR type as weights to arrive at the overall weighted average cost of hip/knee-TJR surgery. For each of the mentioned surgery types, I used per diem rate multiplied it by the ELOS for each year, and finally multiplied it by the average Canadian CPWC, all reported in CIHI hospitalization cost database for hip/knee procedures (144). Finally, I projected the costs for TJR hip/knee surgery for the period between 2010 and 2031 (without inflation) for three age categories and for primary and revision surgery, separately. Details of calculating the above calculations are discussed in Appendix A3. To calculate the inflation rate of a hip /knee TJR surgery, I used separate inflation rates for each cost component, and then calculated the average weighted inflation rate 117 using the weight of each cost component in the average cost of TJR surgery. To project the inflation rate into future years, I used CIHI estimates for the hospital staff wages price inflation (151), in addition to annual increase of the orthopedic surgeon fee from 2005-2010 data (154), and the literature estimates for hip/knee implant price inflation (52). Table 4.1 presents the inflation rates for cost components of the hip/knee TJR and other OA cost categories. Since the CIHI hospitalization cost database (144) does not report the orthopedic surgeon cost, I used data from the St. Paul’s Hospital in Vancouver, Canada (145) and calculated the weight of each cost component for the TJR surgery in 2003. Details for inflation rate calculations for hip/knee TJR are discussed in Appendix B1. I calculated the weighted average inflation rate for hip/knee TJRs to be 4.6% in 2010. I assumed the same annual inflation rate for TJR surgery from 2010-2031 (4.6%). Finally, using the average inflation rate projected from 2010 to 2031 and the (non-inflated) hip/knee TJR cost as discussed above, I calculated and extrapolated the average cost of hip/knee surgery for the next 20 years (in 2010 $CAD). This surgery cost was calculated separately for primary and revision surgeries across three age categories (18-59, 60-79, 80 and older) as shown in Table B4 in Appendix B. Although in POHEM-OA, I had only people with 20 years of age and more, the age categories from CIHI hospitalization cost database (144) started from 18 years of age and, therefore, I assumed the same cost assignment. The hip/knee TJR surgery costs presented in Table B4 include inflation rates for cost components (procedure, inpatient stay and implant costs) as well as the projected decreasing trend for ELOS. For hospital procedures other than hip and knee TJR surgeries, I assumed same inflation rate as the 118 hospital consumer price index reported by CIHI hospital Price Index Report (151), which was 3.6% for the average of the last 5 years from 2007-2011 (Table 4.1). Drugs, Physician Visits, and Out-of-Pocket Costs For the inflation in the price of drugs, I used the average rates of the price index of drugs for over-the-counter (OTC) and prescription drugs from 2009 to 2012, as reported in the CIHI Drug Expenditure in Canada Report (152). As shown there, during the last 5 years, inflation rates for prescription drugs had been lower compared to OTC drugs (152). For the physician fee price index, including physiotherapists and other alternative care categories I used the CIHI-physician price index curve for physicians’ fees for the period 2003-2008 from the National physician Data Base (NPDB) as reported by CIHI Physician Expenditure Report (153). I assumed the same inflation rates for laboratory and other tests including X-rays, MRIs, blood work, and other outpatient costs. For both drugs and physicians price indices, I projected the price index from 2010 to 2031 using average of the recent years data (Table 4.1). I used data from 2003 to 2008 for the physician price index and from 2009 to 2012 for the drug price index, as there was a major shift in CPIs in years prior to 2003 and 2009 for drugs and physicians, respectively (152, 153). For out-of-pocket cost inflation occurred before the TJR surgery, including formal caregiver wages and general healthcare commodities, such as medical aids, transportation cost because of OA, community costs, and home remodeling due to OA (all of which were included in the formal care category), I used the overall economy CPI as reported by Statistics Canada for years 2005 to 2011 (134). For other out-of-pocket 119 cost calculations, I projected the overall economy CPI into the future using an exponential trend model for annual inflation rates in the next 20 years, as it produced a good fit to the previous 6 years of data (2005 to 2011). Details of the CPI projection are shown in Figure B1 and results of CPI from 2010 to 2031 are shown in Table B2, Appendix B. For the side effects of drugs (CVD, GI, dyspepsia and stroke), I also used the overall economy CPI (Table B2). Projecting the rehabilitation cost: I assumed that the probabilities for rehabilitation routes would remain the same in the future years. As discussed in Chapter 3.2, these routes include discharging the patient after the hip/knee TJR surgery to a rehab facility, moving from a rehab facility to home-care, and/or moving from home-care to self-care (Figure 3.1). Additionally, I assumed that both home-care and self-care discharge destinations are associated only with out-of-pocket costs, where self-care is a small amount only due to transportation and community costs for OA patients (115), while home-care includes higher costs for formal caregivers, physiotherapists and other out- of-pocket costs (136). For home-care and self-care costs (out-of-pocket), I used the projected overall economy CPI, as reported by Statistics Canada (134) and projected it into the future according to an exponential trend model for the years 2010-2031 as shown in Table B2, Appendix B. However, for the hospital-based rehabilitation cost component, I used the hospitalization inflation rate, which was 3.6% according to the CIHI Hospital Price Index Report (151) as shown in Table 4.1. 120 Table 4-1. Annual inflation rates for each cost component Cost Category Cost Subcategory Cost component Data Source (year) (average annual (annual Inflation rate %) [reference] inflation rate) 1. Hospitalization Hip/Knee TJR Surgery1 Orthopedic surgeon fee-for CIHI-PSF2 costs (4.6%) service price index (2.1%) (2005-2010) Surgical Implant inflation rate Weinstein et al., (9.8%)3 (2013) [52] Hospital employee wage CIHI-HPI index (3.6 %) (2007-2011)4 Weights for each cost St. Paul’s hospital component used in average costing model inflation calculation (2003) [145] Inpatient procedures other Hospital employee wage CIHI-HPI than hip/knee TJR index (2007-2011)4 surgery (3.6%) (3.6%) 2. Physician and Physician visits and lab Physician fee-for service price NPDB- PPI outpatient care tests (3.2%) index (3.2%) (2003-2008)5 costs 3. Drugs Prescription (0.9%) Prescription-based generic CIHI-DER 6 and brand name drugs price (2009-2012) index (0.9%) Over-the-counter (1.6%) Over-the-counter generic and CIHI-DER 6 brand name drugs price index (2009-2012) (1.6%) 4. Rehabilitation and Healthcare system paid Hospital employee wage CIHI-HPI home care costs rehabilitation and home index (2005-2010) (after hip/knee care costs (3.6 %) surgery)7 Out-of-pocket-cost during Consumer price index (CPI) CANISM rehab (2.1%)8 for overall economy 8 (2005-2011) 9 5. Formal caregiver Formal caregiver, and Consumer price index (CPI) CANISM and other costs community costs for overall economy 8 (2005-2011) 9 (before hip/knee including transportation, surgery)7 medical aids and home remolding (2.1%) 6. Alternative care Physical therapy, Physician NPDB- PPI costs Chiropractic’s, fee-for (2003-2008)5 acupuncture and other service alternative care price index professionals (3.2%) 7. Side effect of Cardiovascular diseases Consumer price index (CPI) CANISM drug’s (CVD’s) for overall economy (2005-2011) 9 Stroke, Gastrointestinal (GI), Dyspepsia 121 Table 4.1 notes: 1 per diem cost weights for different procedures related to TJRs calculated from CIHI hospitalization cost database, CMG client tables (144) for procedures, nursing costs, imaging, indirect and other functional areas (Section 4.2.1); 2 CIHI-PSF: CIHI report on physician and surgeon fees from 2005- 2010 (154), 3 Implant inflation is due to combined effect of technological improvement and general inflation. 4 CIHI-HPI: CIHI report on hospital price index (151) for full time hospital employee wage index from 2007-2011; 5 CIHI-PPI: CIHI report on physician fee-for-service price index from 2003-2008 that used data from NPDB (153). 6 CIHI- DER: CIHI-Drug Expenses Report (152) that I used the average of generic and brand name price indices; 7Out-of-pocket cost during rehab and formal caregiver costs are reported separately as the former reflect the out-of-pocket cost after the TJR surgery (rehab) and the latter is the cost before the surgery. 8 CPI for overall economy was calculated based on an exponential extrapolation from historical data (2005-2011) from Statistics Canada CPI Report (134). Every year has different CPI, but the average over the study period was 2.1%; extrapolation results are in Appendix B, Figure B1 and Table B2. 4.2.2 Discount Rate and the Overall Economy Inflation Rate The total cost of OA in future years (2011-2031) was discounted back to 2010 Canadian dollars. For the discount rate, I used the social discount factor of 3.5%, as reported in the recent article (Boardman et al. (2011). According to Boardman et al. (155), the correct method is to discount future economic impacts based on the rate of social time preference (STP). This has been estimated to be d=3.5%, i.e., lower than the usual 5% recommended in previous guidelines (e.g., ISPOR-2003 (162)). The current dollar value for each year can be reported only for the past years and not the future years and therefore, for future years, I need to discount the dollar value back to the current or previous years dollar value. As a result, in contrast to Chapter 3 where I 122 also reported the cost in current dollars, in this Chapter, I reported all the total cost outcomes for years 2010 to 2031 in 2010 $CAN. For calculating the interest rate, I also accounted for the overall economy inflation according to the Fischer equation (155). In fact, I deducted the discount rate from the overall economy inflation rate and calculated the adjusted interest rate in order to discount the future cost back to 2010 $CAD. For the overall economy inflation rate, I used the overall economy CPI from 2005-2011 as reported by Statistics Canada (134) and used an exponential trend model to project it from 2010 to 2031 as shown in Appendix B, Figure B1. The average overall economy inflation rate from 2010-2031 was therefore calculated to be (π=2.1%). CPI from 2010 to 2031 is listed in Table B2, Appendix B. Although all total direct cost outcomes in this Chapter have been discounted back to 2010, I have performed three scenarios for analyzing the effect of discount and inflation rates on the total direct cost of OA. The first goal of designing these scenarios were to compare the effect of health-care-sector inflation rates associated with each cost component on the overall total direct cost of OA. The second goal of designing the scenarios was to compare the total direct cost of OA in case of actual and nominal discount rate. In this section, I denote the inflation rate with π and the nominal interest rate with i. According to the Fisher equation, the real interest rate ( or discount rate, r), is equal to r = i-π. In fact, the nominal interest rate does not include the overall economy inflation in future years, while actual interest rate does, by deducting the interest rate from the overall economy inflation rate (155). For example, in my setting, the overall average 123 economy inflation rate will be 2.1% for 2010-2031 and therefore, nominal interest rate will be i=3.5%, while the actual (real) interest rate will be r= i-π = 3.5%-2.1%=1.4%. The total direct cost of OA between 2010-2031 will be estimated for three scenarios: 1) no inflation scenario: no health-care-sector inflation used for cost components, while costs were discounted using the nominal interest rate in 2010 $CAD; 2) nominal cost scenario: health-care-sector inflation was used for cost components, while costs were discounted using the nominal interest rate in 2010 $CAD; 3) actual cost scenario: health-care-sector inflation was used for cost components, while costs were discounted using the actual interest rate in 2010 $CAD. As mentioned above, for the actual interest rate, I take the overall economy inflation rate into account according to Fisher equation (155). For each scenario, the formula involves five cost categories as follows: 1. TJR surgery (s), physician (p), drugs (d), rehabilitation (rh), out-of-pocket cost (op), and side effects of drugs (se). Scenario 1. No Inflation Assuming c denotes the cost per OA patient for each category and u denotes the number of OA patients to which the cost is assigned, I would have the following equation (Eq.4-1) for the total cost, Ct for year t in Scenario 1 (in 2010 $CAD). Also, i in the denominator represent nominal interest rates and t is the year for the analysis. cs × us + cp × u p + cd × ud + crh × urh + coc × uoc + cse × use C = Equation 4-1 t (1+ i)t 124 Scenario 2. Nominal Cost Using the same notation as above, with π X denoting the health-care-sector inflation rate for cost component X, I have the following equation (Eq.4-2) for the total nominal cost of OA in each year (all in 2010 $CAD): c × u × (1+ π )t + c × u × (1+ π )t + c × u × (1+ π )t + c × u × (1+ π )t + c × u × (1+ π )t + c × u × (1+ π )t C = s s s p p p d d d rh rh rh oc oc oc se se se t (1+ i)t Equation 4-2 As mentioned earlier, in this scenario, health-care-sector inflation is included (numerator of Eq. 4-2), while the total cost outcome does not reflect the overall economy inflation (denominator of Eq. 4-2); therefore the interest rate is a nominal rate and the outcome of scenario 2 will be in (2010 $CAD), but it is not adjusted for the overall economy inflation rate. Scenario 3. Actual Cost The total direct cost of OA for this scenario is defined according to (Eq.4-3). In this scenario, I include health-care-sector inflation for each cost component (numerator of Eq.4-3), in addition to using actual interest rate (denominator of Eq.4-3). As a result, the cost is reported in (2010 $CAD) and adjusted for overall economy inflation rate: c × u × (1+ π )t + c × u × (1+ π )t + c × u × (1+ π )t + c × u × (1+ π )t + c × u × (1+ π )t + c × u × (1+ π )t C = s s s p p p d d d rh rh rh oc oc oc se se se t (1+ i − π )t Equation 4-3 4.2.3 Internal Validation I performed an extreme sensitivity analysis (SA) (163) for achieving internal validation (verification) of the direct cost module of the POHEM-OA model developed in this study. I have investigated the integrity of the results of the model through two types 125 of internal validation techniques according to two scenarios for performing the extreme SA: 1) Assigning zero unit costs to state-based cost categories, i.e., physician, drugs, and hospital procedures (other than hip/knee TJR surgery, and 2) Setting the probability of event cost categories to zero for hip/knee TJR surgery, rehabilitation, formal caregiver, side effect of drug costs and alternative care costs. I first performed simulation runs separately for each validation scenario by assigning zero unit costs for each cost category and zero probability for events; then, I investigated if the summation of the total direct cost of all scenarios would add up to the result of the base case for the years 2010, 2021 and 2031. This type of analysis aided us in understanding error in cumulating the costs over all individuals at different times such as double counting. For example, after assigning zero unit costs to physician cost component, the overall total direct costs in 2031 was reported to be lower than the base case. This discrepancy aided us in finding the coding error in cost calculations. In another type of internal validation examination, I have randomly selected 1000 individual cost profiles from all individuals whose life histories were simulated by the POHEM model (~25 million individuals in 2010), and evaluated the change in the value of cost components for each. I repeated this analysis for each of the above validation scenarios to investigate if the change in unit cost affects the cost profile of individuals appropriately. Two cost profiles for randomly selected patients are shown in Figure B4 and B5 in Appendix B. This type of analysis assisted us in understanding errors in coding at the individual level for each cost component. For example, when the drug cost component of the patient is still the same after the age of the patient has changed, I have found out that the code related to the individual-level drug cost component were 126 not properly written, since whenever the age of the patient enter a new category, it should trigger the change in drug cost. 4.3 Results According to the POHEM-OA projection results, the size of the adult Canadian population (20 years of age and older) is projected to increase from 26.3 million in 2010 to 32.2 million in 2031, a 22.5% increase as shown in Table 4.2. During the same period, however, the OA population should rise from 3.6 million to almost 6 million, a 64% increase based on the POHEM-OA projections (Table 4.2). The prevalence of OA is much higher for females than males. As shown in Figure 4.1 for the results of the POHEM-OA model, OA prevalence for females is projected to increase from 0.16 to 0.21, while male prevalence is projected to increase from 0.11 to 0.16 throughout the study period. Table 4-2. Characteristics of OA patients and of general population 2010 2020 2031 Average age (OA population) 67.4 68.9 70.7 Average age (General population) 47.8 49.9 52.3 Average BMI** (OA population) 27.2 27.7 28.2 Average BMI (General population) 26.1 26.6 27.1 OA population size* 3,653,787 4,736,082 5,991,299 General population size (>20 years)* 26,260,862 29,677,839 32,161,109 General population size *** 31,211,883 35,338,058 38,303,539 * The POHEM-OA model only include adult Canadian population (20 years of age and older); ** BMI unit: (kg/m^2); *** projected by census (not in POHEM-OA) for overall population at any age (164) 127 The higher rate of increase in the OA population compared to that of the general population can be attributed to an increase in the age and BMI of the population. As shown in Table 4.2, the mean age of the general population will increase from 48 to 52 years during the period between 2010 and 2031, and there is a 12% increase in the 60+ population. The mean BMI of the general population is increasing by 4% (from 26 to 27 Kg/m2). While the mean age of the OA population will increase from 67 to 71, people aged 70-80 will become a significant proportion of those with OA, comprising 34% of that population by 2031. However, while the general population’s share of adults over the age of 80 will increase by 2%, the share of this age group among persons with OA will decrease by 3% (figure not shown). Figure 4-1. Prevalence of OA for males and females from 2010 to 2031 * Vertical axis represents the prevalence of OA (1 means 100%) 128 Three scenarios representing the change in the total cost of OA from 2010 to 2031 according to different inflation and discounting rates are depicted in Figure 4.2– scenarios were discussed in Section 4.2.2. All the scenarios reflect the discounted cost for total cost of OA by backdating it to 2010 $CAD. While no-inflation scenario (Scneraio1) does not include the inflation in prices of healthcare resources, nominal cost (Scenario 2) and actual cost (Scenario 3) scenarios incorporate the inflation rates specific to each cost component as shown in Table 4.1. According to the results of the actual cost scenario presented in Figure 4.2, the total direct cost of OA escalates from $2.9 billion to $7.6 billion dollars from 2010 to 2031 (in 2010 $CAD), a 162%, or almost 2.6-fold increase in total OA-related direct costs. The actual cost scenario reflects inflation rates for all resources in addition to utilizing the actual discount rate that incorporates the overall economy inflation rate as well. The projected inflation of the health-care-sector is one of the cost drivers of the direct cost for OA that is reflect in the actual scenario results. The extra direct cost associated solely with inflation in 2031 was $2.4 billion, of which $1.8 billion was related to health care system costs and CAD$ 600 million to out-of-pocket costs (all in 2010 $CAD) (figure not shown). On the other hand, the nominal cost scenario shows a rise in total costs from $2.9 to $4.4 billion in 2031. This scenario includes the health-care- sector inflation rates and uses the nominal discount rate of d=3.5%, which does not incorporates the overall economy inflation rates and therefore the total direct cost results in 2031 are lower in nominal cost scenario compared to actual cost scenario. The difference between these two scenarios shows the different direct costs associated with OA under two economical growth scenarios, conceptual one with no overall 129 economy inflation vs. realistic one with projected rate of overall economy inflation. The nominal cost scenario result, i.e., $4.4 billion in 2031 (2010 $CAD), is in fact the lowest possible cost associated with direct cost of OA in that year which is independent of the overall economy inflation. In fact, no matter how the overall economy grows over the next decades, overall direct cost of OA would be at least $4.4 billion in 2031 according to the POHEM-OA results. While both of the above scenarios, nominal and actual cost, incorporated the inflation for health-care-sector prices, in the no-inflation scenario I assume all health- care-sector prices remain fixed, and there is no inflation in the overall economy (nominal discount rate). According to the results of no-inflation scenario shown in Figure 4.2, the increase in the total direct cost of OA will be leveled off by the decrease in value of the money from 2010 to 2031. In other words, no-inflation scenario reveals that if around 3 billion dollars can be invested in 2010, it would suffice to pay the direct cost of OA in 2031 just due to the interest rate of the investment (assumed to be equal to the discount rate of 3.5%). The three scenarios reflect the change in total direct cost of OA from different perspectives. However, since the actual cost scenario (Scenario 3) is the closest to a real economic situation, the results in next sections will be according to the actual cost scenario. 130 Figure 4-2. Total direct cost of OA for different inflation scenarios* $11.00$ $10.00$ Billions$ $9.00$ $8.00$ $7.00$ $6.00$ $5.00$ $4.00$ $3.00$ $2.00$ $1.00$ $0.00$ 2010$ 2011$ 2012$ 2013$ 2014$ 2015$ 2016$ 2017$ 2018$ 2019$ 2020$ 2021$ 2022$ 2023$ 2024$ 2025$ 2026$ 2027$ 2028$ 2029$ 2030$ 2031$ no4infla7on$ Nominal$cost$ Actual$cost$ *All costs are discounted back to 2010 $CAD using nominal discount rates (in Scenario 1, 2) and actual discount rate (Scenario 3). Three scenarios represent the increase in total direct cost of OA according to different inflation and discount rates. Health-care-sector price indices represent inflation for different cost components that were incorporated in both of (Scenario 2. nominal cost) and (Scenario 3. actual cost), but were not included in (Scenario 1. no inflation)- scenarios are defined in details in Section 4.2.2. 131 Total Direct Cost by Sex and Age Category Direct cost for females with OA is $ 0.71 billion and $1.85 billion higher than those for males in 2010 and 2031, respectively (Figure 4.3). This is a result of different pattern of resource utilization in addition to different cases of males and females with OA. As shown in Figure 4.3, the share of female cost in both out-of-pocket and healthcare system costs was higher than that of males. The out-of-pocket cost for females and males associated with OA in 2010 was $0.52 billion and $0.14 billion that increased to over $1.12 and $0.47 billion in 2031 (all in 2010 $CAD). The percentage increases in the out-of-pocket and healthcare system costs of all patients with OA were 121% and 172%, in 2031 compared to 2010. Figure 4-3. Sex-specific healthcare system and out-of-pocket cost of OA and effect of inflation $9.0$ Billions' $8.0$ $7.0$ $6.0$ $5.0$ $4.0$ $3.0$ $2.0$ $1.0$ $0.0$ 2010$ 2011$ 2012$ 2013$ 2014$ 2015$ 2016$ 2017$ 2018$ 2019$ 2020$ 2021$ 2022$ 2023$ 2024$ 2025$ 2026$ 2027$ 2028$ 2029$ 2030$ 2031$ Out'of'Pocket'Cost'(Female)' Health'System'Cost'(Female)' Out'oF'Pocket'Cost'(Male)' Health'System'Cost'(Male)' * All costs in the vertical bar are in 2010 $CAD and presented in billion dollars 132 While patients with the highest OA cost in 2010 are those aged between 60-70 years, the costs for patient in other age categories tend to increase to such a degree that, in 2031, OA patients aged between 70-and 80 years will have become the highest cost category. For females, the change in the highest age category from 60-70 to 70-80 occurs in 2020 and for males in 2026. As shown in Figure 4.4, between 2025 and 2030, the costs for those aged 70-80 and 80-90 will increase at the highest rate for both females and males. In fact, by the end of the study period, the overall cost for these age categories will have tend to increase more,!while the cost for the other age groups will reach a steady state. As a result, the direct cost associated with patients 70+ would become the major contributing factor to the total cost of OA in the years 2025-2031 and beyond. 133 Figure 4-4. Total direct cost of OA by age category for males and females* * Results are in actual costs (Scenario 3) reflecting inflation and are in (2010 $CAD). The cost categories are non-overlapping: (0,49), (50, 59), (60-69), (70,79) and (80-89) and over 90. Females 70-79 years of age have the highest share of the total cost in 2031 with $1.6 billion (2010 $CAD). Total Direct Cost by Cost Component While the cost of hip/knee TJR surgeries increases from $700 million dollars in 2010 to $2.8 billion dollars in 2031, the overall costs for physician, OTC and prescription drugs increases from about $1 billion to $2.3 billion (all in 2010 $CAD). At the same time, the cost of side effects from OA drugs rises from $ 305 million to $ 683 million and 134 the rest of the costs including alternative care, rehab, formal caregivers, and other out- of-pocket costs goes from $767 million to $1.9 billion during the same period. As shown in Figure 4.5, while the dollar amounts of all the cost categories increased during the study period, each cost category’s share of the total direct cost has shown a more complex trend. The share of overall costs for physician, OTC drugs, and prescription drugs, in addition to alternative care and formal caregiver costs have decreased while hospitalization costs have increased. The share of rehabilitation and side effect costs has stayed almost the same since 2010 (Figure 4.5). Figure 4.6 represents the out-of-pocket and healthcare system share of the total direct cost of OA and their cost components associated with each in 2010 and 2031. As shown in Figure 4.6, although all cost categories that make up the out-of-pocket costs increase in terms of dollar value from 2010 to 2031, the share of drugs and rehab costs increases, while the formal caregivers’ share of the out-of-pocket cost decreases during the same time period. On the other hand, for the healthcare system costs, the share of all cost categories other than hospitalization cost decreases (or stays the same), while hospitalization costs increase by 12% during the study period (Figure 4.6). The total cost of side effects due to OA drugs increases from $305 million dollars in 2010 to $689 million dollars in 2031 as shown in Figure 4.6 (all in 2010 $CAD). This corresponds to the increase in cost associated with dyspepsia, CVD, stroke and serious GI complications. The total number of cases of all of the mentioned side effects rose from 1300 to 2200 cases, an almost 70% increase, due to the side effects of drugs consumed by OA patients during the study period. In 2031, almost 75% of all of these adverse cases are related to dyspeptic symptoms, 10% to serious GI complications, 8% 135 to CVD and 7% to stroke. However, due to high lifetime cost per case of CVD and stroke compared to dyspepsia and GI complications, 62% of the total cost of side effects is due to cardiovascular disease ($427 million), 8% stroke ($125 million), 16% dyspepsia ($110 million) and 4% GI complications ($27 million) in 2031. The total cost of stroke and CVD due to OA drugs are projected to increase by almost 144%, from $228 million to $551 million, while total cost of dyspepsia and serious GI complications due to OA drugs are projected to increase by 81%, from $76 million to $138 million during the study period (all in 2010 $CAD). 136 Figure 4-5. Share of cost categories of total direct cost from 2010 to 2031 40%# 35%# 30%# 25%# 20%# 15%# Percentage)of)Total)Cost) 10%# 5%# 0%# 2010# 2011# 2012# 2013# 2014# 2015# 2016# 2017# 2018# 2019# 2020# 2021# 2022# 2023# 2024# 2025# 2026# 2027# 2028# 2029# 2030# 2031# Hospitaliza6on#cost*## Ongoing#cost#(#drugs,#physician#and#ambulatory#cost)# Cost#due#to#sideHeffect#of#drugs# Alternate#care#and#formal#caregiver#cost#**# Rehabilita6on#cost#and#homecare#serivces# ! *Hospitalization costs include those related to hip/knee total joint replacements (TJRs) and other OA-related hospital procedures such as arthroscopy, shoulder TJR; ** Alternative care costs include those related to physiotherapy visits, chiropractor visits, and other alternative care professionals (the complete list is presented in Appendix A, Table A12); Formal caregiver cost includes paid caregiver cost in addition to community services cost such as home remodeling and transportation as listed in Gupta et al. (115). 137 Figure 4-6. Total direct cost of OA by cost categories in 2010 and 2031: out-of-pocket and healthcare system cost Out$of$pocket+cost+2010+ Rehab%(out%of% Out$of$pocket+cost+2031+ Rehab%***%% pocket%)***% $314,630,847% $68,056,910% 12%% 8%% Drug%%(over%the% Drug%(over%the% counter)% counter)% $196,927,136% $625,532,112% 24%% 25%% AlternaAve%care% $224,310,992% AlternaAve%care%% 28%% $716,595,239% 28%% Formal%caregiver%^% Formal% $323,714,041% caregiver^% 40%% $885,936,937% 35%% Healthcare+system+cost+2010+ Healthcare+system+cost+2031+ HospitalizaAon%%*% SideQeffect%of% $743,739,126% SideQeffect%of% drugs% 38%% drugs%% HospitalizaAon%,% $683,925,117% Other% $305,221,026% $2,986,689,923% Other% 12%% $456,607,711% 16%% 50%% $1,155,948,681% 20%% Rehab%% Physician%% Rehab%(%healthcare% (%healthcre% $538,399,749% system)**% system)**% 27%% $151,386,685% Physician%% $472,023,564% $1,258,320,062% Drug%% 8%% Drug%% 8%% 21%% (prescripAon%)% (prescripAon)% $208,124,406% $507,599,764% 11%% 9%% ! * The hip/knee TJRs and hospital procedures other than hip/knee TJRs; ** rehabilitation and home care cost that is paid by the government after the hip/knee TJR. *** rehabilitation and home care cost that is paid by patients that includes formal caregiver, home remodeling and transportation cost during the year after the surgery; ^ formal caregiver cost and community care services (meals on wheels, transportation and home remodeling) paid by the patients before surgery. I assumed all prescription drugs are paid by the healthcare system. However, in real life, not all prescription drugs are paid for by healthcare system insurance in Canada. 138 4.4 Discussion I have performed a COI study integrated with a microsimulation model to predict the total cost of OA over the next 20 years (2010-2031). According to Canadian Health Expenditure reports in 1994, the total direct OA cost in Canada was projected to reach around $1.1-$1.6 billion dollars in 2010 (95). This increased to $2.1 billion in 2006 according to Anis et al. (94) that used the expenditure data in 2006. My estimate for the total cost of OA in 2010 was $2.7 billion dollars in 2010 $CAD. This result was close to two previous Canadian studies’ estimates using population-based database (and not sample surveys) for the overall direct cost of OA (94,71). However, it should be noted that none of the previous studies included out-of-pocket costs for patients with OA. In this study, I have shown that the overall direct cost of OA will rise from $2.9 billion in 2010 to almost $7.6 billion in 2031, if inflation is included (in 2010 $CAD). I have also shown that the cost for hip/knee TJR surgery makes up the largest part of direct cost of OA. In this study, I used the results of a previously developed micro-costing COI study discussed in Chapter 3 as an input to the simulation model, projected the costs of each category for OA into the future, and used a microsimulation model to predict the number of Canadians with OA by age, sex, and disease state over the next 20 years. Although most of the input parameters of the cost model were estimated from data over a relatively short period of time, my results can be interpreted as a bottom-up incidence- based COI study due to the use of a longitudinal microsimulation model. While use of estimates from top-down, prevalence-based COI studies as the basis for decision making have been disputed in the literature (57,62), this caveat may not apply to studies 139 using microsimulation. For example, although I used only 1 year of data from the MSP data for physicians fee, the model predicts the increase in number of OA patients as mean age and BMI of the population increases, on one hand, and I also included the inflation rate for the price of physicians’ services on the other hand, using historical changes in the physicians’ fees as reported by CIHI (153,154). For hip/knee TJRs, I estimated the increase in the number of hip/knee replacements surgery over time using CJRR from 1990 to 2010 (124), in addition to the increase in its unit costs, and estimated its future trend. However, other than the case for hip and knee TJR surgeries, I did not capture the potential future increase in per patient-year cost due to increases in healthcare resource utilization, such as physician visits, or drug use that may happen because of changes in patient behavior or introduction of new technologies. The limitations and advantages of my approach will be discussed further in Chapter 6. In this Chapter, I have compared the total direct cost of OA across three different scenarios with respect to different discount and inflation rates. The three scenarios defined in Section 4.2.2 reflect the change in total direct cost of OA according to different health-care-sector and overall economy inflation rates, while all report the costs in 2010 $CAD. For example, in the first scenario, no inflation scenario, I showed that the overall increase in the total direct cost of OA would be stable around $2.9 billion from 2010 to 2031. In this scenario, the increase in the cost of OA will be offset by the decrease in the value of money. This points out the significant effect of the cost components’ price inflation over the study period. In fact, if there were no health-care- sector inflation in the price of cost components of OA, the interest rates of the $2.9 140 billion investment cumulated during the study period would suffice to pay off the total cost of OA in 2031 (assuming 3.5% interest rate annually from 2010-2031). Result of scenario 2, the nominal cost scenario, depicts the change in the total direct cost of OA in a situation where no inflation were assumed for the overall economy and therefore a nominal discount rate was used to backdate the cost to 2010, i.e., d=3.5% every year between 2010-2031. This scenario shows the change in total cost of OA in a conceptual economy with no overall economy inflation. However, in this scenario, I assumed prices for each cost component would be increased because of other reasons such as technological improvement or change in demand. I showed that in this scenario, the total cost would rise from $2.9 billion to $ 4.4 billion (2010 $CAD). In fact, independent of any changes in the overall economy growth, the total direct cost of OA would be at least $4.4 billion in 2031 (2010 $CAD). Finally, in scenario 3, the actual cost scenario, actual discount rate was calculated according to Fischer equation, which was on average 1.4% during the study period (Section 4.2.2). I showed that the direct cost of OA increased from $2.9 billion to $7.6 billion dollars. The difference between costs results in scenario 2 and scenario 3 shows the different direct costs associated with OA according to the effect of the overall economy inflation or a realistic projected rate of inflation. Therefore, the nominal cost scenario result, i.e., $4.4 billion in 2031 (2010 $CAD), is in fact the lowest possible cost associated with direct cost of OA, and is independent of the overall economy inflation. On the other hand, the actual cost scenario result, i.e., 7.6$ billion in 2031 (2010 $CAD), is the realistic estimate of total direct cost of OA if the overall inflation rate would 141 be continued to increase in a linear fashion according to the historical data for the overall economy CPI (the CPI projection model is provided in Figure B1, Appendix B). In this Chapter, I showed that how for chronic diseases such as OA, the use of microsimulation models could be used in conjunction with an individual-level COI could be used to project the direct cost-of-illness into the future and investigate the effects of different scenarios on the cost burden. The approach discussed in this study can be used for the projection of the cost burden for other chronic diseases where heterogeneity among patients is significant and could affect the total cost (13). The detailed discussions regarding advantages and disadvantages of the approach used in this Chapter are provided in Chapter 6. 