American Journal of Epidemiology Vol. 187, No.

6
© The Author(s) 2018. Published by Oxford University Press on behalf of the Johns Hopkins Bloomberg School of DOI: 10.1093/aje/kwx360
Public Health. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com. Advance Access publication:
November 16, 2017

Practice of Epidemiology

Accounting for Time-Varying Confounding in the Relationship Between Obesity

and Coronary Heart Disease: Analysis With G-Estimation

Downloaded from https://academic.oup.com/aje/article-abstract/187/6/1319/4636595 by Universitas Indonesia user on 29 May 2020

The ARIC Study

Maryam Shakiba, Mohammad Ali Mansournia*, Arsalan Salari, Hamid Soori, Nasrin Mansournia,
and Jay S. Kaufman
* Correspondence to Dr. Mohammad Ali Mansournia, Department of Epidemiology and Biostatistics, School of Public Health, Tehran
University of Medical Sciences, Pour-Sina Street, Tehran, Iran (e-mail: mansournia_ma@yahoo.com).

Initially submitted January 28, 2017; accepted for publication November 13, 2017.

In longitudinal studies, standard analysis may yield biased estimates of exposure effect in the presence of time-varying
confounders that are also intermediate variables. We aimed to quantify the relationship between obesity and coronary heart
disease (CHD) by appropriately adjusting for time-varying confounders. This study was performed in a subset of partici-
pants from the Atherosclerosis Risk in Communities (ARIC) Study (1987–2010), a US study designed to investigate risk
factors for atherosclerosis. General obesity was defined as body mass index (weight (kg)/height (m)2) ≥30, and abdominal
obesity (AOB) was defined according to either waist circumference (≥102 cm in men and ≥88 cm in women) or waist:hip
ratio (≥0.9 in men and ≥0.85 in women). The association of obesity with CHD was estimated by G-estimation and com-
pared with results from accelerated failure-time models using 3 specifications. The first model, which adjusted for baseline
covariates, excluding metabolic mediators of obesity, showed increased risk of CHD for all obesity measures. Further
adjustment for metabolic mediators in the second model and time-varying variables in the third model produced negligible
changes in the hazard ratios. The hazard ratios estimated by G-estimation were 1.15 (95% confidence interval (CI): 0.83,
1.47) for general obesity, 1.65 (95% CI: 1.35, 1.92) for AOB based on waist circumference, and 1.38 (95% CI: 1.13, 1.99)
for AOB based on waist:hip ratio, suggesting that AOB increased the risk of CHD. The G-estimated hazard ratios for both
measures were further from the null than those derived from standard models.
abdominal obesity; coronary heart disease; epidemiologic methods; general obesity; G-estimation; obesity; time-
varying confounding

Abbreviations: AOB, abdominal obesity; ARIC, Atherosclerosis Risk in Communities; BMI, body mass index; CHD, coronary heart
disease; CI, confidence interval; GOB, general obesity; HR, hazard ratio; SNAFT, structural nested accelerated failure time.

Coronary heart disease (CHD) is the leading cause of death
and disease burden throughout the world (1, 2). Despite the large
body of evidence regarding the association of obesity with various
adverse health outcomes, its relationship with the risk of CHD
remains controversial. One source of discrepancy between stud-
ies is the type of anthropometric index used to define obesity sta-
tus as either general obesity (GOB), which is based on body
mass index (BMI), or abdominal obesity (AOB), which is based
on waist circumference or waist:hip ratio (3–9). The other sources
of heterogeneity are the type of analysis and the covariates used
for adjustment in the model. Although most observational studies
have shown an association between obesity and CHD, the
association has decreased or disappeared after adjustment for
metabolic mediators of obesity, such as hypertension, diabetes
mellitus, and lipid levels (10–13). For estimation of the total asso-
ciation between obesity and CHD, standard regression methods
model CHD as a function of baseline obesity and other covari-
ates. However, in longitudinal studies, there may be time-varying
covariates whose postbaseline values predict both future exposure
and outcome within strata defined by baseline covariates and prior
obesity. These covariates are referred to as time-varying confound-
ers, and standard regression methods cannot correctly adjust for
them when they are affected by previous exposure (14–17).
Adjustment for such confounders-mediators with standard

