public - CT 3

A v a i l a b l e o n l i n e a t w w w . s c i e n c e d i r e c t . c o m

V A L U E I N H E A L T H 1 9 ( 2 0 1 6 ) 6 0 2 – 6 1 3

1098-3015$36.00 – s

Published by Elsevie

http://dx.doi.org/10

E-mail: ddkim62

* Address correspo NE Pacific Street, Bo

j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / j v a l

Systematic Review

Estimating the Medical Care Costs of Obesity in the United States: Systematic Review, Meta-Analysis, and Empirical Analysis David D. Kim, MS1,*, Anirban Basu, PhD1,2

1Department of Health Services, University of Washington, Seattle, WA, USA; 2The National Bureau of Economic Research, Cambridge, MA, USA

A B S T R A C T

Background: The prevalence of adult obesity exceeds 30% in the United States, posing a significant public health concern as well as a substantial financial burden. Although the impact of obesity on medical spending is undeniably significant, the estimated magnitude of the cost of obesity has varied considerably, perhaps driven by different study method- ologies. Objectives: To document variations in study design and meth- odology in existing literature and to understand the impact of those variations on the estimated costs of obesity. Methods: We conducted a systematic review of the twelve recently published articles that reported costs of obesity and performed a meta-analysis to generate a pooled estimate across those studies. Also, we performed an original analysis to understand the impact of different age groups, statistical models, and confounder adjustment on the magnitude of estimated costs using the nationally representative Medical Expenditure Panel Surveys from 2008- 2010. Results: We found significant variations among cost estimates in

ee front matter Copyright & 2016, International S

r Inc.

.1016/j.jval.2016.02.008

@uw.edu.

ndence to: David D. Kim, MS, Department of Health x 357660, Seattle, WA 98195.

the existing literature. The meta-analysis found that the annual medical spending attributable to an obese individual was $1901 ($1239-$2582) in 2014 USD, accounting for $149.4 billion at the national level. The two most significant drivers of variability in the cost estimates were age groups and adjustment for obesity- related comorbid conditions. Conclusions: It would be important to acknowledge variations in the magnitude of the medical cost of obesity driven by different study design and methodology. Research- ers and policy-makers need to be cautious on determining appropriate cost estimates according to their scientific and political questions. Keywords: economic burden, medical care costs, obesity, United States.

Copyright & 2016, International Society for Pharmacoeconomics and Outcomes Research (ISPOR). Published by Elsevier Inc.

Introduction

The prevalence of obesity, which is defined as a body mass index (BMI) of greater than 30, has increased dramatically in the United States since the late 1990s [1]. So much so that recently obesity has been officially recognized as a disease by the American Medical Association, an action that could put more emphasis on the health condition by doctors and insurance companies so as to minimize its adverse effects. Currently, rates of obesity exceed 30% in most sex and adult age groups, whereas its prevalence among children and adolescents, defined as a BMI of more than 95th percentile, has reached 17% [2].

The alarming rates of the high prevalence of obesity have posed a significant public health concern as well as a substantial financial burden on our society because obesity is known to be a risk factor for many chronic diseases, such as type 2 diabetes, cancer, hypertension, asthma, myocardial infarction, stroke, and

other conditions�[3,4]. To understand the economic burden of obesity, several studies have attempted to estimate the attribut- able costs of obesity, following the burden-of-illness literature on other disease areas [5–9]. A previous cost-of-illness study esti- mated that health care spending attributable to the rising prevalence of obesity has increased by 27% between 1987 and 2001 [10]. In gross terms, the annual medical costs of obesity were estimated to be $40 billion in 2006 [11]. The latest study using an instrumental variable (IV) approach even showed that the esti- mated medical costs related to obesity could reach $209.7 billion, which is twice higher than the previous estimate, $86 billion [12].

As evidenced by the aforementioned estimates, although the impact of obesity on the medical care spending is undeniably significant, the estimated magnitude of the medical care costs attributable to obesity has varied considerably, perhaps driven by different study methodologies, including data, statistical models, confounder adjustment, and target populations. In this article,

ociety for Pharmacoeconomics and Outcomes Research (ISPOR).

Services, School of Public Health, University of Washington, 1959

V A L U E I N H E A L T H 1 9 ( 2 0 1 6 ) 6 0 2 – 6 1 3 603

we approach these issues systematically with two goals: 1) to conduct a systematic review and meta-analysis of recently published articles that estimated the medical costs associated with obesity between 2008 and 2012 and to document the variations in study methodologies and 2) to demonstrate the importance of study methodologies by performing an original analysis to examine the impact of age group, confounder adjust- ment, and statistical methods on the cost estimates of obesity through the empirical analysis of a nationally representative US population. Especially, we also examined the impact of obesity- related diseases (ORDs) on the medical costs of obesity to show that most, if not all, of those costs are attributable to ORDs.

We believe that it would be important to recognize significant variations among estimates of obesity-attributable costs in the existing literature and to understand the impact of study meth- odology on the magnitude of these estimates so that researchers and policymakers are able to determine the appropriate estimate and methods according to their scientific and political questions.

Methods

A Systematic Review and Meta-Analysis

Literature search We searched the MEDLINE and Cochran database to identify articles related to medical costs of obesity using keywords “obesity AND (cost OR expenditure) AND healthcare)) AND “united states.” To account for the unique health care system and the impact of costs attributable to obesity in the United States, we limited the search to studies conducted in the US settings. We initially identified 567 articles from the search, then narrowed down to 16 articles for in-depth reviews. Following the extensive reviews, we excluded three studies that did not provide explicit methods and/or aggregate annual costs per person, in addition to a previously conducted systematic review [13–16]. Finally, we included 12 studies in this study for the systematic review [17–26]. Appendix Figure A in Supplemental Materials found at http://dx.dor.org/10.1016/j.jval.2016.02.008 provides details on search strategies for identifying studies included in this review.

Improve comparability across studies To improve comparability across heterogeneous studies, we performed appropriate adjustments to convert estimates from each study into annual per-person costs among all obese pop- ulation (BMI Z 30).

First, we converted cost estimates to 2014 USD to adjust for the inflation over time using annual average consumer price index for medical care [27]. One study reported the quarter-per- person medical costs, and we annualized the cost estimate [17]. All the 12 studies reported direct medical costs, including the out- of-pocket costs for inpatient, noninpatient (outpatient, emer- gency room, and other), and prescription drug spending.

