Modeling the clinical and economic implications of obesity using microsimulation

Abstract

Objectives:

The obesity epidemic has raised considerable public health concerns, but there are few validated longitudinal simulation models examining the human and economic cost of obesity. This paper describes a microsimulation model as a comprehensive tool to understand the relationship between body weight, health, and economic outcomes.

Methods:

Patient health and economic outcomes were simulated annually over 10 years using a Markov-based microsimulation model. The obese population examined is nationally representative of obese adults in the US from the 2005–2012 National Health and Nutrition Examination Surveys, while a matched normal weight population was constructed to have similar demographics as the obese population during the same period. Prediction equations for onset of obesity-related comorbidities, medical expenditures, economic outcomes, mortality, and quality-of-life came from published trials and studies supplemented with original research. Model validation followed International Society for Pharmacoeconomics and Outcomes Research practice guidelines.

Results:

Among surviving adults, relative to a matched normal weight population, obese adults averaged $3900 higher medical expenditures in the initial year, growing to $4600 higher expenditures in year 10. Obese adults had higher initial prevalence and higher simulated onset of comorbidities as they aged. Over 10 years, excess medical expenditures attributed to obesity averaged $4280 annually—ranging from $2820 for obese category I to $5100 for obese category II, and $8710 for obese category III. Each excess kilogram of weight contributed to $140 higher annual costs, on average, ranging from $136 (obese I) to $152 (obese III). Poor health associated with obesity increased work absenteeism and mortality, and lowered employment probability, personal income, and quality-of-life.

Conclusions:

This validated model helps illustrate why obese adults have higher medical and indirect costs relative to normal weight adults, and shows that medical costs for obese adults rise more rapidly with aging relative to normal weight adults.

Keywords::

• Obesity
• Burden of illness
• Microsimulation
• Economic analysis

Introduction

Approximately 35% of US adults are obese1, which increases their risk for cardiovascular disease, type 2 diabetes, stroke, certain cancers, and other chronic health problems2. Previous studies suggest obesity raises average, annual medical costs by $1429 (2008 estimate) to $3508 (2010 estimate), and $316 billion in US medical costs in 2010 is attributed to obesity3^,4. Furthermore, chronic health problems associated with obesity are reported to reduce productivity and labor force participation5–7, reduce quality-of-life8^,9, and increase mortality10. The World Health Organization reports that excess body weight is responsible for 44% of diabetes-related costs, 23% of ischemic heart disease-related costs, and 7–41% of medical costs associated with certain cancers11.

Previous studies on the economic burden of obesity analyzed medical expenditures to isolate the contribution of excess body weight after controlling for demographics and other factors correlated with obesity and medical expenditures3^,4. While these studies offer a historical snapshot of how obese individuals differ from their non-obese peers, such retrospective analyses provide limited ability to model the relationships between excess body weight and morbidity, and the relationships between morbidity and medical costs.

Health economic models that simulate the link between patient risk factors and outcomes such as medical expenditures have been used to help guide policy and evaluate the value of programs and treatment for a variety of diseases12–17. Before such models can be used to inform policy or clinical practice, however, they should be validated and described in sufficient detail for peer review18^,19.

A recently published microsimulation model estimated the potential value of weight loss and improved blood glucose levels in helping to prevent diabetes onset and sequelae among a pre-diabetic population receiving intensive lifestyle intervention20. The purpose of this paper is to describe the addition of 25 obesity-related comorbidities to this microsimulation model and to describe validation activities for these added comorbidities. We then use the model to quantify the burden of obesity as it relates to morbidity, medical expenditures, mortality, productivity, and quality-of-life. We describe how the human and economic burden of obesity differs across segments of the population and how this burden changes with aging.

Methodology

Modeling approach overview

The Markov-based microsimulation model described here expands on a previously published ‘disease prevention model’ that mapped the relationships between patient risk factors, diabetes, diabetes sequelae (hypertension, ischemic heart disease [IHD], congestive heart failure [CHF], stroke, myocardial infarction, chronic kidney disease [CKD], peripheral vascular disease [PVD], renal failure, amputation, and diabetic retinopathy), medical expenditures, indirect economic outcomes (employment, income, work absenteeism, disability), mortality, and quality-of-life20. The disease prevention model publication and accompanying technical appendix (available as Supplementary material) provide a detailed description of the model data, methods, assumptions, and limitations; and a summary of validation activities and sensitivity analyses as it pertains to diabetes and sequelae. Diabetes and its sequelae are comorbidities of obesity.

