1. health and fitness
2. addiction
3. smoking addiction

American Journal of Epidemiology
Copyright O 1998 by The John3 Hopkins University School of Hygiene and Public Health
All rights reserved
Vol. 147, No. 8
Printed in U.SA
Mortality and Optimal Body Mass Index in a Sample of the US Population
Ram6n A. Durazo-Arvizu, Daniel L McGee, Richard S. Cooper, Youlian Liao, and Amy Luke
body mass index; mortality; obesity
between BMI and health status is almost certainly
bidirectional, and the directionality is likely to" vary
across the range of BMIs. Thus, illness can cause
weight loss and weight gain can cause illness. Second,
many potential forms of confounding can be identified, particularly for lifestyle factors. Smoking and
heavy drinking are more common in lean individuals,
while the obese have greater caloric intake and engage
in less physical activity (19-22). A mixture of these
confounding influences can exist in the same BMI
range. Among persons who are lean, one is likely to
find an excess of the health-conscious as well as the
unhealthy. Finally, standard analytic methods have not
been well established, and a variety of strategies have
been employed, further compromising direct comparisons of the results. The use of arbitrarily defined
quantiles based on the amount of data available and
their empirical characteristics is a particularly common approach, but it has inherent limitations. An
important additional shortcoming of previous studies
has been the use of unrepresentative samples as the
basis for inference to the general population.
In this study, the biethnic US population sample
available through the First National Health and Nutrition Examination Survey (NHANES I) Epidemiologic
Follow-up Study was used to characterize mortality by
level of BMI. In an attempt to define a more robust
analytic method, we fitted the BMI-mortality relation
The impact of overweight on health risk, and on
mortality in particular, is an issue of great public
interest at the present time. Unfortunately, current
recommendations often make reference to contradictory evidence. Although a monotonic relation between
relative body weight, usually expressed as body
mass index (BMI) (weight (kg)/height2 (m2)), and
coronary heart disease and diabetes mellitus have been
consistently observed (1, 2), the relation with total
mortality is less well established. Epidemiologic studies have reported five major types of outcome—either
no relation (3-6), a direct association (7-10), an inverse association (11), a J-shaped relation (3-5, 7,
12-15), or a U-shaped relation (6, 16-18). Not unexpectedly, the interpretation of these findings has been
the subject of ongoing debate.
Given the virtual impossibility of conducting randomized trials on this question, it is necessary to rely
on the findings of observational studies. The interpretation of these findings is complex, however, for several reasons. First, the cause-and-effect relationship
Received for publication November 8, 1996, and in final form
November 10, 1997.
Abbreviations: BMI, body mass index; BMI,,*,, BMI of minimum
mortality; ICD-9, International Classification of Diseases, Ninth Revision; NHANES I, First National Health and Nutrition Examination
Survey.
From the Department of Preventive Medicine and Epidemiology,
Stitch School of Medicine, Loyola University, Maywood, IL
739
Downloaded from http://aje.oxfordjournals.org/ by guest on October 6, 2014
In this paper, the authors model the nonmonotonic relation between body mass index (BMI) (weight
(kgyheight2 (m2)) and mortality in 13,242 black and white participants in the NHANES I Epidemiologic
Follow-up Study in order to estimate the BMI at which minimum mortality occurs. The BMI of minimum
mortality was 27.1 for black men (95% confidence interval (Cl) 24.8-29.4), 26.8 for black women (95% Cl
24.7-28.9), 24.8 for white men (95% Cl 23.8-25.9), and 24.3 for white women (95% Cl 23.3-25.4). Each
confidence interval included the group average. Analyses conducted by smoking status and after exclusion of
persons with baseline illness and persons who died during the first 4 years of follow-up led to virtually identical
estimates. The authors determined the range of values over which risk of all-cause mortality would increase
no more than 20% in comparison with the minimum. This interval was nine BMI units wide, and it included 70%
of the population. These results were confirmed by parallel analyses using quantlles. The model used allowed
the estimation of parameters in the BMI-mortaltty relation. The resulting empirical findings from each of four
race/sex groups, which are representative of the US population, demonstrate a wide range of BMIs consistent
with minimum mortality and do not suggest that the optimal BMI is at the lower end of the distribution for any
subgroup. Am J Epidemiol 1998; 147:739-49.
740
Durazo-Arvizu et al.
with a mathematical model. The primary focus of this
analysis was to determine the BMI of minimum allcause mortality risk and the range of values associated
with an increase in risk of 20 percent or less.
sified as either a never smoker or an ever smoker,
which included both current and former smokers (31,
32). We did not control for variables which are in the
pathway between obesity and illness, such as hypertension, diabetes mellitus, and hypercholesterolemia.