1319 Am J Epidemiol. 2018;187(6):1319–1326

1320 Shakiba et al.
regression analysis eliminates part of the effect of obesity operat-
ing through those variables and may also introduce collider-
stratification bias (18–21). In such settings, G-estimation of a
structural nested accelerated failure-time model overcomes this
issue by adequately adjusting for time-varying confounders (16,
22–29). In this study, we aimed to quantify the relationships of
GOB and AOB with the risk of CHD, using G-estimation of a
structural nested failure-time model, and to compare the results
with those of standard regression methods.
METHODS
Study population and outcome
This study was performed in a subset of participants from the
Atherosclerosis Risk in Communities (ARIC) Study. Details on
study procedures and outcome ascertainment have been pub-
lished previously (30). Briefly, the ARIC Study is an ongoing
prospective cohort study of 15,792 participants aged 45–64 years
residing in 4 US communities, designed to investigate risk factors
for atherosclerosis. The study included 4 repeated examina-
tions that started in 1987–1989 (visit 0), followed by 3 exam-
inations at 3-year intervals (visits 1–3). At each examination,
all participants were given an interviewer-administered ques-
tionnaire and underwent an extensive physical examination.
Outcome events were ascertained by annual questionnaires
administered through telephone interviews, 3-year follow-up
examinations, and community-wide surveillance procedures. In
this study, outcome events ascertained through December 31,
2010, were included, and incident CHD events were defined as
definite or probable myocardial infarction or fatal CHD accord-
ing to the ARIC Study criteria. To adjust for lagged confound-
ers, we included a total of 12,902 subjects who had complete
data at visits 0 and 1 in our analysis.
Exposure
In this study, the main exposure of interest was obesity status.
Three anthropometric measures were used to define GOB and
AOB. Trained and certified technicians took anthropometric
measurements, including height and the circumferences of the
waist (umbilical level) and hip (maximum buttock) to the nearest
centimeter. Weight was measured using a zeroed daily scale that
was calibrated quarterly. All anthropometric indices were mea-
sured in fasting participants wearing standard hospital scrub
suits. Body mass index (BMI) was calculated as weight in
kilograms divided by the square of height in meters. GOB was
defined as BMI ≥30. AOB was defined on the basis of 2 anthro-
pometric measures: waist circumference ≥102 cm in men and
≥88 cm in women and waist:hip ratio ≥0.9 in men and ≥0.85 in
women (31).
Confounders
The potential confounders that were measured only at the base-
line visit included age, sex, race (black or white), educational level
(basic (≤11 years), intermediate (12–16 years), or advanced
(17–21 years)), calorie intake, percentage of daily energy intake
from saturated fat, physical activity level (hours/week) at work
and during leisure time, and study center (Washington County,
Maryland; Forsyth County, North Carolina; Jackson, Mississippi;
or Minneapolis, Minnesota). The covariates measured at all visits,
including the baseline visit, were hypertension (systolic blood
pressure ≥90 or diastolic blood pressure ≥140 mm Hg or use of
any medication for high blood pressure); diabetes mellitus (blood
glucose concentration ≥200 mg/dL and fasting blood glucose
concentration ≥126 mg/dL or use of any medication for diabetes);
plasma lipid concentrations, including levels of cholesterol, high-
density lipoprotein cholesterol, and triglyceride; smoking status
Downloaded from https://academic.oup.com/aje/article-abstract/187/6/1319/4636595 by Universitas Indonesia user on 29 May 2020
(defined as currently smoking cigarettes) and drinking status
(defined as currently drinking alcoholic beverages); and occur-
rence of stroke. Lipid profile was categorized on the basis of cut-
points from the Adult Treatment Panel III guidelines (32): 240
mg/dL for total cholesterol, 200 mg/dL for triglyceride, and
40 mg/dL for high-density lipoprotein cholesterol.
Causal diagram and notations
Figure 1 depicts the causal structure of the relationship between
obesity and CHD (33–35). In this graph, there are 4 visits labeled
by K = 0, K = 1, K = 2, and K = 3. The main exposure of inter-
est is obesity status. Ak denotes obesity status as a dichotomous
variable (0, 1), and Lk corresponds to a vector of time-varying co-
variates at visit k. A0 is obesity status at the baseline visit, and L0
includes variables measured only at the baseline visit and baseline
values of time-varying variables. Uk corresponds to unmeasured
risk factors at visit k, such as comorbidity. The arrows from Ak to
Lk suggest that time-varying covariates are affected by previous
exposure. Adjusting for such variables using conventional regres-
sion methods may lead to biased effect estimates. The overbars
are used to denote the history of time-varying variables through
visit K—for example, Lk = [L 0, L1, L 2, …, Lk ]. The outcome
of interest is CHD or death from CHD and is denoted by Y1, Y2,
and Y3 for each follow-up visit. A subject’s observed failure time
is denoted by T, and counterfactual failure time under the expo-
sure histories A¯ = a¯ is denoted by Ta¯ . In particular, the counter-
factual failure time if the subject was never exposed throughout
the follow-up period is denoted by T0¯ .
U0 U1 U2 U3
L0 L1 L2 L3
A0 A1 A2 A3
Y1 Y2 Y3
Figure 1. Causal diagram for the effect of obesity (A) on coronary
heart disease (CHD) or death from CHD (Y) in the Atherosclerosis
Risk in Communities Study.
Am J Epidemiol. 2018;187(6):1319–1326