Modeling efforts described here reflect addition to the microsimulation model of 16 obesity-related cancers, major depression, pneumonia, pulmonary embolism, osteoarthritis, obstructive sleep apnea (OSA), gastroesophageal reflux disease (GERD), and non-alcoholic fatty liver disease (NAFLD)21. These medical conditions are listed as comorbidities of obesity by organizations such as the Centers for Disease Control and Prevention, the American Heart Association, and the American Diabetes Association. Incidence of all diseases, both the previously included and the newly added, was modeled based on each disease’s unique natural history and risk factors.

The model simulated disease onset over 10 years among each adult in a nationally representative sample of obese adults and a demographically-matched normal weight sample. The model used current health risk factors to predict annual onset of disease, with this process repeated for each person over 10 years unless death occurred sooner. Risk factors consisted of demographics (age, sex, race [white, black, other], and Hispanic ethnicity); biometrics including body mass index (BMI), systolic (SBP) and diastolic (DBP) blood pressure, total cholesterol, high-density lipoprotein cholesterol [HDL-C], hemoglobin A1c (A1c), and fasting plasma glucose (FPG); current smoking status; and the presence of the previously listed comorbidities.

Figure 1 lists disease states in the model, with each line connecting disease states representing a relationship included in the model’s prediction equations. Body weight (usually as reflected by BMI) is linked to all condition categories (endocrine, cardiovascular, cancers, respiratory, and all other). Many of these health states are risk factors for other states, such that there are often first, second, and third order linkages in the model. For example, as illustrated in Figure 2, BMI is linked to cholesterol level, SBP, CHF, and diabetes, which are all linked to risk for myocardial infarction. Cholesterol, SBP, and diabetes are linked to CHF risk (which is linked to myocardial infarction risk), representing seven paths in the simulation model by which BMI is linked to myocardial infarction risk. In addition to the disease states mentioned, atrial fibrillation and left ventricular hypertrophy (LVH) are modeled as risk factors for stroke and myocardial infarction. Health states were then used to predict annual medical expenditures, labor force participation and missed work days, quality-of-life, and mortality.

Figure 1. Model overview. Connecting lines show the items in the model that are linked. BMI, body mass index; CHF, congestive heart failure; CKD, chronic kidney disease; DBP, diastolic blood pressure; GERD, gastroesophageal reflux disease; HbA1c, hemoglobin A1c; HDL, high-density lipoprotein; IHD, ischemic heart disease; LVH, left ventricular hypertrophy; NAFLD, non-alcoholic fatty liver disease; NHL, non-Hodgkin’s lymphoma; OSA, obstructive sleep apnea; PVD, peripheral vascular disease; SBP, systolic blood pressure.

Figure 2. Example of how BMI affects myocardial infarction risk. BMI, body mass index; CHF, congestive heart failure; SBP, systolic blood pressure; MI, myocardial infarction.

Population

The combined 2005–2012 National Health and Nutrition Examination Survey (NHANES)22 files contain data for 5221 adults who are obese (BMI ≥ 30)22^,23, aged 20–85, with valid metrics for BMI, SBP, total cholesterol, HDL-C, and HbA1c or FPG. These NHANES files also contain data for 6534 normal weight (BMI < 25) adults with the same valid metrics needed for modeling. Repeated sampling from both the obese and normal weight samples, using NHANES sample weights to determine selection probability, produced nationally representative samples of 100,000 obese adults for each population modeled: all obese adults, obese I (30.0 ≤ BMI ≤ 34.9), obese II (35.0 ≤ BMI ≤ 39.9), obese III (BMI ≥ 40), obese aged 20–44, obese aged 45–64, and obese aged ≥65. For each obese population analyzed, we created a matched (same demographics and same insurance type [Medicare, Medicaid, commercial, self-pay]) normal weight population of 100,000 adults. The burden of obesity reflects differences in simulated outcomes between the normal weight and obese populations.

Model parameters and data sources

This model builds on a previous model whose data sources and prediction equations describing the complex interaction of patient demographics, biometrics, disease presence, and disease onset for diabetes and diabetes-related comorbidities are described in detail elsewhere20. Data sources and parameters for the additional obesity-related health outcomes newly added to the microsimulation model are described below. Supplementary Appendix 1 summarizes the data sources for each health risk factor and outcome modeled.