142 Chapter 5: Uncertainty Analysis in Population-Based Disease Microsimulation Models 5.1 Introduction9 Uncertainty analysis (UA) plays an important role in the validation of simulation models and interpretation of their results (2,3,32, 33,165). The purpose of UA is to provide uncertainty intervals around the mean estimate of one or more outcomes (2). In UA, the model analyst attempts to quantify the uncertainty around the outcomes that is propagated through the model from different sources of uncertainty (32). These sources include Monte Carlo (MC) error (i.e., random error), parameter uncertainty (i.e., uncertainty resulting from the fact that parameters used in simulation models are estimated and not truly known, structural uncertainty that is the uncertainty associated with the choice of the statistical models and possibly other sources, such as the choice of the starting population and sources of data that inform the model (3,32). In this Chapter, I provide an overview of the main issues and techniques in UA that are applicable to MS models. I illustrate these techniques and provide a step-wise approach to performing sample-size UA that was adapted from a method developed by O’Hagan et al. (4). While the results of O’Hagan et al. (4) was only developed for Probabilistic Sensitivity Analysis (PSA) in decision analytic models, in this Chapter I discuss how their approach can be used for outcomes common in PDMS models such as incidence, prevalence, total cost, and total number of people with a disease. I finally 9 A version of this chapter has been published in the journal Epidemiology Research International (35): Sharif B, Kopec J a., Wong H, Finès P, Sayre EC, Liu RR, et al. Uncertainty Analysis in Population- Based Disease Microsimulation Models. Epidemiology Research International. 2012;2012:1–14. ! 143 illustrate the sample-size approach with two examples in which UA is performed for prevalence and total direct cost of OA projected over the next two decades. My approach is illustrated using the POHEM-OA model (1), a PDMS model of osteoarthritis in Canada. 5.2 Uncertainty Analysis in Simulation Models Several methods have been developed for capturing uncertainty in the analysis of environmental risk assessment models (3), cost-effectiveness models in health economics (166), and technology assessment models (167,168). In environmental risk assessment models (3), prediction of quantities such as risks, exposures, and their effect on human health is critical for public policy decision-making. Prediction is also a major objective of PDMS models (169,170). Providing the distribution for the outcomes or uncertainty intervals around the outcome estimates is an important step in ascertaining predictive validity of surveillance and risk assessment models (171,172). In cost-effectiveness decision analytic models (166), as opposed to predictive risk assessment models, policy decisions are made through comparing cost and effectiveness of multiple scenarios by pre-defined objective functions, such as incremental cost effectiveness ratio (ICER) or incremental net benefit (166,167). An important difference between these two types of models with regards to UA is that in predictive models the goal is to estimate the uncertainty around a projected health outcome under a given scenario, while in decision analytic models one is usually interested in the uncertainty around a relative measure comparing different scenarios (162,168). 144 Probabilistic sensitivity analysis (PSA) is the term used for UA in decision analytic models (167), while quantitative UA is the term used in risk assessment and environmental modeling (3). UA and PSA are considered a necessary component of a model-based analysis in almost all risk assessment modeling guidelines (3) and recent decision analytic modeling guidelines (173). For instance, sensitivity analysis (deterministic or probabilistic) was cited as a requirement in the Principles of Good Practice for decision analytical modeling in healthcare evaluations, developed by a task force from the International Society for Pharmacoeconomics and Outcomes Research (ISPOR) and published in 2007 (162). UA is also required for health technology assessment modeling according to the guidelines provided by the National Institute for Health and Clinical Excellence (NICE) in the UK (166). Performing UA is important because, in certain situations, it may change the decision regarding the optimal scenario (174). For example, Griffin et al. (174) showed that in complex MS models with non-linear relations between the variables, interpretation of the outcomes such as ICER could be biased if uncertainty around the estimates is not considered. UA is most informative when applied to complex models that combine data from different sources (175--178). In addition to providing uncertainty intervals around the outcomes, UA is also used as a prerequisite for value of information analysis in decision analytic models (167) or sensitivity analysis (SA) in environmental models (179). In both of these approaches, the goal is to ascertain the contribution of different sources of uncertainty (e.g., different parameters) to the total uncertainty of the model (166, 167,180,179). As a result, critical regions of the parameter space can be located and 145 future research can be prioritized to obtain better estimates of the key parameters and reduce the uncertainty (180). Differences in the structure and intended applications between PDMS models and other types of simulation models may affect the methodology of UA. As the sources of uncertainty in individual-level models differ from those in macro-level models, the UA methods developed for macro-level models (181,182) should be modified for use in PDMS models (183). Studies by O’Hagan et al. (4) and Stevenson et al. (184) discuss UA for individual-level cost-effectiveness models. However, their approaches are not suited for models used for predictive purposes, such as PDMS models of chronic diseases used in health impact assessment studies (169,185-187), or those used for population projections (188). A guideline for conducting UA in the context of predictive models for risk assessment studies has been published by the Environmental Protection Agency (EPA) (186), but this guideline does not address the specific problems that may arise while performing UA for PDMS models. In sum, no standard approach for UA in the context of PDMS models has been developed (32, 33). In this Chapter and next one, the term “parameter” refers to all possible model inputs. Examples of parameters in PDMS models are those used to define the probability of events, such as disease onset, death or birth, or to describe the probability of changes in the risk factors at the individual level, such as changes in body mass index (BMI) or blood pressure. Parameters can be classified into those derived directly from data, those from literature estimates and those obtained from experts. Data-based parameters can be further categorized into those estimated from a sample and, therefore, subject to sampling uncertainty, and those obtained from a population census 146 (no sampling uncertainty). Expert-derived parameters are characterized by subjective uncertainty (32). Further in the text, I will discuss how to deal with each type of parameter in UA. The uncertainty associated with any given outcome in a disease simulation model is typically a reflection of the complex interplay among the model variables (18). Due to this complexity, estimating uncertainty with analytical approaches is often impossible, and numerical methods, such as the Monte Carlo (MC) method, are almost always required. The MC method involves running the model many times using randomly selected samples from the input parameter space (i.e., parameter joint distribution) (180). While the MC method is the most common numeric approach to estimating parameter uncertainty (171,180) in simulation models, in its simplest (standard) form it is not suitable for MS models due to the resulting computational burden (168). Various techniques have been proposed to reduce the computational time of UA in PDMS models. Cronin et al. (189) used a response-surface approach, which approximates the simulation outcome as a linear function of the input parameters. Others employed more complex approximate functions, such as Gaussian and radial basis functions (184,189). The major limitation of these techniques is that the assumptions of a specific response surface function may not hold in all types of MS models or even for different outcomes within one model (189). Another approach is to perform a guided sampling procedure, for example Latin hypercube sampling (LHS) (189,190) or orthogonal sampling (191). However, these sampling approaches do not reduce the model execution time, which is the bottleneck of the total computational time 147 in UA (182). O’Hagan et al. presented a method based on the analysis of variance and discussed sample size calculations for UA in MS models (4). 5.3 Overview of Uncertainty Analysis in PDMS Models 5.3.1 Sources of Uncertainty The sources of uncertainty vary with the type of simulation model and the objectives and design of a given study. Several sources have been discussed in the literature (168, 189), with the emphasis on: 1) MC error, or first order uncertainty (168,171,173,174), and 2) parameter uncertainty, or second order uncertainty (4,184,192, 193). MC error is introduced when starting values of the variables in the model and the simulated events are assigned to each individual using a stream of random numbers (173). As such, MC error reflects unexplained variation between the units of a simulation model (171,180). Parameter uncertainty arises from random errors in the estimation of the input parameters (171,193). Parameter uncertainty exists in all manner of simulation models. MC error, on the other hand, is not a concern in macro- level models, which do not attempt to reflect stochastic variability between individuals (193). Other types of uncertainty, including uncertainty in the starting population (18) and structural model uncertainty (171,189), have been studied. Structural uncertainty is associated with the choice of model structure, including the statistical models used, variables in the model, and the specified relationships between the variables (189). Examples of structural uncertainty commonly discussed in the literature are uncertainty arising from alternative model types, such as linear vs. non-linear dose-response in environmental risk assessment models (186) or the use of discrete event versus state- 148 based Markov models (194). Structural uncertainty is not considered further in the UA approach discussed in this thesis. 5.3.2 Uncertainty Analysis in Microsimulation Models UA in individual-level simulations needs to account for MC error in addition to parameter uncertainty. This can be accomplished by performing a nested analysis that iteratively samples from the parameter space, calls the base model (which itself is subject to MC error as it samples individuals) with a new set of input parameters, and records the outcomes (174). Throughout this thesis, this approach is referred to as “the MC method” in the context of PDMS models. It should be noted that MC error is reducible by increasing the number of simulated units or the numbers of simulation runs with different random seeds. However, parameter uncertainty is fixed and cannot be reduced in this way (4). There are two major issues when applying the MC method to PDMS models: 1) integrating MC and calibration algorithm(s) inside the simulation model and 2) reducing the computational burden associated with the MC method. Calibration algorithms are used extensively in PDMS models (32,193). Ideally, the calibration algorithm should be integrated into the UA, as discussed in Section 5.4. Importance of Uncertainty Analysis in Microsimulation Models PSA in decision analytic models and UA in environmental models have been described as necessary steps in the validation of such models (171, 181). Several guidelines outlined the necessity of performing UA and PSA using MC analysis in aggregate-level cost-effectiveness models (126,171). Nonetheless, few studies 149 discussed the necessity of PSA in patient-level cost-effectiveness studies (4,174) despite its high computational effort. Decisions based on expectations are appropriate only if the consequences of the uncertainty surrounding the decisions are also considered (19). In decision analytic models that aim to select the best option in terms of both cost and effectiveness, the uncertainty in parameters may bias the outcome. It has been discussed that in cost- effectiveness studies, even in simple situations such as risk–neutral decision makers and linear net benefit functions, the inference on the distribution of outcome is needed (174,180,193). For example, the uncertainty in parameters may bias the outcome when the alternatives under study have very close cost and effectiveness (174). In PDMS models (decision analytic or prediction models), uncertainty must be taken into account for several reasons. First, as in aggregate-level models such as Markov cohort models, UA and sensitivity analysis have been discussed as a necessary step in validation of microsimulation model (32,165). Because microsimulation models often use several different types of input data sets to estimate their parameters, UA would reveal the precision and cohesion of those datasets (32). Moreover, sensitivity analysis following UA would reveal the contribution of each parameter to the total uncertainty, thereby prioritizing future research in order to reduce the uncertainty of the outcome. Second, situations arise in which the (mean) outcome in microsimulation models may become biased if uncertainty has not been taken into account (e.g., uncertainty between individuals or parameter uncertainty). In other words, the mean estimator of an outcome would be different before and after the uncertainty has been taken into account. Examples of this situations- where UA is needed for the mean 150 outcome- are when complex nonlinear or multi-collinear models are used within the simulation models (174). Computational Time of Uncertainty Analysis in Microsimulation Models In MS models, a major burden in applying the MC method for UA is the computational time. UA should be performed using an outer loop of (n) model runs for the parameter uncertainty and an inner loop for a population of (m) units for the MC error (174). To obtain an accurate estimate of parameter uncertainty in macro-level models, it has been suggested that the number of model runs (n) (i.e., MC simulation runs) should be large enough so that the uncertainty estimates converge (174,184). However, applying this recommendation to MS models would make the UA computationally intractable, as one run of such a model may take hours (4). Two approaches have been discussed in the literature for reducing the computational time of UA in MS models: emulator-based (184, 189) and sample-size based (4). Emulator-based approaches approximate the solution space of the simulation model by pre-assumed functions mapping the input parameters to the outcome. These approximate functions predict the outcome of the main simulation model instead of running the time-consuming PDMS at each instance of the MC process (184,189). The types of approximate functions used in MS disease models include linear response function used by Cronin et al. (189) for a model of prostate cancer, with the objective of investigating the effect of screening on prostate cancer mortality rates, and Gaussian function for a patient-level model of osteoporosis in Stevenson et al. (184). The emulator-based method would save a significant amount of computation time at a cost of losing precision due to outcome approximation. Stevenson et al. (184) 151 reported that the required time for a single run of their model was reduced from 150 minutes to virtually instantaneous. However, this method requires the analyst to assume a certain distribution for the outcome in order to be able to fit a response surface. For example, Cronin et al. (189) used a linear approximate function and assumed Poisson distribution for the number of prostate cancer-related deaths, which was the outcome of choice for performing UA in their study. Complex PDMS models would need a specific response function for each outcome and there is no a priori reason to expect that all response functions have the same functional form. In other words, one needs to use a different approximate function for each type of outcome in a single model. The sample-size approach involves finding a combination of values for the simulated population size and number of MC simulation runs that achieve a specified precision for the outcome within a fixed computational time (38, 40). O’Hagan et al. (4) used analysis of variance (ANOVA) to determine sample size in the case where two types of uncertainty, MC error and parameter uncertainty, are considered. Although their approach was originally designed for decision analytic models with incremental net benefit as the outcome, it can be used for other outcomes, including those typically observed in PDMS models. O’Hagan et al. (4) reported that the computational demand was reduced by a factor of almost 36 (they mentioned that this claim is for the incremental net benefit outcome and the resulting efficiency gain could be different for other outcomes). The sample-size approach is simple to implement and does not require an approximation function. UA in MS models has been discussed in detail for cost-effectiveness studies with decision analytic purposes that use specific type of outcome (i.e., cost effectiveness 152 ratio or incremental net benefit) (180). However, systematic presentations of UA for predictive PDMS models with multiple outcomes are lacking. In particular, there has been little discussion of UA in the context of chronic disease models with epidemiological or burden-of-disease outcomes, such as those predicting the incidence and prevalence of a disease in a population (195,196), its impact on mortality, quality of life, or costs (170, 176), in addition to PDMS models evaluating screening programs for preventing the onset of chronic diseases such as colorectal cancer screening (197) or type 2 diabetes (198). In the following section, I provide an overview of the key steps in UA for such models. 5.4 Steps for Performing Uncertainty Analysis Step1. Selecting the Outcomes UA should be performed separately for each outcome of interest. Due to their predictive nature, most of PDMS models are dynamic, i.e., the outcomes are time- dependent. Typical examples for such dynamic outcomes in public health applications include the distribution of health determinants, disease incidence or prevalence, mortality, quality of life, health care utilization, and costs. Step2. Selecting Parameters After selecting the outcome of interest (e.g., the prevalence of disease X over the next N years), the next step is to screen the parameters and select the relevant ones (171, 182). By relevant parameters, I mean those that are likely to have significant influence over the selected outcome, as determined by the model analyst. It is important to note that the uncertainty intervals from UA will be conditional on the assumption that all parameters not selected for UA do not introduce additional uncertainty. Although 153 ideally one would include all model parameters in the UA, this is often impractical in complex models, such as many PDMS disease models, that may include hundreds of parameters (1, 175-177,197,198). Tools for determining parameter inclusion range from qualitative approaches, such as scrutinizing the model’s conceptual diagram and investigating its major components, to more quantitative approaches such as parameter prioritizing (179, 182) and tornado diagrams (32,171,179). Scrutinizing the model, beginning with its major components, helps in identifying the most important parameters, eliminating the irrelevant ones, and identifying possible dependencies between the parameters. Different components of the simulation model, such as demographic, disease and intervention components, need to be examined to ascertain which parameters to include in the UA. The data sources for the parameters should be considered as well. For example, the demographic component may include the parameters describing the distribution of age and sex in the population, as well as birth and death rates over time. These parameters are often derived from vital statistics or large population databases and may be excluded from UA due to minimal uncertainty. However, if any of the above processes are modeled from a sample, then sampling variability should be incorporated into the UA. For example, a CISNET model for breast cancer (194) used Surveillance, Epidemiology and End Results (SEER) (199) to estimate all-cause and cancer-specific mortality rates by age and sex. Because the SEER database is very large, the sampling variability of the SEER-derived mortality rates is likely small, but may not be negligible. Further analyses, discussed later in this section, may be needed to determine the inclusion of such parameters in the UA. 154 In making decisions about which parameters to include, the analyst needs to take into account possible discrepancies between the population being simulated and the base population from which the data have been gathered. If the calibration has been performed to fill in this discrepancy, the parameters used in the calibration should be included in the UA (unless they are estimated from a large population data or census data). In addition, the calibration algorithm itself should be integrated into the UA process. For example, in the POHEM-OA model (1), a PDMS model of osteoarthritis in Canada, the baseline incidence of osteoarthritis (OA) was estimated from a population- based administrative database in British Columbia (PDBC), Canada (84), that includes healthcare data for almost all residents of the province (Table A1, Appendix A). To simulate the incidence of OA in Canada, a calibration algorithm is needed to obtain the incidence of OA in the reference category of BMI, given the BMI hazard ratios and the distribution of BMI in the population. The calibration algorithm uses a direct search method to match the marginal distribution of incidence in the province of BC to that of the simulated population of the Canada (84) by changing the baseline distribution until the target value is reached for every age/sex category within a small acceptable distance. I will discuss the details of implementing calibration into UA in Section 5.3. Parameters in disease components of the PDMS models are mostly coefficients in statistical models used to determine the distribution of risk factors in the population, both the baseline distribution and the dynamics of how risk factors change through time and space, in addition to the coefficients used in modeling the probability of event, mostly through relative risks (or hazard ratios) associated with the risk factors. For example, in the POHEM-OA model (46), the baseline distribution of body mass index 155 (BMI), a risk factor for OA, was derived from the CCHS (51). The dynamics of how BMI changes in time and in different provinces was modeled from the longitudinal NPHS (146). Parameters in disease component can be excluded from UA if they are obtained from full population data, such as the census or vital statistics. However, this is often not the case for the parameters used in modeling probability of events (e.g., relative risks associated with risk factors) as they are mostly estimated from sample data (200). For example, the coefficients in Framingham equations, commonly used for modeling the risk of cardiovascular disease in simulation models, come from the Framingham sample (201) and, therefore, are subject to sampling variation. Another group of parameters that should be considered for UA are those describing the effects of health interventions (202). For example, in the Chronic Disease Prevention (CDP) model (203), a PDMS model for evaluating interventions for reducing obesity, the effects of physician counseling was translated into reductions in the percentage of total cholesterol intake, BMI and systolic blood pressure. Another example are the parameters describing the effects of a screening program on QALYs and excess mortality in a CISNET model of breast cancer (194). Whenever the outcome chosen for the UA is related to the intervention, the best practice is to include the intervention parameters into UA, as most of these parameters are estimated from sample data and the uncertainty associated with them significantly affect the outcome uncertainty. The tornado diagram is a useful tool for prioritizing the parameters in the initial steps of UA (32). A tornado diagram represents the results of multiple univariate sensitivity analyses in which one parameter is changed at a time and the highest and 156 lowest value of the outcome are determined. The parameters are then sorted according to their impact and presented graphically, often as a horizontal bar diagram (32). Tornado diagrams do not consider the interactions and correlations between parameters. Figure 5.3 shows an example of the tornado diagram for the POHEM-OA model discussed in greater detail later in Step 3. Step3. Assigning Distributions to Parameters (or Using Bootstraps) For the selected parameters, the model analyst should either assign a distribution or use a bootstrapping approach to be able to apply the MC method and get the uncertainty interval for the outcome as discussed in previous sections. Parameters in the simulation models may define the risk of an event, such as hazards ratios, relative risks, odds ratios, transition probabilities of events, or estimate the dynamic changes in a continuous variable (e.g., regression coefficients for modeling BMI changes over time (1)). Parameters may be estimated directly from data, drawn from literature, or based on expert opinions. For parameters estimated from data using parametric (and semi- parametric) statistical models, the distribution can be obtained from the relevant statistics. The common regression models, such as the parametric multiple linear, logistic and semi-parametric Cox proportional hazards models utilize a specific distribution for the outcome (192,193). Incorporating uncertainty into the risk of an event depends on the structure of the model. In some predictive MS models, risk functions for events use event-history modeling (204, 205) as opposed to memory-less (Markovian) states (173, 206). Event- history modeling equations represent the behavior of an individual through dependence on his/her simulated past. For example, in the LifePaths model (207), a PDMS model of 157 the Canadian labour force (208), the timing of a child’s birth is determined by the timing of parents’ marriage, the time since the previous births in the family and the age of the parents. In continuous-time event history models with hazard functions, an asymptotic lognormal distribution can be assigned to hazard ratios (1) while a normal (Gaussian) distribution can be assigned to the log transformation of the relative risk or odds ratio, under the large sample assumption (209). Regression coefficients used in modeling changes of risk factors over time are another types of parameter in PDMS models. For instance, coefficients in the dynamic models of BMI changes in POHEM include age, sex, place of birth (time-invariant covariates) and socio-economic status (time-dependent covariate) as well as previous levels of BMI (1). For such complex models, the uncertainty analyst needs to estimate the variance of the coefficients of the regression models (or use bootstrap weights for the outcome) (208). In MS models utilizing longitudinal data, events are often defined based on hazard functions and waiting time distributions (209-212). The waiting time distribution can be exponential (for fixed hazards) (1), Weibull (for time-dependent hazards) (47, 192,209) or Gompertz (for time-dependent hazards with additional complexity, such as those for tumor growth modeling in cancer models (194)). Transition probabilities can be converted into hazards for use in continuous-time models (1,45, 209,210). Other examples of parameter distributions often used in health simulation models include the beta distribution (e.g., for modeling cost data), the gamma distribution (e.g., for modeling quality-adjusted life years) and the Poisson distribution for modeling event counts (193). 158 For parameters drawn from the literature, only the mean and its 95% confidence interval are often reported, without the actual distribution. In such cases, method of moments can be used (192,193). Alternatively, the analyst needs to assume a specific distribution appropriate for a given parameter. The variance of the parameter can then be estimated from the confidence interval reported by the authors (192,193). Parameters estimated by expert opinion are often ignored in UA. This may result in a substantial underestimation of the true uncertainty (181). It is often possible to assign upper and lower bounds to the estimated parameters using experts’ judgment, for example, where there are biological limits to a disease process (e.g., tumor growth) or limits to an epidemiological parameter such as prevalence or case-fatality. In these cases, the distribution could be quantified by a uniform, triangular or trapezoidal distribution by assigning lower and upper bounds in addition to the mode of the distribution, or lower and upper bound on the mode in case of a trapezoidal distribution (213). General distributions such as the generalized beta distribution family and the Johnson translation system of distributions are among the best choices in case several data points have been selected in the literature or gathered from expert opinions (213). Another approach to assigning distribution to parameters estimated from expert opinion is the Bayesian method (214). A key task in applying this method is the expression of expert opinion in the form of a (prior) probability distribution (214). An example is the Sheffield elicitation framework (SHELF) developed by Garthwaite et al. (215), elaborating steps to construct a probability distribution for the parameters expressed by a group of experts. For parameters estimated from non-parametric models, the uncertainty analyst 159 should use bootstrap methods instead of sampling from a distribution. Bootstrap method provides an easier approach for estimating uncertainty for those parameters whose distribution cannot be identified. It can also be used to reduce the computational difficulties in sampling from a parametric distribution for parameters estimated from surveys that use a complex sampling design involving stratification and clustering (216). In both cases, bootstrap methods sample with replacement from the original data (217). As an example, the NPHS (146), a longitudinal household survey conducted by Statistics Canada every two years since 1994 until 2010, provides bootstrap weights for users to estimate the variance of their estimators, such as regression coefficients. I will implement this type of bootstrapping in Step 4. Accounting for correlations between parameters: After selecting the parameters for UA, the analyst needs to examine the relationship between the parameters to ensure their correlations are taken into account. The analyst can identify the correlations between parameters by examining the model diagram and reviewing the parameter estimation process (8). For example, relative risks associated with different levels of an ordinal risk factor are often correlated. As discussed in Chapter 3, in the POHEM-OA model, the hazard ratios of OA incidence for four different BMI categories are interdependent. Such dependence should be reflected in this step of the UA, through assigning a joint distribution using covariance matrices (often estimated from the regression model). When bootstrap methods are used, parameter interdependence will be inherently reflected in the distribution of the outcome. 160 Step 4. Monte Carlo Method This step involves applying the MC method to calculate the uncertainty intervals for the mean outcome of interest. This step involves sampling from the assigned distribution or using bootstrap methods Sampling approach: For those parameters that have been assigned a distribution, I need to specify the sampling approach, as it will affect the uncertainty intervals of the outcome. The sampling approach should be both effective, to encompass the entire distribution of the parameter, and efficient, to sample only the required number of points. There are several sampling approaches for the MC method (38). One of the most effective is the LHS (189, 190). LHS is a particular type of stratified sampling, which samples from the equally probable segments of the parameters distribution. It is designed to provide a comprehensive coverage of the parameter space. More complex sampling approaches include orthogonal sampling and adaptive kriging sampling (184, 190). Although LHS is not the most efficient sampling approach, as it samples more points than necessary, it is commonly used due to its effectiveness. It samples enough points within a reasonable computational time and is more efficient than random uniform sampling. A common sampling technique is to generate a uniform random number (between 0 and 1) through one of the existing random number generator algorithms (209) (e.g., pseudorandom number generators are among the most used methods that deterministically perform series of tasks to generate statistically sound (uniform) random numbers (18)). Then, using the inverse of the cumulative density function (CDF), known as the quantile function, one finds a real value of the parameter. 