MATERIALS AND METHODS
The NHANES I Epidemiologic Follow-up Study
Statistical methods
Preliminary descriptive analyses were performed on
these data. First, crude and age-adjusted mortality
rates were calculated by race and sex for all causes of
death, as well as for specific causes. The age-adjusted
rates were obtained by direct standardization, with
10-year strata of the entire race/sex-specific cohort as
the reference. Second, the BMI range was divided into
group-specific quintiles, and all-cause age-adjusted
mortality risks and relative risks were computed for
each BMI quintile. Third, point estimates of the BMI
of minimum mortality ( B M I ^ J were calculated as the
midpoint of the BMI interval of minimum relative
risk. The BMI interval of minimum risk was defined
as that interval obtained by concatenating all of the
intervals with age-adjusted risk estimates that fell
within the 95 percent confidence interval of the estimated lowest risk. These values were computed primarily to lend credence to the estimates of the BML^n
established by our models.
Logistic regression analysis (33) was used to assess
the association between BMI and mortality, adjusting
for age and smoking history. We used the logistic
regression model primarily because it allowed us to
visualize the relation between BMI and mortality well.
However, the model also provided the basis for seeking transformations to account for the asymmetric
nonmonotonic relation. We examined the sensitivity
of our analyses to the choice of this model by repeating all analyses using the proportional hazards method
(34) and by grouping the data and using Poisson
regression (35). Virtually identical results were
achieved in these reanalyses.
As has been demonstrated in other cohorts (36), the
BMI values had a right-skewed distribution. Furthermore, although the relation between mortality and
BMI has frequently been described as U-shaped, it is
generally asymmetric, as was the case in this large,
representative population sample. Fitting a quadratic
equation to these data can result in an estimated
BMI^n that is either too high or too low, depending on
the nature of the asymmetry. As a consequence, some
authors have suggested that no attempt be made to
model the relation (16), and to our knowledge mathematical functions have not been used for this purpose.
One method of obtaining a fit to these data is to
transform BMI values so that the resultant values for
the variable are normally distributed (37-39). To disAm J Epidemiol
Vol. 147, No. 8, 1998
Downloaded from http://aje.oxfordjournals.org/ by guest on October 6, 2014
As made available from the National Center for
Health Statistics, data from the NHANES I Epidemiologic Follow-up Study (23-27) were used to examine
the relation of BMI to mortality. Briefly, these data
provide follow-up information on morbidity and mortality among 14,407 individuals, initially aged 25-74
years, who received complete medical examinations
during NHANES I, which was conducted from 1971
to 1975 as previously described (2, 28, 29). Follow-up
surveys were carried out in 1982-1984, 1986 (among
persons aged ^ 5 5 years at baseline), and 1987 (26,
27). Our analysis was restricted to the 737 black men,
1,243 black women, 4,644 white men, and 6,618 white
women who were present during at least one of the
three follow-up cycles of the study and for whom BMI
was measured. The small percentage of persons whose
ethnicity was neither black nor white were omitted
(n = 141; 1 percent).
Mortality was defined in terms of the 1987 followup (24, 25). At each follow-up, the subjects (or their
proxies) were interviewed, death certificates were
gathered for subjects who had died, and hospital and
nursing home records were obtained for overnight
stays that had occurred since the most recent contact.
A subject's death had to be confirmed by either a death
certificate or a proxy interview. Death certificates
were coded using the International Classification of
Diseases, Ninth Revision (ICD-9) (30). Coronary
heart diseases were denoted by ICD-9 codes 410-414
and 429.2, cardiovascular diseases by ICD-9 codes
390-448 (stroke constitutes ICD-9 codes 430-434
and 436-438), and cancers by ICD-9 codes 140-208;
all other codes denoted all other causes of death.