Relationships between modeled risk factors and disease state transition probabilities come from published clinical trials, meta-analyses, observational studies, government statistics, and original analyses of NHANES data. Priority was given to recent meta-analyses and studies based on randomized clinical trials, longitudinal studies, and the US population. Key sources described later include the National Cancer Institute’s SEER database, Framingham Heart Study, UK Prospective Diabetes Study (UKPDS), NHANES, and Medical Expenditure Panel Survey (2008–2012 files).

Cancer incidence

Obesity is associated with increased risk for 16 types of cancers—breast, cervical, colon, endometrial, esophageal, gallbladder, kidney, leukemia, liver, multiple myeloma, non-Hodgkin’s lymphoma, ovarian, pancreatic, prostate, stomach, and thyroid24–38. Published relative risks (RR) for cancer by weight category are available, but no identified studies show how cancer risk changes with BMI within a body weight category (which is needed to model weight change). For each cancer type, The National Cancer Institute’s SEER database provided underlying incidence rates by 5-year age band, sex, and race for the general adult population, regardless of body weight category39. Combining these incidence rates with published relative risks by weight category, we extrapolated continuous incidence rates to predict annual risk of each cancer type by a person’s age, sex, race, and BMI. Technical details of this process are in Supplementary Appendix 2.

Major depression episode

Probability of a major depression episode was associated with obese II (odds ratio = 1.90, 95% confidence interval = 0.79–4.60), obese III (4.63, 2.06–10.42), female (2.62, 1.76–2.32), and current smoking status (2.24, 1.32–3.81)40; and with diabetes (1.24, 1.09–1.40)41. The estimated association between major depression episode and diabetes was based on a published meta-analysis involving 11 studies covering 172,521 participants, including 48,808 people with type 2 diabetes. Eaton et al.42 reported the median duration of a major depression episode was 8–12 weeks. Because this model’s cycle length is 1 year, which exceeds the length of an episode, annual incidence and prevalence of major depression episodes are equivalent in the model.

We used the above published odds ratios to estimate depression risk associated with each risk factor. We derived a baseline population from NHANES with no risk factors (i.e., male, not obese II or III, non-smoker, and non-diabetic) and calculated risk (risk_B) for major depression among this population. For example, calculation of depression risk associated with obese II status used the following steps:

1. Compute baseline (B) odds of depression from depression risk as:

2. Calculate odds of depression among obese II: (where 1.9 is the published odds ratio for obese II population40); and

3. Calculate risk for depression among obese II:

Obstructive sleep apnea

Viner et al.43 reported the risk of OSA, defined as apnea-hypopnea Index (AHI) >10, among a group of patients suspected to have OSA. Most of these patients were obese and experienced loud snoring. OSA risk was associated with sex, age, BMI, and loud snoring status (yes/no). Rate of loud snoring in a general population was reported by Philips et al.44. Combining the information above, we calculated OSA risk among a group of people suspected to have OSA. To apply this to the general population, we calibrated the calculation based on numbers reported by Newman et al.45, who estimated average 5-year incidence of OSA was 11.9% in males and 4.9% in females.

Other obesity comorbidities

Other comorbidities modeled are chronic back pain, gallstone disease46, gastroesophageal reflux disease47^,48, non-alcoholic fatty liver disease49^,50, osteoarthritis51^,52, pneumonia53, and pulmonary embolism54. Published incidence rates and relative risks associated with BMI category (and often by demographic) were identified in the literature for each condition, and we estimated disease onset risk applying the same approach described above for modeling cancer incidence.

Economic outcomes

The prediction equation for annual medical expenditures came from regression analysis of the 2008–2012 Medical Expenditure Panel Survey (MEPS) using a generalized linear model with gamma distribution and log link. Explanatory variables were age group, sex, race/ethnicity, insurance status, body weight category (normal, overweight, obese), continuous BMI for obese individuals, and presence of modeled diseases (including interactions terms for diabetes and diabetes sequelae). The estimation process for direct medical expenditures and other economic outcomes (personal income, employment probability, Supplemental Security Insurance for disability, and lost productivity from missed work days) has been described in detail in the diabetes prevention publication20 and in Supplementary Appendix 3. Monetary values for economic outcomes are expressed in 2013 dollars, and per-person measures are presented based on the number of people living in a given year (rather than the number of people at the beginning of the simulation).