Then, we aggregated all BMI-specific estimates into a single composite estimate of costs attributable to all obese individuals. Among the 12 studies, 8 studies defined obesity as a BMI of greater than 30 whereas 4 studies implemented more comprehensive obesity categories, defined as class I obesity (30 o BMI r 35), class II obesity (35 o BMI r 40), and class III obesity (40 o BMI) [21,22,24,26]. Two of the four studies combined class II and class III obesity into a single category because of the sample size issue [22,26]. To generate comparable cost estimates, we calculated a weighted average among subgroup-specific estimates on the basis of the number of each subgroup reported in each of the four studies.

In addition, three studies estimated sex-specific costs of obesity [18,20,23], and one study provided race (non-Hispanic

whites vs. blacks) stratified results [26]. Another study reported both sex and race (non-Hispanic whites vs. blacks) stratified estimates [19]. Based on the sample size of each stratum presented in each study, only the weighted average estimates for aggregating sex and race categories are presented in Table 2.

Evaluating quality of studies We evaluated the quality of studies on the basis of four criteria: the use of nationally representative samples, longitudinal data sets, analysis of adults of all ages, and appropriate confounding factor adjustments. A previous systematic review also used a similar set of criteria for evaluating cost-of-illness studies of obesity [13].

Meta-analysis To generate a pooled estimate of medical costs of obesity across different studies, we conducted a meta-analysis using the metaan command in STATA 12 (StataCorp., College Station, TX) [28]. The metaan command is used to conduct random-effect meta-anal- ysis for one-variable relationship. Because the meta-analysis for one-variable relationship requires both the effect size estimate and the standard error, we were able to include only eight estimates of annual incremental costs of obesity from seven studies (Table 2). Because of the presence of extremely high heterogeneity between studies (I2 ¼ 96.61%; τ2 ¼ 5.6 � 105), the random-effect model is used in the final analysis.

Empirical Analysis: The Role of Alternative Statistical Models in Estimating Costs of Obesity

Study data The medical costs of obesity were estimated using regression analysis and the 2008-2010 Medical Expenditure Panel Surveys (MEPS). The MEPS is a nationally representative survey of the civilian noninstitutionalized population, collecting detailed information on health care expenditures and utilization, health insurance, health status, and sociodemographic factors. Nationally representative estimates were obtained by using MEPS sampling weights.

Variables As a dependent variable, medical care costs (which include costs for office-based visits, hospital outpatient visits, emergency room visits, inpatient hospital stays, prescription drugs, dental visits, and home care) are defined as the sum of direct payments from all parties (out-of-pocket, private insurers, government, and other payers) for care provided during the year. For a primary inde- pendent variable, we identified obesity status on the basis of the constructed BMI through self-reported height and measure [29]. (Please note that because of confidentiality concerns and restric- tions, the self-reported weight and height variables were not available from the public-access MEPS data sets.) Also, we categorized potential confounding factors into four categories to examine the impact of confounder adjustments on the magni- tude of the cost estimates: 1) Demographic factors or cov1 (age, sex, and race/ethnicity), 2) Socioeconomic factors or cov2 (edu- cation, household income based on the federal poverty line, smoking status, and marital status), 3) Additional factors or cov3 (census region and insurance status), and 4) comorbidity conditions or cov4. Comorbidity conditions are defined as a con- tinuous variable ranging from 0 to 10 by summing up 10 potential health consequences that can be caused be obesity. These conditions, called ORDs, which are defined by the Centers for Disease Control and Prevention, include hypertension, heart diseases (coronary heart disease, angina, myocardial infarction, others), stroke, cancer, diabetes, arthritis, and high cholesterol [30]. In this data set, children or adolescents (age o 18 years) do not have any information on comorbidity conditions and

Table 1 – Characteristics of 12 studies included in this review.

Study Data Sample size

Statistical methods*

Obesity class

Target population

Variable adjusted for Quality evaluation (score out of 4)

Medical care cost of obesity—Annual cost per person (common confounders†) Wolf et al. [17],

2008 US PROCEED 1,300 Log-linear Aggregate Adults (aged

35–75 y) (� race/ethnicity, income, marital status)

(þ alcohol use, comorbidities, insurance) 2: Longitudinal data, confounder

adjustment Finkelstein

et al. [11], 2009

MEPS (2006) 21,877 Two-part (logit-GLM)

Aggregate All adults (age Z 18 y)

(þ census region, insurance status) 3: Nationally representative sample, all adults, confounder adjustment

Cai et al. [20], 2010

MCBS (1991– 2000)

5,043 Unadjusted Aggregate Adults aged 35– 55 y in the period 1971– 1975

Unadjusted 1: Nationally representative sample

Finkelstein et al. [21], 2010

MEPS (2006) 8,875 Two-part (logit-GLM)

Obesity class I, II, and III

All adults (age Z 18 y) Full- time employees only

(þ age2, census region, insurance status) 3: Nationally representative sample, all adults, confounder adjustment

Bell et al. [23], 2011

MEPS (2000– 2005)

80,516 Two-part (logit-log (Y) OLS)

Aggregate Children and adults aged 6–85 y

(� smoking, marital status) (þ age2, age3, region, insurance status, survey year)

3: Nationally representative sample, all adults, confounder adjustment

Onwudiwe et al. [22], 2011

MCBS (2002) 7,706 One-part GLM

Obesity class I and II/ III

Medicare beneficiaries (age Z 65 y) Not in HMO plan

(þ insurance status) 2: Nationally representative sample, confounder adjustment

Alley et al. [25], 2012

MCBS (1997– 2006)

29,413 One-part GLM

Aggregate Medicare beneficiaries (age Z 65 y)

(þ census region, metropolitan status, mortality variable)

3: Nationally representative sample, all adults, confounder adjustment

Cawley and Meyerhoefer [12], 2012

MEPS (2000– 2005)

23,689 IV with two- part (logit- GLM) IV: a weight of biological relative

Aggregate Adults (aged 20–64 y with biological children aged 11–20 y)

(� income, smoking, marital status) (þ census region, MSA, household composition, survey information, employment status, fixed effects for year, the sex and age of the oldest children)

3: Nationally representative sample, confounder adjustment (IV)