Statistical analysis
To quantify the relationship between obesity and CHD, G-
estimation of a structural nested accelerated failure-time (SNAFT)
model was used (16, 21–27, 29, 36). G-estimation is one of
Robins’ G-methods for adjustment of time-varying confounders
affected by previous exposure (37). We used the following
SNAFT model:
3
T0¯ = ∑ exp (ψ*Ak )Δtk ,
k=1
where Δtk is the time between visit k and the earliest of the next
visit, an outcome event, or the end of 2010. The SNAFT model
implies that exp(−ψ*) is the ratio of survival time if the partici-
pant is continuously exposed to survival time if the participant is
unexposed throughout. G-estimation is a 2-step iterative process.
At the first step, the value of counterfactual failure time is calcu-
lated using the SNAFT model mentioned above based on TA¯ ,
observed exposure history Ak, and a candidate value of ψ for ψ*:
3 same probability as that included in the denominator, but with-
T (ψ) = ∑ exp (ψAk )Δtk .
k=1
The T(ψ) under the true value of ψ = ψ* is T0¯ , that is, the
counterfactual failure time that would have been observed had
the individual never been exposed throughout the study. In the
second step, the probability of obesity status at each visit K was
modeled as a function of previous obesity and all fixed and time-
varying covariates, as well as T(ψ), using the following pooled
logistic regression model:
logit (Pr (Ak = 1)) = β0k + β1A 0 + β2Ak− 1 + β3L 0
+ β4L k− 1 + β′5L k + β6T (ψ) .
The G-estimate of the ψ* parameter is the value of ψ for which
the coefficient of counterfactual failure time (β6) would be zero
(or equivalently P = 1), meaning that counterfactual survival time
is independent of observed exposure (obesity) given past expo-
sure and covariate history. Thus, the 2 steps mentioned above are
iterated for different values of ψ until the G-estimate is found.
The 95% confidence interval for ψ* was determined by boot-
strapping using 1,000 resamples. To convert the survival time
ratio to a hazard ratio, we assumed a Weibull distribution for sur-
vival time. Then, using the G-estimate of ψ (ψ̂) and the shape
parameter (p) of the Weibull distribution, the hazard ratio com-
paring always obese to never obese was calculated as follows
(21, 23):
HR = exp (ψ̂ × P ) .
The variable T(ψ) can be derived using the SNAFT model
above for participants who experience the events. To account
for censoring by the end of follow-up, the value of T(ψ) was re-
placed by Δ(ψ), a function of T(ψ) and C (the time from visit 1
to the predefined end of follow-up of December 31, 2010) that
takes the value 1 if the subject’s survival time would have been
observed under any possible exposure regime: Δ(ψ) is 1 if T(ψ) <
C(ψ) and Δ(ψ) is 0 if T(ψ) ≥ C(ψ), where for negative values
Am J Epidemiol. 2018;187(6):1319–1326
G-Estimation of the Obesity-CHD Relationship 1321
of ψ, C(ψ) = C* exp(ψ) and for positive values of ψ, Ci (ψ) = C
(21). Thus, Δ(ψ) is zero for all participants who are censored by
the administrative end of follow-up, and it may also be zero for
some of those who did experience an event. There are 2 key
points about Δ(ψ). First, like T(ψ), Δ(ψ) is independent of
observed exposure given past exposure and covariate his-
tory, as it is a function of T(ψ) (and C, which is a baseline co-
variate). Second, unlike T(ψ), Δ(ψ) can be computed for all
subjects (21, 26).
Downloaded from https://academic.oup.com/aje/article-abstract/187/6/1319/4636595 by Universitas Indonesia user on 29 May 2020
We adjusted for selection bias due to competing risks (i.e.,
death from causes other than CHD) and loss to follow-up by
applying visit-specific weights to uncensored individuals equal
to the inverse of the conditional probability of being uncensored
from that visit to the time of the subject’s last visit before an out-
come (CHD or death from CHD) or the administrative end of
follow-up, whichever came first, given past exposure and covar-
iate history. The probabilities were estimated using the pooled
logistic regression model for censoring due to competing risks
and loss to follow-up, with covariates and obesity from the cur-
rent visit, the previous visit, and the baseline visit as predictors.
The weights were stabilized by including in the numerator the
out conditioning on past exposure and covariate history. G-
estimation was then applied to the pseudo-population created by
inverse-probability-of-censoring weighting (38).
We also compared the results from G-estimation with those
obtained from standard regression analysis. To do this, we esti-
mated the associations of GOB and AOB with risk of CHD
using Weibull accelerated failure-time models with 3 different
modeling strategies to adjust for confounding factors. The first
model adjusted for A0 (i.e., obesity status at baseline) and L0,
excluding metabolic mediators of obesity. The second model
adjusted for the variables in the first model plus baseline values
for hypertension, use of antihypertensive medication, diabetes,
and lipid profiles as metabolic mediators of obesity (12). The
third model was a standard time-dependent survival model
and adjusted for A0, L0, Ak−1, Lk−1, Ak, and LK. As for G-
estimation, all 3 standard analyses started at visit 1 (the second
visit) and were performed using Stata, version 13 (StataCorp
LP, College Station, Texas). G-estimation was conducted using
the SNAFTM macro (24, 26) in SAS, version 9.2 (SAS Insti-
tute, Inc., Cary, North Carolina).
RESULTS
Of the 13,580 participants in the ARIC Study with no history
of symptomatic CHD, 678 subjects with missing data at base-
line were excluded. The 12,902 included subjects were fol-
lowed for a total of 242,161 person-years (median duration of
follow-up, 21.2 years), during which 1,481 new cases of CHD
occurred. The incidence rate of CHD was 61.1 per 10,000
person-years (95% confidence interval (CI): 58.1, 64.3). Table 1
shows the baseline characteristics of study participants accord-
ing to GOB status. Obese participants were more likely to be
female, to be black, and to have lower education and annual
family income, and they were less likely to have health insur-
ance compared with nonobese individuals. The prevalence of
drinking and smoking were lower in obese participants than
nonobese participants. In contrast, obese participants had
1322 Shakiba et al.

higher levels of hypertension, antihypertensive medication use,
diabetes mellitus, total cholesterol, and triglyceride than non-
obese individuals and lower high-density lipoprotein cholesterol
than the nonobese. Nonobese participants had higher physical
activity levels during sport and leisure time and lower per-
centages of daily energy intake from saturated fat compared
with obese individuals.
During follow-up, 2,998 subjects died from causes other than
CHD, and 5,456 subjects were lost to follow-up. The mean value
Table 2 shows the value of ψ and the corresponding survival
time ratio as exp(−ψ) for the associations of GOB and AOB
with CHD. The hypothesized value for estimating ψ ranged
between −1.5 to 1.5. The point estimate of exp(−ψ) for GOB
was 0.89, suggesting an 11% decrease in G-estimated survival
time to CHD (95% CI: 0.75, 1.14). AOB based on waist circum-
ference and AOB based on waist:hip ratio decreased G-estimated
survival time to CHD by 33% (95% CI: 21, 41) and 22% (95%
CI: 9, 43), respectively.