Height was measured with the examinee wearing
disposable foam rubber slippers. To minimize observer and recording errors, height was recorded by
Polaroid camera (Polaroid Corp., Cambridge, Massachusetts). Weight was measured using a Toledo selfbalancing scale (Toledo Guild Products, Inc., Toledo,
Ohio) that mechanically printed the person's weight
with an accuracy of lA pound (0.1 kg). Smoking information was collected at baseline on only approximately half of the participants, and these data were
supplemented retrospectively. Because of the difficulty of separating the effect of previous smoking
from that of current smoking, a participant was clas-
Mortality and Body Mass Index
741
(B)
(A)
talit}^ Rate
0.60:
i_
0.50
0.40-
0.40
0.30-
0.30
0^0-
0.20
0.10<
<-21.4
21.5-23.9
24.0-26.1
26.2-292
29.3+
<-22.2
22.3-25.5
28.8-33.0
33.1 +
25.5-29.1
29.2+
(D)
(C)
0.40-
0.40
0.35-
0.35
0.30-
0.30
0^
0.25
0^0
0.20
0.15]
0.15
0.10
0.10<-22.4
25.6-28.7
22.5-24.6
24.7-26.5
26.6-28.8
28.9+
<-20.9
21.0-23.0
23.1-25.4
Body Mass Index
(Quintiles)
FIGURE 1. Observed (—) versus predicted (
) all-cause 10-year mortality rates by quintile of body mass Index (weight (kgyheight2 (m2))
for four race/sex groups, NUANES I Epidemiologic Follow-up Study (1971-1987). (A), black men; (B), black women; (C), white men; (D), white
women.
cern the necessary transformation to normality, we
used Tukey's "ladder of powers" (40) method. This
method consists of transforming the variable of interest, X, by raising it to a power. These powers are
chosen from the subset {-3, - 2 , - 1 , 0, 1, 2, 3}. The
logarithmic transformation is applied to X in addition
to these power transformations. A test for normality
based on skewness and kurtosis is performed on each
of the transformed variables (41) to determine the best
transformation. Application of the "ladder of powers"
method suggested Y = 1/BMI as the best transformation candidate for each of the race/sex groups under
consideration. Recently, Nevill and Holder (36), using
data from the Allied Dunbar National Fitness Study,
demonstrated that the reciprocal of BMI, which they
refer to as "lean body mass index," was also normally
distributed in their cohort, and was more closely related to percentage of body fat than was BMI.
Following the 1/BMI transformation, Bartlett's test
for equality of variance was calculated for decedents
compared with survivors. The p values were all sigAm J Epidemiol
Vol. 147, No. 8, 1998
nificant, suggesting real differences in the variances of
the two groups. As noted by Cornfield et al. (37), this
inequality of variances implies the necessity of including a squared term in the model.
For each of the four race/sex groups, we first derived the logistic regression model with 1/BMI and
1/BMI2, adjusting for age and smoking status. The
goodness of fit of the model was assessed by dividing
the BMI range into subgroups determined by quintiles
and then comparing the observed number of deaths
with the predicted number of deaths in each of the
BMI intervals (figure 1). Observed and predicted
probabilities were extremely close, with the possible
exception of the midpoint among black women. A
formal statistical goodness-of-fit test was performed
using Monte Carlo simulations (42).
The BMI corresponding to minimum mortality was
calculated on the basis of the quadratic form of the
logit derived for each of the four groups. Once the
logistic model containing terms for 1/BMI and 1/BMI2
had been derived, the value of 1/BMI corresponding
Downloaded from http://aje.oxfordjournals.org/ by guest on October 6, 2014
10-jfear A ll-cause fv
o
0.5&
742
Durazo-Arvizu et a).
baseline and those who died during the first 4 years of
follow-up. Stratified analyses including never smokers
and ever smokers were also completed.
RESULTS
The baseline characteristics of the analytic sample
are summarized in table 1. The average BMI was
similar in all groups, with the exception of higher
values among black women. Among the decedents, the
median survival time of women was longer than that
of men for both blacks and whites. As reported by
many other studies, with the exception of black men,
the average BMI among ever smokers was lower than
that among never smokers. Age-adjusted all-cause
mortality was 35 percent higher among black participants than among white participants (men and women
combined), and it was higher for all major categories
of mortality except coronary heart disease (table 2).
Table 3 presents the age-adjusted relative risks of
mortality by quintile of BMI, as well as age-adjusted
mortality rates. In these analyses, the quintile with the
lowest mortality rate was selected as the reference
category for calculation of relative risks. The middle
quintile was associated with the lowest mortality rate,
except for white men, among whom it occurred in the
next-to-highest quintile. Point estimates of the
TABLE 1. Data on key baseline variable* in a study of body mass index and mortality risk, NHANES I
EpMemtotoglc Follow-up Study, 1971-1987*
Mean
age
(yeare)
Total
sample
Body masslndext
Ever
smokers
Nonsmokers
of ever
smokers}
Mortally
rate§
Median
survival^
(days)
361(15)
2,783(168)
266(11)
3,169(208)
284 (5)
2,849 (58)
179 (4)
3,277 (70)
Black man
(737 participants and 328 deaths)
54 (15)#
25.6 (4.9)
25.6 (5.0)
25.5 (4.8)
67
Black women
(1,243 participants and 308 deaths)
48(15)
27.9(6.7)
27.1 (6.9)
28.5 (6.6)
43
White men
(4,644 participants and 1,463 deaths)
52 (15)
25.7 (4.0)
25.6 (4.0)
26.1 (4.1)
72
White women
(6,618 participants and 1,066 deaths)
48(15)
25.3 (5.4)
24.5 (5.2)
25.8 (5.4)
43
* Age-adjusted mortality rates vwre obtained by the direct method, with 10-year strata of the entire saxVracespecrfic cohort used as reference groups.