Ma et al. [26], 2012

MEPS (2006) 15,164 Unadjusted Obesity class I and II/ III

All adults (age Z 18 y)

Unadjusted 2: Nationally representative sample, all adults

Moriarty et al. [24], 2012

Mayo Clinic Database (2001– 2007)

30,529 GEE Obesity class I, II, and III

Adults (18–65 y) vs. adults (465 y)

(� education, income) (þ comorbidity conditions for an additional analysis)

3: Longitudinal data, all adults, confounder adjustment

V A L U E

I N

H E A L T H

1 9

( 2 0 1 6 ) 6 0 2 – 6 1 3

6 0 4

M ed

ic a l ca re

co st

of ob

es it y — Li fe ti m e co st

p er

p er so n (a ll a d ju st ed

fo r su

rv iv a l)

F in k e ls te in

e t a l. [1 9 ],

2 0 0 8

M E P S (2 0 0 1 –

2 0 0 4 )

6 6 ,1 6 1

T w o -p

a rt

m o d e l

(l o g it -G

L M )

O b e si ty

cl a ss

I a n d II /

II I

A ll a d u lt s (a g e

Z 1 8 y )

(� in co

m e ) (þ

in su

ra n ce

st a tu

s, ce

n su

s re g io n ,

p o p u la ti o n

d e n si ty , a g e 2 , a g e 3 )

3 : N a ti o n a ll y re p re se

n ta ti v e

sa m p le , a ll a d u lt s,

co n fo u n d e r

a d ju st m e n t

Y a n g a n d H a ll

[1 8 ], 2 0 0 8

M C B S (1 9 9 2 –

2 0 0 1 )

2 8 ,9 0 6

T w o -p

a rt

m o d e l

(l o g it -l o g

(Y ) O L S )

A g g re g a te

M e d ic a re

b e n e fi ci a ri e s

(a g e Z

6 5 y )

(þ fu

n ct io n a l st a tu

s, e x is ti n g ch

ro n ic

d is e a se

s, a cu

te m e d ic a l e v e n ts , u rb a n /r u ra l)

2 : N a ti o n a ll y re p re se

n ta ti v e ,

co n fo u n d e r a d ju st m e n t

C a i e t a l. [2 0 ],

2 0 1 0

M C B S (1 9 9 1 –

2 0 0 0 )

5 ,0 4 3

O n e -p

a rt

G L M

A g g re g a te

A d u lt s a g e d 3 5 –

5 5 y in

th e

p e ri o d 1 9 7 1 –

1 9 7 5

(� e d u ca

ti o n , in co

m e , m a ri ta l st a tu

s) (þ

ti m e

to d e a th

) 2 : N a ti o n a ll y re p re se

n ta ti v e ,

co n fo u n d e r a d ju st m e n t

G E E , g e n e ra li z e d

e st im

a ti n g e q u a ti o n ; G L M , g e n e ra li z e d

li n e a r m o d e l; H M O , h e a lt h

m a in te n a n ce

o rg a n iz a ti o n ; IV

, in st ru

m e n ta l v a ri a b le ; M C B S , M e d ic a re

C u rr e n t B e n e fi ci a ry

S u rv e y ; M E P S ,

M e d ic a l E x p e n d it u re

P a n e l S u rv e y s;

M S A , M e tr o p o li ta n

st a ti st ic a l a re a ; O L S , o rd

in a ry

le a st

sq u a re s;

P R O C E E D , P ro

sp e ct iv e O b e si ty

C o h o rt

o f E co

n o m ic

E v a lu a ti o n

a n d D e te rm

in a n ts .

* A ll G L M

a n d G E E m o d e ls

u se

d th

e lo g li n k a n d th

e g a m m a d is tr ib u ti o n .

† C o m m o n

co n fo u n d in g co

v a ri a te s:

A g e , se

x , ra ce

/e th

n ic it y , e d u ca

ti o n , in co

m e , sm

o k in g st a tu

s, a n d m a ri ta l st a tu

s.

V A L U E I N H E A L T H 1 9 ( 2 0 1 6 ) 6 0 2 – 6 1 3 605

smoking status, and only very few individuals reported education (N ¼ 25) and marital status (N ¼ 6).

Study design and data analysis To study the impact of various study designs, we examined three different aspects of estimating medical costs in a regression analysis: age groups, statistical models, and confounder adjust- ments. The age groups were 1) children/adolescents (aged 0–17 years), 2) all adults (aged Z 18 years), 3) adults aged 18 to 65 years, and 4) older adults (aged Z65 years). Considering the nature of cost data, such as non-negative observations, a large number of observations at zero, and positive skewness, we examined five different statistical models that have been widely used to estimate the medical costs: 1) linear regression, 2) log- linear model (a simple ordinary least square for ln(y)), 3) one-part log-gamma generalized linear model (GLM), 4) two-part model with a logistic regression and a log-gamma GLM, and 5) extended estimating equation (EEE) that used both a flexible link and a flexible variance function estimated directly from the data to capture the underlying nonlinearity in the data so that it can produce efficient estimates [31–34]. Also, we tested the goodness of fit (GoF) of each statistical model to examine how well the model fits a data set using Pearson correlation, Pregobon’s link test, and Hosmer-Lemeshow test. In regard to confounder adjust- ment, four sets of confounders were studied: 1) Demographic factors (cov1), 2) Demographic þ Socioeconomic factors (cov1 þ cov2), 3) Demographic þ Socioeconomic þ Additional factors (cov1 þ cov2 þ cov3), and 4) All three factors þ ORDs (cov1 þ cov2 þ cov3 þ cov4). Using different combinations of the target population, statistical models, and confounder adjustment, we estimated medical care costs of nonobese individuals (so called costs of normal) as well as incremental costs of obesity through recycled predictions to estimate the counterfactual mean costs if all individuals in the data set were suddenly to have obesity while retained all other characteristics as compared with being nonobese for all individuals. We also tested the GoF of five different statistical models with each of four different sets of confounding factors for the sample population of all adults. All standard errors and confidence intervals (CIs) were estimated from 1000 bootstrap replicates.

Results

A Systematic Review and Meta-Analysis

Descriptive results Among the 12 studies included in this systematic review, 9 studies reported annual medical care costs per person while 2 studies provided lifetime medical care costs per person and 1 study reported both estimates. Six studies used the MEPS data- base, whereas four studies used the Medicare Current Beneficiary Survey database. The remaining two studies used the Prospective Obesity Cohort of Economic Evaluation and Determinants data set, which is a multinational, observational, prospective Internet- based cohort study and the Mayo clinic employment database, respectively (Table 1).