Model validation

Model validation activities follow the International Society for Pharmacoeconomics and Outcomes Research (ISPOR) practice guidelines for validation and transparency18^,19. During conceptualization, the model specification was reviewed by experts in obesity, endocrinology, health services research, modeling, and health economics. The experts critically reviewed the model structure and logic, and reviewed model outcomes to ensure face and clinical validity. Programming code was peer-reviewed internally, and extensive sensitivity analysis was conducted.

Internal simulation validation exercises with the NHANES sample, described elsewhere20, used adults in the 2003–2004 NHANES to simulate the effects of aging on health outcomes over 6 years and were then compared with the population in the 2009–2010 NHANES samples. Simulation results were consistent (i.e., within the margin of error) with NHANES-reported prevalence of hypertension, IHD, diabetes, and stroke prevalence. The exception was simulated IHD prevalence for age group 65–74, which was 40% higher than actual 2009–2010 data. This is likely due to a cohort effect that the wider adaptation of cholesterol-lowing medications has reduced IHD incidence between 2003–2010. Simulated BMI was similar to actual data across all age groups for both genders, with the largest deviation being 3%.

Additional unpublished validation exercises simulated results (prevalence of chronic diseases, and incidence of adverse health events) for each health state for comparison with external public data. Simulated results were consistent with published results for the overall adult population for BMI55, cancer39, chronic kidney disease prevalence56, hypertension and IHD prevalence57, myocardial infarction prevalence58, diabetic retinopathy59, and stroke prevalence58. There was race and gender difference in how well the simulated stroke prevalence matched the actual data. Simulated stroke incidence for age 55–64 was ∼40% lower than the actual levels among the black population, while close to actual levels for the caucasian population. Among the caucasian population aged 45–54, the difference between simulated and actual incidence for males is ∼4-times larger than that for females, while gender difference in other age groups is similar. Even though race and gender are part of the stroke risk equation, the actual difference may be larger than previously expected, and will be a topic for future research. Simulated incidence of renal failure was lower for adults aged 45–64 (0.03% vs 0.06%), but higher for the age 75+ population (0.22% vs 0.18%)60. Simulated mortality was similar to the published estimates used to construct the mortality rates—e.g., mortality related to CHF61, stroke62, cancers63, and overall mortality64.

Transition rates from pre-diabetes to diabetes were estimated under natural history of disease and lifestyle intervention scenarios for comparison with published studies such as the Diabetes Prevention Program and Outcomes Study that tracks outcomes for more than 10 years after participating in the Diabetes Prevention Program trials20^,65.

An important external validation simulated longitudinal health outcomes for patients in the GE Centricity electronic health records (EHR) database to compare simulated to actual outcomes. We derived a patient sample (n = 1357) between age 20–85 that appeared in the EHR data in 2004 or later. This sample had the health risk factors and valid biometric data needed to run the simulation model (e.g., BMI, SBP, total cholesterol, HDL cholesterol, and HbA1c) in years 0 and 5—the time period analyzed for validation. Disease indicators were constructed from medical claims data using ICD-9 diagnosis codes. We simulated how patients’ characteristics evolved over time, and compared the trends against those in the EHR data. The comparison focused on BMI, SBP, HDL-C, total cholesterol, and HbA1c. Results are presented in the Supplementary Appendix, and the model simulated the evolution of patient characteristics that were similar to the actual outcomes. Differences between actual patient outcomes and simulated outcomes are consistent with expectations that patients in the EHR data are receiving professional care (and have increased use of medications and counseling that contribute to better health outcomes than would occur in the absence of such care).

While we have made the validation as comprehensive as possible, validating a complex model presents numerous technical challenges. Additional information on validation activities and external validation using electronic health records is covered in more detail in Supplementary Appendix 4. Overall results suggest that the model outcomes are in line with the expected clinical and economic outcomes.

Results

We simulated health and economic outcomes for a nationally representative obese population and a matched normal weight population over a 10-year horizon. In the initial year, the mean age in both populations was 48.3 years (Table 1). Compared with the matched normal weight sample, the obese sample had slightly fewer males (48.6% vs 50.4%). The obese population was sicker, as indicated by higher prevalence of pre-diabetes, diabetes, hypertension, CHF, IHD, and history of myocardial infarction and stroke. The starting prevalence of other obesity comorbidities (history of cancer by type, osteoarthritis, OSA, NAFLD, GERD, and pulmonary embolism) among the population at the start of the initial year was unknown, but incidence of these comorbidities (as well as major depression episode and pneumonia) was simulated over the 10-year horizon.