161 For correlated parameters, an appropriate joint distribution should be specified, as discussed in Step 3, and the sampling technique should be adapted accordingly (193). In general, if the marginal univariate distributions of the parameters and the correlation between them are known, sampling from a multivariate distribution can be performed using the multivariate normal copula with matrix algebra (e.g., Cholesky decomposition) (193). Calibration in UA: Calibration is a technique commonly used in population-based models with predictive goals when direct estimation of a given parameter is not possible and model parameters are selected so that the model reproduces observed results (218). At each step of the MC sampling, calibration should be performed with the new (random) sample of the parameters, generated in previous step. After the calibration procedure converges, the simulation model should be run with the calibrated parameter. As a result, the calibration procedure should be implemented inside the UA and be executed at each run of the MC method. Appendix C5 discusses the calibration algorithm used in the POHEM-OA model, the microsimulation model of osteoarthritis that I used in this Chapter as the example to present how to perform UA. Determining the number of MC runs and population size: As discussed in Section 5.2, there are two method in the literature to reduce the computational time of UA in simulation models: emulator-based and sample-size approaches. Since the sample-size approach does not impose any distributional assumption for the outcomes of the model, I used this approach to reduce the computational time for PDMS models. According to the sample-size approach, computational burden of UA can be reduced by appropriate choice of the simulated population size (m) and number of 162 simulation runs (n). As discussed in O’Hagan et al. (4), the variance of the mean outcome based on the simulation results is given by the formula: σˆ 2 σˆ 2 V(µˆ) = ( 1 ) + 2 Equation 5-1 mn n 2 2 In (Eq.5-1) above, σˆ1 reflects the MC-error and σˆ 2 reflects the parameter uncertainty (and it is assumed that patients outcomes are independent). The goal of the sample-size approach is to find values of (m) and (n) such that a desired precision level d is achieved, i.e., (σˆ 2 = d ) is achieved given a fixed amount of computation time. The computational time is assumed to be linear in terms of both population size (m) and number of MC runs (n) (i.e., K=mn, where K is measured in terms of number of 2 2 individuals that can be simulated in the available time). Since σˆ1 and σˆ 2 are unknown, the solution is obtained in two steps. In the initial step, a small number of simulation runs are completed and ˆ 2 and ˆ 2 are estimated from the outcome while ˆ 2 is σ σ1 σ 2 obtained by subtraction using (Eq.5-1). Next, in the main step, (Eq.5-1) is solved for (n) with (mn) replaced by K and σˆ 2 replaced by (d) to obtain the lowest number of computational runs, (n*), that satisfies the computational time and precision constraints. The corresponding value for the population size, (m*), is then obtained from the constraint K=m*n*. Note that this procedure typically does not generate the optimal choices for (m*) and (n*). The final estimated number of MC runs is the smallest number that can produce the desired precision level for a given computational time. As discussed in O’Hagan et al. (4), in order to gain the maximum information for estimating uncertainty intervals in a fixed amount of computational time, one should use a relatively small population size and a large number of simulation runs. In other words, 163 the optimal design is to simulate a small population over and over again using a large number of sample points from the joint distribution of the parameters. Details of the O’Hagan et al. (40) solutions for sample-sizes are shown in Appendix C1. Average and aggregate type outcomes in sample-size algorithm: The outcomes of PDMS model can be categorized into two types with regard to the UA procedure: 1) average type outcomes that are calculated by dividing the total outcome over all or certain number of individuals in the model such as average cost, average QALY, prevalence or incidence of a disease; 2) aggregate type outcomes that are calculated by summation of individual outcomes over all or certain number of individuals is the model such as count of people with a disease or overall cost. As discussed above, to reduce the computational time of the UA procedure approach in the sample-size approach, I reduce the population size (m*) to be able to increase the number of MC simulation runs. However, if the number of individuals were reduced for aggregate type outcomes, the new outcome would not be related to the actual outcome. For example, for the case of total cost of a disease among Canadian adults, if the population size is reduced, the total cost is no longer representing the total cost of all Canadians. To overcome this problem, I propose to transfer the aggregate outcome into average type outcome, by dividing the outcome by the population size in each year (m). Next, after the UA is performed, I need to transfer the mean and variance of the average outcome to the aggregate outcome by multiplying the resultant mean by (m) and resulting variance by m2. The details of this approach are discussed in Appendix C3. 164 Step 5. Constructing the Empirical Distribution of the Outcome In the final step of the UA, the analyst constructs the cumulative distribution function (CDF) of the outcome after applying the MC method with a population size of m and number of MC simulation runs of n calculated using the sample-size approach discussed in the previous section. The 95% uncertainty interval of the outcome is constructed using the quantile function based on the cumulative distribution function (CDF) of the outcome being analyzed. The results of the UA, often represented by 95% uncertainty intervals, should be presented as conditional on the assumptions of the model. 5.5 Results of Uncertainty Analysis for Prevalence of OA I provide an example of the UA approach outlined in the previous section, using a simplified version of the POHEM-OA model, a PDMS model of osteoarthritis in Canada (1) to estimate uncertainty associated with the prevalence of OA. The POHEM-OA is a complex model, featuring a large number of parameters as described in Table A1, Appendix A (1). In my example, the focus is on modeling the incidence and prevalence of OA. Table 5.1 lists the parameters used in this example and the data source for each. Table 5.2 describes the steps of the proposed UA approach and the corresponding steps of the POHEM-OA example. In the first step of the proposed UA (Section 5.4), I selected the gender-specific prevalence of OA in Canada from 2001 to 2021 as the outcome of interest. 165 Table 5-1. Components of a simplified version of the POHEM-OA model used in UA example Demographic components Age distribution in 2001** Sex distribution in 2001** Province of residence distribution in 2001** Birth and Mortality rates by age and sex over time# Migration## Disease components Baseline parameters BMI ¥ distribution in 2001** Incidence of OA in 2001 by sex and 5-year age groups¥¥ Reference (“baseline”) hazard rates of OA§ Prevalence of OA in 2001 by sex and 5-year age groups§§ Risk factors model Change in BMI over time§* Disease incidence §# Effect of BMI on incident OA by sex 166 Table 5-1 notes: * POHEM-OA: Population Health Microsimulation model for osteoarthritis (OA) (1); **Observed in Canadian Community Health Survey (CCHS-2001)(54); # Projected mortality in Canada by age and sex from Statistic Canada; ## Migration data obtained from Statistics Canada projections; ¥ BMI = weight / height2 ; ¥¥From administrative data in BC (PDBC)(84), OA is defined as at least 2 visits to a health professional within 2 years or 1 hospitalization with the ICD-9 code 715. Incident cases in 2001 are identified after excluding prevalent cases prior to 2001 using a 10 year run-in period (1); §Obtained numerically using calibration to match the marginal distribution of incidence in BC administrative data;§§ Obtained as the final stable prevalence from a simulation of the Canadian population over a 50-year horizon, under constant age-specific incidence rates;§* Obtained from a linear regression model including age, sex, province, education, and prior BMI;§# BMI was categorized into 4 standard categories (see Table 5.3). The effect (hazard ratio) for each level of BMI is obtained from a survival regression model using longitudinal NPHS data (two cycles: 2000 and 2002) (146), separately for men and women. In the second step of the UA (Section 5.4), I first investigate the conceptual model diagram shown in Figure A2, Appendix A. Most of the parameters listed in Table A1 are not included in the UA. Parameters estimated from the CCHS in 2001 (54), a national survey with a very large sample size (N=130,000), and parameters obtained from provincial administrative data in BC (84) are excluded because their variance is considered to be small. 167 Table 5-2. Proposed steps for UA in PDMS and corresponding steps for the POHEM-OA model Steps for performing uncertainty analysis in Steps for uncertainty analysis of population-based microsimulation models prevalence of OA in POHEM-OA Step 1. Selecting the outcome Sex-specific prevalence in (2001-2021) Step 2. Establishing the UA list of parameters 1. Hazard ratios for each BMI- category 2. BMI progression Step 3. Assign a probability density function to 1. Lognormal distribution for parameters (or use bootstrap sampling) hazard ratios with sex specific correlation matrices 2. Using 8 alternative regression models using bootstrap weights from NPHS§ (1996-2006) Step 4. Applying Monte Carlo method 4.1. Select a sampling approach (or use 1. Latin hypercube sampling bootstrap sampling) 2. Random sampling with replacement (among eight alternate set of parameters) 4.2. Implement calibration into MC method Calibrate on incidence by age and sex; Implement the calibration algorithm into the MC method, used squared error criteria for convergence 4.3. Calculate n*, m* for precision level D* m0= 20, 000,000 population size; n0=5 (for the initial run); m*= 500,000 people and n*=785 MC simulation runs for D=0. 01; for 12 hours run of a PC with 12GB memory, CPU=i7-980 Intel, 3.3 GHz. Step 5. Construct the outcome distribution. Please see Figure 5-2. § NPHS: National Population Health Survey, a longitudinal household survey by statistics Canada with a sample of size of over 17,000 persons that started in 1994 (146). * d=D2 168 For the selected parameters, a Tornado diagram has been constructed as shown in Figure 5.1. Two types of parameters have shown the highest impact on uncertainty in univariate analysis and were included in the final UA analysis. These are: 1) the hazard ratios (HR’s) for the effect of different levels of BMI on OA incidence; 2) the regression coefficients used in modeling BMI progression in the simulated individuals over time. The HR’s for BMI categories have been estimated from the NPHS (146), a longitudinal survey in Canada with a sample of size of over 17,000 persons that started in 1994. 169 Figure 5-1. Tornado diagram for prevalence of osteoarthritis among females in year 2021* *Results are predicted from the POHEM-OA model: Population Health Microsimulation model for osteoarthritis (OA) (1); ** Each of the hazard ratios parameters on the left have been varied based on the upper and lower limits of the estimated 95% CI shown in Table 5.3. Two cycles of NPHS were used for the analysis of HR’s in the POHEM-OA model. The HR’s have been estimated using an exponential survival regression model (1). The point estimates and 95% confidence intervals for the parameters are shown in Table 5.3 and the correlation matrices are shown in Table 5.4. The model for BMI progression is based on 6 cycles of the NPHS (1996-2006) (146) and includes BMI history (up to 4 past BMI values) and other covariates such as age, sex, region, education and income. The model was stratified by age, sex and BMI, resulting in 112 regression models (28 age-sex-BMI strata and 4 autoregression models). 170 Table 5-3. Point estimates and 95% CI for hazard ratios of OA diagnosis implemented in the POHEM-OA model* Male (95% CI) Female (95%CI) Under weight (BMI<18.5) - 0.33 (0.02-0.4) Normal Weight (18.5≤BMI<25) 1 (REF) 1 (REF) Overweight (25≤BMI<30) 1.07(0.68,1.74) 1.76 (1.14-2.62) Obese (BMI≥30.0) 1.69 (1.03,2.81) 2.03 (1.34,2.96) * Body Mass Index (BMI), reference category: Normal weigh (18.5≤BMI<25); estimated from the NPHS (2000-2002), National Population Health Survey, a longitudinal household survey by statistics Canada with a sample of size of over 17,000 persons that started in 1994 (52). In the third step of the proposed UA (Section 5.4), I used a multivariate lognormal distribution for the hazard ratios. In the fourth step of the proposed UA (Section 5.4), Latin hypercube sampling (LHS) has been used to sample from the multivariate HR’s distribution with the given correlation matrices (Table 5.4), translated into covariance matrices for the LHS. 171 Table 5-4. Correlation matrix for hazard ratios of osteoarthritis in the POHEM-OA model BMI<18.5 18.5≤BMI<25 25≤BMI<30 BMI≥30.0 BMI<18.5 1 0.5125 0.5362 (Ref) 18.5≤BMI<25 0.6234 1 0.5405 - 25≤BMI<30 0.6325 0.5812 1 - BMI≥30.0 (Ref) - - * Body Mass Index (BMI): Under weight (BMI<18.5), Normal Weight (18.5≤BMI<25), Over weight (25≤BMI<30), Obese (BMI≥30.0) Reference category: Normal weigh (18.5≤BMI<25); estimated from the NPHS (2000-2002), National Population Health Survey, a longitudinal household survey by statistics Canada with a sample of size of over 17,000 persons that started in 1994 (146). The second sets of parameters are those used in the BMI progression models. To avoid the complexity of assigning distributions to a large number of models, I used the bootstrap weights developed for the NPHS by Statistics Canada (146). For each BMI progression models (age and sex-specific), I derived eight sets of parameter estimates, meaning that each set of parameters represents a sample from the underlying joint multivariate distribution for the covariates of the BMI model. At each run of the UA, I randomly selected one of these samples based on each person's age/sex, in addition to randomly sampling the hazard ratio of the OA event from a multivariate lognormal distribution. The POHEM-OA model calibrates the baseline incidence of OA using population data from the province of British Columbia (84). As discussed in the 172 calibration step of Section 5.4, I have automated the calibration algorithm and integrated it with the MC method, so that for each sample of HR’s, calibration is performed to get the (calibrated) baseline incidence rate; then, repetitive runs of the simulation model are performed by applying the MC method. To determine the number of MC simulation runs and the population size in the fourth step described in Section 5.4, I used the method discussed by O’Hagan et al. (4). The number of MC simulation runs is calculated as a function of precision level (set to 0.01), computational time (set to 12 hours), and initial estimates of the MC-error and parameter uncertainty. The initial estimates were obtained by running the model 5 times with a sample size of 20 million (each POHEM run that simulates the entire Canadian adult population of around 20 million takes about 35 minutes on a PC with 12 GB RAM and 3.33 GHz. CPU, Intel i7-X980). Applying the formulas from O’Hagan et al. (4), as described in Section 5.4, I obtained (n=758) and (m=500,000). Approximately 10-15 iterations were needed for calibration before each main run, and each run with population size of (m=500,000) took about 50 seconds. It should be noted that without applying the sample-size approach by O’Hagan et al. (4), to get the same precision level with a sample size of 20 million, I would need to run the model about 400 times and it would take about 12 days of run- time (an almost 25-fold increase in time efficiency). Although increasing the population size (m) or increasing the number of MC simulation runs (n) increases the precision level of the mean estimate (and consequently the precision of the resultant uncertainty interval), increasing (n) has a much higher impact on the precision level than increasing (m). The reason is that the 173 variance of the mean estimator (Eq. 5-1) is a function of (m) and (n), where (n) affects both terms on the right hand side of the equation, and (m) only affects one term, as discussed in O’Hagan et al. (4). That is why running the model 400 times with a sample of 20 million produces the same precision as running the model 785 times with a sample of 500,000. Result of the UA is shown in Figure 5.2. The estimated 95% uncertainty intervals for the prevalence of OA in Canada in 2021 are 0.09 to 0.18 for men and 0.15 to 0.23 for women if I include both MC error and parameter uncertainty of the hazard ratios of OA associated with different categories of BMI as well as BMI trajectory model parameters. These results are based on the assumption that uncertainty in other parameters shown in Table 5.1 can be ignored. This assumption would be reasonable for parameter estimated from large population-based datasets, such as provincial administrative data (84), but not for parameters estimated from smaller studies, as discussed in Section 5.4. The uncertainty contributed by the parameters estimated from the CCHS (sample size 130,000) can be assumed to be small, compared to those estimated from the NPHS (146) (sample size of 17,000), and has been ignored in this example. If the uncertainty associated with other parameters were included in the analysis, the UA interval would have been equal or larger. 174 Figure 5-2. Results of uncertainty analysis: sex-specific prevalence of OA in Canada from 2001 to 2021 as predicted by the POHEM-OA model Prevalence of OA(Females) 0.25 0.2 0.15 0.1 0.05 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 Year (2001−2021) Prevalence of OA (Males) 0.25 0.2 0.15 0.1 0.05 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 Years (2001−2021) * Curves inside the uncertainty intervals represent the mean estimate surrounded by 95% uncertainty intervals. 175 5.6 Results of Uncertainty Analysis for Total Direct Cost of OA To estimate the uncertainty associated with the total direct cost of OA from 2010 to 2031, I applied the step-wise approach discussed in Section 5.4 in this Chapter for the POHEM-OA direct cost outcome. Since the details regarding the POHEM-OA model and the first two steps of the UA for the prevalence outcome in Section 5.5 are similar to those for the cost outcome, here I discuss steps 3 to 5 according to Section 5.4 and first describe the parameters used in the UA for the direct cost outcome. 5.6.1 Parameters Used in the UA for Total Direct Cost of OA The two major parameter groups used in modeling of OA cost comprising the diagnosis and resource utilization modules have been included in the UA-list of parameters for the direct cost outcome (Table 5.5). First, I included the hazard ratios (HR) for OA incidence- the same parameters included in the UA-list for the prevalence outcome (Section 5.5). HR’s for OA incidence contained eight parameters corresponding to sex and BMI categories of population. The mean and 95% confidence intervals for these parameters, in addition to their covariance matrices are shown in Table 5.3 and Table 5.4, respectively. As discussed in the previous section, the multivariate lognormal distribution is used for hazard ratios and the LHS sampling approach has been used to sample from HR’s distribution considering the covariance matrices. The second parameter type used in the UA-list was parameters related to direct cost of OA. I have chosen input parameters for per patient-year cost of four cost categories: 1) total joint replacement (TJR), 2) rehabilitation, 3) out-of-pocket including those related to alternative care and formal caregiver, and 4) side effect of drug cost. 176 These four cost categories were included as they were constituting more than 80% of total direct cost of OA in 2031 based on the POHEM-OA results as discussed in Section 4.5, Chapter 4.Table 3.1 and 4.2 represent the details of how these cost categories and the related parameters were estimated. The cost categories not chosen in the UA procedure included those hospitalization procedures other than hip/knee TJR surgery, physician visits and drugs costs, in addition to the probabilities of drug use as they all have been estimated from the population-based administrative data (very large sample) (84) and assumed not to contribute to uncertainty. Categories included in UA, in addition to the assumed distributions, confidence intervals and the data sources used for their estimation are shown in Table 5.5. 177 Table 5-5. List of parameters, data sources and their distributions used for UA of direct cost in the POHEM-OA model Parameter type Parameter Data source (year) Distribution of (no. of parameters) subcategory [reference] parameters (approach) 1. OA incidence Hazard ratios by NPHS (2002-2006) Lognormal (estimated (6 parameters) sex and BMI [146] from data) category 1 2. Hospitalization Hip/knee TJR CIHI-HP2 (2010) Normal cost Surgery by age [144] (method of moments)2 (10 parameters) groups and sex 3. Physicians and Physician visit and PDBC 3(2003) Not in UA4 outpatient care costs Medical tests (X- [84] ray, MRI, blood work and other outpatient care) 4. Drugs Prescription drugs PDBC (2003) Not in UA [84] Over-the-counter NPHS (2002) Not in UA drugs [146] 5. Rehabilitation and Probability of Oldmeadow et al. (2003) Dirichlet (estimated from home care cost (after discharge [135] data) hip/knee TJR destination surgery) 4 (8 parameters) Unit cost for Coyte et al. (2000) Beta (method of moments) discharge [136] destination5 6. Formal caregiver Odds ratio by age Gupta et al. (2005) Lognormal (method of and others (before groups and sex [115] moments) hip/knee surgery) 6 (7 parameters) Unit cost7 Gupta et al. (2005) Beta (method of moments) [115] 7. Alternative care Odds ratio for each MOH-OA (2007) Lognormal (estimated cost 8 type of alt. care by [147] 9 from data) (7 parameters) age groups and sex Unit costs MSP (2003) Not in UA11 [148] 10 8. Side effect of Odds ratio of side Systematic reviews [139] Lognormal (method of drug’s effects for each moments) (8 parameters) type of drug Unit cost of CVD12 [140], [156] Uniform Unit cost of stroke13 [141], [157], [158] Triangle Unit cost of [142], [143] Beta (method of moments) dyspepsia and GI14 178 Table 5.5 notes: 1 mean, 95 % CI and covariance matrix are reported in Table 5.3, 5.4; 2 CIHI-HOSP: 95% CI for per diem rates of hip/knee TJR surgery reported from CIHI hospital cost database (144), method of moments used when 95% CI are reported from the literature as discussed in Section 5.3, Step 3; 3 PDBC: Population Data BC (84); 4not used in UA because the parameters were estimated from population-level data; 4 out-of-pocket cost for rehab and all of alternative care costs in addition to formal caregiver category were included in out-of-pocket cost category in the results;5 healthcare system paid rehabilitation and home care cost; 6 formal caregiver cost is before the surgery only; 7 unit costs is given for average of formal caregiver and other out-of- pocket cost of patient before the surgery and its mean and 95 %CI were given in Gupta et al. (115) for patients that have non-zero costs associated with formal caregiver ; 8Alternate care includes cost of physiotherapy, chiropractic’s and other complementary care categories listed in Table A12, Appendix A; 9 MOH-OA : BC ministry of Health OA survey Data (147), 10 MSP fee: Medical Services Plan (MSP) Payment Information File in 2003; 11Assumed fixed fee-per-visit (148);12 as discussed in Appendix A5 for side effect costs I used Birnbaum et al. (2003) [140), Kok et al. (2009) [156] for CVD cost in UA ; 13 for UA of cost of stroke , I used a triangular distribution for average results ( over all age groups) from Taylor et al. (1999) [141], de Oliveira (2012) [157] and Kolominsky-Rabas (2006) [158]; 14 for UA of cost of GI and dyspepsia , I used the 95% CI reported in Appendix A4 from 14Rahme et al. (2001) [142] , Moayyedi et al. (2002) [143] for GI and dyspepsia , respectively. 5.6.2 Sample-size Approach for Total Direct Cost of OA As discussed in Section 5.4, the POHEM-OA model calibrates the baseline- incidence of OA using BC population data (84). I have automated the calibration algorithm, so that after each MC-simulation run, calibration procedure is performed to get adjusted baseline-incidence rate for the new parameter replicate; next, the simulation model is performed with the calibrated parameters- details of calibration procedure are discussed in Appendix C4. I have used a 15-hour run of a 12G memory, 179 CPU=i7-980 Intel, 3.3 GHz for direct cost outcome of POHEM-OA. I used outcome in 2031 in the initial step to calculate m*and n*. Figure 5.3 depicts the uncertainty intervals for total direct cost of OA from 2010 to 2031. Here I discuss details of performing the sample size algorithm for the direct cost outcome in POHEM-OA (i.e. step 4, the MC method, as discussed in Section 5.4). In the initial step, I used the fractional factorial design (FFD) as described in Appendix C1 and performed three simulation runs with selected parameters replicates. I set the parameters at their mean value, mean minus the standard error (low value) and mean plus the standard error (high value) as recommended by the FFD method in Appendix C1. This has been done to cover the overall parameter space that results in 2 ˆ 2 high precision for the initial estimates of σˆ1 and σ 2 and will in turn provide acceptable (n*) and (m*) (i.e., n*>1). Table 5.6 shows the results of the initial step. Additionally, as discussed in the Step 4 of Section 5.3 for the aggregate outcomes such as total cost in this section, I first transfer the total outcome to average outcome by dividing the total cost by a fixed population size in each year, which was 32 million in 2031. As a result, 2 the initial σˆ1 averaged over the 3 runs of the initial step was calculated to be $107,293. 2 To calculate the initial estimate for σˆ 2 , I first calculated the overall uncertainty between 2 MC simulation runs and then deducted it from the σˆ1 as described in Step 4, Section 5.4. 180 Table 5-6. Results for the initial step for UA of total cost in the POHEM-OA model * MC-run 1 MC-run 2 (High) MC-run 3 (Low) Z (total cost) $5,144,469,327 $22,513,350,742 $678,638,047 ** zi (average cost) $159.60 $698.45 $21.053 ˆ 2 σ1 0.129 0.288 0.045 * Results are for year 2031; ** m=32,233,230 was the population size in 2031 which was used to transfer the total costs to average cost in 2031; ** zi represent average costs and were used to 2 calculate the initial estimate for parameter uncertainty (σˆ 2 ) and (m) and (n*) as discussed in Section 5.6. Next, I adjusted the results for parameter variance according to the fractional factorial design (FFD) for n0=3 runs as described in Appendix C2. This has been done due to the fact that in order to cover the overall parameter space in the initial step and provide reasonable initial estimates, instead of randomly selecting the parameter replicates, I selectively designed the experiment in the initial step according to the FFD (4). Therefore, I needed to adjust the resulting parameter variance according to the values presented in Appendix C6. As a result, by using (Eq.5-1) and resulting values for overall variation around the average cost and first order uncertainty as shown in Table 2 5.6, the resulting initial estimate for parameter uncertainty was σˆ 2 =0.36. Finally, to calculate n*, I replace (mn) by K or the computational time available K in (Eq.5-1). As it takes 50 minutes to run around a 40-million population size, i.e., full population at baseline including future birth and migration in the POHEM-OA model, I 181 calculated how many individuals can be simulated within a 12.5 hours run10 in the main run of the sample-size algorithm. The results was K= 602,885,550. The precision level (d) representing the V (µˆ) in the above calculations was chosen to be 0.0003 (Eq.5-1). This value needs to be very small and was recommended to be (0.0001-0.001) in Section 5.4, step 5. However, any number between 0.0001 up to 0.0003 would result a negative n *- that showed this level of precision was not possible to achieve within 12.5 hours of computational time, or it resulted in a very small (m*) which makes the 2 calculated σˆ 2 negative as discussed in Section 5.4. Therefore, the smallest possible value for d was 0.0003. I then calculated (n*), by replacing the resultant initial estimates 2 ˆ 2 σˆ1 , σ 2 and K into (Eq. 5-1), which resulted in n*= 2910, m*=207,131. In the main step, I used m*=207,000 for the population size and preformed n*=2910 MC simulation runs in the available.12.5 hours of computational time. I rounded down the population size to achieve at least the precision level (d=0.0003) for the resulting estimate for the variance of the overall mean (Eq.5-1). As discussed in step 4 of Section 5.4 for the aggregate outcomes, I transformed the outcome to average type by dividing the aggregate outcome by a fixed number in the UA procedure. In this example, I divided the total cost by the overall population size in each year (e.g., 207,000 in 2031), the size of the overall Canadian population. After the main runs is performed, I then, transformed the results back into total cost by multiplying the average outcome by the actual full population size at each (e.g., 32 million). This transformation needs to be performed since in the sample-size algorithm, I reduce the population size 10 We used 15 hours for computational time from which 2.5 hours was used in the initial step. 182 and therefore, the final total cost, for example, would not be correct if I use reduced population size rather than the actual size (e.g., total cost over the population size of 207,000 instead of 32 million). After the main run was performed, I translated the average results back to total cost for all (n*=2091) results in each year from 2010 to 2031 5.6.3 Uncertainty Intervals for the Total Direct Cost of OA As illustrated in Figure 5.3, I have shown that the total direct cost of OA will increase from $ 2.9 billion with 95 % UI of [$2.4, $3.1] billion dollars, to $7.6 billion dollars with 95% UI of [$6.2, $9.1] billion dollars (in 2010 $CAN actual cost), an almost 2.6 times increase in total OA-related direct costs. The uncertainty intervals were calculated according to the distribution of the total direct cost outcome, which was shown to be best approximated by the normal distribution. This is due to the fact that total cost outcomes appeared symmetrical around the mean and the cumulative distribution function (CDF) closely corresponded to that of a normal distribution. As a result, I have estimated the standard deviation of the outcome for each year from 2010 to 2031 and then calculated the 95% uncertainty intervals according to the intervals of the normal distribution. The resulted uncertainty intervals around the mean total cost are increasing from 2010 to 2031. This represents the decreasing certainty about the future events in the model. The 95% UI of [$6.2, $9.1] for the total cost in 2031 shows that conditional on the assumptions of the model and the parameters that were included in the UA procedure, the total direct cost of OA will be within this interval with 95% probability. 183 Figure 5-3. Results of uncertainty analysis: total direct cost of OA in Canada from 2010 to 2031 as predicted by the POHEM-OA model * Y-axis shows the total direct cost of OA in billion dollars (2010 $CAD); curves inside the graphs represent the mean estimate surrounded by 95% uncertainty intervals. ! ! 184 The resulted uncertainty intervals for total cost in Figure 5.3 are symmetrical, while those in Figure 5.2 for the prevalence of OA are not symmetrical. As described above, I have used different approaches to estimate these intervals; for the prevalence results, I have used the CDF of the results in each year and calculated the 2.5th and 97.5th percentile of the distribution (empirical distribution). On the other hand, for the total cost outcomes, I have calculated the uncertainty intervals according to the normal distribution’s confidence interval equations. This was done since the prevalence distribution that resulted from several runs was not a symmetrical distribution (i.e., an unknown distribution), while the total cost outcome appeared symmetrical and could be approximated with the normal distribution. Additionally, I have not used the BMI bootstraps for the BMI changes in the total cost outcomes. The second method used for the total cost outcome can be used to estimate standard deviations related to the parameter uncertainty (4). 5.7 Discussion Complex PDMS models are increasingly used in modeling chronic diseases (1, 175,176,195-198). UA is critical to the validation of such models and interpretation of their results (32,165). The goal of UA in PDMS models is to provide uncertainty intervals around each outcome. Applying the MC method to estimate the variance of the mean outcome is the core of the UA process. Several guidelines discuss the steps in performing UA to capture the uncertainty in risk assessment models (3,171) and cost- effectiveness models in health economics (2,162). However, there is a gap in the literature regarding the application of MC methods to UA in predictive PDMS models. To my knowledge, no UA guidelines have been developed specifically for PDMS models 185 that incorporate health outcomes other than cost-effectiveness ratios and simulate real populations rather than hypothetical cohorts. In this Chapter, I suggested an approach for estimating the uncertainty intervals for PDMS models that implicitly takes into account the complexity of the model, number of parameters, model calibration, and correlations among the parameters. I have outlined the steps involved in performing UA, including a practical approach to lowering the computational demand, developed by O’Hagan et al. (4). The key idea that led to the sample-size approach of O’Hagan et al. (4) is the trade-off between the number of MC simulation runs and the population size. In fact, by increasing the number of MC simulation runs while decreasing the population size, one can increase the precision level of the resultant uncertainty interval for a fixed computational time (4). This result has been discussed in both Rutter et al. (33) and Griffin et al. (174) for the optimal design of the UA. The major assumption is that the computational time is linear in m and n. In contrast to the optimal design, UA in PDMS models is often performed by running the model a small number of times due to the computational time constraint (32, 174). However, this practice is not acceptable due to the low precision of the resultant intervals. On the other hand, ignoring first order uncertainty (i.e., MC error) would result in a biased, and too small, uncertainty estimate of the outcome in MS models. The approach applied in this Chapter can be used to reduce the computation time while taking into account both the first and the second order uncertainty. Alternatively, one can use response surface approaches such as linear regression surfaces (189) and Gaussian process modeling (184) to reduce the required number of simulation runs. 186 The response surface methods can be also combined with the sample-size approach applied in this thesis from O’Hagan et al. (4). Future studies can compare these approaches in terms of their performances in estimating uncertainty intervals for MS models within a fixed computational time. It is worth emphasizing that the results of UA should be interpreted as conditional on the assumptions involved in performing this analysis. For example, ignoring the structural uncertainty should be reported in conjunction with the resulting uncertainty intervals. In addition, the resulting sample sizes of the UA approach for population size, number of MC simulation runs, and the precision level used in sample size calculations should be reported. In the examples of UA provided in Section 5.5 and 5.6, I have implemented the proposed approach into the POHEM-OA model for prevalence and total cost of OA. The resultant 95% uncertainty intervals for prevalence of OA from 2001-2021 incorporated the uncertainty associated with the hazard ratios for the effect of BMI on OA risk and use bootstraps to include the BMI trajectory model parameters. For the total cost of OA, I included the uncertainty associated with hazard ratios for effect of BMI on OA diagnosis, in addition to the uncertainty of major cost components. By adapting a sample-size approach developed by O’Hagan et al. (4), I were able to reduce the computational time from almost 12 days to 12 hours and 15 hours for prevalence and direct cost of OA, respectively. The estimated uncertainty interval approaches presented in this Chapter were not free of limitations. Several parameter types in the POHEM-OA model were not included in the UA approach for total direct cost and the prevalence outcomes. These include parameter types not related to these outcomes such as those involved in 187 estimating quality of life of patients in addition to those from large population-based databases including initial population parameters for birth, death and immigration processes, baseline incidence and prevalence of OA and those used for calibration procedures. In addition, I assumed no uncertainty for transition between different OA states (after diagnosis, surgeon visit, primary surgery, revision surgery) as they were estimated from population-based data (84). Further disadvantages and advantages of the sample-size approach used in estimating uncertainty intervals for prevalence and total cost of OA in future years are provided in Chapter 6. 