Quality evaluation Based on four quality criteria, no studies met all criteria, because all the studies that used a nationally representative data set were not a longitudinal study or vice versa. Among 10 studies that reported annual costs per person, 5 studies were designated as “high-quality (score ¼ 3)” study (Table 1). Because Cawley’s study was the only study that estimated medical costs of obesity using the instrumental variable IV approach to account for unobserved

Table 2 – Systematic review—Medical costs of obesity (2014 USD).

Study Cost of normal*

Cost of obesity*

Incremental cost of obesity*

Year of cost reporting

Note

Medical care cost of obesity—Annual cost per person Wolf et al. [17], 2008 $2,541 $6,611 $4,070† 2004 Finkelstein et al. [11],

2009 $4,087 $5,783 $1,696† 2008

Cai et al. [20], 2010 $5,750 $13,019 $7,269 2000 Unadjusted Finkelstein et al. [21],

2010 NA NA $1,024 2006 Obese I, medical costs only

$1,944 Obese II, medical costs only $2,215 Obese III, medical costs only $1,397† Aggregate, medical costs only

Bell et al. [23], 2011 $3,629 $5,488 $1,859† 2005 Onwudiwe et al. [22],

2011 $5,666 $5,578 �$88 2002 Obese I, uncorrected for the height

loss $6,637 $971 Obese II/III, uncorrected for the height

loss $5,892 $227† Aggregate obese, uncorrected for the

height loss Alley et al. [25], 2012 $8,781 $7,338 �$1,443 2006 Annual spending in 1997 converted to

2008 USD $157 $191 $34 Expenditures increased per year from

1997 to 2006 Cawley and

Meyerhoefer [12], 2012 NA NA $877† 2005 Without using the instrumental

variable $3,665† With using the instrumental variable

Ma et al. [26], 2012 $4,797 $6,152 $1,356 2006 Obese I, unadjusted $8,408 $3,611 Obese II/III, unadjusted $7,082 $2,285 Aggregate, unadjusted

Moriarty et al. [24], 2012 NA NA $2,278 2007 Obese I, no comorbidity adjustment $3,759 Obese II, no comorbidity adjustment $6,794 Obese III, no comorbidity adjustment $3,302† Aggregate obese, no comorbidity

adjustment Medical care cost of obesity—Lifetime costs per person (all adjusted for survival)

Finkelstein et al. [19], 2008

NA NA $19,892 2007 From age 20 y, obese I—Discounted lifetime costs

$28,441 From age 20 y, obese II/III—Discounted lifetime costs

$23,123 From age 20 y, aggregate—Discounted lifetime costs

$15,641 From age 65 y, obese I—Discounted lifetime costs

$24,589 From age 65 y, obese II/III—Discounted lifetime costs

$19,022 From age 65 y, aggregate—Discounted lifetime costs

Yang and Hall [18], 2008 $288,934 $332,838 $43,904 2001 From age 65 y—Not discounted Cai et al. [20], 2010 $194,013 $269,628 $75,615 2000 From age 45 y—Not discounted

NA, not applicable/available. * All costs were converted to 2014 US dollar using Consumer Price Index—Medical Care. † Represents the cost estimates included in the meta-analysis to calculate the pooled incremental cost of obesity.

V A L U E I N H E A L T H 1 9 ( 2 0 1 6 ) 6 0 2 – 6 1 3606

confounding factors, the study received an extra score for con- founder adjustment criteria [12].

Cost estimates The annual incremental costs of obesity per person ranged from $227 to $7269 depending on the study designs and research methods. Tables 2 and 3 provide details on different study methodologies and reported cost estimates. From the meta- analysis with the random-effect model, the pooled estimate of

annual medical costs of obesity was $1910 (95% CI $1239–$2582). A forest plot from the random-effect model is shown in Figure 1.

Three studies reported the lifetime costs associated with obesity after adjusting for survival. One study estimated the lifetime costs of obesity with a 3% annual discount on the dollar value, resulting in $23,123 from age 20 years and $19,022 from age 65 years [19], whereas another study without using a discount rate reported the lifetime costs of $43,904 from age 65 years [18].

Table 3 – Systematic review—Medical costs of obesity by study methodologies (2014 USD).

Confounding factors Unadjusted Demographic factors

Demographic þ socioeconomic factors

Demographic þ SES þ additional factors

All factors þ comorbidities

Statistical methods

Age groups

Linear regression Children/adolescents All adults (age Z 18 y) $2,285 (Ma et al. [26], 2012)

Adults (18–65 y) $7,269 (Cai et al. [20], 2010) Older adults (age Z 65 y)

Log-linear Children/adolescents All adults (age Z 18 y)

Adults (18–65 y) $4,070 (Wolf, 2008) Older adults (age Z 65 y)

One-part GLM Children/adolescents All adults (age Z 18 y)

Adults (18–65 y) Older adults (age Z 65 y) $227 (Onwudiwe, 2011)

�$1,443 (Alley, 2012) Two-part GLM Children/adolescents

All adults (age Z 18 y) $1,696 (Finkelstein, 2009) $1,359 (Finkelsten, 2010)

$1,859 (Bell, 2011) Adults (18–65 y) $877 (Cawley, 2012) Older adults (age Z 65 y)

Other methods Children/adolescents All adults (age Z 18 y) $3,302 (Moriarty, 2012)—

GEE Adults (18–65 y) $3,665 (Cawley, 2012)—IV Older adults (age Z 65 y)

V A L U E

I N

H E A L T H

1 9

( 2 0 1 6 ) 6 0 2 – 6 1 3

6 0 7

Fig. 1 – Meta-analysis: Medical costs of obesity (random-effect model). CI, confidence interval; IV, instrumental variable.

V A L U E I N H E A L T H 1 9 ( 2 0 1 6 ) 6 0 2 – 6 1 3608

Stratified results Among the studies that provided stratified estimates of the medical costs attributable to obesity, the annual costs increased in higher obese categories [21,22,24,26] and the lifetime costs of obesity were also positively associated with an increasing BMI [19]. The magnitude of the annual costs related to obesity was higher among women than among men [18–20,23]. Also, obese blacks were found to spend less medical costs than obese whites, mainly due to more use of relatively inexpensive types of care (office-based visits, outpatient care, medications) rather than more costly ones (inpatient, emergency room) [19,26].