Downloaded from https://academic.oup.com/aje/article-abstract/187/6/1319/4636595 by Universitas Indonesia user on 29 May 2020

of stabilized censoring weights was 0.997 (standard deviation, Figure 2 shows the hazard ratios estimated by G-estimation in
0.34; range, 0.43–5.98). comparison with 3 standard modeling strategies for all obesity

Table 1. Baseline Characteristics of 12,902 Participants According to General Obesity Status, Atherosclerosis Risk
in Communities Study, 1987–2010

Obese (n = 3,424) Nonobese (n = 9,478)

Characteristic No. of No. of P Value
% Mean (SD) % Mean (SD)
Persons Persons

Age, years 54 (5.7) 54 (5.8) 0.377

Female sex 2,125 62 5,019 53 0.001
Black race 1,332 39 1,991 21 0.001
Educational level
Basic (≤11 years) 1,011 29 1,854 20 0.001
Intermediate (12–16 years) 1,359 40 3,932 41 0.001
Advanced (17–21 years) 1,046 31 3,685 39 0.001
Family income, dollars/year
<16,000 982 31 1,637 18 0.001
16,000–50,000 1,643 51 4,729 53 0.001
>50,000 573 18 2,608 29
Possession of health insurance 2,961 86 8,727 92 0.001
Family history of CVD 1,886 55 5,240 55 0.105
Current alcohol drinking 1,555 45 5,882 62 0.001
Current tobacco smoking 659 19 2,728 29 0.001
Hypertensiona 1,717 50 2,585 27 0.001
Antihypertensive medication use 1,297 38 1,755 18 0.001
Diabetes mellitusb 729 21 690 7.3 0.001
Physical activity level, hours/week
Work 2.21 (0.98) 2.19 (0.92) 0.717
Sports 2.28 (0.72) 2.50 (0.81) 0.001
Leisure time 2.27 (0.58) 2.40 (0.56) 0.001
Total energy intake, kcal/day 1,615 (615) 1,625 (607) 0.748
Saturated fatty acid, % of kcal/day 12.2 (2.9) 11.9 (3.0) 0.001
Lipid profilec
High total cholesterol level (>240 924 27 2,227 23 0.001
mg/dL)
High triglyceride level (>200 mg/dL) 618 18 1,050 11 0.001
Low HDL cholesterol level (<40 mg/dL) 1,141 33 2,127 22 0.001

Abbreviations: CVD, cardiovascular disease; HDL, high-density lipoprotein; SD, standard deviation.
a
Hypertension was defined as systolic blood pressure ≥90 or diastolic blood pressure ≥140 mm Hg or use of any
medication for high blood pressure.
b
Diabetes mellitus was defined as blood glucose concentration ≥200 mg/dL and fasting blood glucose concentra-
tion ≥126 mg/dL or use of any medication for diabetes.
c
Lipid profile was categorized on the basis of cutpoints from the Adult Treatment Panel III guidelines (32).

Am J Epidemiol. 2018;187(6):1319–1326
 G-Estimation of the Obesity-CHD Relationship 1323

Table 2. Estimates of the Causal Effects of General and Abdominal 3.0

Obesity on Risk of Coronary Heart Disease Using G-Estimationa,
Atherosclerosis Risk in Communities Study, 1987–2010
2.0
G-Estimate Survival
Obesity Status 95% CIc 95% CIc
of ψb Time Ratio

Hazard Ratio
1.5
General 0.11 −0.13, 0.28 0.89 0.75, 1.14
obesityd
Abdominal 1.0
obesity

Downloaded from https://academic.oup.com/aje/article-abstract/187/6/1319/4636595 by Universitas Indonesia user on 29 May 2020