Table 1. Simulation sample characteristics in initial year.

Patient characteristics	Obese sample	Matched normal weight sample
Mean (SD)
Age (years)	48.3 (15.6)	48.3 (16.1)
BMI (kg/m^2)	35.6 (5.6)	22.4 (2.0)
SBP (mmHg)	125.4 (17.0)	121.4 (18.4)
HDL-C (mg/dl)	47.0 (13.1)	61.0 (17.6)
T-C (mg/dl)	197.4 (41.0)	196.3 (39.3)
Hemoglobin A1c (%)	5.8 (1.1)	5.4 (0.7)
Prevalence (%)
Male	48.6	50.4
Pre-diabetes	50.4	40.9
Diabetes	20.7	5.1
Congestive heart failure	3.5	1.5
History of myocardial  infarction	20.7	2.8
History of stroke	3.4	2.3
Hypertension	50.7	30.9
Ischemic heart disease	7.6	4.3

Source: Constructed sample of 100,000 obese adults in the 2005–2012 National Health and Nutrition Examination Survey files and a demographically matched sample of 100,000 normal weight adults.

Simulated onset of diabetes within 10 years for these obese and normal weight cohorts was, respectively, 28.4% and 23.7% (Table 2). Prevalence of diabetes was 4-times higher among the obese population in the initial year, which limited the incidence of new diabetes cases among the obese population. Over 10 years, the obese population was projected to experience 4215 new cases of cancer per 100,000 population, while among the normal weight population this number was 3373. The 842 simulated excess cancer cases per 100,000 population over 10 years associated with obesity consisted of endometrial (+434 cases), breast (+271), kidney (+67), multiple myeloma (+48), liver (+43), non-Hodgkin’s lymphoma (+37), thyroid (+29), leukemia (+28), ovarian (+27), pancreas (+26), stomach (+23), esophagus (+21), colon & rectal (+7), and gall bladder (+6) cancer. Simulated cancer incidence was lower among the obese population for cervical (−8) and prostate (−217) cancer.

Table 2. Simulated outcomes for burden of obesity: Cumulative over 5 and 10 years.

Outcomes	All obese population		Normal weight population		Burden associated with obesity
	5-year	10-year	5-year	10-year	5-year	10-year
Cancer onset (per 100,000 population)	2065	4215	1683	3373	382	842
Disease onset (relative % vs normal)
Type 2 diabetes	15.2	28.4	10.9	23.7	4.2	4.7
Ischemic heart disease	4.3	10.2	2.9	7.0	1.5	3.2
Myocardial infarction	3.9	9.4	2.3	6.1	1.5	3.3
Congestive heart failure	9.5	22.7	7.2	17.4	2.3	5.3
Stroke	6.0	14.1	4.7	11.0	1.3	3.1
Obstructive sleep apnea	34.2	70.2	11.8	25.2	22.4	44.9
Major depression	39.3	80.6	21.9	44.4	17.4	36.2
Osteoarthritis	1.8	3.6	1.0	2.1	0.8	1.5
Non-alcoholic fatty liver disease	13.8	24.9	4.6	8.9	9.3	16.0
Gastroesophageal reflux disease	4.9	9.8	1.5	3.1	3.4	6.8
Gallbladder disease	2.6	5.5	1.2	2.5	1.5	3.0
Pneumonia	3.6	7.7	2.7	5.8	0.9	1.9
Pulmonary embolism	0.5	1.1	0.2	0.6	0.2	0.5
Economic outcomes ($ per capita)
Medical expenditures	70,800	159,500	50,600	116,700	20,200	42,800
Indirect economic metrics
Personal income	196,400	391,300	205,500	408,400	−9100	−17,100
Earnings from higher employment rate	119,100	231,800	122,000	238,600	−2900	−6800
Supplemental Security Income payments	1900	4100	1300	2800	600	1300
Loss due to sick days for employers	3600	7500	2500	5300	1100	2200
Total economic burden					33,900	70,200
QALYs/person ($ per capita)†	300,000	540,000	350,000	660,000	−50,000	−120,000

†Calculated using an incremental cost-effectiveness ratio (ICER) of $100,000 per QALY. Historically, $50,000 was often used as an ICER, but $100,000 is often used in recent literature.