Different study methodology Among the 10 studies that estimated annual medical care cost attributable to obesity, a two-part model was the most popular method used by four studies. Three of those four studies imple- mented a logistic regression with a log-gamma GLM [11,12,21], whereas another study used a logit model with a log-linear model [23]. Two studies used a one-part GLM method with a log link and a gamma distribution [22,25]. A generalized estimating equation [24] and a log-linear regression model [17] were used in other two studies, respectively, and the remaining two studies reported unadjusted annual costs of obesity [20,26]. Because each study used distinctive data sets to various study methodologies, it is hard to predict the actual impact of a range of study methodologies on the magnitude of the costs of obesity. However, one study using both a two-part model and an IV approach provided that the effect of obesity on medical care costs was much greater with the IV method than was previously appreciated with the two-part model. The instrument used in the study was the weight of a biological relative [12].

Different confounder adjustment We defined confounding factors that were widely used to adjust for the causal inference among these studies, including demo- graphic factors, socioeconomic factors, additional factors, and obesity-related comorbidity conditions. However, none of these studies selected the same set of confounding factors. Table 1

presents a significant variability in choosing confounding factors, and the substantial variability poses an essential problem of comparing these estimates directly. A study that examined two models with and without adjusting for comorbidity conditions as a confounding factor found that incremental costs of obesity dropped significantly when adjusted for comorbidities [24].

Estimating Medical Costs of Obesity

With possible combinations of the four age groups, the five statistical models, and the four sets of potential confounders, 80 estimates of costs of normal and incremental costs of obesity were generated with 95% CIs. All possible cost estimates are provided in Table 4.

Characteristics of individuals in the data set Individual characteristics used in estimating medical costs of obesity are presented in Appendix Table A in Supplemental Materials found at http://dx.dor.org/10.1016/j.jval.2016.02.008. In this analysis, the data set includes 15,176 children and adoles- cents and 69,382 adults aged 18 years and older, including 10,382 older adults (age Z 65 years) with existing obesity status. Among the obese, there was a significantly higher proportion of females, blacks, Hispanics, individuals in lower household income level (o125% federal poverty line), only high school or equivalent degree holder, those who were married, individuals with public insurance, and those living in the South region, compared with the nonobese. Also, as expected, individuals with obesity have a higher number of obesity-related comorbidity conditions than do those without obesity.

Effect of different target populations For children/adolescents, regardless of different statistical mod- els and confounding factor adjustment, there was no significant difference between costs of the nonobese and the obese, except only one scenario with a one-part GLM controlling for all con- founding factors available (no comorbidities and smoking) that reported $1085 ($92–$2377) for the incremental costs of obesity for children/adolescents. However, for the adult population, the

Table 4 – Factors affecting costs of normal (nonobese) and incremental costs of obesity.

Confounding factors Unadjusted Demographic factors1 Demographic þ socioeconomic factors2

Demographic þ SES þ additional factors3

All factors þ comorbidities4

Statistical methods

Age groups Cost of normal

Incremental cost of obesity

95% CI Cost of normal

Incremental cost of obesity

95% CI Cost of normal

Incremental cost of obesity

95% CI Cost of normal

Incremental cost of obesity

95% CI Cost of normal

Incremental cost of obesity

95% CI

Linear regression

Children/ adolescents

1,851 �62 (�455 to 333)

1,265 182 (�198 to 563)

1,591 621 (�285 to

1528)

1,580 647 (�269 to

1563)

NA NA NA

All adults (age Z 18 y)

4,574 1,766 (1477 to

2056)

4,252 1,411 (1096 to

1689)

4,436 1,453 (1139 to

1738)

4,443 1,406 (1101 to

1700)

4,792 153 (�156 to 456)

Adults (18–65 y)

3,524 1,795 (1502 to

2088)

3,345 1,283 (995 to 1570)

3,494 1,312 (1006 to

1617)

3,559 1,153 (851 to 1456)

3,855 93 (�211 to 397)

Older adults (age Z 65 y)

9,558 2,290 (1385 to

3195)

9,247 2,898 (1975 to

3820)

9,380 2,747 (1794 to

3701)

9,362 2,717 (1762 to

3673)

9,954 496 (�421 to

1413) Log-linear Children/

adolescents 1,975 90 (�413

to 826) 1,640 366 (�178

to 1186)

2,145 1,112 (�67 to 2533)

2,074 1,191 (�36 to 2652)

NA NA NA

All adults (age Z 18 y)

5,186 1,562 (1318 to

1823)

5,394 1,375 (1089 to

1705)

5,546 1,412 (1101 to

1753)

5,452 1,261 (950 to 1561)

6,236 61 (�250 to 358)

Adults (18–65 y)

4,175 1,614 (1372 to

1858)

4,298 1,015 (759 to 1294)

4,407 1,031 (764 to 1312)

4,321 819 (544 to 1086)

4,935 �78 (�347 to 197)

Older adults (age Z 65 y)

9,468 2,283 (1522 to

3012)

9,439 3,528 (2556 to

4456)

9,602 3,431 (2461 to

4371)

9,654 3,378 (2425 to

4354)

10,849 425 (�435 to

1273) One-part

GLM Children/

adolescents 1,851 �62 (�396

to 349) 1,376 140 (�140

to 532) 1,571 715 (�240

to 1567)

1,556 1,085 (92 to 2377)

NA NA NA

All adults (age Z 18 y)

4,574 1,766 (1462 to

2056)

4,279 1,450 (1139 to1720)

4,446 1,506 (1210 to

1822)

4,511 1,397 (1091 to

1665)

5,048 429 (103 to 728)

Adults (18–65 y)

3,524 1,795 (1509 to

2089)

3,383 1,088 (843 to 1336)

3,517 1,143 (890 to 1407)

3,580 970 (702 to 1216)

4,115 258 (�14 to 535)

Older adults (age Z 65 y)

9,558 2,290 (1415 to

3193)

9,235 3,124 (2090 to

4046)

9,358 2,992 (1918 to

3977)

9,342 2,944 (1858 to

3859)