e
WC 0.39 0.23, 0.51 0.67 0.59, 0.79
f 0.7
WHR 0.25 0.09, 0.56 0.78 0.57, 0.91

Abbreviations: CI, confidence interval; WC, waist circumference; WHR, 0.5

waist:hip ratio.
a Body Mass Index
Adjusted for study visit (indicator variable), age, sex, race, educa-
tional level, physical activity, calorie intake, and percentage of daily
energy intake from saturated fat at baseline, and for baseline and cur-
rent tobacco smoking status, alcohol drinking status, hypertension,
diabetes mellitus, antihypertensive medication use, high cholesterol
level (level >240 mg/dL), high triglyceride level (>200 mg/dL), low
high-density lipoprotein cholesterol level (<40 mg/dL), and occurrence
of stroke.
b
ψ is the causal parameter in the structural nested accelerated
failure-time model.
c
Bootstrap-based 95% CI.
d
General obesity was defined as body mass index (weight (kg)/
height (m)2) ≥30.
e
Abdominal obesity was defined as WC ≥102 cm in men and ≥88
cm in women.
f
Abdominal obesity was defined as WHR ≥0.9 in men and ≥0.85 in
women.
measures. G-estimation suggested that obesity based on BMI,
waist:hip ratio, and waist circumference increased the hazard of
CHD by 15%, 38%, and 65%, respectively, though the 95%
confidence interval for the BMI hazard ratio included the null
value (Figure 2). The results from the first standard model
adjusting for baseline covariates, excluding metabolic medi-
ators of obesity, showed an increase in the hazard of CHD
for all measures of obesity. Further adjustment for metabolic
mediators of obesity resulted in a decrease in hazard ratios to
0.95 (95% CI: 0.83, 1.08) for GOB, 1.04 (95% CI: 0.88, 1.14)
for AOB based on waist circumference, and 1.04 (95% CI: 0.86,
1.25) for AOB based on waist:hip ratio. The third model, which
adjusted for baseline, lagged, and current values of time-varying
covariates, showed reduced risk of CHD associated with all
obesity measures, though all 95% confidence intervals included
the null value. The hazard ratios from G-estimation were higher
than those from standard time-dependent survival models and
suggested increased risks of CHD for general obesity (hazard
ratio (HR) = 1.15, 95% confidence interval (CI): 0.83, 1.47),
waist circumference (HR = 1.65, 95% CI: 1.35, 1.92), and
waist:hip ratio (HR = 1.38, 95% CI: 1.13, 1.99) (Figure 2).
DISCUSSION
In the present study, the association of 3 different obesity mea-
sures with risk of CHD was assessed using the G-estimation
method. AOB decreased the time to CHD by 33% based on waist
Waist Circumference Waist:Hip Ratio
Obesity Index
Figure 2. Hazard ratios for coronary heart disease (CHD) or CHD
mortality according to general obesity and abdominal obesity in the
Atherosclerosis Risk in Communities Study, 1987–2010. Hazard ratios
were based on G-estimation as well as on 3 standard modeling strate-
gies. The vertical axis is on the log scale. ●, standard model adjusted
with no metabolic mediators; ■, standard model adjusted with metabolic
mediators; ▲, standard time-varying adjusted model; ♦, G-estimation.
Bars, 95% confidence intervals.
circumference and 22% based on waist:hip ratio. Previous studies
attempting to address time-varying confounding in the association
between obesity and cardiovascular events had inconsistent re-
sults (29, 39–43). In 2 Mendelian randomization studies using
genetic score as an instrument, no association of BMI with CHD
was found (40, 41). The effect of obesity was also assessed
through weight-loss strategies as an intervention in randomized
trials or using the parametric G-formula in observational studies
(39, 42, 43). The findings were inconclusive in that weight loss
was associated with improvement in cardiovascular disease risk
factors but no association was found with risk of cardiovascular
events. In a previous study by Tilling et al. (29) using G-
estimation, no evidence of increased CHD risk was found with
high BMI. However, in previous studies that used causal model-
ing approaches, no measures of AOB were used. Therefore, to
our knowledge, this is the first study that has examined the
G-estimated relationship between AOB (based on 2 different
measures) and risk of CHD. It has been suggested that visceral
fat captured by AOB, in contrast to peripheral or subcutaneous
fat, is proatherogenic and that central distribution of adipose tis-
sue confers a greater risk of cardiovascular disease than total
amount of adiposity (3, 5, 44–47).
In accordance with most previous observational studies (10–
12, 48–50), the results of the current study using different model-
ing strategies showed that the higher risk associated with obesity
in standard models adjusting for baseline variables disappeared
after adjustment for metabolic mediators of obesity. This was due
to overadjustment bias introduced by conditioning on variables
that lie on the causal pathway between obesity and CHD. More
importantly, adjusting for time-varying confounders using G-
estimation showed an increase in risk for all measures of AOB
in comparison with the standard time-dependent accelerated
failure-time model, suggesting that the association of obesity