The obese population had higher starting prevalence of cardiovascular disease and higher incidence of adverse cardiovascular outcomes over the projection horizon. For instance, the simulated 10 year total incidence rates of IHD, myocardial infarction, and CHF were, respectively, 3.2%, 3.3%, and 5.3% higher among the obese population. Total quality adjusted life years (QALY) averaged 1.2 years higher over 10 years for the normal weight population. Using a commonly accepted incremental cost-effectiveness ratio (ICER) of $100,000 per QALY66–68 (in general, a health technology is considered cost effective when ICER is between $50,000–$100,00069–73, but $100,000 has been increasingly advocated in recent studies74), this survival benefit is equivalent to $120,000 over 10 years, or $12,000 per year.

Total simulated medical costs over 10 years were $42 800 higher per person for the obese population relative to a matched normal weight population. In the initial year, the obese population averaged $3900 in higher medical expenditures and, by year 10, this difference was $4600. In the initial year, simulated excess medical expenditures were $2560 for the obese I population, $4550 for the obese II population, and $8040 for the obese III population (Figure 3). These findings suggest that medical costs rise more rapidly with aging for people with obesity compared with their normal weight peers, and costs among the obese III population are rising with age faster than the rise in costs among the obese II and obese I populations.

Figure 3. Simulated excess medical expenditures associated with obesity.

The 10-year results do not use a discount rate (reflecting projected medical expenditures rather than the present value of projected expenditures), but we show the sensitivity of results if a discount rate were applied. Excess medical expenditures over 10 years associated with obesity decline from $42,800 (undiscounted), to $37,400 (3% discount rate), to $34,400 (5% discount rate), with the discounted numbers showing the present value of excess future medical costs. These results are based on adults who are still living in each simulation year (i.e., survivors). If one looks at population medical expenditures over 10 years for the starting cohort, then the average difference in annual medical expenditures between the obese and normal weight populations is $3430 (vs $4280), reflecting higher mortality among the obese population.

Additionally, over 10 years the simulated indirect burden of obesity averaged $6800 per person in lost income from not participating in the labor force, $17,100 in lower earnings, despite participating in the labor force, an additional $1300 in Supplemental Security Income payments, and $2200 in lost productivity to employers associated with missed work days (Table 2). The average, total economic burden attributed to obesity, therefore, was projected to be $70,200 over 10 years ($7020 annually).

Sub-sets of the obese population modeled were obese I, obese II, obese III, obese aged 20–44, obese aged 45–64, and obese aged 65 and over. Each group was compared with a normal weight group with similar age, sex, race, and medical insurance status distributions. All outcomes presented in this section are simulated 10-year outcomes (Table 3). As expected, the obese III population had the highest projected excess medical costs attributed to obesity ($87,100), followed by the obese II ($50,100) and obese I ($28,200) populations. Other economic outcomes followed the trend of increasing costs as excess body weight increases, and clinical outcomes such as cancers and cardiovascular conditions followed this trend.

Table 3. Ten-year burden of obesity (relative to matched normal weight population).

	All obese	Obese I	Obese II	Obese III	Obese 20–44 years	Obese 45–64 years	Obese 65+ years
Cancer onset (per 100,000 population)	842	451	866	2 230	390	1 395	1 373
Disease onset (relative % vs normal)
Type 2 diabetes	4.7	5.2	6.4	4.5	12.4	0.1	−1.3
Ischemic heart disease	3.2	2.4	3.5	5.4	1.2	3.9	12.5
Myocardial infarction	3.3	3.3	3.9	4.1	1.7	4.2	11.3
Congestive heart failure	5.3	5.3	5.5	7.7	3.8	9.0	7.6
Stroke	3.1	2.0	2.6	5.3	2.8	3.4	4.9
Obstructive sleep apnea	44.9	37.3	50.4	64.4	39.4	51.8	49.3
Major depression	36.2	0.3	41.8	152.3	41.5	32.1	26.8
Osteoarthritis	1.5	1.8	1.8	0.7	2.1	1.2	0.8
Non-alcoholic fatty liver disease	16.0	13.0	20.5	20.0	15.1	15.8	19.1
Gastroesophageal reflux disease	6.8	5.3	8.5	8.9	5.5	7.7	8.7
Gallbladder disease	3.0	2.7	3.7	3.4	1.6	3.3	8.6
Pneumonia	1.9	0.8	3.1	3.7	0.6	2.5	7.4
Pulmonary embolism	0.5	0.4	0.6	0.8	0.2	0.3	2.8
Economic outcomes ($ per capita)
Medical expenditures	42,800	28,200	50,100	87,100	24,800	55,700	112, 600
Indirect economic metrics
Personal income	−17,100	−13,600	−20,000	−26,800	−25,300	−25,400	−5300
Earnings from higher employment rate	−6800	−1,100	−10,300	−22,900	−6800	−16,600	−2700
Supplemental Security Income payments	1300	800	1400	3000	800	2300	300
Loss due to sick days for employers	2200	2 100	2400	2 00	2300	2700	100
Total economic burden	70,200	45,800	84,200	142,600	60,000	102,700	121,000
QALYs ($ per capita)†	−120,000	−110,000	−120,000	−150,000	−130,000	−130,000	−100,000