188 Chapter 6: Discussion The chapters comprising this thesis are unified by the common goal of gaining a better understanding of the amount and distribution of direct cost burden of OA in the past and future years in addition to identifying major sub-populations who bear this cost. This thesis represents a comprehensive cost-of-illness program of research in conjunction with the use of a microsimulation model. I use a microsimulation model previously developed at Statistics Canada, named the Population Health Microsimulation Model (POHEM) to estimate the direct cost of osteoarthritis (OA) for the Canadian population. First, in Chapter 3, I estimated the average cost associated with OA between 2001 and 2010. Next, in Chapter 4, I predicted the future total direct cost of OA in Canada for the years 2010-2031. Finally, I provided a framework for performing uncertainty analysis (UA) in PDMS models for estimating the 95% uncertainty intervals (UI) for the mean and parameter uncertainty in Chapter 5. I implemented the proposed approaches in POHEM-OA to estimate the 95% uncertainty intervals (UI) for the prevalence of OA and the total direct cost estimates for individuals with OA from 2001 to 2031. Methodologically, this study provides evidence on how a microsimulation model can be used to estimate direct cost-related outcomes from both healthcare system and patient perspective. In this program of research, I have developed direct cost modules inside the simulation model, POHEM-OA, that reflect direct cost components from the societal perspective, including the cost borne by healthcare system, such as hospitalization cost, medications cost, and physicians cost, in addition to those endured by patients, such as over-the-counter drugs cost, alternative care cost and formal 189 caregiver cost. To model direct cost burden of OA, I first performed a bottom-up cost-of- illness (COI) study, using a BC administrative population-based database that covers almost all BC residents (PDBC) (84), BC Ministry of Health survey data collected in 2007 (147), Canadian Institute of Health Information (CIHI) data sources (144,151-154), Canadian Joint Replacement Registry (CJRR) data on the number of hip and knee replacement surgeries (124), St. Paul’s Hospital Costing Model (145) and literature estimates (140-43) to estimate per patient-year average cost for each cost component as a function of age, sex and disease stage. Next, I implemented them into the POHEM-OA model, a microsimulation model of OA previously developed by Kopec et al. (1). To study the uncertainty associated with my results, I developed a methodological framework for performing UA in PDMS models and then adapted the ANOVA-based sample-size approach previously developed for probabilistic sensitivity analysis (PSA) in decision analytic models by O’Hagan et al. (4). The goal of this approach was to reduce the computational time needed to perform UA in PDMS models. Using this approach, I have provided methods for calculating the uncertainty intervals around the mean outcome of the simulation model for prevalence and total cost of OA over the future decades. 6.1 Key Findings To address the overall goal of investigating the amount and distribution of direct cost of OA in Canada, its past and future trend, required synthesis of several different data sources, previous literature estimates, and development of new methodologies and statistical models. I have used microsimulation (MS) modeling as a tool to integrate and 190 unify different data sources and include all direct cost components associated with OA. In this thesis, I have used POHEM-OA, an MS model previously developed by Kopec et al. (1). The POHEM-OA model is a validated MS model that used several inter-related statistical models to project demographics of Canadian population in addition to the OA- related risk factors and health outcomes (1). In the first section of this thesis, I described methods to estimate the direct cost of OA using a traditional micro-costing approach integrated with an MS model. I further provided the results for average cost of OA from 2003 to 2010 in Chapter 3, and projection of total cost of OA in future from 2010 to 2031 in Chapter 4. In Chapter 3, I demonstrated how MS models could be used to fill in the gap in the literature in terms of OA cost data in Canada. To estimate the increasing trend for average cost of OA from 2003 to 2010, I imputed the direct cost data by the means of an MS model, i.e., POHEM-OA. I have first applied a micro-costing approach and estimated the average per patient-year direct costs of OA as a function of patients’ age, sex, and disease state from PDBC (84). I then used the POHEM-OA model that provided us with the number of OA cases according to patients’ age, sex and disease state, and integrated it with the result of the micro-costing approach. In the results section of Chapter 3, I showed that the average direct cost of OA was increased by almost 40% in 2003-2010, from $578 in 2003 to $811 in 2010 in current dollars. I provided the results for the average cost of OA across different age, sex and disease state categories. I also indicated that hospitalization cost due to hip and knee TJR surgery constituted the highest proportion of the average cost in 2003 and 2010, i.e., 21%, 32%, respectively. I reported the average cost of OA patients in different OA 191 states including: state 1) OA diagnosis, after OA diagnosis and before the orthopedic surgeon (OS) visit; state 2) OS visit, after visit to an orthopedic surgeon and before the replacement surgery; state 3) primary surgery during and after the hip/knee total joint replacement (TJR); state 4) revision surgery, during and after revision hip/knee TJR surgery. While patients in state 1 had the lowest average cost throughout the study period, they comprised the highest number of patients with OA (around $450 on average). Additionally, the average cost of patients in states 3 and 4 was almost 7 and 7.5 times higher than those in state 1, respectively (around $3,000 on average). The highest proportion of the average cost for patients in states 1, 2 and 3 was due to physician visits, out-of-pocket and surgery cost, respectively. The rehabilitation cost, including the patients’ out-of-pocket portion and the healthcare system’s portion, was 18% in 2003 and increased to 23% in 2010. Additionally, in Chapter 3, I performed a sensitivity analysis to investigate the effect of different cost drivers on the change in average cost of OA during the study period. I have identified the following cost drivers for average cost of OA from the highest to lowest, respectively: increase in the number of hip/knee TJRs, healthcare resources’ price inflation, effect of other cost drivers such as increase in the mean age and BMI of the population and increase in the life expectancy of population (due to a decrease of mortality rates). The integration of a micro-costing approach and POHEM-OA together with the projections of unit costs for direct cost components and number of hip/knee surgeries were carried out in Chapter 4 and results were first presented in the American College of Rheumatology (ACR) and published in proceedings of ACR, 2012 (219). For the first 192 time in Canada, in Chapter 4, I have used a validated individual-level simulation model and demonstrated that the total cost of OA would increase by almost 160% from 2010- 2031, from $2.9 billion to $7.6 billion in 2010 $CAD, while total number of people with OA would increase by almost 60% in the same time period from 3.6 million to almost 6 million OA cases. In addition, I described how the proportion of total cost would change across different cost components. The share of hospitalization cost was the highest among all other cost components and increased from 25% to 37%, while the next highest proportions were physician visit, drug and out-of-pocket cost, respectively. Consequently, I were able to identify sub-populations associated with high total costs in terms of age, sex and OA states. For example, female patients between the ages of 70- 80 showed to be the most expensive sub-population compared to other age and sex groups, with almost 1.6 billion dollars in total cost of OA in 2031 (in 2010 $CAD). A second component of this thesis was about providing a new methodology for performing uncertainty analysis (UA) in PDMS models. By adapting common concepts from published UA guidelines, I developed a comprehensive, step-by-step approach to UA in PDMS models In Chapter 5. The steps for performing UA in PDMS models were as follows: step 1) selecting the outcome for the UA approach; step 2) selecting the parameters to be included in the UA-list; step 3) assigning distributions to parameters (or using bootstraps); step 4) calculating the sample sizes for population size and number of Monte Carlo (MC) simulation runs and performing the MC method; and finally step 5) constructing the empirical distribution for the outcome and calculating the 95% uncertainty interval (UI) around the mean outcome. In step 4, I described the sample- size approach, a version of ANOVA-based approach first discussed in a 2007 study by 193 O’Hagan et al. (4) for which the UI’s around the mean estimate was calculated for any give precision level. I have discussed that due to the computational time constraint, and in order to the calculate a high precision level of the mean estimate, I need to reduce the population size as much as possible while increase the number of MC-runs (4). Finally, due to the nature of PDMS models, I need to provide inference for outcomes that are of any types. This includes outcomes in the form of: 1) average types such as average cost of a disease, prevalence, incidence; and 2) aggregate-based or summation types that are linear functions of average outcomes such as summation outcomes including total cost or total number of people with a certain disease or condition. I provided a two-step algorithm for the proposed sample-size approach in Chapter 5. As an illustration of the approach proposed in Chapter 5, I used the POHEM-OA model with prevalence of OA as the outcome and included hazard ratios for OA incidence as a function of individual’s sex, and BMI categories that were estimated according to National Population Health Data (NPHS) data (146). In addition, parameters related to the BMI progression model were also included in the UA-list of parameters. With the goal of estimating the mean outcome with the precision level of the variance of the mean outcome to be within (D=0.01) distance of its true value, i.e., d=0.0001 for the standard error, I calculated the sample sizes according to equation given in Chapter 5, Step 4. The size of the simulated population was then calculated to be 500,000 and the number of MC-runs was 785 for a 12 hours computational time. The estimated 95% UI’s for the prevalence of OA in Canada in 2021 were 0.09 to 0.18 for men and 0.15 to 0.23 for women and the uncertainty surrounding the sex-specific 194 prevalence of OA increased over time. The proposed approach to UA in PDMS models considers the challenges specific to such models, such as selection of parameters and calculation of sample sizes to reduce computational burden. The example of UA showed that the proposed approach is feasible and estimation of uncertainty intervals should become a standard practice in reporting results from PDMS models. As an illustration of the sample-size approach for the aggregate-based outcomes, I used the POHEM-OA model with total direct cost of OA as the second example in Chapter 5. I have included around seven different types of parameters including hazard ratios for OA incidence, in addition to the unit costs of TJR surgeries, side effect of drugs, alternative care, formal caregivers’ and rehabilitation services. The size of the simulated population was calculated to be m*=207,000 and n*=2910 for a 15 hour computational time. I have shown that the total direct cost of OA will increase from $ 2.9 billion with 95 % UI of [$2.4, $3.1] billion dollars, to $7.6 billion dollars with 95% UI of [$6.2, $9.1] billion dollars (in 2010 $CAN actual cost), an almost 2.6 times increase in total OA-related direct costs. 6.2 Integration and Implications of the Research Addressing the overall goal of this thesis called for synthesis of a wide body of literature, development of a novel approach in cost-of-illness studies by integrating a micro-costing COI with a microsimulation model, estimating the trend of average cost of OA over the previous years, projection of total cost of OA in future years, development of a framework for performing UA in PDMS models, estimating uncertainty associated with prevalence and total direct cost of OA from 2010 to 2031 and interpretation of findings for policy stakeholders, research and patient populations. In turn, as stand- 195 alone studies or as a collective work, this thesis offers potential contributions across several health-related disciplines including health economics, health administration, and policy analysis. Among other costly arthritic diseases in Canada, OA can be singled out in terms of its total cost burden (7,220). The increase in OA cost burden in recent years was in part due to the high rise in both of its major risk factors: age and obesity (5,89). It has been established that burden of OA and other chronic diseases will be significant in the future years in Canada and other industrialized societies as baby boomers are entering their final stages of life and obesity epidemics vastly propagates (10). However, there has not been a valid and precise model that provides a clear picture of how age, BMI and other risk factors affect the cost burden of OA (131, 220). The POHEM-OA cost model developed in this thesis has provided a broad view of how the future direct cost burden of OA would increase in Canada over the next 20 years in addition to the uncertainty around these estimates. Although more and more individual-level simulation models of OA have been developed in recent years (50,52,53), neither has been used to project the cost burden of OA. This thesis is the first study that uses MS modeling to estimate and project the direct cost burden of OA at the population-level. In this thesis, I have developed the cost modules within the POHEM-OA model, a population-based MS model that constructs the initial population according to actual data from adult population of Canada (1) using CCHS 20001 (54). As other methodological advancement in statistics and mathematics such as regression modeling and factor analysis (230) have aided the advancement of 196 epidemiology throughout its history, predictive PDMS models can be used to enhance studies in descriptive epidemiology. This thesis sheds light on the benefits of using PDMS models in descriptive epidemiology by providing applications of PDMS models in estimating and projecting the direct cost of OA. 6.2.1 Cost-of-Illness Methodology Although several COI studies for OA have used large population-based surveys or patients questionnaire to estimate the direct cost of OA associated with, for example, total joint replacement surgery (128), no study used a comprehensive list of cost components to estimate the national direct cost burden of OA in Canada that includes both patients’ and healthcare system’s costs. The main reason is the lack of a single database that includes utilization data for all direct resources associated with OA. For example, administrative databases such as PDBC (84) that were used for patients’ medical utilization and cost data in this thesis, do not include out-of-pocket costs of patients or alternative care costs such as those associated with physiotherapist visits. The only studies that include a comprehensive cost lists are those with a rather small sample sizes (114, 115). However, use of these studies to estimate the national burden of OA is extremely limited. Therefore, to be able to synthesize different types of databases including use of large population-based survey’s available in Canada (linked provincial databases for physicians, drugs and hospitalization cost), in addition to literature estimates and patient questionnaires for out-of-pocket costs, I need a tool that is capable of combining all data sources. In this thesis, I described how an MS model could be used to accomplish this. 197 Although several recent simulation models performed cost-effectiveness analysis of different OA-related interventions (47, 48, 50, 52), no COI study of OA has yet utilized an individual-level simulation model. As the systematic review of OA cost studies revealed in Chapter 2, traditional COI studies have been the target of criticism since they use single source of data and they are not capable of projecting the cost into future (57,62). Chapters 3 and 4 described a new advancement in COI studies, as they show how integrating a bottom-up type COI study within an MS model could address the traditional disadvantages of COI methodologies (57). In Chapters 3 and 4, I performed a bottom-up COI study in two phases; first, for the three publicly-funded cost categories of the total direct cost of OA, I calculated the per-patient year cost for each category using all individuals inside the PDBC (84) for both OA and non-OA cases (around 3 million individuals). Second, for other cost categories such as out-of-pocket cost and rehabilitation, I calculated the utilization rate and unit costs based on patients’ characteristics from OA surveys and literature estimates (see data sources in Chapter 3, Section 3.2.1). Finally, I used the POHEM-OA model as an integrating vehicle that allows the use of multiple data sources to both project the cost into future (Chapter 4) and fill the gap for cost data and estimate the average cost of OA in past years (Chapter 3). Recent individual-level simulation models of OA have been developed for estimating the current or future burden of OA and hip/knee TJRs (1, 51-53). However, except for the POHEM-OA model (1), rest of the simulation models used hypothetical population as their initial population (51-53). In a very recent study by Holt et al. (53), the OA Policy model (OAPol), a Markov individual-level model developed in US (52), 198 was used for forecasting the burden of advanced knee osteoarthritis over a 10-year period in a cohort of 60-64 year-old US adults (53). In that study, the distribution of the initial population according to patients’ age and sex were estimated from the population- based data. However, due to the structure of their model, they were not able to use the population-based data directly to populate the model and assign characteristics to individuals inside the initial population (e.g., using bootstrap methods as used in PDMS models such as the POHEM-OA model). Therefore, none of the models contained the tools to reflect the actual heterogeneity between patients (168) and their inherited correlated characteristics (221). As a result, POHEM-OA is empowered with the ability to predict the burden of OA into the future years with a higher predictive power compared to other individual-level models. Additionally, multiple complex statistical models can be utilized in such PDMS models to project several inter-related risk factors and their effects on disease burden – this increase the predictive power of MS models as compared to state-based Markov models or other aggregate-level (group-based) models (206). 6.2.2 Direct Cost of OA Results For the first time in Canada, in this thesis, I have analyzed the average and total direct cost of OA according to their distribution across different sub-populations, cost components and disease states. Direct cost of OA consists of all the cost components related to patients’ direct use of health care resources due to OA and its comorbidities including drugs, physician visits and other outpatient services, hospitalization, rehabilitation, out-of-pocket costs, and alternative care services. In Chapter 3, I provided the average cost of OA and its trend from 2003 to 2010, in addition to the change in its 199 distribution across patients’ age, sex and disease state. I showed that patients in state 2, those visited by orthopedic surgeon but in the wait list for the surgery, have the highest average out-of-pocket cost compared to patients in other OA states. Previous studies have also focused on estimating the cost burden associated with patients waiting for the surgery (128,137); but none has provided an overall view of costs associated with patients in terms of their flow into the healthcare system (223). My findings have implications for policy analysts, as it partly reflects the benefits associated with reducing the waiting time for TJR surgeries (222). Additionally, these results can be further used in decision analysis studies for assigning resources to OA patients awaiting the surgery. While recent studies delineate the significant share of the out-of-pocket cost of OA (137), In Chapter 3, I have provided and compared the out-of-pocket cost share of patients in different OA states. I showed that the out-of-pocket cost share of the patients in state 3, i.e., in the post-surgery state for the primary surgery, have increased from 18% to 23%, which was the highest increase compared to patients in other states. These results have significant implications for policy analysis, in terms of public coverage of services such as physiotherapists and other alternative care cost burden to reduce the out-of-pocket costs for patients at post-TJR surgery state (131,222,223). As the systematic review in Chapter 2 revealed, one of the main disadvantages of COI studies is their lack of ability to provide cost estimates for different sub- populations and identify cost drivers. In Chapters 3 and 4, I showed that by integrating the micro-costing COI approach with an MS model, I were able to provide cost estimates across different sub-populations. For example, in Chapter 4, I showed that 200 people between 70 and 80 years of age would have the highest direct cost after 2025, while patients between 60 and 70 years of age have the highest cost in 2010. This has major implications for policy analysts in terms of expanding resources for patients over 70 years old in the future years (after 2025) to decrease the cost burden of OA in these years (131,224). Additionally, in Chapter 3, I have analyzed and identified major cost drivers for the increase in average cost of OA in 2003-2010. I have shown that the average cost increased from $735 to $811 between 2003 and 2010 (in 2010 $CAD), for which 33% was due to the increase in number of surgeries, 30% to inflation, 10% to increase in the life expectancy of the population (due to mortality rates) and 27% to other factors including increase in mean age of the population (due to increase in birth and migration rates) and other factors such as increase in mean BMI of the population. My results were in accord with the CIHI study for the cost drivers of the health expenditure in 2010 (161). Authors of this report used a simulation model to calculate the effect of each cost driver on the total health care expenditure in Canada between 1998 and 2008 (161). To explain the increase in the health expenditure, the authors of the CIHI report cite such factors as the increase in healthcare prices, life expectancy and mean age of the population (161). In Chapter 4, I have projected the total cost burden of OA in (2010-2031), considering the inflation rate for each cost component separately. I have shown that, the total cost burden of OA would increase by almost 2.5 times, after discounting the total cost back to 2010. If the cost were reported according to 2031 $CAD, the total cost would increase by more than this amount (the total cost would be almost quadrupled in 2031 (in 2031 $CAD) as compared to 2010 (in 2010 $CAD) (228)). However, from the 201 economic point of view, presenting the cost in 2031 dollar value is nor reasonable, as it is in the future, and therefore, I presented the results all in 2010 $CAD and provided thee different scenarios according to different discount rates and economic growth rates for future years. In Chapter 4, I showed that the hospitalization share of the total cost was the highest compared to other cost components, i.e., drugs, physicians and out-of- pocket; hospitalization share of the total cost was around 28% in 2010 and was projected to increase up to 39% in 2031. Almost 95% of the hospitalization cost was due to hip and knee TJR surgeries and 5% due to other hospital procedures for OA patients (detailed list for all procedures are provided in Table A5, Appendix A). Although the LOS after hip and knee TJRs was decreasing from 2010 to 2031, I showed that, the cost of TJR would still be increasing mostly due to its very high inflation rate. This result shows that, to reduce the cost burden associated with OA patients, researchers need to focus on ways to reduce the procedure cost of hip/knee TJRs that includes prosthesis and materials costs. Using data from the Canadian Joint Replacement Registry (CJRR) data for the past decade, I have projected that, in 2025, LOS for knee and hip TJR will become as short as two days (I assumed no further decrease after 2 days). The very short LOS after the hip/knee TJRs would result in an increase of the number of hip/knee TJRs as patients would discharge faster form hospitals after the surgery. However, this would increase the cost associated with post-surgery rehabilitation as patients may not receive appropriate care at the facility or at home after the surgery. In spite of these findings and those mentioned in other studies (222), new researches are focused on reducing the LOS after the hip/knee TJRs, e.g., recent non-invasive day-surgery TJR (223), 202 without much attention to the cost and quality of life of patients in the post-surgery stages. The results of this thesis show the benefits of investing in research studies to develop new TJR technology to further reduce the ever-increasing cost of TJR procedures and prosthesis, while providing cost-effective care during the post-stages of the surgery - rather than investing in TJR surgeries with high prosthesis costs and very low LOS, without considering the post-surgical state and cost burden of patients. One of the unique findings of this study was the increase in terms of the costs associated with the side effects of drugs consumed by OA patients during the study period from 2010 to 2031. According to the POHEM-OA cost model, the total cost of side effects due to OA drugs will increase from $305 million dollars in 2010 to $689 million dollars in 2031 (all in 2010 $CAD). This is due to the increase in the cost associated with dyspepsia, CVD, stroke and serious GI complications. Additionally, the total number of cases of all of the aforementioned side effects increased by almost 70%. I have shown that due to a high lifetime cost per case of CVD and stroke compared to dyspepsia and GI complications, the total cost of stroke and CVD due to OA drugs are projected to increase by almost 144%, from $228 million to $551 million, while the total cost of dyspepsia and serious GI complications due to OA drugs are projected to increase by 81%, from $76 million to $138 million during the study period (all in 2010 $CAD). 6.2.3 Uncertainty Analysis and Methodological Findings As mentioned in recent guidelines for decision analytic and environmental models (166,169, 186), to provide inference on health measures of a population, MS models should be equipped with tools to estimate the uncertainty associated with the 203 outcome. Uncertainty Analysis (UA) should be performed in a systematic fashion, often by implementing the numerical Monte Carlo (MC) approach after the basic mean- estimate simulation model is built (182, 186). Brute-force MC-analysis, as applied in probabilistic sensitivity analysis (PSA) of decision analytic models, is not applicable for PDMS models due to burdensome computational time of individual-level models (174). Sources of variation that I included in the analysis in Chapter 5 were the variability between individuals and the parameters (189). In addition to variance reduction techniques in sampling approaches, such as Latin Hyper-cube Sampling (LHS), and orthogonal sampling (225), several other approaches have been devised for reducing the time of UA for simulation models. Among the non-parametric methods, ANOVA-based approach of O’Hagan et al. (4) was the only non-parametric approach discussed for PSA in individual-level decision analytic models. However, their approach is not designed for PDMS models; for example, it is not suited for all types of outcomes such as prevalence of a disease or its total cost. I adapted the ANOVA-based approach of O’Hagan et al. (4) and developed a sample-size approach in PDMS models to provide an unbiased minimum-variance estimator for the mean outcome and its 95% UI’s. The solution to the sample-size approach was obtained from a method developed in O’Hagan et al. (4). According to this solution, I reduced the population size as much as possible, while increasing the number of MC-runs within a fixed computational time. However, for practicality of this approach, I showed in Chapter 5 how to calculate the sample sizes for any given precision level of the variance. 204 6.3 Strengths and Limitations of the Research Each chapter of this thesis was accompanied with its own study-specific discussion. However, this section will focus on the collective thesis work, and highlight the key strengths and limitations as they apply across studies. 6.3.1 Limitations of the Research While the implementation of standardized cost categories has been established in guidelines for economic evaluations, there is no guiding principle for this type of standardization in COI studies (62). Although, in the last decade, COI studies have been moving toward using a more comprehensive cost listings in OA and arthritis studies, no standardized cost listings for these diseases in Canada have been published yet (other than that for overall musculoskeletal diseases (61)). Therefore, in Chapter 2, I performed a literature review of previous COI studies of OA, and integrated all cost components used for direct cost in these studies11. Also, in Chapters 3, 4, I used a comprehensive list of cost components from societal perspective that included medical and non-medical cost categories. However, I did not estimate direct cost of OA associated with nursing home services and health research. Data regarding OA patients entering nursing homes is not captured in provincial databases, and limited data is available in other sources. More importantly, it is not possible to assign nursing home costs to OA patients as it is unknown whether the admission of a patient into the facility was only due to OA - as high level of comorbidity exists among elderly with OA. On the other hand, health research expenditure data within OA is very limited and it is 11 We used all direct cost components that were common in at least two COI studies of OA. 205 cumbersome to predict the research expenditure in future years. Since it is an aggregate variable and is not related to OA burden at the individual level, it is best to be treated as an external variable. In 2003 expenditure report for musculoskeletal diseases, total research cost was reported to be around 1 billion dollars in 2003 $CAD (123). However, it is unknown how much of this amount was spent for OA research. One of the limitations in my study was missing data in the MOH-OA survey that I used to estimate the cost associated with alternative care professionals’ services prior to hip/knee TJR surgery. The cost associated with physiotherapy and other alternative care professionals after the hip/knee TJR were estimated within the rehabilitation cost category (Chapter 3). The cost associated with alternative care professionals included visits to physiotherapists, chiropractors or other complementary care providers. While the data was provided for ~1700 OA patients, several patients reported to have rheumatoid arthritis (RA) or other type of arthritis (~70 patients). I have deleted these subjects, as I needed to estimate the parameters related to alternative care visits due to OA. I have also reduced the sample size to include only OA patients prior to hip/ knee replacement surgeries (~250 cases had surgeries). Among the remaining 1284 patients, around 10% did not respond to one or more questions regarding the visits to any alternative care professionals. As a result, the final sample used for the model included n=1124 OA patients who did not have hip/knee surgery and responded to all questions for physiotherapy and chiropractic visits, and n=1118 patients who responded to questions regarding other complementary care visits. In this study, I did not implement a micro-costing approach for side effects of OA- related drugs. In fact, I assigned an aggregate-level cost that represented a lifetime cost 206 for CVD and stroke, and an incidence-based cost for GI and dyspepsia. As discussed in Chapter 3.2 for drug costs, I have provided a comprehensive list of all drugs associated with OA, and categorized them within 4 major drug categories including NSAIDs, opiods, coxibs and acetaminophen; but these do not include detailed models for stroke, CVD, GI or dyspepsia medications (such as proton-pump inhibitors). Rather, I assigned an aggregate amount that included the weighted average for all costs associated with these diseases from literature estimates. This, in return, reduced the predictive power of the model in terms of costs related to side effect of drugs. As I discussed in Chapter 2, external validation in cost evaluation studies was defined in terms of the generalizability of the data source used for each cost component, meaning that the data sources should be representative of all cost centers (62). In our micro-costing study, in Chapter 3, I used the results from a prior COAST study that used administrative data, PDBC (84), for estimating the per patient-year unit costs which then used as an input to the POHEM-OA model. The linked PDBC data was used to calculate the utilization rate of patients in terms of health resources including physicians, drugs and hospital procedures within the province (84). Although in Canada the provincial administrative data sources enable researchers to partially calculate public expenditures associated with chronic diseases, these data sources do not provide the full picture, such as out-of-pocket costs or alternative care costs. For example, PDBC includes almost 18 years of health care system utilization (1998-2012) for almost all BC’s residents of drugs, physicians, and hospitalization (84). However, it does not provide any data regarding the alternative care utilization by OA patients, such as physiotherapy visits, or out-of-pocket resources such as patients’ transportation and 207 formal caregivers. Therefore, I used other datasets for cost calculations related to out- of-pocket resources and other cost categories. For example, inpatients procedure cost was calculated from one hospital within the province (St. Paul’s costing model in 2003 (145)), while CIHI data was used for calculating hip and knee TJR surgery that represented the overall Canadian variation for cost (I used the average CPWC for Canada (144)). Except for the latter CIHI data used for TJRs unit costs, the rest of cost unit calculations were performed for resources in one province of Canada (i.e., on hospital in BC), and cannot be generalized. Therefore, as I used these input parameters into the simulation model to predict the overall Canadian cost, lack of generalizability of input parameters calculations for the mentioned unit costs would prevent the external validation of the costing methodology (62). On the other hand, due to the use of simulation modeling that accounted for the heterogeneity within Canadian population, the internal validity of costing methods can be established (62) (I discuss this in the next section). An important limitation in my studies regarding cost input parameters calculations was the fact that I used data from different sources at different time points (Table 3.1). However, to project the cost into future or impute cost data for previous years, I used the POHEM-OA simulation model that provided us with the change in demographic structure of the population such as age of the population, change in risk factors such as BMI that affects the OA diagnosis and/or change in distribution of disease severity across the population (1). Nonetheless, some of the changes at the population level were not reflected due to the use of only 1-year cost and utilization data for some cost components (physician visits, drugs and hospitalization procedures other than hip/knee 208 TJRs). For example, change in population use of resources such as increase in the utilization rate of physician visit or change in utilization of drug use was not reflected in the model. As reported in the 2010 Healthcare Cost Drivers Report by CIHI in Canada (161), one of the cost drivers in terms of drug expenditure over the last decade was the increase in utilization rates of medication - this may be due to higher number of visits to physicians in recent years, or change in physicians practices and treatments guidelines (161). However, it should be mentioned that I used inflation rates from a recent CIHI reports on change in health-care-sector prices (151-153) and projected the change in price of each cost components, separately, from 2010 to 2031. Conversely, the POHEM-OA cost model does not reflect the change in utilization of health resources by OA patients other than those for hip and knee TJR surgery that I modeled using CJRR historical data (124). I projected the time to surgery (after visit of an orthopedic surgeon) in future years using the model to estimate the LOS after hip/knee TJR from the CJRR historical data (124). Since the number of surgeries in POHEM-OA was estimated according to data from 1998-2003, I used CJRR trend from 2003 to 2009 and calibrated the future rates (124). In fact, I used historical data and by use of a regression model, estimated a time trend for change in number of TJR surgeries (Appendix B, Figure B2). I then performed calibration so that the number of surgeries in the model would reflect the result of the regression model. This is another limitation of my model. In fact, instead of incorporating a causal model that links the change in number of TJR surgeries to policy changes, either due to the changes in guidelines of TJR surgery, or capacity of hospitals, I used calibration and directly included the trend for number of TJR surgeries 209 as a variable in the model. This would be contrary to the goal of PDMS models that should be causal-based and model the behavior of individuals (221), instead of calibrating the outcomes. Although, use of the calibrated trends may increase the predictive ability of the PDMS models in the short term, their flexibility and applicability in terms of performing what-if scenarios would be deceased. For example, in the model, I won’t be able to examine the effect of change in the capacity of hospitals on the number of hip/knee TJRs, and on the overall burden of OA in future. In other words, there is a tradeoff between predicative power of PDMS models, on one hand, and their applicability to perform what-if scenarios according to their causal nature, on the other hand (226). In Chapters 5, I provided the methodology for performing UA in PDMS models for the mean and parameter uncertainty. In this Chapter, I addressed first and second order uncertainty, i.e., MC-error and parameter uncertainty. However, other aspects of model uncertainty such as those associated with assumptions, alternative databases, model structure, statistical models and alternative definitions were not included. In the proposed UA approach in Chapter 5, I have assumed that the results were conditional on the true values for all sources of uncertainty and parameters not included the UA. I, then, discussed that other sources of model uncertainties such as alternative structures and statistical models or databases can be investigated with what-if scenarios (35). Although including all parameters into the proposed UA would be ideal for uncertainty procedure, in Chapter 5, I have discussed approaches to select the most significant parameters in terms of their effect on the overall uncertainty to reduce the computational time and complexity of the UA. 210 With the frequentist statistical point of view of Chapter 5, I were not able to provide confidence intervals for population-based estimates across the reference Canadian population (and over time). For example, the overall cost per patient-year parameters that were used for physicians, drugs and hospitalization procedures other than hip/knee TJRs were calculated from PDBC (Table A7-A11, Appendix A). However, I were not able to assign confidence intervals, distribution or covariance measures between these costs parameters. However, in the Bayesian perspective, I would be able to assign a prior belief in the form of distributions to all the cost parameters based on expert opinions, for example, by comparison between BC data and overall Canadian cost estimates as described in Oakly et al. (216). As a result, the final uncertainty interval estimates for the total direct cost of OA are conditional on the given values for these parameter types. 