Am J Epidemiol. 2018;187(6):1319–1326
1324 Shakiba et al.
with CHD might be biased by adjustment for time-varying
confounders through standard methods.
While standard regression methods provide estimates that
are conditional on the confounders, G-methods, including G-
estimation, produce marginal estimates. Marginal and condi-
tional estimates are not directly comparable, because of 1) biases
caused by conditioning on a variable affected by exposure and
2) noncollapsibility of certain effect measures (19, 51). Assuming
correct model specification, we suspect that the main reason for
the divergence between the results of G-estimation and those of
the corresponding conventional model (model 3) is the bias
induced by conditioning on the time-varying confounders
affected by the previous exposure, as the outcome is uncommon
and thus noncollapsibility of the hazard ratio is not an issue (52).
It is notable that the estimates from a standard analysis without
adjustment for metabolic mediators were roughly similar to the es-
timates derived from the G-estimation method. It was only when
adjusting for metabolic mediators (with or without accounting for
their changing values over time) that the standard regressions
gave null results (Figure 2). This pattern of results is consis-
tent with the interpretation that in this application there was
essentially no direct effect, and that confounding of the total
effect by the metabolic mediators was not an important source of
bias. Another explanation for this pattern is that there is an unmea-
sured shared cause of the metabolic mediator(s) and the out-
come, which would cause collider stratification bias in the
adjusted regression.
The validity of inference from G-estimation relies on several
assumptions (15, 53). The first assumption, which is untestable,
is comparability of exposed and unexposed subjects at any visit
k given the previous exposure and confounder history. We used
many time-varying and time-fixed confounders to justify the
assumption of exchangeability between obese and nonobese in-
dividuals, but there might still have been some other, unmea-
sured confounders, such as diet quality and use of statins. In the
ARIC Study, as in other interval cohorts, confounders are not
measured at the times of exposure change and, in fact, concur-
rent or lagged time-varying confounders are the mismeasured
versions of the true values of the confounders. Moreover, resid-
ual confounding bias can arise from using dichotomous instead
of continuous measures of time-varying confounders such as
blood pressure and serum lipid levels, as well as measurement
errors in crudely self-reported confounders (e.g., physical activ-
ity, smoking, and alcohol consumption). The second assumption
is specifying a suitable model for linking observed failure time
and exposure history with counterfactual failure time. Similar to
the work done by Witteman et al. (21), we used the current value
of exposure (i.e., obesity) status and assumed that obesity at visit
t represented obesity during the interval (t, t + 1). Another
assumption for causal methods is positivity—that is, both obese
and nonobese individuals are observed within every stratum of
confounders. However, unlike inverse-probability-of-treatment
weighting, G-estimation can result in asymptotically unbiased
estimates even under violation of positivity if one is willing to rely
on extrapolation to the empty cells, assuming that confounders
are not effect modifiers (24, 53). The final assumption that is
dubious in our study is well-defined intervention or consis-
tency. This assumption states that for each subject, the coun-
terfactual outcome under the observed value of exposure is
equal to the observed outcome. Since there are different ways in
which an obese subject could have been nonobese (more exer-
cise, less food intake, bariatric surgery, etc.) that may correspond
to different counterfactual outcomes, the consistency assumption
cannot be considered credible. In fact, this assumption is best
achieved for well-defined experiments and medical treatments.
Therefore, interpretation of the causal effect in an observational
study for metabolic or biomarker exposures such as obesity is not
straightforward and should be made cautiously (54, 55).
In summary, our study used G-estimation to quantify the rela-
Downloaded from https://academic.oup.com/aje/article-abstract/187/6/1319/4636595 by Universitas Indonesia user on 29 May 2020
tionship between obesity and risk of CHD by appropriately ad-
justing for time-varying confounding affected by previous
obesity. Shorter survival time to CHD was associated with
higher levels of AOB, whether measured as waist circumfer-
ence or as waist:hip ratio. This highlights the importance of
considering AOB rather than BMI as a risk factor for CHD
or CHD mortality.
ACKNOWLEDGMENTS
Author affiliations: Cardiovascular Diseases Research
Center, Guilan University of Medical Sciences, Rasht, Iran
(Maryam Shakiba, Arsalan Salari); Road Trauma Research
Center, Guilan University of Medical Sciences, Rasht, Iran
(Maryam Shakiba); School of Health, Guilan University
of Medical Sciences, Rasht, Iran (Maryam Shakiba);
Department of Epidemiology and Biostatistics, School of
Public Health, Tehran University of Medical Sciences,
Tehran, Iran (Mohammad Ali Mansournia); Department of
Epidemiology, School of Public Health, Shahid Beheshti
University of Medical Sciences, Tehran, Iran (Hamid Soori);
Department of Endocrinology, AJA University of Medical
Sciences, Tehran, Iran (Nasrin Mansournia); and
Department of Epidemiology, Biostatistics and
Occupational Health, Faculty of Medicine, McGill
University, Montreal, Quebec, Canada (Jay S. Kaufman).
This article was prepared using ARIC research materials
obtained from the National Heart, Lung, and Blood Institute.
We thank Dr. Sally Picciotto for providing the SNAFTM
macro and for her helpful advice on G-estimation analysis.
Conflict of interest: none declared.
REFERENCES
1. Lozano R, Naghavi M, Foreman K, et al. Global and regional
mortality from 235 causes of death for 20 age groups in 1990
and 2010: a systematic analysis for the Global Burden of
Disease Study 2010. Lancet. 2012;380(9859):2095–2128.
2. Murray CJ, Vos T, Lozano R, 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;380(9859):2197–2223.
3. Canoy D, Boekholdt SM, Wareham N, et al. Body fat
distribution and risk of coronary heart disease in men and
women in the European Prospective Investigation Into Cancer
and Nutrition in Norfolk cohort: a population-based
prospective study. Circulation. 2007;116(25):2933–2943.
4. Fujimoto WY, Bergstrom RW, Boyko EJ, et al. Visceral
adiposity and incident coronary heart disease in Japanese-
Am J Epidemiol. 2018;187(6):1319–1326