Notes: Obese I is body mass index (BMI) 30.0–34.9; Obese II is BMI 35.0–39.9; Obese III is BMI ≥ 40.

†Calculated using an incremental cost effectiveness ratio of $100, 000 per QALY Historically, $50, 000 was often used as an ICER, but $100, 000 is often used in recent literature.

Average starting BMI in the initial year for the obese, obese I, obese II, and obese III populations were, respectively, 35.6, 32.2, 37.2, and 45.5. When considering the average excess weight of each population (relative to BMI of 25), model results suggest that each excess kilogram of body weight increased annual medical expenditures by $140, on average (Figure 4). The simulated average medical cost associated with one excess kilogram ranged from $136 for the obese I population to $152 for the obese III population.

Figure 4. Average annual medical costs per excess kilogram.

Similarly, 10-year burden of obesity increased with age. The obese population aged 65 and above was forecast to have $112,600 higher average medical expenditures relative to their normal weight peers of similar age. Few in this age group were in the workforce, so the economic burden of obesity for this population consisted primarily of excess medical expenditures. The obese aged 45–64 population was projected to be the most costly to employers, averaging $2700 loss due to more work days missed because of higher obesity-related illness. The obese population aged 65 and above was forecast to experience 3241 fewer new cases of diabetes per 100,000 people over 10 years relative to a matched normal weight population, but initial prevalence of diabetes among obese adults in this age group was already high, so there were fewer opportunities among the obese population to experience diabetes onset. Furthermore, higher projected mortality in the obese aged 65 and above population limited incidence of new disease relative to their normal weight peers.

Discussion

The obesity epidemic has raised considerable public health concerns, but there are few validated longitudinal simulation models examining the human and economic cost of obesity and how costs change with respect to aging, disease onset, and other changes in health risk factors. Use of a novel microsimulation tool in the current study highlights the tremendous cost associated with excess body weight. Specifically, our projections indicate that obesity is associated with $3900 higher per capita medical expenditures in the initial year, growing annually to reach $4600 higher expenditures in year 10. Cumulative over 10 years as the obese population ages, projected excess medical expenditures per person associated with obesity is $42,800. Simulated annual incidence of cancer, diabetes, cardiovascular conditions, stroke, and many other conditions are higher among the obese population relative to a normal weight population. Poor health associated with obesity includes higher work absenteeism and mortality, and lower employment probability, earnings, and quality-of-life.

Compared to their normal weight peers, our simulated average annual medical costs for obese adults over the subsequent 10 years were projected to be 37% higher. This percentage is close to a study based on the 2006 MEPS dataset3, from which it estimated the annual per capita medical spending for obese individuals was 41.5% higher than the healthy-weight individual. A similar 37% increase was reported in an earlier study by Thorpe et al.75 using the 2001 MEPS dataset. The initial year average excess cost for all obese adults ($3900) is similar to recently published estimates by Cawley et al.4 (average $3508 higher expenditures in year 2010 dollars [$3820 in year 2013 dollars] per obese adult) despite using quite different approaches.

An expert panel convened by American Heart Association, American College of Cardiology, and The Obesity Society identified the need for more information on clinical and economic outcomes associated with weight loss and whether outcomes differ by patient characteristics such as demographics or BMI76. This is an area of ongoing research using the microsimulation model, but understanding why medical costs are higher among an obese population compared with a normal weight population is a first step to understanding how to reverse the economic burden associated with obesity.