9,981 600 (�272 to

1496) Two-part

GLM Children/

adolescents 2,162 �62 (�359

to 455) 1,953 140 (�146

to 653) 2,063 475 (�175

to 1289)

2,066 750 (�2 to 1894)

NA NA NA

All adults (age Z 18 y)

5,483 1,834 (1553 to

2170)

5,358 1,524 (1247 to

1803)

5,498 1,578 (1284 to

1848)

5,535 1,481 (1186 to

1754)

5,949 399 (112 to 665)

Adults (18–65 y)

4,369 1,880 (1581 to

2206)

4,391 1,190 (938 to 1428)

4,528 1,234 (957 to 1467)

4,579 1,070 (813 to 1313)

4,974 269 (8 to 506)

Older adults (age Z 65 y)

9,894 2,328 (1481 to

3280)

9,628 3,191 (2264 to

4187)

9,749 3,060 (2040 to

4017)

9,740 3,014 (1991 to

3993)

10,315 555 (�320 to

1451) continued on next page

V A L U E

I N

H E A L T H

1 9

( 2 0 1 6 ) 6 0 2 – 6 1 3

6 0 9

T a b le

4 – co

n ti n u ed

C o n fo u n d in g fa ct o rs

U n a d ju st e d

D e m o g ra p h ic

fa ct o rs

1 D e m o g ra p h ic

þ so

ci o e co

n o m ic

fa ct o rs

2 D e m o g ra p h ic

þ S E S þ

a d d it io n a l fa ct o rs

3 A ll fa ct o rs

þ co

m o rb

id it ie s4

E E E

C h il d re n /

a d o le sc

e n ts

1 ,8 5 1

� 6 2

(� 3 7 6

to 3 8 9 )

N A

N A

N A

1 ,5 9 3

6 5 3

(� 3 3 to

1 5 5 6 )

N A

N A

N A

N A

N A

N A

A ll a d u lt s (a g e

Z 1 8 y )

4 ,5 7 4

1 ,7 6 6

(1 4 4 5

to 2 0 4 2 )

4 ,2 5 0

1 ,3 4 6

(1 0 6 5

to 1 6 0 4 )

4 ,4 4 1

1 ,3 5 6

(1 0 7 6

to 1 6 4 4 )

4 ,4 1 4

1 ,3 4 3

(1 0 7 6

to 1 6 2 1 )

4 ,7 6 4

2 0 9

(� 2 1 to

4 3 4 )

A d u lt s (1 8 – 6 5

y )

3 ,5 2 4

1 ,7 9 5

(1 5 0 9

to 2 0 8 9 )

3 ,3 5 8

1 ,1 3 5

(9 0 4 to

1 3 8 3 )

3 ,5 0 5

1 ,1 7 0

(9 2 7 to

1 4 4 5 )

3 ,4 9 2

1 ,0 9 4

(8 5 9 to

1 3 5 9 )

3 ,8 0 4

1 8 4

(� 4 1 to

3 9 8 )

O ld e r a d u lt s

(a g e Z

6 5 y )

9 ,5 5 8

2 ,2 9 0

(1 4 1 5

to 3 1 9 3 )

9 ,2 4 6

2 ,8 4 4

(1 8 1 9

to 3 7 7 6 )

9 ,3 9 7

2 ,6 9 9

(1 4 2 7

to 3 5 2 9 )

9 ,4 4 7

2 ,6 9 8

(9 7 8 to

4 4 1 8 )

9 ,9 7 3

3 4 4

(� 3 8 7

to 1 2 3 1 )

C I, co

n fi d e n ce

in te rv a l; E E E , e x te n d e d e st im

a ti n g e q u a ti o n ; G E E , g e n e ra li z e d e st im

a ti n g e q u a ti o n ; G L M , g e n e ra li z e d li n e a r m o d e l; IV

, in st ru

m e n ta l v a ri a b le ; N A , n o t a p p li ca

b le /a v a il a b le .

1 . D e m o g ra p h ic

fa ct o rs : A g e , se

x , a n d

ra ce

/e th

n ic it y ; 2 . S o ci o e co

n o m

ic fa ct o rs : E d u ca

ti o n , h o u se

h o ld

in co

m e , sm

o k in g st a tu

s, a n d

m a ri ta l st a tu

s; 3 . A d d it io n a l fa ct o rs : C e n su

s re g io n

a n d

in su

ra n ce

st a tu

s; 4 . C o m o rb id it ie s:

T h e p re se

n ce

o f 1 0 o b e si ty -r e la te d d is e a se

s (0 – 1 0 )

V A L U E I N H E A L T H 1 9 ( 2 0 1 6 ) 6 0 2 – 6 1 3610

incremental costs of obesity were significantly higher than for the nonobese. Among adults, in most of the combinations with statistical models and confounding factors, the incremental costs of obesity for older adults (age Z 65 years) were significantly higher than those for adults aged 18 to 65 years (Fig. 2 and Table 4). From the EEE model controlling for demographic, socioeconomic, and additional factors (cov1 þ cov2 þ cov3), the incremental costs of obesity for adults aged 18 to 65 years were reported as $1094 ($859–$1359) while the costs of obesity for older adults were $2668 ($978–$4418).

Effect of different statistical models Medical care costs attributable to obesity did not differ signifi- cantly by using different statistical models, although the GoF tests showed that an EEE model fitted the data most thoroughly, followed by a one-part GLM. For all adults, controlling for demographic and socioeconomic and additional factors (cov1 þ cov2 þ cov3), the EEE model reported $1343 ($1076–$1621) for the incremental costs of obesity, whereas the estimates ranged from $1261 ($950–$1561) in the log-linear model to $1481 ($1186–$1754) in the two-part GLM (Fig. 2; Table 4). Also, compared with other models, the EEE model provided the most stable estimates over the different sets of confounding factor adjustment, varying only from $1346 with cov1 to $1356 with cov1 þ cov2 to $1343 with cov1 þ cov2 þ cov3.

Confounding factor adjustment For children, regardless of statistical models, the point estimates of the incremental costs of obesity were increased as we adjusted with more sets of possible confounding factors (up to cov1 þ cov2 þ cov3), despite a huge CI that made those estimates statistically insignificant. For adults, controlling for demographic, socioeco- nomic, and additional confounding factors (cov1 þ cov2 þ cov3) did not make any substantial impact on point estimates as well as statistical significance. However, by controlling for ORDs (cov4) that were available only for adults in this data set, the incre- mental costs of obesity reduced to one-fourth to one-seventh of the original estimates. Among all adults, the EEE model esti- mated the costs attributable to obesity as $1343 ($1076–$1621) controlling for all confounding factors except the comorbidity conditions, whereas after adding obesity-related comorbidities in the model the estimates were decreased to $209 (�$21 to $434), which was statistically insignificant, compared with the costs of the nonobese (Fig. 2; Table 4).