6.3.2 Strengths of the Research Through studies presented in this thesis I have revealed the advantages of PDMS models with regard to performing COI studies. These advantages can be summarized as the followings: integrating different COI approaches and different data sources, imputing missing data and extrapolate beyond data, and finally performing what-if scenarios to disentangle the effects of different cost drivers in COI studies. Similar advantages have been identified by Weinstein et al. (14) for benefits of modeling in cost evaluation studies. Flexibility of PDMS models allows the analyst to use several different approaches for different cost components in addition to integrate different data sources. For example, in a bottom-up, incremental cost study, I can include both types of regression- 211 based and matched cohort approaches (57) into one COI study with the use of a PDMS model. In Chapters 3 and 4 of this thesis, on one hand, I included the result of the matched cohort study using administrative data for cost of physician visits (84) and CIHI estimates for hospital procedure costs (144), and on the other hand, I included estimates from regression-based models for physiotherapist visits using MOH-OA survey data (147). Simulation of individuals’ characteristics and behaviors in PDMS models would allow analysts to estimate health-related outcomes in the past and future and, therefore, impute missing data, while extrapolate beyond data (14). In terms of performing COI within a PDMS model, I were able to impute the average cost of OA in the past decade from 2003 to 2010 using administrative and OA-related survey data and other literature estimates. Additionally, use of microsimulation modeling in a COI study allowed us to identify distribution of cost across sub-populations, cost components and disease states. On the other hand, although I only used specific input data, use of a simulation modeling aided us to project the cost burden of OA among the Canadian population over the future decades. As a result of using the POHEM-OA model, results of the COI analysis reflect the distribution of the direct cost of OA among the Canadian population. Another strong aspect of the cost analysis in Chapters 3 and 4 was the fact that internal validations were established for the costing methodology and simulation model. As mentioned in Larg et al. (62), two types of reliability measures were used in the literature to evaluate the reliability of costing methods in terms of internal validity (62): methods and case mix criteria. According to methods criteria, the costs for all relevant resources should be included in the cost listings (including overhead costs) and these 212 costs should be appropriately valued (excluding any excess profits or unpaid allowances). According to the case mix criteria, the cost output units, i.e., the objects in terms of which the costs are expressed, should be classified so that they reflect the resource intensity of the services being measured (62). As discussed in Chapter 3, all the parameters were estimated according to patient characteristics including age, sex, and disease state. For cost of TJR surgery and other hospitalization procedures for OA, I used CIHI data (144) that estimated cost of surgery according to case mix approaches to reflect patients’ heterogeneity. While both measures for interval validation were taken into account in the analysis, I did not include variation of unit cost across different resources such as use of administrative and cost data from different provinces and hospitals, which prevented us from establishing the external validity of the costing methodology. This was discussed in the limitation section of this Chapter. A key feature of the UA approach developed in Chapter 5 is its applicability to PDMS models in which two level of uncertainty exist. As mentioned in Chapter 5, I differentiate between predictive disease models and decision analytic models. In terms of the language used, the term UA is typically applied in predictive models, while probabilistic sensitivity analysis (PSA) is used in decision analytic models (173). According to Briggs et al. (193) in an aggregate-level (group-based) models such as cohort Markov models, the expected value of cost and utilities for each treatment arm is calculated without introducing any randomness. In a second stage of the analysis, the PSA is performed and the parameters are presented with distributions (193). In individual-level predictive models, on the other hand, random sampling is performed, even in the base case model for estimating the mean outcome (226). This type of 213 sampling is based on drawing a random value for assigning an event to each individual (for example, occurrence of a disease) (4). In this sense, in individual-level predictive PDMS models, I simulate the intrinsically uncertain nature of events over time (4), and, therefore, two levels of uncertainty exist in such models. A practical step-wise approach that address two level of uncertainty for the predictive PDMS models was first elaborated in this thesis (35). Another feature of the UA approach developed in Chapter 5 is its applicability to any type of outcome in the predictive PDMS models. In decision analytic models, the outcome is in the form of cost-effectiveness ratio, while in predictive PDMS models any type of outcome such as prevalence, incidence, and cost or cost-effectiveness ratio needs to be estimated related to the disease in the model. In decision analytic models, PSA is often performed after the base-case model is constructed and according to different willingness to pay thresholds, the optimal scenario or intervention is chosen (180). For example, in a cost-effectiveness analysis, cost-effectiveness acceptability curves (CEAC) are reported where they represent the uncertainty surrounding the adoption of the specific policy at different willingness-to-pay threshold (180). On the other hand, in predictive-based models, where the focus is not only on the decision between several alternatives, different types of disease-related outcomes are measured. The UA developed in Chapter 5 of this thesis can be used for any type of outcomes in PDMS models including average types, such as prevalence or incidence of a disease, or aggregate types such as total cost of a disease or total number of individuals with a disease. 214 6.4 Recommendations and Future Research Results of the studies performed in this thesis can be used as a basis for policy developments and future researches to reduce the high cost associated with OA management. 6.4.1 Recommendations The results of the simulation-based COI studies performed in Chapters 3 and 4 translate the adverse effects of OA into dollar terms, the universal language of decision makers and the policy arena (93). These estimates was used to define the magnitude of the OA in dollar terms and can be further used to: 1) justify intervention programs; 2) assist in the allocation of research dollars on specific interventions; 3) provide a basis for policy and planning relative to prevention and control initiatives; and 4) provide an economic framework for program evaluation (93). Given the magnitude of OA costs and the related impact on other chronic conditions, developing polices to cut the cost associated with OA by 10-20%, for instance, would make an extraordinary reduction in the increasing rate of the healthcare expenditures in US and Canada (131). With the aid of the POHEM-OA simulation model in this thesis, I have shown that increase the TJR surgeries and inflation in the cost of healthcare resources for patients with OA are among the major drivers of the average OA cost. Policies developed to target the cost drivers identified in this study, can reduce the OA cost burden significantly. As shown by the studies in this thesis, the cost burden of OA will be high in the elderly population over the next 20 years. Developing polices targeted at integrating OA management for elderly could reduce the cost burden associated with OA care. 215 Developing these policies over the next decades becomes very important in Canada and US as the number of elderly (over 60 years of age) will increase by almost 78 million by 2027 in US and more than 9 million by 2021 in Canada (131). In the elderly population, it has been estimated that nearly half would show signs of OA at 65 years of age or older and are at increased risk for other comorbid chronic conditions (131). Development of novel, integrated health delivery programs for OA management among the elderly in recent years has shown significant improvements in terms of increasing the quality of life of patients and decreasing the cost compared to the current OA management guidelines (229). Examples of such novel care paradigm for elderly include those developed in small settings in US including patient-centered medical home, where there is an assigned coach in charge of patient care and communicating among multiple provider (131), or Partners healthcare and Care group that worked to decrease the number of emergency rooms and hospital stays by elderly by adding a health coach integrated into the OA management (131). Recent OARSI guidelines for OA care and management recommend conservative care in early and mild stages of OA to limit the use of highly invasive surgery in severe OA stages (86). Results of this study can be used to justify the effect of providing publicly funded paid insurance coverage for early OA treatments that are included in the conservative OA management guidelines; example of these treatments are physical therapies after TJR, aerobic and strengthening exercises, braces, shoe wedges (224). As shown in this thesis, cost of alternative care and caregiver is going to increase with a fast rate in future decades, specifically for patients who are on the wait- lists for hip and knee TJR surgeries (Chapter 3). However, several early stage OA 216 interventions are not currently offered by healthcare system in Canada or not covered by insurance programs for OA patients in other countries (131). Also, those that are covered are only for a predetermined number of visits such as few number of physiotherapy visits after hip/knee TJR surgeries (222). As shown in this thesis, the out-of-pocket cost for OA patients is going to rise in future years by almost 100% from 2010 to 2031 and the overall out-of-pocket cost will be around $1 billion $CAD in 2031 in Canada (in 2010 $CAD). Compared to other cost components, the out-of-pocket cost is going to be very high in 2031, specifically, among patients at post-surgical stage (Chapter 3). Development of new delivery systems to implement conservative OA management guidelines (86), in addition to providing cost- effective rehabilitation interventions such as physical therapies, exercise therapies, and home-based treatments (224) have shown to increase the quality of care and reduce the out-of-pocket cost of OA for patients at post-surgical stages (131). Results of this study calls for a comprehensive guideline for physicians and orthopedic surgeons to manage OA-related number of TJR surgeries as the ever increasing rate for hip and knee TJRs would reach a non-manageable level by 2031 in Canada (around 160,000). In Chapter 4, I have projected the number of OA-related hip and knee TJR surgeries according to its recent increasing rate (124). I have shown that there is an excess burden of TJR surgeries exerted into the system by policy makers in addition to the number of surgeries that is caused by increasing number of OA patients – this is also mentioned in other studies (52). As a result, the total number of hip/knee TJR surgeries could reach a level beyond the capacity of all hospitals in Canada by 2031. Results of this thesis calls for developing policies to manage the capacity of hip 217 and knee TJRs, and to reduce the variation in terms of policies addressing the excess burden of TJR surgeries. 6.4.2 Future studies Future studies can use the models developed in this thesis to investigate interventions and new OA management strategies to reduce its cost burden. In terms of UA methodology in PDMS models, future studies can further extend the methodologies of the study in order to develop stochastic sensitivity analysis methods within PDMS models to find significant parameters that affect the uncertainty of outcomes. The POHEM-OA simulation model that was enhanced in this thesis can be used to further investigate the cost-effectiveness of such strategies such as exercise therapy and physical therapy during pre-surgery stages to improve the outcome of surgeries (52, 224). Although improvements have been made in surgical practice, TJR is associated with attenuated risk of fracture, and is not considered appropriate for many individuals with concomitant comorbidities (224). Additionally, it requires months of rehabilitation services and augments a high cost burden to the overall OA cost. According to recent studies, exercise and strengthening interventions during the pre- surgery stages of OA are supporting forces that considerably increase the success of the total knee replacement, lower patients suffering and pain, and enable on-going quality of life function (131). As a result, there is a need to focus on policies promoting pre-surgical non-pharmacological treatment strategies to improve the outcome of surgeries (131, 220, 224). The results of Chapter 4 have significant implications in terms of future research on ways to reduce the cost burden associated with TJR among OA patients. As majority 218 of hospitalization cost is due to hip and knee TJRs, the cost modules developed in this Chapter can be used to investigate different strategies to reduce the cost associated with new types of TJR in the future (220). Although LOS decreased over the last decade, which in part caused the cost of surgery to decrease, cost of prosthesis was increasing due to new emerging surgery technologies and more complex methods for replacement surgeries (223). Accordingly, the rate of inflation that I calculated from the literature was very high for hip/knee TJRs compared to other cost components (Chapter 4, Table 4.1). More importantly, the less the patients stay in the hospitals, the higher the cost of post-surgery rehabilitation would become - since patients would not receive appropriate care after release from hospitals and this may cause additional out-of- pocket costs, alternative care costs or even increase the chance for a revision surgery (222). Therefore, decrease of LOS alone would not aid in decreasing the cost associated with TJR. Future research needs to focus on developing ways to improve the care of patients after their release from the hospitals (220). The models developed in this thesis can be used to investigate the cost-effectiveness of new technologies for TJR surgery with regard to their shorter LOS complemented with rehabilitation strategies such as home-based rehabilitation (131). To estimate the contribution of each parameter on the total uncertainty of the outcome, methodologies have been developed for stochastic sensitivity analysis (SA) in environmental models (179,225). Same type of approaches has been developed under the value of information methodology in decision analytic literature (180). However, to my knowledge, there has not been any method developed for stochastic SA in PDMS models. Future studies can further extend the sample-size approach developed in 219 Chapter 5 and develop a framework for performing stochastic sensitivity analysis in PDMS models. Estimating the effect of each parameter on the overall uncertainty would benefit the modeler on several aspects. First, as mentioned in Saltelli et al. (179), in almost all simulation models, there is a small set of key parameters that are responsible for majority of the uncertainty in the outcome. Therefore, if these key parameters were identified, analyst would gain an important knowledge about the operations and workings of the model that would further aid in reducing the uncertainty and establishing the external validity of the model (225). Second, future research can be justified according to the value of information analysis (193). In the value of information analysis, one tries to distinguish sensitive (important and influential) parameters for future research in terms of the contribution of each parameter on the uncertainty of the overall outcome. If the expected value of future research exceeds its cost, additional research for that parameter is justified (193). Future studies can develop methodologies to objectively measure the uncertainty produced by different types of model structure in PDSM. As mentioned in the limitations section of this Chapter, due to the frequentist perspective throughout this thesis, I were not able to estimate the uncertainty associated with the model structures including data sources and statistical models. Additionally, I were not able to assign correlation coefficients to parameters estimated from separate data sets, or assign uncertainty to parameters calculated from population-based data. For example, PDBC (84) was used for calibrating the OA incidence and prevalence of the overall Canadian population and, therefore, it was not possible to assign a level of uncertainty to the results (estimates) from population-based data. Future studies can incorporate Bayesian perspective into 220 the UA methodology developed in this thesis. Within the Bayesian perspective, one could assign a degree of truth according to a prior knowledge of the data source (e.g., population-based data) by using a distribution for the estimates or assume a correlation between cost estimates (190). Bayesian approaches for UA in decision analysis literature have been previously described in the literature (190, 216). 6.5 Summary of Thesis Recent availability of population-based data in Canada in addition to the advent of high computing power provides a unique opportunity to develop and utilize large population-based disease microsimulation(PDMS) models. These types of simulation models provide a virtual replica of a population and allows researchers and policy analysts to generate birth cohorts according to individuals’ characteristics, age them with regard to their health-related risk factors and project measures of disease, health determinants, and cost of an illness at the population-level (1, 32). Taken collectively, in this thesis, three applications of microsimulation modeling in a cost-of-illness (COI) study were presented using POHEM-OA, a PDMS model developed by Kopec et al. (1). First, in Chapter 3, I estimated the historical trend of the average direct cost of OA from 2003 to 2010 in Canada, in addition to identifying its cost drivers. I showed that the average cost of OA would increase from $735 to $811 in 2010 $CAD during the study period. Additionally, increase in the number of hip/knee TJR surgeries and price inflation of health resources were identified as the major cost drivers. Second, I projected the total direct cost of OA from 2010 to 2031 and described how it is predicted to increase by almost 160% from $2.9 billion dollars to $7.6 billion dollars in 2010 $CAD under a real cost scenario. Additionally, I analyzed the total cost 221 distribution according to individuals’ age, sex and disease state. Finally, I provided a framework for performing uncertainty analysis (UA) in PDMS models. To reduce the computational time associated with UA in PDMS models, I adapted the sample-size approach in Chapter 5 from a method developed in O’Hagan et al. (4). By incorporating the proposed approach into the POHEM-OA model, I have demonstrated that the total direct cost of OA would increase from $2.9 billion with 95% UI of [$2.4, $3.1] billion dollars, to $7.6 billion dollars with 95% UI of [$6.2, $9.1] billion dollars (all in 2010 $CAD). The results of this study calls for investments in developing methods to reduce the cost burden of OA in future years through novel OA management strategies. 222 References 1. Kopec JA, Sayre EC, Flanagan WM, Fines P, Cibere J, Rahman MM, et al. Development of a population-based microsimulation model of osteoarthritis in Canada. Osteoarthritis and Cartilage. 2010 Mar;18(3):303–11. 2. Bilcke J, Beutels P, Brisson M, Jit M. Accounting for methodological, structural, and parameter uncertainty in decision-analytic models: a practical guide. Medical Decision Making!. 2011;31(4):675–92. 3. Poulter SR. Monte Carlo simulation in environmental risk assessment – science, policy and legal Issues. Risk: Health, Safety & Environment. 1998;7 (Winter):7– 26. 4. O’Hagan A, Stevenson, Madan. Monte Carlo probabilistic sensitivity analysis for patient level simulation models: Efficient estimation of mean and variance using ANOVA. Health Economics. 2007;16(1):1009-1023. 5. Zhang Y, Jordan JM. Epidemiology of osteoarthritis. Clinical Geriatrics Medicine. 2010 Aug;26(3):355-69. 6. Buckwalter JA, Saltzman C, Brown T. The impact of osteoarthritis: implications for research. Clinical Orthopaedics and Related Research. 2004 Oct(427 Suppl):S6- 15. 7. Badley EM. The effect of osteoarthritis on disability and health care use in Canada. The Journal of Rheumatology Supplement. 1995 Feb;43:19-22. 8. Kopec JA, Rahman MM, Sayre EC, Cibere J, Flanagan WM, Aghajanian J, et al. Trends in physician-diagnosed osteoarthritis incidence in an administrative database in British Columbia, Canada, 1996-1997 through 2003-2004. Arthritis and Rheumatism. 2008 Jul 15;59(7):929–34. 9. The Impact Of Arthritis In Canada!: Today And Over The Next 30 Years. [internet]. Arthritis Alliance of Canada reports. 2011.[cited 2012 Sep 01]. Available from: http://www.arthritisalliance.ca/docs/20111022_2200_impact_of_arthritis.pdf 10. Badley EM, Wang PP. Arthritis and the aging population: projections of arthritis prevalence in Canada 1991 to 2031. The Journal of Rheumatology. 1998 Jan;25(1):138-44. 11. Teichtahl AJ, Wang Y, Wluka AE, Cicuttini FM. Obesity and knee osteoarthritis: new insights provided by body composition studies. Obesity (Silver Spring). 2008 Feb;16(2):232-250. 223 12. Murray CJ, Vos T, Lozano R, Naghavi M, Flaxman AD, Michaud C, Ezzati M, Shibuya K, Salomon JA, Abdalla S, Aboyans V, Abraham J, Ackerman I, Aggarwal R, Ahn SY, Ali MK, Alvarado M, Anderson HR, Anderson LM, et al. Disability-adjusted life years (DALYs) for 291 diseases and injuries in 21 regions, 1990-2010: a systematic analysis for the Global Burden of Disease Study 2010. Lancet. 2012 Dec 15;380(9859):2197-223. 13. Yelin E. The economics of osteoarthritis. In: Brandt K, Doherty M, Lohmander LS, editors. Osteoarthritis. New York: Oxford University Press; 1998. p. 23-30. 14. Weinstein MC. Recent developments in decision-analytic modelling for economic evaluation. Pharmaco Economics. 2006;24(11):1043-53. 15. Vanness DJ, Tosteson AN a, Gabriel SE, Melton LJ. The need for microsimulation to evaluate osteoporosis interventions. Osteoporosis International. 2005 Apr;16(4):353–8. 16. Fone D, Hollinghurst S, Temple M, Round A, Lester N, Weightman A, et al. Systematic review of the use and value of computer simulation modelling in population health and health care delivery. Journal of Public Health Medicine. 2003 Dec;25(4):325-35. 17. Karnon J, Goyder E, Tappenden P, McPhie S, Towers I, Brazier J, et al. A review and critique of modelling in prioritising and designing screening programmes. Health technology assessment 2007; 11: 1-145. 18. Wolf, DA.The role of microsimulation in Longitudinal data analysis. [Internet]. Papers in Microsimulation Series Paper No. 6, Center for Policy Research, The Maxwell School. Syracuse, NY Syracuse University, Feb 2001, 33 pp. [cited 2012 Feb 01]. Available from : http://www.maxwell.syr.edu/cpr/publications/microsimulation/Microsimulation_Pap er_No__6/. 19. Greenhalgh D, Hay G. Mathematical modelling of the spread of HIV/AIDS amongst injecting drug users. IMA Journal of Mathematics Applied in Medicine and Biology. 1997 Mar;14(1):11-38. 20. Sendi PP, Craig B a., Pfluger D, Gafni A, Bucher HC. Systematic validation of disease models for pharmacoeconomic evaluations. Swiss HIV Cohort Study. Journal of Evaluation in Clinical Practice. 1999 Aug;5(3):283-95. 21. Brewer TF, Heymann SJ, Harris JB. Tuberculosis control in Asia and the western Pacific: a role for computer modelling. Annals of the Academy of Medicine, Singapore. 1997 Sep;26(5):642-6. 224 22. Mather D, Crofts N. A computer model of the spread of hepatitis C virus among injecting drug users. European Journal of Epidemiology. 1999 Jan;15(1):5-10. 23. Ramsey SD, McIntosh M, Etzioni R, Urban N. Simulation modeling of outcomes and cost effectiveness. Hematology Clinics of North America. 2000 Aug;14(4):925-38. 24. Retsky MW, Demicheli R, Swartzendruber DE, Bame PD, Wardwell RH, Bonadonna G, et al. Computer simulation of a breast cancer metastasis model. Breast Cancer Research and Treatment. 1997 Sep;45(2):193-202. 25. Ross KS, Carter HB, Pearson JD, Guess H a. Comparative efficiency of prostate- specific antigen screening strategies for prostate cancer detection. Journal of the American Medical Association. 2000 Sep 20;284(11):1399-405. 26. Urban N, Drescher C, Etzioni R, Colby C. Use of a stochastic simulation model to identify an efficient protocol for ovarian cancer screening. Controlled Clinical Trials. 1997 Jun;18(3):251-70. 27. Paltiel a D, Fuhlbrigge a L, Kitch BT, Liljas B, Weiss ST, Neumann PJ, et al. Cost- effectiveness of inhaled corticosteroids in adults with mild-to-moderate asthma: results from the asthma policy model. The Journal of allergy and Clinical Immunology. 2001 July;108(1):39-49 28. Tomar RH, Lee S, Wu SY, Klein R, Klein BE, Moss SE, et al. Disease progression and cost of insulin dependent diabetes mellitus: development and application of a simulation model. Journal of the Society for Health Systems. 1998;5(4):24-37. 29. Naidoo B, Thorogood M, McPherson K, Gunning-Schepers LJ. Modelling the effects of increased physical activity on coronary heart disease in England and Wales. Journal of Epidemiology and Community Health. 1997 Apr;51(2):144-50. 30. Zethraeus N, Johannesson M, Jönsson B. A computer model to analyze the cost- effectiveness of hormone replacement therapy. International Journal of Technology Assessment in Health Care. 1999 Spring;15(2):352-65. 31. Roersma ME, Veldhuis GJ. Proposal and evaluation of a Monte Carlo model for hair regrowth following plucking. Skin Research and Technology. 2001 Aug;7(3):176-83 32. Kopec JA, Finès P, Manuel DG, Buckeridge DL, Flanagan WM, Oderkirk J, et al. Validation of population-based disease simulation models: a review of concepts and methods. BMC Public Health. 2010 Jan;10(1):710. 225 33. Rutter CM, Zaslavsky AM, Feuer EJ. Dynamic microsimulation models for health outcomes: a review. Medical Decision Making. 2011;31(1):10–8. 34. Gupta A, Harding A. Modelling our Future - Population Ageing, Health and Aged Care. Bingely, UK: Emerald Group Publishing Limited; 2007. 35. Sharif B, Kopec JA, Wong H, Finès P, Sayre EC, Liu RR, et al. Uncertainty Analysis in Population-Based Disease Microsimulation Models. Epidemiology Research International. 2012;2012:1–14. 36. Wolfson MC. Social Proprioception: Measurement, Data, and Information from a Population Health Perspective. In: Evans RG, Barer ML, Marmor TR (Editors). Why Are Some People Healthy and Others Not? The Determinants or Health of Populations. Aldine de Gruyter, New York;1994, p. 287-316. 37. Ellwood P. Outcomes Management, a Technology of Patient Experience. The New England Journal of Medicine.1988,318(23): 1549-1556. 38. Coombs JW. The report of the National Task Force on Health Information--a summary. Health Reports. 1991;3(4):349-61. 39. Habbema JDF, Schechter CB, Cronin K a, Clarke LD, Feuer EJ. Modeling cancer natural history, epidemiology, and control: reflections on the CISNET breast group experience. Journal of the National Cancer Institute. Monographs. 2006;(36):122- 6 40. Hoogendoorn M, Rutten-van Mölken MPMH, Hoogenveen RT, van Genugten MLL, Buist a S, Wouters EFM, et al. A dynamic population model of disease progression in COPD. The European Respiratory Journal. 2005 Aug;26(2):223- 33. 41. Paltiel a D, Weinstein MC, Kimmel AD, Seage GR, Losina E, Zhang H, et al. Expanded screening for HIV in the United States--an analysis of cost- effectiveness. The New England Journal of Medicine. 2005. Feb 10;352(6):586- 95. 42. Unal B, Capewell S, Critchley JA. Coronary heart disease policy models: a systematic review. BMC Public Health. 2006 Aug 18;6:213. 43. Computer modeling of diabetes and its complications: a report on the Fourth Mount Hood Challenge Meeting. Value in Health. 2013 Jun;16(4):670-85. 44. Wolfson MC. POHEM--a framework for understanding and modelling the health of human populations. World Health Statistics Auarterly. 1994;47(3-4):157-76. 226 45. Will BP, Berthelot J, Nobrega KM, Flanagan W, Evans WK. Canada ’ s Population Health Model ( POHEM ): a tool for performing economic evaluations of cancer control interventions. European Journal of Cancer. 2001;37:1797–804. 46. Will BP, Berthelot JM, Le Petit C, Tomiak EM, Verma S, Evans WK. Estimates of the lifetime costs of breast cancer treatment in Canada. European Journal of Cancer. 2000 Apr;36(6):724-35. 47. Busija L, Bridgett L, Williams SRM, Osborne RH, Buchbinder R, March L, et al. Osteoarthritis: Best practice & research. Clinical Rheumatology. 2010 Dec;24(6):757–68. 48. Wielage RC, Bansal M, Andrews JS, Klein RW, Happich M. Cost-utility analysis of duloxetine in osteoarthritis: a US private payer perspective. Applied Health Economics And Health Policy.2013 Jun; 11(3):219–36. 49. Reichmann WM, Wright EA, Holt HL, Solomon DH. Cost-effectiveness of Total Knee Arthroplasty in the United States. Archives of Internal Medicine. 2009;169(12):1113–21. 50. Higashi H, Barendregt JJ. Cost-effectiveness of total hip and knee replacements for the Australian population with osteoarthritis: discrete-event simulation model. PLoS One. 2011 Jan;6(9):e25403 51. Losina E, Daigle ME, Suter LG, Hunter DJ, Solomon DH, Walensky RP, et al. Disease-modifying drugs for knee osteoarthritis: can they be cost-effective? Osteoarthritis and Cartilage. 2013 May;21(5):655–67. 52. Weinstein AM, Rome BN, Reichmann WM, Collins JE, Burbine S a, Thornhill TS, et al. Estimating the burden of total knee replacement in the United States. The Journal of Bone and Joint Surgery. 2013 Mar 6;95(5):385–92. 53. Holt HL, Katz JN, Reichmann WM, Gerlovin H, Wright E a, Hunter DJ, et al. Forecasting the burden of advanced knee osteoarthritis over a 10-year period in a cohort of 60-64 year-old US adults. Osteoarthritis and cartilage; 2011 Jan; 19(1):44–50. 54. Statistics Canada. Canadian Community Health Survey (CCHS). 2001 Microdata File User Guide .Ottawa, Ontario: Health Statistics Division, Statistics Canada. [cited 2010 June 01]. Available from: http://www.statcan.gc.ca/imdb- bmdi/document/3226_D7_T9_V4- 55. Van Imhoff E, Post W. Microsimulation methods for population projection. Population: an English selection. 1998 Jan;10(1):97–138. 227 56. Clabaugh G, Ward MM. Cost-of-illness studies in the United States: a systematic review of methodologies used for direct cost. Value In Health. 2008;11(1):13–21. 57. Akobundu E, Ju J, Blatt L, Mullins CD. Cost-of-illness studies!: a review of current methods. PharmacoEconomics.2006 Jan;24(9):869–90. 58. Bitton R. The economic burden of osteoarthritis. The American journal of managed care. 2009 Sep;15(8 Suppl):S230–5. 59. Leardini G, Vaccaro E. Osteoarthritis: Socioeconomic Problems. Seminars in Arthritis and Rheumatism. 2005 June;34(6 Suppl 2):35–7. 60. Kotlarz H, Gunnarsson CL, Fang H, Rizzo JA. Insurer and Out-of-Pocket Costs of Osteoarthritis in the US: Evidence From National Survey Data. Arthritis & Rheumatology. 2009;60(12):3546–53. 61. Mittendorf T, Merkesdal S, Huelsemann JL, von der Schulenburg J-M, Zeidler H, Ruof J. Implementing standardized cost categories within economic evaluations in musculoskeletal diseases. The European Journal of Health Economics. 2003;4(1):43–9. 62. Larg A, Moss JR. Cost-of-illness studies: a guide to critical evaluation. PharmacoEconomics. 2011 Aug;29(8):653–71. 63. Helmick CG, Felson DT, Lawrence RC, Gabriel S, Hirsch R KC. Estimates of the prevalence of arthritis and other rheumatic conditions in the United States. Part 1. Arthritis & Rheumatology. 2008;58(1):15–25. 64. Hunter DJ. Imaging insights on the epidemiology and pathophysiology of osteoarthritis. Rheum Dis Clin North Am. 2007;35(3):447–63. 65. Bauman A. Updating the evidence that physical activity is good for health: an epidemiological review 2000-2003. J Sci Med Sport. 7(1 Suppl):6–19. 66. Buckwalter JA, Mankin HJ, Groodzinsky HG. Articular cartilage and osteoarthritis. Instr Course Lect. 2005;54:465-80 67. Haydon E, Roerecke M, Giesbrecht N, Rehm J, Kobus-matthews M. Chronic Disease in Ontario and Canada!: Determinants , Risk Factors and Prevention Priorities. [internet].2006 March.[cited 2012 Feb 01]. Available from: http://ocdpa.echidnadev.ca/sites/default/files/publications/CDP-FullReport- Mar06.pdf 68. WHO Scientific Group on the Burden of Musculoskeletal Conditions at the Start of the New Millennium. The burden of musculoskeletal diseases at the start of the 228 new millennium. [Internet]. Geneva: World Health Organization,. World Health Organ Tech Rep Ser. 2003;919:i-x, 1-218, back cover.[cited 2013 Feb 01]. Available from: http://www.boneandjointdecade.org 69. Woolf AD, Pfleger B. Burden of major musculoskeletal conditions. Bulletin of the World Health Organization. 2003 Jan;81(9):646–56. 70. Lawrence RC, Felson DT, Helmick CG, Arnold LM, Choi H, Deyo RA, Gabriel S, Hirsch R, et al. Estimates of the prevalence of arthritis and other rheumatic conditions in the United States. Part II. Arthritis and Rheumatism. 2008 Jan;58(1):26–35. 