Strengths and limitations

While other studies have simply compared historical medical expenditures between obese and normal weight populations to quantify the burden of obesity, the strength of microsimulation is that it can help describe why people who are obese have higher medical expenditures and worse indirect economic outcomes via the impact of obesity on poor health. Models such as the one described here can be forward looking to project future disease onset among a population, and are highly flexible for sub-group analysis on people with specified characteristics. Modeling also allows one to look at economic outcomes from both a societal perspective, as well as the perspective of payers and employers.

Models are simplified representations of complex real-world systems, and, like all modeling studies, this study made assumptions and used imperfect data to inform complicated relationships. There are many potential sources of error in the model, but calculating confidence intervals for study findings is not possible because standard errors were unavailable for some published parameters used in the model and some model assumptions were subjective. Consequently, any standard errors derived would represent only a portion of any overall error.

Validation of the model followed ISPOR practice guidelines—including review by subject matter experts, internal and external quantitative evaluations, and sensitivity analysis. Even though various biases exist in EHR data, comparison of simulated longitudinal results to actual health outcomes were consistent with expectations. Overall, the validation exercises suggest a robust model, although the model sometimes over-predicts or under-predicts disease incidence for select demographic groups.

One limitation was the lack of a single longitudinal data source of the US population that covers a sufficient time period and is of sufficient size to quantify the relationships between disease onset and patient characteristics. Consequently, model parameters and equations came from multiple sources.

A second limitation is that older data sources were sometimes used (e.g., Framingham), and standards of care such as statin use have evolved over time. This may lead to a cohort effect that biases the risk estimation of certain health conditions. For example, data from the Look-AHEAD trial and other studies report that statin use has increased over time and is associated with decreased risk of adverse cardiovascular events77–79. The constructed population files used for simulation are based on recent data, so cholesterol levels and other health risk factors likely reflect characteristics of the current obese population. Another example is the definition of OSA. International Classification of Sleep Disorders published updated criteria on the diagnosis of OSA as AHI ≥ 580. Since AHI can be higher than 60 for the most severe OSA cases, the diagnosis criteria (AHI ≥ 10)43 used in the study will lead to slight under-estimation of the mildest cases of OSA.

A third limitation is that poor health can lead to reduced performance at work (presenteeism), with the economic toll of presenteeism possibly exceeding the cost to employers due to sick days (absenteeism). From this perspective, the model possibly under-estimates the negative economic impact of obesity from the perspectives of employers and society, because presenteeism is excluded from this analysis5. Another uncertainty around work-related economic estimations is there are often multiple confounding variables and interactions with other socioeconomic factors. Our approach adopted a simplified algorithm that did not account for all confounding factors.

The simulations used prediction equations for personal income to model how obesity (whether directly, or indirectly through higher disease incidence) affects earnings. The implications of obesity on family income (rather than personal income) might represent a more accurate portrayal of the indirect burden of obesity to society or families—as the labor force decisions of other family members can be influenced by the labor force decisions and earnings of the obese family member. Regression analysis using family income as the dependent variable found a much larger negative obesity impact relative to personal income, but endogeneity concerns lead to modeling using personal income. The implications of using personal income rather than family income are unclear and an area for future research. The drop in family income might exceed any drop in personal income if a family member reduced their labor force participation to help care for an obese family member with health problems. Alternatively, if the obese family member’s personal income declined, then other family members might increase their labor force participation to compensate for any loss in family income.

Conclusions

This paper describes development and validation of an innovative model that simulates disease onset and associated health and economic implications. The microsimulation approach used here offers a comprehensive and flexible framework that longitudinally projects onset of obesity-related comorbidities and their associated costs in various sub-populations, from individual, employer, and societal perspectives. Model results indicate that, among the obese population, each excess kilogram of body weight is associated with $140 increase in annual medical costs. Findings suggest that people who are obese currently have $3900 higher average annual medical expenditures relative to their normal weight peers, but over 10 years this difference grows to $4600, suggesting that medical costs rise more rapidly with aging for people with obesity compared with their normal weight peers.

Transparency

Declaration of funding

Funding of this research was provided by Novo Nordisk, Inc.

Declaration of financial/other relationships

JH and MM are employees of Novo Nordisk, Inc. The other study authors conducted paid consulting services to Novo Nordisk, Inc. for this and other research. The study sponsor and LP approved the analysis plan designed by WS, FC, IW, and TMD. Data collection, analysis, and manuscript writing were done by WS, FC, IW, and TMD. All authors contributed to interpretation of study findings. JME peer reviewers on this manuscript have no relevant financial or other relationships to disclose.