Discussion

This article provided a systematic review and meta-analysis of the 12 recently published articles that reported the medical care costs associated with obesity, and also performed an original analysis to understand the impact of study methodology on the magnitude of these estimates. From the meta-analysis, the pooled estimate of annual medical costs attributable to obesity was $1901 ($1239–$2582) in 2014 USD, accounting for $149.4 billion at the national level. The extremely high heterogeneity score from the meta-analysis signified the presence of hetero- geneity between different studies due to the use of different data sets from multiple time periods, various statistical methods, and adjustment for a wide range of confounding factors to estimate the costs. Compared with the findings from the previously conducted systematic review that reported the incremental costs of obesity as $2046 (2014 USD) [13], the estimate from this analysis is very comparable.

From the empirical analysis, not surprisingly, different stat- istical methods did not have a significant impact on the

Note: EEE, extended estimating equation; SES, socio-economic status; GLM, generalized linear model; ORDs, obesity-related diseases

Fig. 2 – Impact of age groups, statistical models, and confounding factor adjustment on the estimates of costs attributable to obesity. EEE, extended estimating equation; GLM, generalized linear model; ORDs, obesity-related diseases; SES, socioeconomic status.

V A L U E I N H E A L T H 1 9 ( 2 0 1 6 ) 6 0 2 – 6 1 3 611

V A L U E I N H E A L T H 1 9 ( 2 0 1 6 ) 6 0 2 – 6 1 3612

variability of the estimates in this analysis. However, we caution that this analysis does not endorse that any statistical model can be used in estimating highly skewed cost data. Ignoring the nature of cost data and misspecification of statistical models may lead to inefficient or sometimes biased estimates [32,33]. For all adults, controlling for the demographic and socioeconomic and additional factors (cov1 þ cov2 þ cov3), the EEE model, which had the best GoF among all models based on the GoF tests, reported $1343 ($1076–$1621) for the incremental costs of obesity. This result provides that the estimate of the national medical care costs attributable to obesity would be $94.3 billion ($75.6– $113.2 billion), which accounts for 3.8% (2.8%–4.3%) of national health expenditures in 2010 [35]. This estimate of medical costs attributable to obesity from the empirical analysis was lower than the pooled estimate from the meta-analysis. The reporting error in the BMI measure through self-reported height and weight is likely to bias the coefficient estimates, although the direction of bias is not clear. Also, the possibility of omitting unobserved confounders or reverse causality of obesity on medical costs is likely to underestimate the true costs attributable to obesity. The IV approach by Cawley et al. could address these problems using a weight of biological relative as an instrument. The estimate from the IV approach, however, may not be generalizable to the entire population because of the restriction of the study popula- tion to only adults aged 20 to 64 years with biological children aged 11 to 20 years.

The two most significant drivers of variability in the cost estimates were age groups and adjustment for obesity-related comorbid condition. First, as expected, there is no significant difference in costs attributable to obesity in children/adolescence population because of the presence of very few ORDs that may take a long time to develop among children. In contrast, the incremental costs of obesity were significantly higher than those for the nonobese for the adult population, and the older pop- ulation reported significantly higher costs associated with obesity than did adults aged 18 to 65 years. However, because we included obesity-related comorbidity as a confounding factor in the model, the medical costs of obesity were not significantly higher that the costs among the nonobese. These findings confirmed that most, if not all, of the costs attributable to obesity are mainly caused by ORDs, and as age increases, the obese population is more likely to develop ORDs, incurring higher costs of obesity for the older population.

The main limitation of estimating costs attributable to obesity is the lack of distinction between costs of obesity caused by ORDs and costs of obesity care itself. If the ORDs are caused by the obesity, then by controlling for them, it estimates only the “partial” effect of obesity alone on the cost. However, by omitting such comorbidities as covariates, it estimates the “total effect” of obesity directly on cost and indirectly through mediators, the ORDs. Although the “partial” effect of obesity alone on medical costs was represented by the estimates controlling for ORDs (cov4) in my analysis, which were not significantly different from costs of the nonobese, the true “total effect” of obesity on costs is not easy to estimate, because the regression model could not capture the true counterfactual costs of obesity by just omitting comorbidities as covariates, ignoring the presence of ORDs in the nonobese population. Future study needs to be directed at estimating true counterfactual costs related to the absence/ presence of obesity and ORDs.

After recognizing obesity as a disease, a national survey found that survey participants are more likely to support the disease classification of obesity, and they believe that this change would bring more attention to weight changes and more access to obesity treatment [36]. However, a recent evaluation of adherence to national obesity clinical practice guidelines found the lack of increase in documentation of diagnosis and planned management

of obesity patients [37]. Thus, recognizing obesity as a disease may not lead to immediate changes in health care utilization or significant policy changes. However, what we can do is produce better evidence of effectiveness and cost-effectiveness of obesity treatment through future research. Then, better research alone will increase obesity treatment and reduce the burden of illness, and we hope the overall medical expenditure will be expected to decrease in the long run as we make more diligent efforts to fight against the obesity epidemic. (We appreciate valuable insights from an anonymous reviewer and David Arterburn.)

However, the utility of published estimates for the medical costs of obesity should be examined carefully, because of their wide variation, and the estimates should be applied cautiously in future research and health policy making.

Source of financial support: This work was supported by the Agency for Healthcare Research and Quality predoctoral training fellowship (grant no. 5 T32 HS 013853-10).

Supplementary Materials

Supplemental material accompanying this article can be found in the online version as a hyperlink at http://dx.doi.org/10.1016/j. jval.2016.02.008 or, if a hard copy of article, at www.valueinhealth journal.com/issues (select volume, issue, and article).

R E F E R E N C E S

[1] Sturm R, Ringel JS, Andreyeva T. Increasing obesity rates and disability trends. Health Aff (Millwood) 2004;23:199–205.