71. Chronic Disease Surveillance and Monitoring Division, Public Health Agency of Canada (PHAC). Life with Arthritis in Canada: a personal and public health challenge [Internet]. Ottawa, Canada; 2010.[cited 2012 March 01]. Available from: http://www.phac-aspc.gc.ca/cd-mc/arthritis-arthrite/lwaic-vaaac-10/index-eng.php 72. Arthritis Community Research, and Evaluation Unit (ACREU). Arthritis in Ontario [Internet]. 2013. [cited March 2013]. Available from: https://www.arthritis.ca/document.doc?id=908 73. Lawrence RC, Felson DT, Helmick CG, Arnold LM, Choi H, Deyo R a, et al. Estimates of the prevalence of arthritis and other rheumatic conditions in the United States. Part II. Arthritis and Rheumatism. 2008 Jan;58(1):26–35. Available from: http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid=3266664&tool=pmcentr ez&rendertype=abstract 74. Kopec JA, Rahman MM, Berthelot J-M, Le Petit C, Aghajanian J, Sayre EC, et al. Descriptive epidemiology of osteoarthritis in British Columbia, Canada. The Journal of Rheumatology. 2007 Feb;34(2):386-93. 75. Cooke TD, Daniel W, Feldman D et al. The American College of Rheumatology criteria for the classification and reporting of osteoarthritis of the hip. Arthritis and Rheumatology. 1991;34(5):505–14. 76. Maheu E, Altman RD, Bloch DA, Doherty M, Hochberg M, Mannoni A, Punzi L, Spector T, Verbruggen G, Osteoarthritis Research Society International Hand OA Task Force, Carr A, Cicuttini F, Dreiser RL, Haraoui BP, Hart D, Pelletier JP, Ramonda R RL. Design and conduct of clinical trials in patients with OA of the hand: recommendation from a task force of the OARSI. Osteoarthritis and Cartilage. 2006;14:303–22. 229 77. Felson DT, Naimark A, Anderson J, Kazis L, Castelli W MR. The prevalence of knee osteoarthritis in the elderly. The Framingham Osteoarthritis Study. Arthritis and Rheumatology. 1994;30(8):914–8. 78. Jordan JM, Helmick CG, Renner JB, Luta G, Dragomir AD, Woodard J et al. Prevalence of knee symptoms and radiographic and symptomatic knee osteoarthritis in African Americans and Caucasians: the Johnston County Osteoarthritis Project. The Journal of Rheumatology. 2001;34(1):172–80. 79. Tepper S HM. Factors associated with hip osteoarthritis: data from the First National Health and Nutrition Examination Survey (NHANES-I). Americal Journal of Epidemiology. 2005;137(10):1081–8. 80. McConnell S, Kolopack P, Davis a M. The Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC): a review of its utility and measurement properties. Arthritis and Rheumatism. 2001 Oct;45(5):453–61. 81. Harrold LR, Yood R a, Straus W, Andrade SE, Reed JI, Cernieux J, et al. Challenges of estimating health service utilization for osteoarthritis patients on a population level. The Journal of rheumatology. 2002 Sep;29(9):1931–6. 82. Ladouceur M, Rahme E, Pineau C a, Joseph L. Robustness of prevalence estimates derived from misclassified data from administrative databases. Biometrics. 2007 Mar;63(1):272–9. 83. Pinto D, Robertson MC, Hansen P AJ. Good agreement between questionnaire and administrative databases for health care use and costs in patients with osteoarthritis. BMC Medical Research Methodology. 2011 Apr ;13:11-45. 84. British Columbia’s pan-provincial population health data services. Population Data BC (British Columbia, Canada)- Multi-university platform [Internet]. [cited 2012 Sep 10]. Available from: http://www.popdata.bc.ca 85. Brandt K, Doherty M LL, editors. Osteoarthritis. New York, New York, USA: Oxford University Press;1998. 86. Zhang W, Moskowitz RW, Nuki G, Abramson S, Altman RD, Arden N, et al. OARSI recommendations for the management of hip and knee osteoarthritis, Part II: OARSI evidence-based, expert consensus guidelines. Osteoarthritis and Cartilage. 2008 Feb;16(2):137–62. 87. Segal L, Day SE, Chapman AB, Osborne RH. Can we reduce disease burden from osteoarthritis? The Medical Journal of Australia. 2004 Mar 1;180(5 Suppl):S11–7. 230 88. Murphy L, Helmick CG. The impact of osteoarthritis in the United States: a population-health perspective: A population-based review of the fourth most common cause of hospitalization in U.S. adults. Orthopedic Nursing. 2012;31(2):85–91. 89. Perruccio AV, Power JD, Badley EM. Revisiting arthritis prevalence projections-- it’s more than just the aging of the population. The Journal of Rheumatology. 2006 Sep;33(9):1856–62. 90. Segel JE. Cost-of-Illness Studies — A Primer. [internet]. RTI International. RTI- UNC Center of Excellence in Health Promotion Economics. 2006;(January):1–39. [cited 2012 March 01]. Available from: http://www.rti.org/pubs/coi_primer.pdf 91. Neumann PJ, Hermann RC, Kuntz KM, Araki SS, Duff SB, Leon J, Berenbaum PA, Goldman PA, Williams LW and WM. Cost-Effectiveness of Donepezil in the Treatment of Mild or Moderate Alzheimer’s Disease. Neurology. 1999;52(6):1138–45. 92. Rice DP, Fox PJ, Max W, Webber PA, Lindeman DA, Hauck WW and SE. The Economic Burden of Alzheimer’s Disease Care. Health Affairs. 1993;12(2):164– 76. 93. Rice DP. Cost of illness studies: what is good about them? Injury Prevention. 2000 Sep;6(3):177–9. 94. Anis AH, Zhang W, Bansback N, Guh DP, Amarsi Z BC. Obesity and overweight in Canada: an updated cost-of-illness study. Obesity Review. 2010;11(1):31–40. 95. Patra J, Popova S, Rehm J, Bondy S, Flint R, Giesbrecht N. Economic Cost of Chronic Disease in Canada 1995-2003. Health Canada. [internet]. 2007(March): 1-37. [cited 2012 April 01]. Available from: http://ocdpa.ca/sites/default/files/publications/OCDPA_EconomicCosts.pdf 96. Yelin E, Murphy L, Cisternas MG, Foreman AJ, Pasta DJ HC. Medical care expenditures and earnings losses among persons with arthritis and other rheumatic conditions in 2003, and comparisons with 1997. Arthritis Rheum.2007; 56(5):1397–407. 97. March LM, Bachmeier CJ. Economics of osteoarthritis: a global perspective. Baillieres Clin Rheumatol. 1997. 11(4):817–34. 98. American Academy of orthopedic Surgeons.The Burden of Musculoskeletal Diseases in the United States: Prevalence, Societal and Economic Cost [Internet]. Rosemont, IL, USA; 2008. [cited 2012 Sep 01].Available from: www.boneandjointburden.org 231 99. National Collaborating Centre for Methods and Tools [Internet]. McMaster University. 2014. [cited 2012 April 01]. Available from: http://www.nccmt.ca/index- eng.html 100. Canadian Research Index [Internet]. Micromedia ProQuest. 2014. [cited 2012 April 01]. Available from: http://libaccess.mcmaster.ca/login?url=http://search.proquest.com/canadianresear ch/advanced?accountid=12347 101. Xie F, Thumboo J, Li S-C. True difference or something else? Problems in cost of osteoarthritis studies. Seminars in Arthritis and Rheumatism.2007 Oct ;37(2):127– 32. 102. Petitti D. Meta-analysis, decision analysis, and cost-effectiveness analysis: methods for quantitative synthesis in medicine. New York, New York, USA: Oxford University Press; 1994. 103. Stafinski T, Menon D. The Burden of Osteoarthritis In Canada [Internet]. Institute of Health Economics, Edmonton, Alberta, Canada. 2003; [cited 2011 April 01]. Available from: http://www.ihe.ca/documents/2001-03paper.pdf 104. Chen a, Gupte C, Akhtar K, Smith P, Cobb J. The Global Economic Cost of Osteoarthritis: How the UK Compares. Arthritis. 2012 Jan;2012:698709. 105. Tarride J-E, Haq M, O’Reilly DJ, Bowen JM, Xie F, Dolovich L, et al. The excess burden of osteoarthritis in the province of Ontario, Canada. Arthritis and Rheumatism. 2012 Apr;64(4):1153–61. 106. Le TK, Montejano LB, Cao Z, Zhao Y, Ang D. Health care costs in US patients with and without a diagnosis of osteoarthritis. Journal of Pain Research . 2012 Jan;5:23–30. 107. Loza E, Lopez-Gomez JM, Abasolo L, Maese J, Carmona L, Batlle-Gualda E. Economic burden of knee and hip osteoarthritis in Spain. Arthritis and Rheumatism. 2009 Feb 15;61(2):158–65. 108. Rabenda V, Manette C, Lemmens R, Mariani A-M, Struvay N, Reginster J-Y. Direct and indirect costs attributable to osteoarthritis in active subjects. The Journal of rheumatology. 2006 Jun;33(6):1152–8. 109. Xie F, Thumboo J, Fong KY, Lo NN, Yeo SJ, Yang KY LS. Direct and indirect costs of osteoarthritis in Singapore: a comparative study among multiethnic Asian patients with osteoarthritis.2008; J Rheumatol. 34(1):165–71. 232 110. Hermans J, Koopmanschap MA, BiermaZeinstra SM, van Linge JH, Verhaar JA, Reijman M BA. Productivity costs and medical costs among working patients with knee osteoarthritis. 2010;Arthritis Care Res (Hoboken). 64(6):853–61. 111. Dibonaventura MD, Gupta S, McDonald M, Sadosky A, Pettitt D, Silverman S. Impact of self-rated osteoarthritis severity in an employed population: cross- sectional analysis of data from the national health and wellness survey. Health and Quality of Life Outcomes. 2012 Jan;10(1):30. 112. Dibonaventura MD, Gupta S, McDonald M, Sadosky A. Evaluating the health and economic impact of osteoarthritis pain in the workforce: results from the National Health and Wellness Survey. BMC Musculoskeletal Disorders. 2011 Jan;12(1):83. 113. White AG, Birnbaum HG, Janagap C, Buteau S, Schein J. Direct and indirect costs of pain therapy for osteoarthritis in an insured population in the United States. Journal of Occupational and Environmental medicine.2008 Sep;50(9):998–1005. 114. Maetzel A, Li LC, Pencharz J. The economic burden associated with osteoarthritis, rheumatoid arthritis, and hypertension: a comparative study. Annals of the Rheumatic Diseases. 2004;63:395–401. 115. Gupta S, Hawker GA, Laporte A, Croxford R, Coyte PC. The economic burden of disabling hip and knee osteoarthritis ( OA ) from the perspective of individuals living with this condition. Rheumatology (Oxford, England). 2005;44(August):1531–7. 116. Levy E, A. Ferme, D. Perocheau, and I. Bono . Socioeconomic costs of osteoarthritis in France. Revue du Rhumatisme.1993;60(6):63S–67S. 117. Le Pen, C. Reygrobellet and IG´. Financial cost of osteoarthritis in France: the “COART” France study. Joint, bone, spine!: revue du rhumatisme. 2005;72(6):567–70. 118. Leardini G, Salaffi F, Caporali R, Canesi B, Rovati L, Montanelli R. Direct and indirect costs of osteoarthritis of the knee. Clinical and Experimental Rheumatology. 2004;22(6):699–706. 119. Lanes SF, Lanza LL, Radensky PW, Yood R a, Meenan RF, Walker a M, et al. Resource utilization and cost of care for rheumatoid arthritis and osteoarthritis in a managed care setting: the importance of drug and surgery costs. Arthritis and Rheumatism. 1997 Aug;40(8):1475–81. 120. MacLean CH, Knight K, Paulus H, Brook RH, Shekelle PG. Costs attributable to osteoarthritis. The Journal of Rheumatology;1998 Nov;25(11):2213–8. 233 121. Woo J, Lau E, Lau CS, Lee P, Zhang J, Kwok T, et al. Socioeconomic impact of osteoarthritis in Hong Kong: utilization of health and social services, and direct and indirect costs. Arthritis and Rheumatism. 2003 Aug 15;49(4):526–34. 122. Murphy LB, Helmick CG, Cisternas MG YE. Estimating medical costs attributable to osteoarthritis in the US population: comment on the article by Kotlarz et al. Arthritis Rheum. 2010;62(8):2566–7. 123. Health Canada. Arthritis in Canada. An ongoing challenge.[internet] Ottawa, Canada; 2003. [cited 2011 Oct]. Available from: http://www.acreu.ca/pdf/Arthritis_in_Canada.pdf 124. Canadian Institute of health Information (CIHI). Canadian Joint Replacement Registry (CJRR) [Internet]. CIHI. 2011. Available from: http://www.cihi.ca/CIHI- ext-portal/xlsx/internet/stats_cjrr2011_clinicaldata_en 125. Carr AJ, Robertsson O, Graves S, Price AJ, Arden NK, Judge A, et al. Knee replacement. Lancet. 2012 Apr 7;379(9823):1331–40. 126. Canadian Institute of health Information (CIHI). Hip and Knee Replacements in Canada!: Canadian Joint Replacement Registry 2013 Annual Report [Internet]. 2013. Available from: https://secure.cihi.ca/free_products/CJRR_2013_Annual_Report_EN.pdf 127. Lapsley HM, March LM, Tribe KL, Cross MJ, Brooks PM. Living with osteoarthritis: patient expenditures, health status, and social impact. Arthritis and Rheumatism. 2001 Jun;45(3):301–6. 128. Hawker G a, Badley EM, Croxford R, Coyte PC, Glazier RH, Guan J, et al. A population-based nested case-control study of the costs of hip and knee replacement surgery. Medical Care. 2009 Jul;47(7):732–41. 129. Bozic KJ, Stacey B, Berger A, Sadosky A, Oster G. Resource utilization and costs before and after total joint arthroplasty. BMC Health Services Research. 2012 Jan ;12(1):73. 130. Phillips. CJ. Health economics: an introduction for health professionals. London, UK: Blackwell Publishing, BMJ Books; 2005. 131. Gruber WH, Hunter DJ. Transforming osteoarthritis care in an era of health care reform. Clinics in Geriatric Medicine. 2010 Aug;26(3):433–44. 132. Gignac M a M, Cao X, Lacaille D, Anis AH, Badley EM. Arthritis-related work transitions: a prospective analysis of reported productivity losses, work changes, and leaving the labor force. Arthritis and Rheumatism. 2008 Dec;59(12):1805–13. 234 133. Kotlarz H, Gunnarsson CL, Fang H, Rizzo J a. Osteoarthritis and absenteeism costs: evidence from US National Survey Data. Journal of Occupational and Environmental Medicine. 2010 Mar;52(3):263–8. 134. Statistics Canada. The Consumer Price Index [Internet]. Ottawa, Canada; 2012. [cited 2012 April]; Available from: http://www.statcan.gc.ca/tables-tableaux/sum- som/l01/cst01/econ46a-eng.htm 135. Oldmeadow MC, GD, Mcburney H, Robertson VJ. Predicting Risk of Extended Inpatient Rehabilitation After Hip or Knee Arthroplasty. Journal of Arthroplast2y. 2003 Sep;18(6):775-9. 136. Coyte PC, Young W, Croxford R. Costs and outcomes associated with alternative discharge strategies following joint replacement surgery: analysis of an observational study using a propensity score. Journal of Health Economics. 2000 Nov;19(6):907–29. 137. March L, Cross M, Tribe K, Lapsley H, Courtenay B, Brooks P, et al. Cost of joint replacement surgery for osteoarthritis!: the patients’ perspective. Journal of Rheumatology. 2002 May;29(5):1006-14. 138. Mitchell C, Walker J, Walters S, Morgan AB, Binns T, Mathers N. Costs and effectiveness of pre- and post-operative home physiotherapy for total knee replacement: randomized controlled trial. Journal of Evaluation in Clinical Practice.2005 Jun;11(3):283–92. 139. Sayre EC, Rahman MM, Aghajanian J, Kang W, Cibere J, Anis AH, Jordan JM, Badley EM, Kopec JA. A comparison of provincial prescription-only pharmaceutical database with self-reported usage of acetaminophen and NSAIDs according to osteoarthritis stage in British Columbia. Value in Health 2008;11:A267-8. 13th Annual International Meeting of the International Society for Pharmacoeconomics and Outcomes Research (ISPOR), Toronto, ON, Canada, May 3-7, 2008. 140. Birnbaum H, Leong S, Kabra A. Lifetime medical costs for women: Cardiovascular disease, diabetes, and stress urinary incontinence. Womens Health Issues 2003;13:204-13. 141. Taylor TN, Davis PH, Torner JC, Holmes J, Meyer JW JM. Lifetime cost of stroke in the United States. Stroke. 1999. 27(9):1459–66. 142. Rahme E, Joseph L, Kong SX, Watson DJ, LeLorier J. Cost of prescribed NSAID-related gastrointestinal adverse events in elderly patients. British journal of clinical pharmacology. 2001 Aug;52(2):185–92. 235 143. Moayyedi P, Mason J. Clinical and economic consequences of dyspepsia in the community. Gut. 2002 May;50 Suppl 4:iv10–2. 144. Canadian Institute of health Information (CIHI). CMG Client Tables 2011 [Internet]. DAD acute inpatient data. 2012 [cited 2011 Jan 10]. Available from: https://secure.cihi.ca/estore/productFamily.htm?pf=PFC1627&lang=en&media=0 145. Borsa J, Anis A. The cost of hospital care in Canada: a comparison of two alternatives. Healthcare Manage Forum. 2005 Spring;18(1):19-27. 146. Statistics Canada. National Population Health Survey data (NPHS) [internet]. Survey information available; [cited 2011 Aug 20]. Available from: http://www.statcan.gc.ca/cgi- bin/imdb/p2SV.pl?Function=getSurvey&SDDS=3225&lang=en&db=imdb&adm=8 &dis=2 147. Li L. British Columbia Osteoarthritis Survey [Internet]. Vancouver, Canada; 2008 pp. 1–43.[cited Sep 2013]. Available from: http://arthritis.rehab.med.ubc.ca/files/2011/08/BCOASurvey.pdf 148. Population Data BC Ministry of Health Services. Medical Services Plan (MSP) Payment Information File [Internet]. 2013. Available from: https://www.popdata.bc.ca/data/internal/health/msp 149. Organization for economic and co-operation and development (OECD). OECD- StatExtracts,OECD iLibrary, PPPs and exchange rates. [cited Feb 2012]. Available at : http://stats.oecd.org/Index.aspx?datasetcode=SNA_TABLE4 150. Population Data BC Ministry of Health Services. Discharge Abstracts Database (Hospital Separations file) [Internet]. 2013. Available from: https://www.popdata.bc.ca/data/internal/health/dad 151. Canadian Institute for Health Information. Hospital Cost Drivers Technical Report; What Factors Have Determined Hospital Expenditure Trends in Canada? [Internet]. Ottawa, Canada; 2010. Available from: http://www.cihi.ca/cihi-ext- portal/pdf/internet/hospital_costdriver_tech_en 152. Canadian Institute of health Information (CIHI). Drug Expenditure in Canada , 1985 to 2011. Spending and Health Workforce [Internet]. Ottawa, Canada; 2012. Available from: https://secure.cihi.ca/free_products/DEIC_1985_2011_EN.pdf 153. Canadian Institute for Health information. Health Care Cost Drivers!: Physician Expenditure — Technical Report Spending and Health Workforce [Internet]. Ottawa, Canada; 2009. Available from: http://www.cihi.ca/cihi-ext- portal/pdf/internet/health_costdriver_phys_tech_en 236 154. Average Payment per Physician Report , Fee-for-Service Physicians in Canada 2004-2005 [Internet]. Ottawa, Canada: National Physician Database, Canadian Institute for Health Information; 2006. Available from: https://secure.cihi.ca/free_products/app_average_payment_per_physician_report _2004_e.pdf 155. Boardman AE, Moore M a., Vining AR. The Social Discount Rate for Canada Based on Future Growth in Consumption. Canadian Public Policy .2010;36(3):325–43. 156. Kok L, Engelfriet P, Jacobs-van der Bruggen MA, Hoogenveen RT, Boshuizen HC, Verschuren MW.The cost-effectiveness of implementing a new guideline for cardiovascular risk management in primary care in the Netherlands. Eur J Cardiovasc Prev Rehabil. 2009 Jun;16(3):371-6. 157. de Oliveira C, Wijeysundera HC, Tobe SW, Lum-Kwong MM, Von Sychowski S, Wang X, Tu JV, Krahn MD. Economic analysis of Heart and Stroke Foundation of Ontario's Hypertension Management Initiative. Clinicoecon Outcomes Res. 2012;4:323-36. 158. Kolominsky-Rabas PL, Heuschmann PU, Marschall D, et al. Lifetime cost of ischemic stroke in Germany: Results and national projections from a population- based stroke registry: The Erlangen Stroke Project. Stroke 2006;37:1179-83. 159. Tarride J-E, Lim M, DesMeules M, Luo W, Burke N, O’Reilly D, et al. A review of the cost of cardiovascular disease. The Canadian journal of cardiology. 2009 Jun;25(6):e195–202. 160. Moore RA.The hidden costs of arthritis treatment and the cost of new therapy-- the burden of non-steroidal anti-inflammatory drug gastropathy. Rheumatology. 2002 Apr;41 Supp 1:7-15. 161. Canadian Institute of health Information (CIHI), Spending and Workforce; Health Care Cost Drivers!: The Facts. Ottawa, Canada; 2010. 162. Weinstein, M. C., O’Brien, B., Hornberger, J., Jackson, J., Johannesson, M., McCabe, C. Principles of good practice for decision analytic modeling in health- care evaluation: report of the ISPOR Task Force on Good Research Practices-- Modeling Studies. Value in health. 2003;6(1), 9-17. 163. Culyer AJ. The Dictionary of Health Economics, Second Edition. Second edi. Massachusetts, USA: Edward Elgar Publishing; 2010. 164. Statistics Canada, Population Projections for Canada, Provinces and Territories 2009 to 2036 - Catalogue no. 91-520-X. [cited 2012 January 10]. Available from: http://www.statcan.gc.ca/pub/91-520-x/91-520-x2010001-eng.pdf 237 165. Sargent RG. Verification and validation of simulation models. In: Proceedings of the 37th conference on winter simulation. WSC ’05: Winter Simulation Conference; 2005;130–143. 166. Claxton K, Schulpher M, McCabe C, Briggs A, Akehurst R, Buxton M, et al. Probabilistic sensitivity analysis for NICE technology assessment: not an optional extra. Health Economics. 2005;14(1):339–347. 167. Hay J, Jackson J. Panel 2: methodological issues in conducting pharmacoeconomic evaluations – modeling studies. Value Health 1999;2:78–81. 168. Koerkamp BG , Weinstein MC, Stijnen T, Heijenbrok-Kal M. H., Hunink M. Uncertainty and Patient Heterogeneity in Medical Decision Models. Med Decis Making 2010;30:194–205. 169. Lhachimi SK, Nusselder WJ, Boshuizen HC, Mackenbach JP., et al. Standard tool for quantification in health impact assessment: a review. Am J Prev Med. 2010;38(1):78-84. 170. Ackerman, E. Simulation of micropopulations in epidemiology: tutorial 1. Simulation: an introduction. A series of tutorials illustrated by coronary heart disease models. International Journal of Biomedical Computing. 1994 Jul;36(3):229-238. 171. Morgan MG., Henrion M. Uncertainty: a guide to dealing with uncertainty in quantitative risk and policy analysis. Cambridge: Cambridge University Press; 1992. 172. Feuer EJ, Etzioni R, Cronin KA, Mariotto A. The use of modeling to understand the impact of screening on U.S. mortality: examples from mammography and PSA testing. Stat Methods Med Res. 2004 Dec;13(6):421-442. 173. Brennan A., Stephen E. Chick, and Ruth Davies. A taxonomy of model structures for economic evaluation of health technologies. Health Economics. 2006; 15(12):1295–1310. 174. Griffin S, Claxton K, Hawkins N, Sculpher M. Probabilistic analysis and computationally expensive models: Necessary and required?. Value in Health. 2006 Jul-Aug; 9(4): 244-52. 175. Struijs JN, Van Genugten ML, Evers SM, Ament AJ, Baan CA, Van Den Bos GA. Modeling the future burden of stroke in The Netherlands: impact of aging, smoking, and hypertension. Stroke. 2005 Aug; 36(8):1648–1655. 176. Mandelblatt J, Schechter CB, Lawrence W et al. The SPECTRUM population model of the impact of screening and treatment on U.S. breast cancer trends 238 from 1975 to 2000: principles and practice of the model methods. J Natl Cancer Inst Monogr. 2006;(36):47-55. 177. Nosyk B , Sharif B , Sun H , Cooper C , Anis AH. The Cost-Effectiveness and Value of Information of Three Influenza Vaccination Dosing Strategies for Individuals with Human Immunodeficiency Virus. PLoS ONE[internet]. 2011.6(12): e27059. [cited January 2012].doi:10.1371/journal.pone.0027059. Available from: http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0027059 178. Shoujun Z , Xu Z, Lu Y. A mathematical model of hepatitis B virus transmission and its application for vaccination strategy in China. International Journal of Epidemiology. 2000 Aug;29(4):744-752. 179. Saltelli A, Ratto M, Andres T, Campolongo F, Cariboni J, Gatelli D, Saisana M, Tarantola S. Global Sensitivity Analysis: The Primer. Chichester, UK: John Wiley& Sons; 2008. 180. Baio G, Dawid P, Probabilistic sensitivity analysis in health economics, Stat Methods Med Res [internet], September 18 2011 [cited December 25, 2011]: [Epub ahead of print], Available from: http://smm.sagepub.com/content/early/2011/09/17/0962280211419832.long 181. Briggs AH, Ades AE, Price MJ, Probabilistic Sensitivity Analysis for Decision Trees with Multiple Branches: Use of the Dirichlet Distribution in a Bayesian Framework. Medical Decision Making. 2003 Jul-Aug; 23(4):341-350. 182. Doubliete P, Begg C, Weinstein M, Brun P, Mcneil B. Probabilistic sensitivity analysis using Monte Carlo simulation. Medical Decision Making. 1985;5(2), 157- 177. 183. Klevmarken, N. Anders. Statistical Inference in microsimulation models: Incorporating external information. Mathematics and Computers in Simulation. 2002 May 10;59(1-3):255-265. 184. Stevenson, Oakley J, Chilcott. Gaussian process modeling in conjunction with individual patient simulation modeling: a case study describing the calculation of cost-effectiveness ratios for the treatment of established osteoporosis. Medical Decision Making. 2004 Jan-Feb;24(1):89-100. 185. Hoogenveen, r. T., van baal, p. H., boshuizen, h. C. & feenstra, t. L. Dynamic effects of smoking cessation on disease incidence, mortality and quality of life: the role of time since cessation. Cost Eff ResourAlloc. 2008 Jan 11; 6: 1. 239 186. Burmaster DE, Anderson PD. Principles of Good Practice for the Use of Monte Carlo Techniques in Human Health and Ecological Risk Assessments. Risk Analysis 1994;14(4):1539-6924. 187. Cole BL, Fielding JE. Health impact assessment: A tool to help policy makers understand health beyond health care. Annual Review of Public Health. 2007;28:393-412. 188. Constance F. Citro, Eric A. Hanushek, Improving information for social policy decisions: the uses of microsimulation modeling. National Research Council, New York: National Academy Press; 1991. 189. Cronin KA, Legler JM, Etzioni RD . Assessing uncertainty in microsimulation modelling with application to cancer screening interventions. Statistics in Medicine. 1998;17(21):2509–2523. 190. Oakley J, O’Hagan A. Probabilistic sensitivity analysis of complex models: a Bayesian approach. Journal of royal statistical society. 2004;66(3):751-769. 191. McKay MD, Beckman RJ, Conover WJ. Comparison of three methods for selecting values of input variables in the analysis of output from a computer code. Econometrics. 1979; 21(1):239-245. 192. Halpern E, Weinstein M, Hunink M, Gazelle S. Representing both first- and second-order uncertainties by Monte Carlo simulation for groups of patients. Medical Decision Making. 2000 Jul-Sep;20(3):314-22. 193. Briggs A, Claxton K, Sculpher M. Decision modeling for health economic evaluation. Handbooks in health economic evaluation. London: Oxford university press; 2006. 194. Berry DA, Cronin KA, Plevritis SK, et al. Effect of Screening and Adjuvant Therapy on Mortality from Breast Cancer. N Engl J Med. 2005 Oct;353(17):1784- 1792. 195. Smolen HJ, Cohen DJ, Samsa GP, Toole JF, Klein RW, Furiak NM, Lorell BH, Development, Validation, and Application of a Microsimulation Model to Predict Stroke and Mortality in Medically Managed Asymptomatic Patients with Significant Carotid Artery Stenosis, Value Health. 2007 Nov-Dec;10(6):489-97. 196. Palmer AJ, Roze S, Valentine WJ et al. Validation of the CORE Diabetes Model against epidemiological and clinical studies. Curr Med Res Opin. 2004 Aug;20 Suppl 1:S27-40. 197. Loeve F, Boer R, van Ortmarssen GJ et al. The MISCAN-COLON simulation model for the evaluation of colorectal cancer screening. Comput Biomed Res. 1999 Feb;32(1):13-33. 240 198. Kahn R, Alperin P, Eddy D, Borch-Johnsen K, Buse J, Feigelman J, et al. Age at initiation and frequency of screening to detect type 2 diabetes: a cost- effectiveness analysis. Lancet. 2010 Apr 17;375(9723):1365-1374. 199. National Cancer Institute, Surveillance epidemiology and end results (SEER) [Internet]. US National Institute of Health; [cited 2011 Nov 12]. Available from: http://seer.cancer.gov/popdata/download.html 200. Salomon JA, Mathers CD, Murray CJL, Ferguson B. Methods for life expectancy and healthy life expectancy uncertainty analysis. Geneva, World Health Organization, GPE Discussion Paper [internet]. Aug 2001 [cited January 29, 2012]: No. 10 [18 p.]. Available from: http://www.who.int/healthinfo/paper10.pdf 201. Adler AI. UKPDS-modelling of cardiovascular risk assessment and lifetime simulation of outcomes. Diabetic Medicine: a Journal of the British Diabetic Association. 2008 Aug;25 Suppl 2:41–46. 202. Lhachimi SK, Nusselder WJ, van Baal P., et al. DYNAMO-HIA: A Dynamic Microsimulation Model-Model specification for a dynamic model for Health Impact Assessment [internet]. User Guide and Manual-3rd EUPHA conference, Amsterdam, the Netherlands - 2010 November 11 [cited 2011 October 12]. Available at: http://www.dynamo-hia.eu 203. Sassi F, Cecchini M, Lauer J, Chisholm D. Improving Lifestyles, Tackling Obesity: The Health and Economic Impact of Prevention Strategies [Internet]. OECD Health Working Papers, No. 48, OECD publishing; 2009 [cited 2011 October 1]. Available from: http://www.dynamo- hia.eu/object_binary/o2925_Simulation Brief.pdf 204. Blossfeld H, Rohwer G. Techniques of event history modeling: New approaches to causal analysis. Second edition. New Jersey, US: Lawrence Erlbaum Associates Ltd.; 2002. 205. Levy D, Mabry P, Wang Y, Gortmaker S, et al. Simulation models of obesity: a review of the literature and implications for research and policy. Obesity Reviews. 2011 May; 12(5):378-394. 206. Hiligsmann M, Ethgen O, Bruyère O, et al. Development and Validation of a Markov Microsimulation Model for the Economic Evaluation of Treatments in Osteoporosis. Value in health. 2009;12(5):678-696. 207. Statistics Canada. The LifePaths Microsimulation Model (Version 1.1): An Overview. Information of the model [Internet]; [cited 2011 Dec 10]. Available from: http://www.statcan.gc.ca/spsd/LifePathsOverview_E.pdf 241 208. Rowe G., Nguyen H. Longitudinal analysis of labour force survey data. Survey Methodology (Statistics Canada). 2004;30(1):105-114. 209. Willekens F. Continuous-time microsimulation in longitudinal analysis. In: A. Zaidi, A. Harding and P. Williamson eds. New frontiers in microsimulation modelling. Surrey, UK: Ashgate Publishing Ltd. 2009: 413-436. 210. Flanagan WM, Le Petit C, Berthelot JM, White KJ, Coombs BA, Jones-McLean E. Population health model-overview. In: Proceedings of the 2nd International Microsimulation Association Conference[Internet], Ottawa, Canada. 2009[cited 2011 June 10]. Available from: http://www.microsimulation.org/IMA/Ottawa_2009.htm 211. Will BP, Nobrega KM, Berthelot JM, Flanagan W, Wolfson MC, Logan DM, Evans WK. First do no harm: extending the debate on the provision of preventive tamoxifen, British Journal of Cancer.2001;85(9):1280–1288. 212. Bender R, Augustin T, Blettner M. Generating survival times to simulate Cox proportional hazard models. Statistics in Medicine. 2005 Jun 15; 24(11):1713- 1723. 213. Kuhl ME, Ivy JS, Lada EK, et al. Univariate input models for stochastic simulation. Journal of Simulation. 2010;4(1):81–97. 214. Kuhl ME, Lada EK, Steiger NM, Wagner MA, Wilson JR. Introduction to modeling and generating probabilistic input processes for simulation. In: WSC ’07: Proceedings of the 39th conference of winter simulation. Piscataway, NJ, USA: IEEE Press. 2007: 63–76. 215. Garthwaite PH, Kadane, JB, O'Hagan A. Statistical methods for eliciting probability distributions. Journal of the American Statistical Association. 2005;100(3): 680-701. 216. Oakley JE, O'Hagan A. Uncertainty in prior elicitations: a nonparametric approach. Biometrika. 2007; 94(2):427-441. 217. Pasta DJ, Taylor JL, Henning JM . Probabilistic sensitivity analysis incorporating bootstrap: An example comparing treatments for the eradication of Helicobacter pylori. Medical Decision Making. 1999 Jul-Sep;19(3):353-63. 218. Dowling R, Skabardonis A, Halkias J, McHale G, Zammit G. Guidelines for Calibration of Microsimulation Models: Framework and Applications. Transportation Research Record: Journal of the Transportation Research Board. 2004;1876 (1):1-9. 242 219. Sharif B, Kopec JA, Bansback N, Rahman M, Anis A. Projecting the Direct Cost Burden of Osteoarthritis in Canada Using a Population-Based Microsimulation Model From [2010-2031]. 2012 American Conference of Rheumatology Meeting (ACR/ARHP), Washington DC, US, November 9-14, 2012. 220. Barr A, Conaghan PG. Osteoarthritis: a holistic approach. Clinical medicine (London, England) . 2012 Apr;12(2):153–5. 221. Spielauer M. Dynamic microsimulation of health care demand, health care , Finance and the economic impact of health behaviours: survey and review. International Journal of Microsimulation. 2007;1(1):35-53. 222. Renkawitz T, Rieder T, Handel M, Koller M, Drescher J, Bonnlaender G, et al. Comparison of two accelerated clinical pathways--after total knee replacement how fast can we really go? Clinical rehabilitation . 2010 Mar;24(3):230–9. 223. Wylde V, Dieppe P, Hewlett S, Learmonth ID. Total knee replacement: is it really an effective procedure for all? The Knee. 2007 Dec;14(6):417–23. 224. Ramsey SD, Spencer a C, Topolski TD, Belza B, Patrick DL. Use of alternative therapies by older adults with osteoarthritis. Arthritis and rheumatism. 2001 Jun;45(3):222–7. 225. Stein, M. Large sample properties of simulations using latin hypercube sampling. Technometrics. 1987 ; 29(2):143-51 . 226. Railsback SF. Agent-based simulation platforms: review and development recommendations. Simulation. 2006 Sep; 82(9):609-23. 243 Appendix A . Details of Methods for Estimating the Direct Cost of OA A.1 Population Health Microsimulation of Osteoarthritis (POHEM-OA) Initially developed by Statists Canada (44), POHEM was a sub-model of the microsimulation model, LifePaths, which was developed using a MODGEN language built within C++. POHEM integrates data from different data sources to model the evolution of such health variables as smoking, body mass index (BMI), cholesterol, blood pressure, and mortality (1). Initializing the population by generating individuals directly from a national population data survey, i.e., Canadian Community health survey 2001 (CCHS), and, using the baseline health variables for each individual, the model employs continuous-time dynamic microsimulation techniques to age the initial population forward in time. At the same time, POHEM models the onset and evolution of a disease, its co-morbidities, and the influence of risk factors, and projects the population’s health determinants and disease-related outcomes, such as incidence, prevalence, change in risk factors, and quality of life across the entire population (POHEM (1). The POHEM model simulates representative populations and allows the comparison of competing health interventions (Figure A1). It has been successfully applied in a series of studies evaluating treatment and preventive strategies in breast and lung cancer, AMI, and osteoarthritis (referred to as POHEM-OA in this study). POHEM-OA simulates the life history of the entire Canadian population in terms of OA-related events. 244 Figure A1. POHEM causal flow diagram (from Kopec et al. (1)) Upstream health Intermediate Intermediate Diseases Treatment Death determinants risk factors diseases TIME (AGE and YEAR) Health-related Quality of life (e.g., HUI) Coronary CAPG, Heart PCI, Alcohol Depression Disease CATH… Ethnicity Stroke ABS Smoking Region Peripheral Vascular Amputation Age Cholesterol Disease Sex Hyper- Diabetes Diabetic Cataract Nutrition tension surgery... Income Retinopathy Death Blood pressure Kidney Dialysis Disease Education Obesity Osteoarthritis Surgery other Physical risk activity factors Surgery, other Radio/Chemo/ diseases 25 Cancers Hormonal therapy Initial state assigned from CCHS (+CHHS) competing risk of death from other causes Demographics-and-Longevity-Components-in-POHEM:OA- POHEM-OA uses the Canadian Community Health Survey (CCHS-2001) to sample individuals for its initial population where each sample represents a group of individuals corresponding to its given survey weighs, to generate the adult population of Canada (~ 20 Million initial population). Main risk factor for incidence of OA in POHEM-OA is age and BMI. Individual’s life history, in terms of OA-related events, is being simulated through “inter-related” stochastic processes; at each calendar year, the events are assigned to individuals, based on their characteristics (e.g. OA risk factors, sex, etc.). For example, BMI is modeled as auto-regression model based on SES, age, sex, region, etc. At the same time, based on the current value of BMI, a proportional hazard model was used to estimate the time of the OA diagnosis. 