American men. The 10-year follow-up results of the Seattle
Japanese-American Community Diabetes Study. Diabetes
Care. 1999;22(11):1808–1812.
5. Litwin SE. Which measures of obesity best predict
cardiovascular risk? J Am Coll Cardiol. 2008;52(8):616–619.
6. Rexrode KM, Buring JE, Manson JE. Abdominal and total
adiposity and risk of coronary heart disease in men. Int J Obes
Relat Metab Disord. 2001;25(7):1047–1056.
7. Rexrode KM, Carey VJ, Hennekens CH, et al. Abdominal
adiposity and coronary heart disease in women. JAMA. 1998;
280(21):1843–1848.
8. Rimm EB, Stampfer MJ, Giovannucci E, et al. Body size and
fat distribution as predictors of coronary heart disease among
middle-aged and older US men. Am J Epidemiol. 1995;
141(12):1117–1127.
9. Terry RB, Page WF, Haskell WL. Waist/hip ratio, body mass
index and premature cardiovascular disease mortality in US
Army veterans during a twenty-three year follow-up study. Int
J Obes Relat Metab Disord. 1992;16(6):417–423.
10. de Koning L, Merchant AT, Pogue J, et al. Waist
circumference and waist-to-hip ratio as predictors of
cardiovascular events: meta-regression analysis of prospective
studies. Eur Heart J. 2007;28(7):850–856.
11. Lakka HM, Lakka TA, Tuomilehto J, et al. Abdominal obesity
is associated with increased risk of acute coronary events in
men. Eur Heart J. 2002;23(9):706–713.
12. Global Burden of Metabolic Risk Factors for Chronic
Diseases Collaboration (BMI Mediated Effects), Lu Y,
Hajifathalian K, et al. Metabolic mediators of the effects of
body-mass index, overweight, and obesity on coronary heart
disease and stroke: a pooled analysis of 97 prospective
cohorts with 1.8 million participants. Lancet. 2014;
383(9921):970–983.
13. Yatsuya H, Toyoshima H, Yamagishi K, et al. Body mass
index and risk of stroke and myocardial infarction in a
relatively lean population: meta-analysis of 16 Japanese
cohorts using individual data. Circ Cardiovasc Qual
Outcomes. 2010;3(5):498–505.
14. Robins J. A new approach to causal inference in mortality
studies with a sustained exposure period—application to
control of the healthy worker survivor effect. Math Model.
1986;7(9–12):1393–1512.
15. Robins J, Hernán M. Estimation of the causal effects of time-
varying exposures. In: Fitzmaurice GM, Davidian M, Verbeke
G, et al., eds. Longitudinal Data Analysis. New York, NY:
Chapman and Hall/CRC Press; 2009:553–599.
16. Robins JM, Blevins D, Ritter G, et al. G-estimation of the
effect of prophylaxis therapy for Pneumocystis carinii
pneumonia on the survival of AIDS patients. Epidemiology.
1992;3(4):319–336.
17. Mansournia MA, Etminan M, Danaei G, et al. Handling time
varying confounding in observational research. BMJ. 2017;
359:j4587.
18. Hernán MA, Hernández-Díaz S, Robins JM. A structural
approach to selection bias. Epidemiology. 2004;15(5):615–625.
19. Mansournia MA, Danaei G, Forouzanfar MH, et al. Effect of
physical activity on functional performance and knee pain in
patients with osteoarthritis: analysis with marginal structural
models. Epidemiology. 2012;23(4):631–640.
20. Robins J. The control of confounding by intermediate
variables. Stat Med. 1989;8(6):679–701.
21. Witteman JC, D’Agostino RB, Stijnen T, et al. G-estimation of
causal effects: isolated systolic hypertension and
cardiovascular death in the Framingham Heart Study. Am J
Epidemiol. 1998;148(4):390–401.
Am J Epidemiol. 2018;187(6):1319–1326
G-Estimation of the Obesity-CHD Relationship 1325
22. Robins JM. Structural nested failure time models. In: Armitage P,
Colton T, eds. Encyclopedia of Biostatistics. Chichester, United
Kingdom: John Wiley & Sons Ltd.; 1997:4372–4389.
23. Sterne JAC, Tilling K. G-estimation of causal effects,
allowing for time-varying confounding. Stata J. 2002;2(2):
164–182.
24. Chevrier J, Picciotto S, Eisen EA. A comparison of standard
methods with g-estimation of accelerated failure-time models
to address the healthy-worker survivor effect: application in a
cohort of autoworkers exposed to metalworking fluids.
Downloaded from https://academic.oup.com/aje/article-abstract/187/6/1319/4636595 by Universitas Indonesia user on 29 May 2020
Epidemiology. 2012;23(2):212–219.
25. Costello S, Picciotto S, Rehkopf DH, et al. Social disparities in
heart disease risk and survivor bias among autoworkers: an
examination based on survival models and g-estimation. Occup
Environ Med. 2015;72(2):138–144.
26. Hernán MA, Cole SR, Margolick J, et al. Structural accelerated
failure time models for survival analysis in studies with time-
varying treatments. Pharmacoepidemiol Drug Saf. 2005;14(7):
477–491.
27. Keiding N, Filiberti M, Esbjerg S, et al. The graft versus
leukemia effect after bone marrow transplantation: a case study
using structural nested failure time models. Biometrics. 1999;
55(1):23–28.
28. Joffe MM, Hoover DR, Jacobson LP, et al. Estimating the
effect of zidovudine on Kaposi’s sarcoma from observational
data using a rank preserving structural failure-time model. Stat