[2] Flegal KM, Carroll MD, Ogden CL, et al. Prevalence and trends in obesity among US adults, 1999–2008. JAMA 2010;303:235–41.

[3] Hu FB. Obesity Epidemiology. .. Oxford, UK: Oxford University Press, 2008. [4] Dixon JB. The effect of obesity on health outcomes. Mol Cell Endocrinol

2010;316:104–8. [5] Colditz GA. Economic costs of obesity. Am J Clin Nutr 1992;55:503S–7S. [6] Wolf AM, Colditz GA. Current estimates of the economic cost of obesity

in the United States. Obesity Res 1998;6:97–106. [7] Kortt MA, Langley PC, Cox ER. A review of cost-of-illness studies on

obesity. Clin Therapeut 1998;20:772–9. [8] Colditz GA. Economic costs of obesity and inactivity. Med Sci Sports

Exerc 1999;31:S663–7. [9] Anis AH, Zhang W, Bansback N, et al. Obesity and overweight in

Canada: an updated cost-of-illness study. Obesity Rev 2010;11:31–40. [10] Thorpe KE, Florence CS, Howard DH, et al. The impact of obesity on

rising medical spending. Health Aff (Millwood) 2004;Suppl. Web Exclusives:W4-480-6.

[11] Finkelstein EA, Trogdon JG, Cohen JW, et al. Annual medical spending attributable to obesity: payer-and service-specific estimates. Health Aff (Millwood) 2009;28:w822–31.

[12] Cawley J, Meyerhoefer C. The medical care costs of obesity: an instrumental variables approach. J Health Econ 2012;31:219–30.

[13] Tsai AG, Williamson DF, Glick HA. Direct medical cost of overweight and obesity in the USA: a quantitative systematic review. Obesity Rev 2011;12:50–61.

[14] Gavard JA. Health care costs of obesity in women. Obstet Gynecol Clin North Am 2009;36:213–26, xii.

[15] Wang YC, McPherson K, Marsh T, et al. Health and economic burden of the projected obesity trends in the USA and the UK. Lancet 2011;378:815–25.

[16] Trogdon JG, Finkelstein EA, Feagan CW, et al. State- and payer-specific estimates of annual medical expenditures attributable to obesity. Obesity (Silver Spring) 2012;20:214–20.

[17] Wolf AM, Finer N, Allshouse AA, et al. PROCEED: Prospective Obesity Cohort of Economic Evaluation and Determinants: baseline health and healthcare utilization of the US sample. Diabetes Obes Metab 2008;10:1248–60.

[18] Yang Z, Hall AG. The financial burden of overweight and obesity among elderly Americans: the dynamics of weight, longevity, and health care cost. Health Serv Res 2008;43:849–68.

[19] Finkelstein EA, Trogdon JG, Brown DS, et al. The lifetime medical cost burden of overweight and obesity: implications for obesity prevention. Obesity (Silver Spring) 2008;16:1843–8.

V A L U E I N H E A L T H 1 9 ( 2 0 1 6 ) 6 0 2 – 6 1 3 613

[20] Cai L, Lubitz J, Flegal KM, et al. The predicted effects of chronic obesity in middle age on Medicare costs and mortality. Med Care 2010;48:510–7.

[21] Finkelstein EA, DiBonaventura M, Burgess SM, et al. The costs of obesity in the workplace. J Occup Environ Med 2010;52:971–6.

[22] Onwudiwe NC, Stuart B, Zuckerman IH, et al. Obesity and Medicare expenditure: accounting for age-related height loss. Obesity (Silver Spring) 2011;19:204–11.

[23] Bell JF, Zimmerman FJ, Arterburn DE, et al. Health-care expenditures of overweight and obese males and females in the Medical Expenditures Panel Survey by age cohort. Obesity (Silver Spring) 2011;19:228–32.

[24] Moriarty JP, Branda ME, Olsen KD, et al. The effects of incremental costs of smoking and obesity on health care costs among adults: a 7-year longitudinal study. J Occup Environ Med 2012;54:286–91.

[25] Alley D, Lloyd J, Shaffer T, et al. Changes in the association between body mass index and Medicare costs, 1997–2006. Arch Intern Med 2012;172:277–8.

[26] Ma S, Frick KD, Bleich S, et al. Racial disparities in medical expenditures within body weight categories. J Gen Intern Med 2012;27:780–6.

[27] Bureau of Labor Statistics. Archived Consumer Price Index Detailed Report Information. Available at http://www.bls.gov/cpi/cpi_dr.htm.

[28] StataCorp. Stata Statistical Software: Release 12. College Station, TX: StataCorp LP, 2011.

[29] Defining Overweight and Obesity, National Center for Chronic Disease Prevention and Health Promotion. Centers for Disease Control and Prevention, 2012. Available at http://www.cdc.gov/obesity/adult/ defining.html.

[30] National Heart Lung and Blood Institute, National Institute of Diabetes and Digestive and Kidney Diseases (U.S.). Clinical Guidelines on the Identification, Evaluation, and Treatment of Overweight and Obesity in Adults: The Evidence Report. Bethesda, MD: National Institutes of Health, National Heart, Lung, and Blood Institute, 1998.

[31] Jones A. Health econometrics. In: Culyer AJ, Newhouse JP, eds., Handbook of Health Economics. Amsterdam: Elsevier, 2000.

[32] Manning WG, Mullahy J. Estimating log models: to transform or not to transform? J Health Econ 2001;20:461–94.

[33] Buntin MB, Zaslavsky AM. Too much ado about two-part models and transformation? Comparing methods of modeling Medicare expenditures. J Health Econ 2004;23:525–42.

[34] Basu A, Rathouz PJ. Estimating marginal and incremental effects on health outcomes using flexible link and variance function models. Biostatistics 2005;6:93–109.

[35] Hartman M, Martin AB, Benson J, et al. National health spending in 2011: overall growth remains low, but some payers and services show signs of acceleration. Health Aff (Millwood) 2013;32:87–99.

[36] Puhl RM, Liu S. A national survey of public views about the classification of obesity as a disease. Obesity (Silver Spring) 2015;23:1288–95.

[37] Barnes ER, Theeke LA, Mallow J. Impact of the Provider and Healthcare team Adherence to Treatment Guidelines (PHAT-G) intervention on adherence to national obesity clinical practice guidelines in a primary care centre. J Eval Clin Pract 2015;21:300–6.