245 Due to the continuous time nature of the POHEM, the time to next event in queue is assigned to all possible events in the near future, based on the competing “hazards” approach. For example, the hazards of both OA diagnosis and death of an individual are randomly sampled as continues time from a distribution of hazard functions (which is a function of individual’s characteristics); and whichever comes first, that event is assigned to that individual (i.e., discrete event simulation). After events are assigned to each individual, the cost module will be updated. Table A1. List of Parameters inside POHEM-OA (from Kopec et al. (1)) Parameter Source and method of derivation Age distribution in 2001 Observed in CCHS (2001) Sex distribution in 2001 Observed in CCHS (2001) Province of residence Observed in CCHS (2001) distribution in 2001 Body Mass Index Observed in CCHS (2001); BMI = weight / height2 distribution in 2001 Health Utilities Index Observed in CCHS (2001) distribution in 2001 (Range: -0.36 to 1) Prevalence of OA in 2001 by Obtained as the final stable prevalence from a simulation sex and 5-year age groups of the Canadian population over a 50-year horizon, under constant age-specific incidence rates. Prevalence of prior hip or From PDBC. Defined as hospitalization with a procedure knee replacement in 2001 code 935 or 934.1 and ICD-9 code 715 between 1986 by sex and 5-year age and 2001 (excluding certain codes, for example, for groups revisions or fractures). Change in BMI over time Obtained from a linear regression model including age, sex, province, education, and prior BMI. The model is based on an analysis of longitudinal data from NPHS (5 cycles) Mortality rates by age and Based on mortality data and using Statistics Canada sex over time projections of mortality. Migration Migration data obtained from Statistics Canada projections Incidence of OA in 2001 by From PDBC. OA is defined as at least 2 visits to a health sex and 5-year age groups professional within 2 years or 1 hospitalization with the ICD-9 code 715. Incident cases in 2001 are identified after excluding prevalent cases prior to 2001 using a 10- year run-in period. Effect of BMI on incident OA BMI was categorized into 4 standard categories (<18.5, by sex 18.5-25, 25-30, 30+). The effect (hazard ratio) for each level of BMI is obtained from a survival regression model using longitudinal NPHS data (two cycles: 2000 and 2002), separately for men and women, and adjusted for 246 Parameter Source and method of derivation age. Reference (“baseline”) Obtained numerically using an iterative algorithm hazard rates of OA (calibration) to match the marginal distribution of incidence in PDBC. Rates of visits to orthopedic From PDBC. The frequency of visits is described by a surgeon for OA by sex and Kaplan-Maier survival curve using data from the PDBC 5-year age groups following (1991 - 2003). A visit is defined based on a combination OA diagnosis of provider and diagnostic codes. Weibull parameters to A piece-wise Weibull model was developed using describe the distribution of iterative numerical methods to produce simulated survival survival times curves that match the observed KM curves (calibration). Rates of first joint Data are obtained from PDBC. First hip or knee replacement surgery (TJRs) replacement was defined as hospitalization with a by sex and 5-year age procedure code 935 or 934.1 and ICD-9 code 715 groups following a visit to (excluding certain codes, for example, for revisions or OS fractures) in 2003 in a person without previous TJRs. Frequency of TJR over time following a visit to OS is described by a Kaplan-Meier survival curve. JR revision (JRR) rates Hip or knee revision is defined as hospitalization with a following JR surgery procedure code for hip/knee revision (excluding injuries and cancer). Frequency of JRR over time following a primary JR is described by a Kaplan-Meier survival (1). Follow-up Health Utilities HUI was obtained from a tobit regression model including Index 3 (pre-treatment) age, sex, BMI, and OA status (including interactions as required). The model is estimated from CCHS (2001, cross-sectional data). OA duration Calculated by POHEM as current time minus OA onset time Use of prescription drugs From PharmaNet data (2004) linked to PDBC. For each drug category, proportion of person-time of use is obtained for each age/sex and OA duration category (the drug categories are acetaminophen, NSAIDs, COX-2, opioids, gastro-protective agents and combinations of drugs). Effect of each drug category Obtained from a comprehensive review of the literature on HUI (RCTs). Defined as a weighted (by sample size) average relative effect on standardized pain score (delta pain / baseline pain) and converted to HUI3 using a procedure based on the known contribution of pain to HUI3 score. Incidence of side effects of Based on risks reported in the literature for drugs gastrointestinal complications, such as bleeding or ulcer, myocardial infarction, stroke, dyspepsia, and death, according to age/sex category. Effects of side effects of Obtained based on published effects of stroke, heart drugs on HUI disease, and ulcer on HUI3 in cross-sectional NPHS data. The effects are based on a regression model with age, sex and co-morbidity. Effect of TJR on HUI Obtained from a regression model (with change in HUI as 247 Parameter Source and method of derivation outcome), using longitudinal data from a cohort of patients followed for up to 3 years at the Vancouver General Hospital. - Step:wise-Algorithm-of-POHEM:OA-Event-Assignment:-an-Example-of-the-Steps-to-Determine-when-OA- Diagnosis-Occurs-- ! First, POHEM selects a record from the Canadian Community Health Survey database in simulation start year 2001, for example: ! Female, aged 49, overweight, no history of OA ! At each birthday in the person’s simulated life, the probability (P) of developing OA is evaluated as follows: 0 Look up the baseline rate (B) for females aged 49 from the input parameter OA_BASELINE_INCIDENCE 1) B = 0.00732 1 Look up the relative risk of OA based on person’s weight category from input parameter OA_BMI_RR 1. R = 1.76 (for overweight female) 2 Probability (P) is combined baseline and relative risk: 2. P = B * R = 0.00732 * 1.76 = 0.012883 Note: many other events, possibly competing, are also evaluated on the birthday, such as change in risk factor status, risk of developing other diseases, risk of dying. 3) We generate a uniform random number (u) between 0 and 1 to determine if the event will occur. We can implement this as a discrete time event or as a continuous time event (preferred). Generate a random u = 0.011 a) In a discrete time implementation, we would compare the probability directly with a uniform random number (u) between 0 and 1 according to the following rule: If ( u < P ) then the person developed OA at the current age u=0.011 is less then P=0.012883 248 therefore, person would develop OA at age 49.0 b) In a continuous time implementation (preferred method), we convert the probability to a hazard and generate a waiting time A value of u P, the waiting time will exceed 1 year and will be discarded at the next birthday when it is re-evaluated. • Convert probability to an annualized hazard (h): o h = -ln(1-P) = -ln(1-0. 012883) = 0.012967 • Generate a waiting time by comparing the hazard to a uniform random number (u) between 0 and 1 o t = -ln(1-u) / h = -ln(1-0.011) = 0.853 years o therefore, person would develop OA at age 49.853 • Competing events like mortality could censor this event 4) Steps 2 and 3 are repeated at every subsequent birthday until the person develops OA (which in this example occurred in the very first year) or dies. When OA is diagnosed, waiting times to subsequent events in the disease process (such as time to surgery) are evaluated (according to models built from empirical data). 5) Steps 1 to 4 are repeated for every individual record in our start-up database CCHS. Simulation results can be tabulated in a variety of ways: number of incidence cases by calendar year, age, sex, income level, educational attainment, and so on. The-Drug-Module-Inside-POHEM:OA- Probabilities for each type of drugs utilization implemented in POHEM-OA (1) are discussed in Sayre et al. (139). Hazards of side effects for each drug calculated from systematic reviews (Sayre et al. (139)) are shown in Table A2. 249 Table A2: Differences in rates of main side effects between persons on drug vs. placebo per 1000 person-years on drug (i.e., extra cases among those taking medication per 1000 p-y) Drug class Serious CVD Stroke Dyspepsia Other/ GI Generic NSAIDS 7.4 3.7 1.4 57.3 1000 Coxibs 1.1 4.5 2.4 8.8 1000 Acetaminophen 0.0 0.0 0.0 0.00 1000 Opioids 0.0 0.0 0.0 117.6 1000 A.2 Random Sample Weights for the Study Population According to the prior COAST study, a stratified random sample of 100,000 individuals was selected for PharmaNet data application. Sample was stratified according to the OA states and age. Details of the weights for the sampling scheme are shown in Table A3 (discussion in Section 3.2). In the random sample, if a few cases were found in the younger and/ or older age groups took all of them and split the rest in equal proportion to the other age groups. Therefore, each individual in the sample represents at least one individual in the population. Appropriate weights were assigned to each of the selected individuals; hence, the sum of the weights equals the BC population. Table A3. Age stratification weights for random sample Age Percent 0-9 5 10-19 5 20-29 5 30-39 5 40-49 10 50-59 15 60-69 15 70-79 15 80-89 15 90+ 10 Total 100 250 A.3 Calculating the Input Parameters for Hospitalization, Drugs and Physician Visit Costs Hip/Knee-Total-Joint-Replacement-(TJR)-Surgeries-- The goal of this step was to report the overall average cost of TJR surgeries for 3 age categories (18-60,60-80 and 80+) for primary and revision, separately. To be able to calculate these values, we performed the weighted average using the volume calculated in Step 1 below (as weights) and then calculated the average cost of each age category for primary and revision surgeries. As discussed in Step 2 below (as the costs), we used three types of primary and revision surgeries as reported by CIHI database (144); primary TJRs used were: bilateral/unilateral, hip /knee Revision TJRs were: bilateral/unilateral, with/without infection (hip and knee). Here, we discuss the steps in details: Step 1. Calculating the volume of TJRs: Total hip and total knee surgery cases were identified using the procedure ICD-9 codes 934.1 and 935 respectively. We excluded all surgeries (principal, primary or emergency), with codes at the time of surgery for all injury or poisoning (ICD-9 800-995, 997 and 999), all neoplasm other than benign (ICD-9 140 - 208 and 235 - 239) and all non-medical external causes of injury (ICD-9 E800-869 E880-E928, E950-E999). Revision procedures were identified having the above two procedure codes with an ICD-9 codes 730.x (infections), 9964, 9965, 9966, 9967 or 998 (complications of certain surgical procedures) in the same hospital discharge record. Hip and knee procedures identified to be related to OA were used separately for primary and revision. Step 2. Calculating the components of TJR cost: For each type of surgery and each age category, we calculated the average cost of surgery using their major components: length of stay at hospital after surgery (LOS), procedure cost and implant cost. Here, we discuss each of the hip/knee TJRs cost components. Length of stay cost calculations: The CJRR observed ELOS was used from 2003-2009 (124). We assumed the same rate as in the 2007-2009 to be used for 2009-2012. Procedure cost calculations: Procedure cost includes medical imaging, surgery cost and staff during surgery. Using functional area percentages from CIHI database (144), we calculated the procedure cost of each surgery type and each age category. We first calculate the total cost of surgery, then by multiplying it by the function area percentage 251 of the procedure sections, we calculate the procure cost of each type of hip/knee surgery. Having the ELOS of each type of surgery (i.e., CMG’s of hip/knee TJR surgeries) and each age category for primary/revision, we then used the per diem reported from CIHI data in 2010 (144). Per diem rates represents per day relative RIW (resource intensity weights) of each type of surgery, compared to the cost per weighted case (CPWC), which is equal to $5484 for the national average (that is average of all acute inpatient typical cases submitted to the 2010 DAD) (144). Multiplying the ELOS by per diem rate and then by the (national average CPWC of 2010), would give us the total cost of hip/knee surgery for each type of and each age category (for the national average over all types of hospitals). However, to take out the procedure cost (fixed cost) out of the total cost, we multiply the total cost by the functional area percentages from CIHI database for the three mentioned types of TJRs (for each three CMG+ group). 252 Table A4. List of Primary and revision surgeries from CMG+ used for cost calculations Case%Mix%Group%(CMG)%assignment%/% Per7Diem% %Surgery%Procedure%functional% OA7related%% %Age%groups% rate**%% area%^%(%)%% hip/knee%% surgery%%% volume%^^% 315.%Bilateral%Hip/Knee%Replacement*% ! 57%! ! 18$59'Years'(Adult)' 0.47! ! 352! 60$79'Years'(Adult)' 0.43834! ! 594! 80+'Years'(Adult)' 0.33874! ! 42! 320.%Unilateral%Hip%Replacement% !! 58%! ! 18$59'Years'(Adult)' 0.4306! ! 6050! 60$79'Years'(Adult)' 0.35123! ! 12623! 80+'Years'(Adult)' 0.30307! ! 2971! 321.%Unilateral%Knee%Replacement% !! 52%! ! 18$59'Years'(Adult)' 0.32878! ! 8293! 60$79'Years'(Adult)' 0.31225! ! 24355! 80+'Years'(Adult)' 0.28606! ! 4202! 316.%Revised%Hip%Replac%w%Inf^^^% !! 42%! ! 18$59'Years'(Adult)' 0.47293! ! 175! 60$79'Years'(Adult)' 0.43178! ! 361! 80+'Years'(Adult)' 0.33609! ! 124! 317.%Revised%Hip%Replac%wo%Inf% !! 49%! ! 18$59'Years'(Adult)' 0.42136! ! 587! 60$79'Years'(Adult)' 0.37883! ! 1367! 80+'Years'(Adult)' 0.34217! ! 687! 318.%Revised%Knee%Replac%w%Inf% !! 42%! ! 18$59'Years'(Adult)' 0.38947! ! 223! 60$79'Years'(Adult)' 0.35814! ! 629! 80+'Years'(Adult)' 0.35076! ! 147! 319.%Revised%Knee%Repl%wo%Inf% !! 57%! ! 18$59'Years'(Adult)' 0.44947! ! 458! 60$79'Years'(Adult)' 0.40606! ! 1198! 80+'Years'(Adult)' 0.34834! ! 277! * List of types of hip/knee TJRs from CMG+ (Case Mix Group assignment), The CMG assignment is a grouping of patient stays with similar clinical and resource utilization for comparison of hospital resource use. **RIW: resource intensity weights are used to monitor utilization of acute health care services. The RIW costing algorithm measures the relative cost of acute care resources by patient types.^ Functional area percentages that related to procedure cost of the hip/knee TJR surgery and is the sum of functional percentages of following areas: medical imaging, operating and recovery room nursing services , clinical lab, professional services during surgery and materials (implant) that is calculated from CIHI database CMG group related to TJR surgeries , CMG+: 315-321 (i.e., TJR surgery types). ^^ This column is the data from the CIHI hospitalization cost database (144).^^^ w-inf: With infection; wo-inf: without infection ! 253 Table A5. Final cost results for knee/hip TJR from 2003-2010 (in current year $) * The costs do not include inflation from 2003 to 2010; to include the inflation for hip/knee surgery Primary--Surgery- 2003- 2004- 2005- 2006- 2007- 2008- 2009- 2010- ------<50- $5,445& $5,759& $6,090& $6,441& $6,812& $7,198& $7,615& $7,795& 50:60------$5,221& $5,522& $5,840& $6,177& $6,532& $6,903& $7,302& $7,467& 60:70- $5,420& $5,732& $6,062& $6,411& $6,780& $7,165& $7,580& $7,755& Revision-Surgery- $9,655& $9,847& $10,051& $10,26 $10,49 $10,73 $10,98 $11,11 ------<50- ---50:60------$9,653& $9,846& $10,049& $10,266& $10,493& $10,720& $10,986& $11,100& 60:70- $11,04 $11,26 $11,497& $11,744& $12,002& $12,279& $12,565& $12,744& as discussed in Section 3.3, each6& year cost5& needs to be added2 by& extra 4.6%2& of that2 &year’s cost3& to 5& represent actual costs with inflation. - Hospital-Procedures-Other-Than-Hip/Knee-TJR-(Inpatient-Procedures)-- For hospital procedures (inpatient procedures) other than hip/knee surgeries, we first matched the non-hip/knee TJRs list from ICD-9 to CMG+ list of inpatient procedures in CIHI hospital cost database (144). The list is shown in Table A3. We used the hospital length of stay from the ICD-9 list of our OA sample and calculate the hospital stay cost, while calculated the procedure cost from the CIHI hospital cost database in 2010 (144) For this purpose, we first calculated the average length of stay for the non-hip/knee TJR procedures from CIHI data for 8 age/sex categories. Using the OA cohort length of stay and dividing them by average length of stay from CIHI data, we calculated the average number of procedures performed by cost factors (i.e., 40 categories for age, sex, OA progression state and years in each state). Having the number of procedures, we need the average procedure cost of all non-hip/knee TJRs to find the total procedure cost by cost factors. To calculate the average cost for procedure proportions, we used 2010 national CPWC (cost per weighted case) that is the average of all acute inpatient typical cases submitted to the 2010 DAD of the average Canadian hospital (equal to $5484 in 2010 $CAD). We multiply the national CPWC by RIW for procedure proportions of each non-hip and knee TJRs (data from CIHI hospital cost database for 2009-2010 year [144]). Finally, we added the procedure cost to hospital stay cost for each cost factor and divided the total cost by 254 number of person years of our OA cohort by each cost factor to calculate the “non-hip/knee TJRs” per person year cost in 2001 $CAD as shown in Table A7. Table A6. List of hospital procedures for OA patients (other than hip/knee total joint replacement (TJR) surgeries) List of OA in-patient procedures PDBC (from OA List of inpatient procedures from study population *) CIHI CMG+ (2010 CIHI database *) 332-Other Repair Bone of Leg except OTHER REPAIR OF KNEE Ankle/Foot 325- Closed Knee Intervention except OTHER EXCISION OF JOINT, KNEE Fixation without Infection 323-Open Knee Intervention except EXCISION OF SEMILUNAR CARTILAGE OF KNEE Fixation without Infection 324-Closed Knee Intervention except ARTHROSCOPY, KNEE Fixation with Infection ARTHROPLASTY OF SHOULDER WITH SYNTHETIC PROSTHESIS 326-Shoulder Replacement 330- Fixation of Lower Limb Except OTHER REPAIR OF HIP Ankle/Foot EXCISION OF BONE FOR GRAFT, OTHER 343-Other Musculoskeletal Intervention SPECIFIED SITE except Soft Tissue 343-Other Musculoskeletal Intervention TARSOMETATARSAL FUSION except Soft Tissue 335-Other Foot Intervention, except Soft OTHER EXCISION OF JOINT, FOOT AND TOE Tissue 346-Other Musculoskeletal Soft Tissue OTHER TENOTOMY Intervention OSTEOTOMY 338-Osteotomy 334-Major Foot Intervention except Soft ANKLE FUSION Tissue without Infection 335-Other Foot Intervention, except Soft OTHER FUSION OF TOE Tissue 332-Other Repair Bone of Leg except OPEN REDUCTION OF FRACTURE WITH Ankle/Foot INTERNAL FIXATION, FEMUR DIAGNOSTIC ULTRASOUND OF LIMBS 362-Osteoarthritis 335-Other Foot Intervention, except Soft OTHER EXCISION OF JOINT, ANKLE Tissue ARTHROPLASTY OF HAND AND FINGER WITH SYNTHETIC PROSTHESIS 327-Other Joint Replacement 323-Open Knee Intervention except EXCISION OF BURSA Fixation without Infection 324-Closed Knee Intervention except SYNOVECTOMY, KNEE Fixation with Infection ARTHROCENTESIS 362-Osteoarthritis 255 List of OA in-patient procedures PDBC (from OA List of inpatient procedures from study population *) CIHI CMG+ (2010 CIHI database *) TENOTOMY OF HAND 337-Hand Intervention TOTAL ANKLE REPLACEMENT 327-Other Joint Replacement OTHER REPAIR OF ANKLE 335-Oth Foot Intv exc Soft Tissue 466 -Intervention Related to Dialysis- HEMODIALYSIS Planned Admission 330- Fixation of Lower Limb Except OTHER PARTIAL OSTECTOMY, FEMUR Ankle/Foot OTHER PARTIAL OSTECTOMY, TIBIA AND 330-Fixation of Lower Limb Except FIBULA Ankle/Foot 324-Closed Knee Intervention except OTHER ARTHROTOMY, KNEE Fixation with Infection TOTAL OSTECTOMY, CARPALS AND 330- Fixation of Lower Limb Except METACARPALS Ankle/Foot ARTHROSCOPY, ANKLE 335-Oth Foot Intv exc Soft Tissue 330- Fixation of Lower Limb Except OTHER PARTIAL OSTECTOMY, PATELLA Ankle/Foot OTHER PARTIAL OSTECTOMY, TARSALS AND 330-Fixation of Lower Limb Except METATARSALS Ankle/Foot ARTHROPLASTY OF ELBOW WITH SYNTHETIC PROSTHESIS 327-Other Joint Replacement OTHER PARTIAL OSTECTOMY, RADIUS AND 330-Fixation of Lower Limb Except ULNA Ankle/Foot SYNOVECTOMY, WRIST 337-Hand Intervention ARTHROSCOPY, ELBOW 340-Elbow Intervention *Study population discussed in Chapter 3.2 that was all OA patients (and some non-OA controls) generated from PDBC (BC linked population data (84)). 256 Table A7. Per person-year costs for inpatient procedures other than hip/knee TJR (all in 2010 $CAD)* oa0.0- oa2.0- os0.0- os2.0- prim0.0- prim2.0- revi0.0- revi2.0- oa5.0+y os5.0+y prim5.0+y revi5.0+y 1.9y** 4.9y 1.9y 4.9y 1.9y 4.9y 1.9y 4.9y Female 00-49*** $0.00 $2.81 $8.26 $582.45 $25.42 $53.70 $98.36 $24.16 $0.00 $0.00 $0.00 $31.39 50-59 $0.00 $0.00 $4.62 $649.71 $33.64 $42.26 $109.08 $7.96 $7.76 $50.57 $0.00 $44.40 60-69 $9.20 $0.00 $0.00 $482.72 $36.46 $50.97 $79.23 $26.65 $12.07 $100.67 $31.75 $0.00 70-79 $0.00 $0.00 $0.00 $372.72 $33.17 $31.23 $54.30 $16.16 $11.80 $34.11 $61.93 $5.45 80-89 $0.00 $0.00 $0.00 $197.98 $20.84 $24.80 $63.85 $3.44 $9.60 $23.37 $0.00 $18.25 90+ $10.70 $0.00 $0.00 $148.42 $0.00 $17.45 $0.00 $0.00 $3.92 $0.00 $0.00 $0.00 Male 00-49 $5.62 $0.00 $0.00 $676.42 $28.68 $41.17 $120.49 $0.00 $0.00 $69.24 $0.00 $0.00 50-59 $0.00 $6.75 $0.00 $659.11 $55.32 $66.72 $149.28 $13.76 $33.00 $48.98 $0.00 $0.00 60-69 $0.00 $0.00 $0.00 $612.72 $23.95 $64.20 $73.52 $20.77 $10.14 $94.24 $0.00 $0.00 70-79 $15.73 $0.00 $0.00 $287.76 $16.86 $64.23 $72.50 $20.17 $20.43 $72.45 $0.00 $14.84 80-89 $0.00 $0.00 $0.00 $115.04 $13.13 $23.44 $91.46 $0.00 $5.04 $0.00 $0.00 $7.50 90+ $0.00 $0.00 $0.00 $140.97 $0.00 $0.00 $0.00 $0.00 $0.00 $0.00 $0.00 $0.00 *Costs were translated from 2003 using CPI of 1.351 from (134);* *first row represent OA states and time in each state as defined in ***Section 3.2 ( oa: State1, os: State 2, prim: State 3, revi: State 4);***age groups ! ! 257 Drug!Cost!Module! Table A8. Per patient-year cost of prescription drug cost as implemented in POHEM-OA (in 2010 $CAD)* oa0.0- prim0.0- prim2.0- revi0.0- revi2.0- oa2.0-4.9y oa5.0+y os0.0-1.9y os2.0-4.9y os5.0+y prim5.0+y revi5.0+y 1.9y** 1.9y 4.9y 1.9y 4.9y Female 00-49*** $115.78 $47.67 $36.32 $153.24 $80.59 $63.57 $287.18 $115.78 $136.21 $331.45 $397.29 $230.43 50-59 $87.40 $52.21 $45.40 $178.21 $86.27 $78.32 $308.75 $126.00 $205.45 $323.51 $288.32 $170.27 60-69 $78.32 $53.35 $53.35 $192.97 $89.67 $87.40 $227.02 $130.54 $132.81 $240.64 $210.00 $275.83 70-79 $86.27 $61.30 $28.38 $157.78 $91.94 $96.48 $198.64 $118.05 $98.75 $175.94 $141.89 $173.67 80-89 $65.84 $34.05 $38.59 $164.59 $89.67 $63.57 $116.92 $81.73 $88.54 $165.73 $54.49 $120.32 90+ $74.92 $39.73 $35.19 $145.29 $81.73 $57.89 $85.13 $85.13 $60.16 $99.89 $89.67 $60.16 Male 00-49 $160.05 $47.67 $62.43 $164.59 $61.30 $59.03 $497.18 $157.78 $261.07 $623.17 $359.83 $245.18 50-59 $99.89 $55.62 $56.76 $160.05 $89.67 $72.65 $248.59 $137.35 $107.84 $271.29 $380.26 $488.10 60-69 $70.38 $36.32 $45.40 $146.43 $62.43 $53.35 $212.27 $93.08 $102.16 $347.34 $240.64 $144.16 70-79 $69.24 $36.32 $45.40 $150.97 $62.43 $55.62 $129.40 $63.57 $86.27 $166.86 $143.02 $103.29 80-89 $49.94 $24.97 $22.70 $112.38 $64.70 $37.46 $97.62 $38.59 $49.94 $103.29 $70.38 $59.03 90+ $54.49 $27.24 $31.78 $118.05 $32.92 $44.27 $7.95 $44.27 $29.51 $0.00 $6.81 $31.78 *Costs were translated from 2003 using CPI of 1.351 from (134);* *first row represent OA states and time in each state as defined in Section 3.2 ( oa: State1, os: State 2, prim: State 3, revi: State 4); ***age groups ! 258 Table A9. Per patient-year cost of over-the-counter drug cost as implemented in POHEM-OA (in 2010 $CAD)* oa0.0- prim0.0- prim2.0- revi0.0- revi2.0- 1.9y** oa2.0-4.9y oa5.0+y os0.0-1.9y os2.0-4.9y os5.0+y 1.9y 4.9y prim5.0+y 1.9y 4.9y revi5.0+y Female 00-49*** $127.13 $46.54 $49.94 $147.56 $82.86 $61.30 $205.45 $139.62 $87.40 $449.50 $91.94 $246.32 50-59 $89.67 $54.49 $56.76 $140.75 $85.13 $72.65 $244.05 $135.08 $116.92 $185.02 $82.86 $135.08 60-69 $72.65 $44.27 $46.54 $215.67 $80.59 $73.78 $204.32 $124.86 $133.94 $180.48 $172.54 $171.40 70-79 $76.05 $46.54 $31.78 $143.02 $79.46 $66.97 $157.78 $106.70 $95.35 $179.35 $129.40 $144.16 80-89 $53.35 $30.65 $36.32 $119.19 $63.57 $48.81 $111.24 $69.24 $71.51 $122.59 $38.59 $79.46 90+ $49.94 $30.65 $42.00 $175.94 $65.84 $46.54 $96.48 $101.02 $61.30 $115.78 $74.92 $40.86 Male $78.32 $43.13 $44.27 $156.64 $76.05 $61.30 $170.27 $112.38 $94.21 $205.45 $98.75 $136.21 00-49 50-59 $131.67 $43.13 $81.73 $107.84 $39.73 $44.27 $215.67 $127.13 $126.00 $511.93 $262.21 $73.78 60-69 $90.81 $38.59 $40.86 $96.48 $48.81 $54.49 $186.16 $68.11 $81.73 $154.37 $124.86 $183.89 70-79 $99.89 $42.00 $51.08 $94.21 $59.03 $46.54 $185.02 $82.86 $81.73 $329.18 $119.19 $88.54 80-89 $64.70 $28.38 $40.86 $148.70 $64.70 $73.78 $149.83 $79.46 $77.19 $173.67 $149.83 $107.84 90+ $46.54 $23.84 $18.16 $95.35 $45.40 $37.46 $107.84 $35.19 $52.21 $129.40 $57.89 $40.86 *Costs were translated from 2003 using CPI of 1.351 from (134);* *first row represent OA states and time in each state as defined in Section 3.2 ( oa: State1, os: State 2, prim: State 3, revi: State 4); ***age groups ! 259 Table A10. List of drugs used in calculation for prescription-based drugs due to OA Category Subcategory Acetaminophen 'ACETAMINOPHEN' 'ACETAMINOPHEN/CAFFEINE' 'ACETAMINOPHEN/PYR MA/PAMABROM' 'CODEINE PHOS/ACETAMINOPHEN' 'CODEINE/ACETAMINOPHEN/CAFFEINE' 'CODEINE/ACETAMINOPHEN/DOXYLAM' 'OXYCODONE HCL/ACETAMINOPHEN ‘ACETYLSALICYLIC ACID', 'DICLOFENAC POTASSIUM' 'DICLOFENAC SODIUM', 'DIFLUNISAL' 'ETODOLAC', 'FENOPROFEN CALCIUM', 'FLURBIPROFEN' 'IBUPROFEN', 'INDOMETHACIN', 'KETOPROFEN' NSIADs 'MEFENAMIC ACID', 'MELOXICAM', 'NABUMETONE' 'NAPROXEN', 'NAPROXEN SODIUM', 'OXAPROZIN' 'PIROXICAM', 'PIROXICAM BETA CYCLODEXTRIN CP', 'SULINDAC' 'TENOXICAM', 'TIAPROFENIC ACID', 'TOLMETIN SODIUM' 'OXYCODONE/ACETYLSALICYLIC ACID' 'DICLOFENAC SODIUM/MISOPROSTOL' 'ACETYLSALICYLIC ACID/CAFFEINE COXIBs ‘CELECOXIB' 'VALDECOXIB' 'ROFECOXIB ‘CODEINE ANHYD', 'CODEINE PHOS', 'CODEINE SULF' 'MORPHINE HCL', 'MORPHINE SULFATE','MORPHINE SULFATE/PF', 'OXYCODONE HCL', 'PROPOXYPHENE HCL', 'PROPOXYPHENE NAPSYL' Opioids 'CODEINE PHOS/ASA/CAFFEINE', 'CODEINE PHOS/ASA/PHENOBARBITAL','CODEINE/ASA/CAFFEINE/BUTALB','C ODEINE/ASA/CAFFEINE/MEPROBAM' 'OXYCODONE HCL/ASA' ‘PROPOXYPHENE HCL/ASA/CAFFEINE' 'PROPOXYPHENE/ASA/CAFFEINE ! 260 Physician)Visit)Cost)Module*) Table A11. Per patient-year cost of physician and other outpatient costs (lab, MRI) as implemented in POHEM-OA (in 2010 $CAD)** oa0.0- prim0.0- prim2.0- revi0.0- revi2.0- oa2.0-4.9y oa5.0+y os0.0-1.9y os2.0-4.9y os5.0+y prim5.0+y revi5.0+y 1.9y*** 1.9y 4.9y 1.9y 4.9y Female 00-49**** $447.48 $32.87 $40.66 $311.95 $9.45 $25.22 $194.42 $5.62 $0.00 $164.15 $23.96 $15.43 50-59 $686.61 $94.04 $119.07 $506.32 $80.27 $56.95 $235.34 $0.00 $4.97 $206.77 $23.88 $37.36 60-69 $678.62 $132.85 $152.90 $592.56 $88.33 $86.03 $218.09 $18.34 $0.00 $145.13 $4.99 $2.51 70-79 $619.29 $128.07 $162.78 $541.85 $49.95 $74.50 $485.33 $35.44 $21.85 $247.08 $0.00 $0.00 80-89 $334.31 $13.28 $60.53 $178.73 $0.00 $0.00 $352.85 $0.00 $0.00 $152.34 $0.00 $0.00 90+ $19.71 $0.00 $0.00 $49.87 $0.00 $0.00 $144.79 $0.00 $0.00 $106.03 $0.00 $0.00 Male 00-49 $522.43 $43.93 $35.73 $303.50 $3.30 $13.42 $169.82 $7.75 $0.00 $229.75 $0.00 $9.94 50-59 $623.06 $61.25 $160.14 $433.04 $42.46 $53.19 $177.55 $0.60 $0.00 $164.58 $19.80 $0.00 60-69 $592.75 $78.67 $163.30 $510.56 $30.06 $50.41 $216.89 $0.00 $0.00 $120.57 $0.47 $0.00 70-79 $640.36 $90.81 $98.21 $501.50 $40.42 $42.23 $450.63 $12.45 $18.83 $203.81 $0.00 $0.00 80-89 $237.66 $0.00 $18.80 $197.05 $0.00 $0.00 $309.41 $0.00 $0.00 $112.81 $0.00 $0.00 90+ $30.82 $0.00 $0.00 $66.99 $0.00 $0.00 $77.50 $0.00 $0.00 $247.96 $0.00 $0.00 * Physician types included in this cost category was all visits to GPs, rheumatologists and all other professional physicians ( except alternative out-of-pocket cost and surgeons); we calculated physician cost for non-OA cost and deducted them from OA-related physician cost. Finally, we divided the overall cost by the time spent in each state ( for age, sex and OA status). **Costs were translated from 2003 using CPI of 1.351 from (134);***first row represent OA states and time in each state as defined in Section 3.2 ( oa: State1, os: State 2, prim: State 3, revi: State 4); ****age groups 261 A.4 Other Cost Categories: Rehabilitation, Formal Caregiver, Alternative Care and Side Effect of Drugs Cost Rehabilitation+and+Home+Care+Cost++ In POHEM-OA, we modeled the year after the surgery based on standard rehabilitation procedures (135). During the first year after surgery, possible target destination is either home or rehabilitation facility (hospital) (134, 135). The cost calculation is discussed in two sections here. First, we discuss how the probability of moving to each discharge destination is calculated (from 135). Then, different cost components (out-of-pocket and healthcare system) involved at each rehab destination are calculated. Finally, we discuss how the parameters were implemented and used in the rehab cost category of POHEM-OA. Probability*of*Discharge*to*Each*Destination* To model rehab procedures, we assumed four possible targets for OA patients after TJR that matches the current clinical pathways at Canadian hospitals (134). These pathways are as follows: 1. Rehab facility stay (average of 10-20 days) and then move to the home care (paid by the healthcare system), 2. Rehab facility stay and then move to home without the home care (i.e., self-care), 3. Discharge to home with paid home care 4. Discharged to home without home care (i.e., self-care after the acute care) (135). The probability of moving to a rehab facility or home (with or without paid care) is depending on several factors including the living situation of the patient, existent of a caregiver at hem for pa patient, age and sex in addition to the year of the study (134). However, in this study, we only used the age, sex odds ratios of discharging to home (vs. rehab facility) reported by (134), controlled for other factors, to calculate the probability of moving to each destination by age, sex. The average rate of moving to each destination has been calculated by average of 10 years data from CIHI data (1994-2004) reported by Coyte et al. (135)- we did not include the time trend (probability of home care have been increased in recent years). Using the average probability for each destination, we used the age and sex distribution of the POHEM population in year 2010, in addition to we the odds ratio of discharging to home by age and sex (estimated by Oldmeadow et al. (134)), to calculate the probability of each individual moving to each of the 262 four destination. We used series of equation, one set for odds ratio translated into probability and others for average probability for each destination using POHEM age/sex proportions. Cost*of*Rehabilitation*After*the*Hip/Knee*Primary*and*Revision*TJR*Surgery* Each target destination after discharge from hospital (after hip/knee surgery) has its own cost depending on the status of a patient. Three cost components are included: (1) Cost of Rehab facility, (2) Home care cost (both paid by healthcare system), and (3) Out-of-pocket cost of patients for (caregivers, aids and equipment’s and home remolding). These are also called the community costs in some references (US studies). We have modeled the year after the hip/knee TJR surgery with more accuracy and 3 different time intervals have been implemented during the first year after surgery First type of government paid (healthcare system) rehabilitation cost has been calculated based on weighed duration of short/long stay during the first year after hip/knee surgery as reported by (Coyte et al. (135)). The dollar amount was propensity adjusted (variation adjusted) number that was transferred into 2010 $CAD which was $6,111.36 (average for all sex, age categories). Second type of cost, home cost paid by government, was $1,166.56 (Coyte et al. (135)). The third type of cost, out-of-pocket cost of patients, were calculated for each 3 months during the year after surgery (expect the first month) as reported d by Australian survey of OA patients after the surgery (Hawker et al. (128). The cost was translated into 2010 Canadian dollars using CPI (reported by statistics Canada for healthcare sector). After first year post-surgery, the only cost would be out-of-pocket for both types of patients discharged to home and rehab. We calculate this cost based on 3 –months period as this cost component converge to be $19.8 for home and $39.5 for rehab patients. Based on the results of (Hawker et al. (128)) study, we fitted a power function and the second year costs were calculated to be $80 and $160 per year for out-of-pocket cost for each year after the first year. Rehab*Cost*Algorithm*in*POHEMDOA** For each patient that undergoes hip/knee TJR surgery (either primary or revision), we first assign them to one of four possible target discharge destinations. Then, according the time period described in the previous section, the mean cost is assigned to the patients, e.g., monthly basis for home care and out-of-pocket cost. The year after TJR surgery is modeled with more accuracy (each month), to reflect the detailed cost results reported in (137, 138). The physiotherapy cost is assigned at the end of each year as calculated in Section 3.2 for physiotherapy cost after surgery. 263 Formal+Caregiver,+Transportation,+and+Community+Cost+ We used available literature estimates for out-of-pocket cost of OA patients for formal caregiver and community services (Gupta et al. (115)). In a survey used in the Gupta et al. (115) study, 1,258 OA respondents have been asked about their out-of-pocket cost relate to arthritis in the past 3 months during 2000-2001 in Toronto, Canada, using standard questionnaire (115). The results were annualized and separated into direct and indirect. For the direct cost, the out-of- pocket cost were due to paid help, transportation, equipment and medical aids and community services (i.e., nurse visits, meals on wheels, transportation and equipment’s). Since the sample has only ~2% undergone surgery, the results of this study were used in calculation of input parameters for pre-surgery cost only. Unit*Cost*of*Formal*Caregiver*and*OutDofDPocket*Cost*for*Uncertainty*Analysis* For the base case (without uncertainty analysis), we assigned a fixed cost due to formal caregiver as reported by Gupta et al. (115), translated back to 2010 $CAD using CPI from (134), which was $2647.3. For uncertainty analysis in Chapter 5, we use lognormal distribution for Log (2647.30)= mean of total; and SD= log (2200). As Gupta et al. (115) reported percentile. We matched that using lognormal distribution. The final results are shown in Chapter 3 for the age and sex specific probabilities of reporting non-zero values. Table 3.5 reports the distribution of the formal caregiver and community costs for those OA patients who reported non-zero cost based on the results of Gupta et al. (115)- we assumed those reported zero cost incurred zero cost. OutDofDPocket*Cost*Algorithm*in*POHEMDOA** We performed a two-step probabilistic approach for out-of-pocket costs; in the first step, based on the age and sex of the person being simulated, POHEM-OA simulates whether the individual incur a non-zero cost or not. If non-zero cost is incurred, then a mean cost is assigned. In uncertainty analysis scenarios, this cost is randomly assigned to the patient based on the distribution shown in Tables 3.5, 3.6. 264 Alternative+Care+Cost+Module+ Table A12. List of Alternatives cost categories Alternative*cost* Baseline*prob.*of*visit* Overall*prob.*Visit* categories! (mean*no.*visit)** (mean*no.*visit)* * * Group&1:**Health*practitioners*(within*last*year)*** ******Physical*therapy* 24%(5.1)! 46%(6.2)! *****Chiropractic*therapy* 9.2%!(4.5)! 24%!(4.7)! Group&2:*Complementary*and*alternative*therapies*(within*last*6*months)*** Massage*therapy* 6.1%!(7.7)! 10.3%!(6.9)! ********Acupuncture* !!!!!!!3.1%!(4.5)! !!!!!!!4.2%!(7.1)! Naturopathic* 0.7%!(1.2)! 1.1%!(2.2)! Homeopathy* 0.5%(1.7)! 1.0%(2.1)! Traditional*Chinese** 0.2%(2)! 0.9%(5.7)! Herbal*remedy** 0.2%(3.4)! 0.9%(6.5)! Ayurvedic*medicine* 0.1%(1)! 0.2%!(1.2)! * Baseline probability is for the youngest age group (<60) prior to OS visit; probabilities defined for at least one visit within the last year. The baseline probabilities were then adjusted for the baseline population in POHEM-OA in 2003, based on distribution of age groups and sex, for which the odds ratio for alt. care was reported –previous tables). ** First group include physiotherapy and chiropractic therapy for which respondents were asked if visited then or not within the last year; second group includes all other alternate care categories and survey respondents were asked if visited each of these 7 categories within last 6 month. The seven alt. categories for group 2 are sorted in terms of highest frequency according to the MOH-OA data (147): message therapy was the highest; Ayurvedic was the lowest in terms of frequency of usage among OA patients. 265 Table A13. Odds ratios for visiting alternative care professionals (at least one time) # Group 1 Group 2 n=1124* n=1118* Odds ratio Mean no. visit Odds ratio Mean no. visit (95% CI) (sd) (95% CI) (sd)^ Sex Female: 6.2 (9.4) Female: 6.4 (9.4) (ref.=male) 2.4 (1.98,3.56) Male: 5.1(10.0) 3.1 (1.5,3.9) Male: 6.1(29.5) OS Before: 5.1 (9.1) Before: 5.9 (10.1) **(ref.