Med. 1998;17(10):1073–1102.
29. Tilling K, Sterne JA, Szklo M. Estimating the effect of
cardiovascular risk factors on all-cause mortality and incidence
of coronary heart disease using G-estimation: the
Atherosclerosis Risk in Communities Study. Am J Epidemiol.
2002;155(8):710–718.
30. The ARIC Investigators. The Atherosclerosis Risk in
Communities Study: design and objectives. Am J Epidemiol.
1989;129(4):687–702.
31. World Health Organization. Waist Circumference and Waist-
Hip Ratio: Report of a WHO Expert Consultation, Geneva,
8–11 December 2008. Geneva, Switzerland: World Health
Organization; 2011.
32. National Cholesterol Education Program (NCEP) Expert
Panel on Detection, Evaluation, and Treatment of High Blood
Cholesterol in Adults (Adult Treatment Panel III). Third
Report of the National Cholesterol Education Program
(NCEP) Expert Panel on Detection, Evaluation, and
Treatment of High Blood Cholesterol in Adults (Adult
Treatment Panel III). Final report. Circulation. 2002;106(25):
3143–3421.
33. Greenland S, Mansournia MA. Limitations of individual causal
models, causal graphs, and ignorability assumptions, as
illustrated by random confounding and design unfaithfulness.
Eur J Epidemiol. 2015;30(10):1101–1110.
34. Mansournia MA, Hernán MA, Greenland S. Matched designs
and causal diagrams. Int J Epidemiol. 2013;42(3):860–869.
35. Mansournia MA, Higgins JP, Sterne JA, et al. Biases in
randomized trials: a conversation between trialists and
epidemiologists. Epidemiology. 2017;28(1):54–59.
36. Joffe MM, Hoover DR, Jacobson LP, et al. Effect of treatment
with zidovudine on subsequent incidence of Kaposi’s sarcoma.
Clin Infect Dis. 1997;25(5):1125–1133.
37. Naimi AI, Cole SR, Kennedy EH. An introduction to g
methods. Int J Epidemiol. 2017;46(2):756–762.
38. Mansournia MA, Altman DG. Inverse probability weighting.
BMJ. 2016;352:i189.
39. Danaei G, Robins JM, Young JG, et al. Weight loss and
coronary heart disease: sensitivity analysis for unmeasured
1326 Shakiba et al.
confounding by undiagnosed disease. Epidemiology. 2016;
27(2):302–310.
40. Hagg S, Fall T, Ploner A, et al. Adiposity as a cause of
cardiovascular disease: a Mendelian randomization study. Int J
Epidemiol. 2015;44(2):578–586.
41. Holmes MV, Lange LA, Palmer T, et al. Causal effects of body
mass index on cardiometabolic traits and events: a Mendelian
randomization analysis. Am J Hum Genet. 2014;94(2):
198–208.
42. Romeo S, Maglio C, Burza MA, et al. Cardiovascular events
after bariatric surgery in obese subjects with type 2 diabetes.
Diabetes Care. 2012;35(12):2613–2617.
43. Look AHEAD Research Group, Wing RR, Bolin P, et al.
Cardiovascular effects of intensive lifestyle intervention in type 2
diabetes. N Engl J Med. 2013;369(2):145–154.
44. Cameron AJ, Zimmet PZ. Expanding evidence for the multiple
dangers of epidemic abdominal obesity. Circulation. 2008;
117(13):1624–1626.
45. Canoy D, Wareham N, Luben R, et al. Serum lipid
concentration in relation to anthropometric indices of central
and peripheral fat distribution in 20,021 British men and
women: results from the EPIC-Norfolk population-based
cohort study. Atherosclerosis. 2006;189(2):420–427.
46. Dalton M, Cameron AJ, Zimmet PZ, et al. Waist
circumference, waist-hip ratio and body mass index and their
correlation with cardiovascular disease risk factors in
Australian adults. J Intern Med. 2003;254(6):555–563.
47. Janssen I, Katzmarzyk PT, Ross R. Waist circumference and
not body mass index explains obesity-related health risk. Am J
Clin Nutr. 2004;79(3):379–384.
48. Keihani S, Hosseinpanah F, Barzin M, et al. Abdominal
obesity phenotypes and risk of cardiovascular disease in a
decade of follow-up: the Tehran Lipid and Glucose Study.
Atherosclerosis. 2015;238(2):256–263.
49. Logue J, Murray HM, Welsh P, et al. Obesity is associated with
fatal coronary heart disease independently of traditional risk
factors and deprivation. Heart. 2011;97(7):564–568.
Downloaded from https://academic.oup.com/aje/article-abstract/187/6/1319/4636595 by Universitas Indonesia user on 29 May 2020
50. Page JH, Rexrode KM, Hu F, et al. Waist-height ratio as a
predictor of coronary heart disease among women.
Epidemiology. 2009;20(3):361–366.
51. Mohammad K, Hashemi-Nazari SS, Mansournia N, et al.
Marginal versus conditional causal effects. J Biostat
Epidemiol. 2015;1(3–4):121–128.
52. Mansournia MA, Greenland S. The relation of collapsibility
and confounding to faithfulness and stability. Epidemiology.
2015;26(4):466–472.
53. Hernán MA, Robins JM. Causal Inference. Boca Raton, FL:
Chapman&Hall/CRC; 2017.
54. Hernán MA. Invited commentary: hypothetical interventions to
define causal effects—afterthought or prerequisite? Am J
Epidemiol. 2005;162(7):618–620.
55. Hernán MA, Taubman SL. Does obesity shorten life? The
importance of well-defined interventions to answer causal
questions. Int J Obes (Lond). 2008;32(suppl 3):S8–S14.
Am J Epidemiol. 2018;187(6):1319–1326