=before) 2.1(1.34,2.59) After: 6.1(10.8) 1.2(0.7,2.6) After: 6.7(13.8) Age <60 (ref.) 1 3.8(7.6) 1 4.8(8.2) 60-70 1.4 (0.92,2.3) 4.9(8.6) 1.9(1.1,2.9) 5.9(10.6) 70-80 2.9 (1.2,3.8) 5.2(9.8) 3.1(1.9,4.3) 6.1(7.8) >80 3.4 (2.6,3.9) 7.1(14.3) 3.9(2.1,5.2) 5.1(11.2) *Group 1 includes physiotherapy and chiropractic therapy and was asked if respondents visited them within last year; (n) was for number of respondents that were pre-surgery. Group 2 includes all other alternate care categories (7 categories shown in next Table A8); ** OS: Orthopedic surgeon. Odds ratio of visiting an alternative car professional before vs. after visit to an orthopedic surgeon (OS); # Odds ratios and baseline probabilities (Table A8) were used to calculate probability for each cell (age/sex and OS status) which were implemented in POHEM-OA (Table 3.5 in Chapter 3). ^ Weighted average for mean number of visits across 7 category of complementary care (list shown in Table A8) 266 Table A14. Parameters related to physiotherapist cost for patients after the hip/knee TJR surgeries (implemented in POHEM-OA) + First+year+ Second+year+ ++ Mean+No.+ Prob.+Of+non@ Mean+No.+ Prob.+of+non@zero+visit+ visits+(SD)*+ zero+visit+ visits(SD)+**+ (95%CI)**+ + (95%CI)+*+ REHAB+ 5.3$(5.2)$ 0.53$ 3.2$(7.9)$ 0.15$(0.03,0.29)$ (0.31,0.69)$ HOME+ 3.6$(5.8)$ 0.35$ 3.2$(7.9)$ 0.15$(0.03,0.29)$ (0.11,0.54)$ * The probability of non-zero visit to physiotherapist (or home and nursing home visits) and mean number of visits to all calculated from (Australia study (2010)) during first year for patients whom discharged to home or rehab facility after the replacement surgery ( either primary or revision); SD: standard deviation and 95% CI is Confidence interval (used in Chapter 5 for uncertainty analysis). ** Probability of non-zero visit and mean number of visits to physiotherapist based on the data from MOH study for those who had hip/knee TJR surgery more than 2 years ago. + Cost+Module+for+Side+effects+of+OA@related+Drugs++ Four types of side effects from drugs have been implemented: gastrointestinal (GI), cardiovascular disease (CVD’s), stroke, and dyspepsia. Probability of occurrence of each side effect has been modeled for each drug type from systematic reviews (139): GI from NSAIDS and Coxibs, dyspepsia from opiods, CVD and stroke from Coxibs and NSAID (no side effects for acetaminophens). Appendix A1 discussed the drugs module inside POHEM developed previously by (1). Probabilities for each type of drugs utilization and hazards of side effects for each drug were shown in Table A2. Here, we discuss details of how we calculated the lifetime and per-event cost of side effects from literature studies. All cost are reported here in 2010 $CAD. For implementation in POHEM-OA, we translated them back to 2003 $CAD and 2010 $CAD using CPI for overall economy (134) in Chapters 3 and 4, respectively. 267 Lifetime*and*PerDcase*Cost*of*Each*Side*Effect** Lifetime cost of stroke.$We used the result for lifetime cost of stroke from a US study (141) and then transfer the rate from US to Canada, using PPP reported from Statistics Canada (149). We used the weighted average of all types of strokes reported in Taylor et al. (141) as they all might happen due to side effects of NSIAD and COXIB. Table A15 reported the final average result for lifetime cost of stroke by age and sex category in 2010 $CAD. According to Taylor et al. (141), direct cost of the stroke were disaggregated into acute-care costs incurred in the first 2 years following a stroke, long-term ambulatory care costs incurred 3 or more years following a stroke, nursing home costs, and costs attributable to stroke recurrence. As reported 42% of the overall cost of stoke was due to direct costs and it included hospitalization, drugs, physician, possibility of recurrence, and nursing home costs. For preforming UA in Chapter 5, we used an age-adjusted estimate for lifetime direct cost of stroke from Kolominsky-Rabas et al. (158), a study from Netherland, which was $72,210 in 2010 $CAD. Lifetime cost of CVD. We used results of Brinbaum et al. (140) that reported the incremental lifetime cost of CVD for women only at different age categories. We used the overall economy CPI from (134) to transfer the results into 2010 $CAD as shown in Table A16. According to Tarride et al. (159), we assumed that the female lifetime costs were [1.0-1.3] times higher than males. For preforming UA in Chapter 5, we used an age-adjusted estimate for lifetime direct cost of stroke from de Oliveira et al. (157), a Canadian study that estimated $26,120 for lifetime cost of CVD in Canada in 2010 $CAD. Additionally, Kok et al. (156) was also estimated the lifetime cost of CVD which was $122,230 in 2010 $CAD (average over all ages). Per-case cost of a serious gastrointestinal (GI) case. We used results of Rahme et al. (142), a Canadian study using population database in Quebec to calculate the average GI cost resulted from NSIAD use among OA patients. The average direct cost of NSAID-related GI events for a day on NSAID therapy was $1.34-$0.40 = $0.94 (95% confidence interval: $0.42, $1.59)- which was used in UA study of Chapter 5. Thus, the average GI iatrogenic cost factor for NSAIDs therapy (142) was estimated at (0.94 + 1.28)/1.28 = 1.73 (95% confidence interval: 1.28, 2.18) which meant that for each Canadian dollar spent on NSAIDs during the years of the study, an additional $0.73 was spent on managing and treating the GI side effect for NSAID. The total number of days on NSAID’s estimated from the (142) for the random sample (10%) of the total population of elderly Quebec patients during the years 1993–97 was 97, 159, 340 days 268 and the number of patients who had a dispensed prescription of NSAIDs during these years was estimated at 490,330. As a result, we have around $198.15 for average number of days on NSAID for an OA patient according to (142). We then multiplied this by $1.73 as calculated above for cost factor (95% CI): 1.28, 2.18), which resulted in an average NSAID -related GI cost of ($1.8* 198.15=) $356.67 in 2010 $CAD. Co-prescription cost for GI due to NSAID. According to a 2002 review study of Moore et al. (160), the median cost related to the prescription drugs for treatment of GI was $90 ($61, $136) in 2010 $CAD. As a result, final GI cost was the mentioned cost (for co-prescription) plus the previous cost: $356.67+90= $446.67 (95% CI: $316, $903). The 95% CI was used in Chapter 5 for UA. Per-case cost of a dyspeptic case. We used Moayyedi et al. (143), a UK study in 2002, for calculating direct cost of dyspepsia. Dyspepsia is very common, with a point prevalence of 25– 40%; however, many sufferers do not seek healthcare. In Moayyedi et al.’s study (143) for almost all UK population at risk, dyspepsia direct cost was £11.25 per event per year. Since a 2- year duration was assumed for the time period in which dyspeptic symptoms can last per-case cost was £22.5. As a result, we calculated $48.3 (95%CI: $25-$67) per person per case, which we translated back to 2010 $CAD using CPI of overall economy used from (134). Table A15. Lifetime direct cost of stroke per person by age, sex, and type of stroke MEN 2010 $CAD, Thousand $* <25 $67.83 25-45 $64.70 45-65 $48.48 65-85 $24.64 FEMALE <25 $69.70 25-45 $67.57 45-65 $51.29 65-85 $24.64 We used the results for average types of all types of strokes according to Taylor et al. (149), and transferred the results to 2010 $CAD using PPP and CPI from (134) 269 Table A16. Lifetime direct cost of CVD MEN 2010 $CAD, Thousand $ <60 $380 60-70 $220 70-80 $95 >80 $54 FEMALE <60 $423 60-70 $302 70-80 $121 >80 $57 ** Results was translated from (140) and (159) using CPI from Statistics Canada (134) 270 Appendix B Details of Methods for Projecting the Direct Cost of OA B.1 Hospitalization Cost Hip/Knee+TJR+Cost+Components++and+Inflation+Rate+ We used the literature estimates and CIHI report in 2010 (144) to calculate the proportion of the implant cost of the overall surgery procedure cost. Based on 2003 St. Paul’s data (145), the implant cost was $1,982.50 and for the procedure cost including nursing, Operating room professionals and medical images was $2,472.50. Additionally, $3579 was due to inpatient or hospital stay cost including nursing wages and indirect hospital cost, which occurred during the 7-day stay in inpatient ward in 2003 (i.e., expected length of stay in 2003 according to St. Paul’s data source). Finally, orthopedic surgeon fee was $696 (2003 $CAN) (145). As a result, after taking the inflation into account form 2003 to 2010 (144), the implant cost was calculated to be 48% of the procedure cost in 2010. This resulted to the final percentage calculations of the cost components of the total cost of hip/knee TJR surgery: 42% of knee/hip TJR cost was due to hospital stay including inpatient nursing and indirect cost (inflation rate: 3.6%), 28% due to procedure cost (inflation rate: 3.6%), 20% due to implant (inflation rate: 9.8%) and 10 % to orthopedic surgeon (inflation rate: 2.1%). To calculate the inflation for hip/knee TJR, we calculated the weighted average for the inflation rate according to the percentages for each cost component mentioned above. This resulted in 4.6%. It should be noted that, since the LOS decrease over the next 20 years, the weight of inpatient nursing decreases. However, we assumed 4.6% would be the average rate of inflation over the next 20 years. B.2 Projecting Consumer Price Index (CPI) from 2010-2031 We used the historical trend for consumer price index (CPI) as reported by Statistics Canada, CANSIM report from 1998-2010 (134), and projected that using an exponential model toward the next 20 years. Table B1, below, shows the historical trend from 2003-2010 used in Chapter 3, while Table B2 and Figure B1 depicts the projected rate for Chapter 4. As it is shown in Table B2, the annual inflation rate (i.e., CPI at each year) it is different in each year from (2010-2031). However, as we move on to the future years, annual rate should either increase or remain fixed; if we use a linear trend for future projection, annual inflation decreases. So, we used exponential and annual rate will become constant (i.e. linear inflation rate); for example: in 2030 (CPI=166.1) and 2031 (CPI=169.4), therefore i=1-(CPI2031/CPI2030)=0.0198=1.98%. 271 We used this rate for most of out-of-pocket cost rates (except for alternative care and over-the- counter drugs); in fact, we used it for all formal caregiver, transportation, community cost, medical aids and self-care for rehabilitation, and side effect of drugs. In addition, we used the projected CPI for the interest rate of the overall inflation rate. Based on this figure, on average for the next 20 years, the CPI will be (2.1%). For inflation in price of health-sector resources and services such as drugs, physicians and hospitals, we used the average rates from CIHI reports for each (151-154). Table B1. Overall economy Consumer Price Index from 2003 to 2011 + Year CPI % Change from previous year 2003 100 2004 101.848249 1.8 2005 104.0450017 2.2 2006 106.0076185 2 2007 108.2074352 2.2 2008 110.5392737 2.3 2009 110.802201 0.3 2010 112.6378653 1.8 2011 115.5563203 2.9 2012 117.0575713 1.5 2013 117.9614332 0.9 + * CPI results from Statistics Canada (134) $ $ $ $ $ $ $ $ $ $ $ 272 Figure B1. Projection of Consumer Price Index 200" Consumer"Price"Index"Pojec@on""" 180" y"="101.74e0.0196x" R²"="0.97536" 160" 140" 120" 100" 80" 60" 40" 20" 0" 2005" 2006" 2007" 2008" 2009" 2010" 2011" 2012" 2013" 2014" 2015" 2016" 2017" 2018" 2019" 2020" 2021" 2022" 2023" 2024" 2025" 2026" 2027" 2028" 2029" 2030" 2031" CPI"A"overal"economy" Expon."(CPI"A"overal"economy)" * Projected from 2010-2031- using an exponential fit to the historical data from 2005-2011 (last 6 years) from Consumer Price Index; Annual inflation at each year from 2010-2031 based on above projection becomes (2.1%) where CPI in 2002=100; – from Statistics Canada report (134) 273 Table B2. Consumer price index projected from 2010-2031 $ $ 2010* 2011* 2012 2013 2014 2015 2016 2017 2018 2019 2020 + + + + + + + + + + + 115. 117.4 119.0 121.4 123.8 126.2 128.7 131.3 133.9 136.5 139.2 1 + + + + + + + + + + + 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 + + + + + + + + + 142.0 144.8 147.6 150.6 153.5 156.6 159.7 162.8 166.1 169.4 1 7 2 .4 + + + + + + + + + $ $ *2010 and 2011 CPI are from actual data from (2003-2011) from Statistics Canada report (134) where 2002=100; ** for 2011-2031 period, we used the resulted rates from our projection using historical data; we set year 2010 as base year (CPI=115.1), year 2011 (CPI=117.1), therefore i=1.7% 274 B.3 Projection of Cost, Length of Stay and Number of Hip/Knee TJR Surgeries Table B3. Number of calibrated primary surgeries + + Year+ Calibrated+number+of+primary+surgeries++ 2010+ 73,873$ 2011+ 77,021$ 2012+ 80,169$ 2013+ 83,317$ 2014+ 86,465$ 2015+ 89,613$ 2016+ 92,761$ 2017+ 95,909$ 2018+ 99,058$ 2019+ 102,206$ 2020+ 105,354$ 2021+ 108,502$ 2022+ 111,650$ 2023+ 114,798$ 2024+ 117,946$ 2025+ 121,094$ 2026+ 124,242$ 2027+ 127,390$ 2028+ 130,539$ 2029+ 133,687$ 2030+ 136,835 $ 2031+ 139,983$ *Extrapolated using Canadian Joint Replacement Registry (CJRR) data (124) for 2010 to 2031. 275 Figure B2. Number of hip/knee primary surgeries in Canada from 1994 to 2009 from CJRR data and extrapolated from 2010 to 2031 using a linear trend model* 160,000.0# 140,000.0# y"="3148.1x"+"20355" R²"="0.92947" 120,000.0# 100,000.0# 80,000.0# 60,000.0# 40,000.0# 20,000.0# 0.0# 94# 96# 98# 2000# 2002# 2004# 2006# 2008# 2010# 2012# 2014# 2016# 2018# 2020# 2022# 2024# 2026# 2028# 2030# Actual#Data#from#CJRR# Projected#>using#linear#trend# * Number of primary surgeries was calculated from CJRR (124) and projected using a linear trend into 2031. 276 Figure B3. Average Length of stay after hip/knee TJR from [1996, 2031]* $ * Observed data from CJRR (2009/210) fiscal year (124), using Hospital Mobility Database, 1996– 1997 to 2005–2006 and Discharge Abstract Database, 2006–2007. ** We assumed after 2027, the LOS will become 2.5 days and will stay there (we corrected the end of the curve) $ 277 Table B4. Cost of hip/knee TJRs for primary and revision by age categories from 2010 to 2031 (in 2003 $CAD)* Hip/Knee(TJR(surgery( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 PRIMARY:((18A59(Years( & & & & & & & & & & & & & & & & & & & & & & 804 $7,795 $7,917 $7,934 $7,965 $8,002 $8,049 $8,104 $8,166 $8,237 $8,314 $8,398 $8,489 $8,628 $8, $8,996 $9,193 $9,391 $9,590 $9,791 $9,992 $10,192 $11,169 60A79(Years(( & & & & & & & & & & & & & & & & & & & & & & $7,467 $7,587 $7,603 $7,633 $7,669 $7,713 $7,765 $7,825 $7,893 $7,967 $8,047 $8,134 $8,226 $8,382 $8,518 $8,668 $8,842 $9,033 $9,226 $9,418 $9,610 $10,481 80+(Years(( & & & & & & & & & & & & & & & & & & & & & & $7,755 $7,877 $7,895 $7,925 $7,962 $8,009 $8,063 $8,125 $8,195 $8,272 $8,356 $8,446 $8,542 $8,644 $8,749 $8,878 $9,010 $9,183 $9,356 $9,555 $9,755 $10,562 REVISION:(18A59(Years( & & & & & & & & & & & & & & & & & & & & & & ,683 $11,110 $11,231 $11,221 $11,229 $11,247 $11,277 $11,319 $11,372 $11,436 $11,509 $11,592 $11 $11,783 $11,891 $12,007 $12,129 $12,258 $12,453 $12,698 $12,943 $13,188 $14,698 & & & & & & & & & & & & & & & & & & & & & & 60A79(Years( 8 $11,104 $11,230 $11,219 $11,227 $11,245 $11,275 $11,317 $11,370 $11,433 $11,506 $11,589 $11,680 $11,780 $11,887 $12,003 $12,125 $12,254 $12,448 $12,693 $12,93 $13,184 $13,429 80+(Years(( & & & & & & & & & & & & & & & & & & & & & & $12,745 $12,845 $12,835 $12,846 $12,868 $12,904 $12,954 $13,016 $13,090 $13,175 $13,271 $13,377 $13,493 $13,618 $13,751 $13,892 $14,041 $14,265 $14,546 $14,828 $15,109 $15,390 * Costs are reported in 2003 $CAD (without inflation); for inflation, cost at each year needs to be multiplied by CPI based on 4.6% according to Table 4.1; Costs are projected based on LOS from CJJR and CIHI data for RIW and CPWC of 2010 ($5300) average for Canada; orthopedic surgeon is included (fee of $848 in 2010); 278 Figure B4. Internal validity results (1): patient cost profile* * A female patient, age 52 entered the model in 2001; her cost profile was verified for diagnosis and OS visit. Direct cost here reflects the drugs, physician and hospital procedures costs (for non hip/knee TJRs) grouped as ongoing costs. 279 Figure B5. Internal validity (2): patient cost profile* *A male, age 61 entered the model in 2001; his cost profile was verified for primary and revision surgery. Direct cost reflects drugs, physician visits and hospital procedures costs (non hip/knee TJRs) grouped as ongoing costs, in addition to the hip/knee TJR and rehabilitation costs (event cost). 280 Appendix C . Details of Equations and Statistical Methods in Chapter 5 C.1 Minimization Problem for the Sample-size Approach 2 2 σˆ σˆ 2 2 In Equation 5-1in Chapter 5, V(µˆ) = ( 1 ) + 2 , σˆ reflects the MC-error and σˆ reflects the mn n 1 2 parameter uncertainty and µ is the grand or overall mean (over all parameter replicates). It is also assumed that the outcomes are independent. On the other hand, the variance of the V(µˆ) σˆ 2 outcome is shown by σˆ 2 = = 1 +σˆ 2 . In fact, while the goal of UA procedure is to find n m 2 the uncertainty intervals around the outcome, we need to minimize the variance of the (grand) mean (Eq.5-1). Given the constraint on the computational time for the simulation model, we will define the following minimization problem in Eq.5-3 to find the optimal parameters of the UA procedure, i.e., m* and n*, to minimize the variance of the mean outcome: ⎧ σˆ 2 σˆ 2 ⎪Min {V(µˆ) = ( 1 ) + 2 } ⎨ mn n ⎩⎪Subject to: mn = K The above minimization problem is subject to the computational entrant of mn=K, where K is the constant representing the computational time available. The computational time is assumed to be linear in terms of both population size (m) and number of MC simulation runs (n) (i.e., K=mn, where K is measured in terms of number of individuals that can be simulated in the available time). This is often the case in population-based microsimulation. We also assume it increases in the same fashion for (m) and (n). For example, if we run our simulation program once with 1000 patients it takes the same time as we run it twice with 500 patients. According to O’Hagan et al. (40), m*=1, n*=K will minimize the value ofV (µˆ) and m*=1 is the optimum value (40). On the other hand, using an optimal value of (m*=1) is not practical because of several reasons. Firstly, simulating only one patient will not allow us to separate first and second order uncertainty (40). Secondly, omitting first order uncertainty is not realistic since it is inherent in the population. If you are able to simulate a patient twice and observe his/her life paths, it should not be the same because of ever-presenting natural errors. Finally, we should note that “m” couldn’t be too low as it may contradict the basic assumption of positive variance. As shown in Chapter 5 for the sample-size approach and to calculate uncertainty intervals for the 281 σˆ 2 σˆ 2 outcome, from σˆ 2 = 1 +σˆ 2 , we need to calculate σˆ 2 by reducing 1 from σˆ 2 (variance of m 2 2 m the outcome). This is due to the fact that we can observe both variance of the outcome and first order uncertainty, but it is not possible to directly calculate the parameter uncertainty. If “m” is 2 too small, this may result in negative value for σˆ 2 which is not acceptable. As a result of above discussion and to provide a practical approach to calculate uncertainty ˆ intervals for the outcome, we define “d” as a measure of precision forV (µ), such that V(µˆ) = d , where d can be any (small) value defining the precision for our uncertainty estimates. The smaller the d, the higher the precision of our uncertainty estimates of the outcome. We recommend using small values within (0.0001-0.001) for precision level d. We suggest starting with a small number such as 0.0001. If it resulted in n*<1 or negative initial estimate for parameter uncertainty, user needs to increase d in 0.0001 increments until a feasible (n*) and (m*) is reached. The negative value or n*<1 reflect the fact that the computational time is not enough for such a low precision or very low precision estimates were used for initial estimates. C.2 Details of Initial Step of the Sample-size Approach Here, we describe the algorithm to estimate (m0,n0) for first step of sample-size approach, in uncertainty analysis of simulation models by considering the minimization problem of variance of the grand mean as shown in (Eq. 5-1) in Chapter 5. In order to be able to calculate (m*), (n*) for the sample-size approach (in Chapter 5) and sample-size approach (in Chapter 5), we need 2 2 to initially estimate σˆ1 , σˆ2 from a pre-determined values which will be calculated by (m0) and (n0). Step1. Choice of m0: As described in Chapter 5, “m0” should be chosen to be high in the initial step. By “high m”, i.e., population size (m), we mean high relative to the initial population size as σˆ 2 it is used for the outcome of interest. Due to constraint for negative variance: σˆ 2 − 1 ≥ 0, (const.2)“m” 2 m can’t be too low as it may cause the left hand side to be negative. As a result, we recommend following the steps below in order producing an acceptable “m0” and “n0”. For population size, we should choose the largest possible number for “m0” in the initial loop, where “m0” is the initial value used for the initial population of the outcome. For example, if the outcome was average cost of a disease among Canadian population, the initial pupation value for the outcome would be the Canadian population for which the microsimulation has been designed for around 40 282 million individuals. This 40 millions represent (survey) weighted number of initial population that were in the survey used for generating the initial population of the simulation model (e.g., CCHS 2001 for POHEM model used as an example in Section 4, Chapter 5), in addition to the number of new births to be entered in the simulation model and new immigrants. Therefore, we would start with m0=40 million based on the above equation. Next, one should set the subsample size to the maximum possible value (e.g. we used 40 sub-samples for initial population representing weights assigned to each individual in the survey used for generating the initial population of the POHEM-OA simulation model). Step 2. Choice of n0: The main criteria for choosing “n0” in this step are to provide a small number of initial samples from the parameter distribution that also cover the parameter space. As in design of experiment, we need to design the approach to generate sample points rather than preforming a random sampling in order to cover the overall space of parameters using small number of sample points in the initial step. Fractional factorial design (227) is one of the approaches that can be used. Here, we propose a fractional factorial design approach for the sampling in the initial step. Fractional Factorial Design (FFD): This method is generalized oorm that of O’Hagan et al. (4). As mentioned in the initial step for the first loop (r=0), we select 3 parameter replicates that covers the overall parameter space according to the fractional factorial design. For this, we assign 3 levels for each parameter (mean, mean–SE, mean+SE) 2 that would result in (n0=3). However, we should correct the estimates of our initial σˆ 2 2SE 2 by dividing it with the maximum of a fraction of variances of three points, i.e., ( ) 3 2SE 2 2SE 2 for each parameter (4, 227). This is equal to max ( for parameter 1, for 3 3 parameter 2, until the last parameter). This formula can be described by the delta method for Var (f(x)) (227). This is done to correct for the biased estimate produced as the result of non-random choice for the sample parameter points in the initial step. This approach is used in initial sample for ANOVA approach of O’Hagan et al (4), where they had 3 parameters in their UA-list. If the time allows to perform additional loops in the initial step (the first loop takes less than 15% of the overall computational time (K)), we would calculate the new (m1), which is lower than (m0), and this time use (n1=5) instead of (n0=3), by assigning 5 levels for each parameter (mean, ,mean-SE, mean+SE, mean–2SE, mean+2SE). However, we 283 2 should correct the estimates of our initial σˆ 2 (in Step 4) by dividing it with maximum of 4SE 2 4SE 2 the variances of these 5 points=max ( *var(parameter 1) , *var(parameter 5 5 4SE 2 2), …, *var(last parameter) (227). For additional loops, if the overall time used in 5 the initial step is not more than 15% of the overall computational time up to the current loop, we would perform additional loop with the resulted (m3) and (n3=7). For calculating the (m2), (m3), etc., we would use the first loop results fro sigma 1, but the new results for sigma 2 will be used in (step 3) to calculate the next population size (m2), (m3), etc. At each loop, (m2), (m3) will be decreasing, that will allow us to increase (n2), (n3), etc. The more the loops in the initial step are performed, the lower the (m*) for the main run of the sample-size algorithm would turn out to be. The higher number of loops would be 2 2 necessary for PDMS models where σ1 is large compared to σ 2 (this would happen in case of small initial population sample in the model). 2 2 Step 3 . Estimating initialσˆ1 : Run the simulation (n0) times for (m0) patients and calculate σˆ1 by averaging over variances within patients’ outcome over all n0 runs. 2 Step 4. Estimating initial σˆ 2 : For each simulation run in Step 3, calculate µˆ and finally, ! 2 calculate ! denoting variance of all µˆ ’s over the n0 runs. Next, use (Eq. C1) to calculateσˆ 2 : σˆ 2 σˆ 2 = σˆ 2 + ( 1 ) (Equation C1) 2 mn 1 n 2 ˆ 2 ˆ 2 ˆ In (Eq. C1), σ is the overall variance: σ = ( )∑(zi − µ) . n −1 i=1 C.3 Transformation of Aggregate Outcomes in the Sample-size Approach To generalize the UA method for any type of outcome in PDMS, here we provide details of how to transform the results of average outcome back to an aggregate type outcomes. We assume that the aggregate type outcome is a linear function of individual outcome; for example, weighted summation of individual outcomes or difference between individual outcomes. We represent the outcome of the microsimulation model as a linear aggregate function of individual’s outcome (to integrate both of these types of outcomes into one variable shown by Lf 284 (.): either mean or aggregate of individual outcomes. The Lf (.) would be the function of all of the individuals’ outcome in the initial population. We refer to zipj as “individual outcome” and by “outcome at each run”, we mean the zi (e.g. average cost of all individuals or total cost of all individuals) for each parameter replicate, while by “overall outcome”, or grand mean, we mean expected value of all outcomes over all MC- runs, i.e., z=E (zi) =Ep,j (Lf (zipj)). In general terms, we assume zip , outcome at each MC-run and for each sub-sample, is the linear function of zipj or the individual outcomes. The general outcome of a simulation model, whether it is of “average type “or the “total type”, can be written as the followings: z ⎧ ∑ ipj ⎪ j ⎪= E(z ipj ) if z i = z i = , average of individual outocme z = E(z i ) = E(Lf (z ipj )) → ⎨ n i p j ⎪ (Equation C2) = mE(z ij ) if z i = sum(zipj ) , total sum of individual outocme ⎩⎪ j As shown in the left hand side of the (Eq.C2), if the outcome is “average type”, the linear function LF (.), becomes equal to “ !i”, which is the overall (average) outcome at each replicate. Finally, mean of zi, over all parameter replicates would be !, which is a fixed value. For any type of outcome, our estimate for the Lf (zip,j), is zˆi , which is calculated according to following: 1 n zˆ z ; zˆ is an estimator for Lf (z ), which is calculated by averaging over all the “n” i = ∑ ip i ipj n i=1 replicates of the parameter uncertainty. Second goal of UA procedure, in addition to estimating overall grand mean, is to estimate parameter uncertainty (*). For the variance of the outcome of summation types, we would use the relation between variance of the average type outcome as the followings: 2 var(LF(µ)) = LF0 (var(µ)) (Equation C3) In (Eq. C3), LF0 (.) is the linear function without any intercept. We will use (Eq.C2) and (Eq.C3) so that for any type of outcome (average or summation types), the proposed ANOVA-based UA approach can be performed. Providing inference for mean and parameter uncertainty of average outcome was discussed in details in Chapter 5. For the summation-type outcomes (e.g., total cost of a disease within a population), we first calculate the 95% UI for the corresponding average outcome (e.g., average cost of a person with the disease); next, we use 285 (Eq.C2) and (Eq.C3) to transfer the results back to the non-average outcomes. Specifically, for parameter variance of summation-type outcomes, we use equation (Eq.C3) to translate the calculated result for average outcomes back to summation ones. C.4 Latin Hypercube Sampling (LHS) in the Sample-size Approach In the main run of the sample-size algorithm in Chapter 5, we assign (m*) for the population size and (n*) for MC-simulation runs or the number of replicates that will be randomly selected according to the (joint) distribution of the parameters in the UA-list. For parameter replicates we will use Latin Hypercube Sampling (LHS) as described in (234): First, we need to generate one sample point according to matching without replacement as the followings: we randomly choose 1 point from the first parameter (out of the n* points); then, we delete this point (in order to avoid this point to be chosen in the subsequent samples). Next, we match it with another point in the second parameter (selected out of n* points sampled for the second parameter). Second, we move to the next parameter and match the selected points to form the first sample point until we get to the last parameter. For the second sample point, we again start from the first parameter and repeat the previous procedure. However, we now have n*-1 points to choose from since we selected one point from each parameter in the previous step. Finally, we have n* sample points, in which some of the points contains a tuple corresponding to multivariate parameter. C.5 Calibration Algorithms in POHEM-OA Step 1. Finding target prevalence: Age and sex-specific incidence rates of OA have been estimated from the PDBC database (84) as discussed in Kopec et al. (1). These rates are adjusted for BMI levels. Assuming in year 2001, we have no OA events , we run POHEM for 50 years and we get the prevalence at 2051 = OA_PrevalenceRate_Target . So, by running POHEM for 50 years and assuming HR of BMI=1, we will get the background prevalence at year 50, which is the estimated baseline prevalence for our reference category at year 2001 and is equal to OA_PrevalenceRate_Target . This is the prevalence of our reference category. In the next step, we want to find the prevalence of OA for l population (including all BMI categories) Step 2. Calibrate Prevalence by risk factors (BMI distribution) : OA-PrevelanceRate_Baseline: To count in the risk factors (e.g. BMI level), we start by OA_PrevelanceRate_Target from previous step and try to find the OA-PrevelanceRate_Baseline that takes the HR's into account. 286 In this step, we run the model and calculate the current baseline levels at year 2050 (in accordance with step 1), then POHEM generates a table called Prevalence at start up that shows the reference category prevalence rate at the basleine and should be equal to OA_prevelance_Target. We keep running the model so that Prevelance_Target= Prevelance_Startup. As a result, the OA-PrevelanceRate_Baseline is derived from this step that represents the baseline prevalence for all the BMI categories. Step 3. Calibrate Incidence by risk factors (BMI distribution): OA-IncidenceRate-Baseline. After calculating the baseline prevalence, repeat the Step 2 for incidence to get OA-IncidenceRate- Baseline . Below is the details steps for POHEM-OA users. Calibration*of*OA*Prevalence*for*Risk*Factors* a. Set parameter OAP_Conditional_on_SROA_at_StartUp == FALSE b. Set parameter SROA_RelativeRisk with the relative risks of self-reporting OA based on BMI and HUI groups, by sex and age group. c. If not already done, set Initial Baseline prevalence rates to Target prevalence rates (copy/paste via Excel) OA_PrevalenceRate_Baseline <- OA_PrevalenceRate_Target d. Set parameters pPopulationStartMethod = Full pNewArrivalsControl = StartPop pMaxYear = 2001 e. If exact number of Scenario Cases is not known, • Run the simulation with 1 case to generate user table utCases i) Open user table utCases to identify number of cases actually required for this run and copy it back into Scenarios \ Settings \ Cases. Save Scenario. f. Run the simulation with exact number of cases required g. Extract new estimated baseline prevalence rates from user table utNEW_OA_PREVALENCE_BASELINE And copy them back into parameter OA_PrevalenceRate_Baseline h. Evaluate convergence by comparing (graphically) POHEM generated prevalence rates in table tOA_Prevalence_at_StartUp 287 to the target rates in parameter OA_PrevalenceRate_Target Also could additionally look at sum of squared errors between these in table utITERATOR_SSQD_SEX or could compare the change in baseline values from one iteration to next) i. Repeat three previous steps until convergence is achieved Calibrate)Prevalence)Conditional)on)Self)reported)OA)(SROA)) After prevalence has been calibrated for risk factors (HUI and BMI), the user can then optionally choose to further control the assignment of prevalent cases, conditional on SROA. j. Set parameter OAP_Conditional_on_SROA_at_StartUp == TRUE k. Re-calibrate, by repeating steps 13-16 (one iteration is likely sufficient). Even without recalibrating, the user can run the model in this mode (ie prevalence conditional on SROA) without much loss on hitting the target rates. Automated*Calibration*Algorithm*Used*in*UA*of*POHEM?OA** The above steps have been automated in MATLAB for UA algorithm. Here is the algorithm performed for the automated calibration procedure. Step 1. Put Observed prevalence rate into baseline-rates Step 2. Sample from HR distribution REPEAT UNTIL CONVERGENCE: 1. Run POHEM-OA with the outcome of prevalence of OA at each year. 2. Get the variable “Next _incidence_Baseline”; Copy it back to Baseline-Incidence (Both of these vectors are adjusted for BMI distribution) 3. Check Sum of Squared errors (SSQ) error of "Incidence at start year- Incidence observed" (Both crude rates); END. 288 C.6 List of Parameters Used in the UA Approach for Total Cost of OA Table C1. List of parameters, distribution and variances Parameters** Distribution* SE* Variance* Hazard)ratio):)OA)incident)) Lognormal) )) )) Female,)Overweight) )) 1.994690286) 3.978789336) Female,Obese) )) 2.027603985) 4.11117792) Male,Overweight) )) 1.767564356) 3.124283753) Male,obese) )) 1.767138424) 3.122778211) Cost)of)hip/knee)TJR)surgery) Normal) )) )) Male,)age)group)1) )) 190.6) 36328.36) Male,)age)group)2) )) 130.1) 16926.01) Male,)age)group)3) )) 150.3) 22590.09) Male,)age)group)4) )) 159.4) 25408.36) Male,)age)group)5) )) 273.2) 74638.24) Female,age)group)1) )) 249.3) 62150.49) Female,age)group)2) )) 193.5) 37442.25) Female,age)group)3) )) 285.5) 81510.25) Female,age)group)4) )) 315.6) 99603.36) Female,age)group)5) )) 303.2) 91930.24) Probability)of)discharged)destination)after)surgery) )) )) Rehab,to,home)Care) Dirichlet) 0.12) 0.0144) distribution) Home)care,to,home)care) )) 0.14) 0.0196) Home)care,to,self)care) )) 0.21) 0.0441) Self)care,to,self)care) )) 0.23) 0.0529) Unit)cost)associated)with)each)discharge) )) )) )) destination)) Rehab,to,)home)care) Beta)distribution) 370.1) 136974.01) Home)care,to,home)care) )) 250.1) 62550.01) Home)care,to,self)care) )) 310.1) 96162.01) Self)care,to,self)care) )) 450.2) 202680.04) Odds)ratio:)annual)payment)to)formal) )) )) )) caregiver)(zero)vs.)non:zero)) Male,)age)group)2) Lognormal)dist.) 1.2) 1.44) Female,age)group)2) )) 2.1) 4.41) Unit)costs)of)formal)caregiver)) Beta)dist.,)mean)) )) )) Annual)cost)of)formal)caregiver),)averaged) )) 732) 535824) over)all)ages/sexes)) ) 289 Parameters** Distribution* SE* Variance* ) )) )) )) Odds)ratio:)Annual)visit)to)a)physical) therapist) Male,)age)group)2) Lognormal)dist.) 1.2) 1.44) Male,)age)group)3) )) 1.4) 1.96) Male,)age)group)4) )) 1.7) 2.89) Female,)age)group)2) )) 1.5) 2.25) Female,)age)group)3) )) 2.1) 4.41) Female,)age)group)4) )) 1.7) 2.89) Odds)ratio:)Annual)visit)to)any)type)of) )) )) )) alternative:care)professionals) Male,)age)group)2) Lognormal)dist.) 1.1) 1.21) Male,)age)group)3) )) 1.6) 2.56) Male,)age)group)4) )) 2.2) 4.84) Female,)age)group)2) )) 2.2) 4.84) Female,)age)group)3) )) 1.7) 2.89) Female,)age)group)4) )) 1.1) 1.21) Odds)ratio:)side)effect)of)OA)drugs)) Lognormal)dist.) )) )) CVD)) )) 2.1) 4.41) Stroke) )) 3.1) 9.61) Dyspepsia) )) 1.6) 2.56) GI)) )) 2.1) 4.41) Life)time)cost)for)CVD)) Uniform) 703) 494209) Life)time)cost)for)stroke)) )Triangular)) 432) 186624) Unit)cost)of)dyspepsia)) Beta) 31) 961) Unit)cost)of)GI)) Beta) 49) 2401) ) ) ) ) Number)of)Parameters:)50) ) ) MAX)(VAR)=) 535824) Adjustment)coefficient)*=2/3),,>) Var=2/3*Max) )=)) 357216) (VAR) *The adjustment coefficient, i.e., 2/3 multiplied by max of variance, is calculated for overall variance (maximum variance of all the parameters’ variances) in the initial step according to the "fractional factorial design" for n0=3 MC runs where all parameters are set to (Mean- SE, Mean, Mean+SE). The resulted initial step solution is shown in Appendix C1. 290