t Weight (kg)/heighti (m»).
X Includes both current smokers and ex-smokers.
§ 10-year age-adjusted mortality per 1,000 subjects.
U Medan survival times for those who died.
# Numbers in parentheses, standard deviation.
Am J Epidemiol
Vol. 147, No. 8, 1998
Downloaded from http://aje.oxfordjournals.org/ by guest on October 6, 2014
to minimum mortality was computed by setting the
derivative of the quadratic form of the logit equal to
0 and solving for 1/BMI. The reciprocal of this value
is the BMI level corresponding to minimum mortality.
A point estimate for this index was computed as
BMImin = - 2/§2//3,. Here /3, and /32 are the maximum likelihood estimates of terms associated with
1/BMI and 1/BMI2 in the logistic regression, respectively. Confidence intervals for the BMI,,,;,, were based
on the delta method (43). The BMI values associated
with relative risks of 1.1 and 1.2 above and below the
minimum were calculated with the model, and the
proportion of the population falling into those intervals
was determined; 1.2 was chosen arbitrarily as the
upper bound in the belief that it represents a level of
risk that most individuals would find acceptable.
Following the derivation of the logistic regression
model using only main effects, we examined the possibility of interaction with age. No effect was seen
(data not shown). In addition, we adjusted for educational level by introducing two indicator variables, the
first of which identified persons who attended school
through the 12th grade and the second of which identified those with post-high school instruction. To control for the potential effect of prevalent illness, we
carried out additional analyses after eliminating all
persons with cardiovascular disease and/or cancer at
Mortality and Body Mass Index
743
TABLE 2. Crude and age-adjusted mortality rates In the NHANES I Epidemlologlc Follow-up Study, by
sex and race, 1971-1987
Cause of death
All
causes
of
death
Coronary
heart dbease
(ICD-9* codes
410-414
and 428.2)
Stroke
(cerebrovascutar
dsaasa)
(ICD-9 codes
430-434 and
438-438)
Oner
cardiovascular
dteease
(CO-9 codes
390-448)
Cancer
All other
(al)
OCO-9 codes
140-208)
(another
CD-8 codes)
CflUSftfl
IOrude mortality ratef
Women
Black
White
175
248
161
51
56
50
20
38
16
18
31
16
39
43
38
47
Man
Black
White
333
445
315
113
106
115
24
37
22
31
57
26
75
71
90
146
81
20
33
17
42
46
42
51
85
45
27
66
83
64
81
121
99
80
41
Age-adjusted mortality rate\
192
266
57
61
179
57
22
41
18
Men
Black
White
294
361
99
83
102
21
28
20
284
46
24
74
* ICD-9, International Classification of Diseases, Ninth Revision.
110-year mortality par 1,000 subjects.
are given as the midpoint of the BMI interval of
minimum relative risk. For instance, for white men,
the BMI min is the midpoint of the interval 22.5-28.8,
since the age-adjusted rates for the second, third, and
fourth quintiles are indistinguishable (the 95 percent
confidence interval around 285 was 256-304). A substantial increase in risk was noted in the lowest quintile for all groups, especially black women. The ageadjusted mortality rates are depicted graphically in
figure 2.
Table 4 presents the estimated BMI^n based on the
described modeling technique, including adjustment
for age and smoking status. The minimum was also
calculated for the strata that included only persons
who had never smoked and current/past smokers. In
each case, the average value for the group is included
in the confidence interval. The goodness-of-fit p values based on 1,000 Monte Carlo simulations yielded p
values in the range 0.3-0.9 for each of the four race/
sex groups.
The BMIn,;,, was also estimated after eliminating
from the analysis individuals who died during the first
4 years of follow-up (table 5). On average, this procedure yielded values 0.3 BMI units lower than those
for the whole cohort; most of this reduction was observed among blacks. Further analyses were restricted
to the subset of individuals who had never smoked,
were free of cardiovascular disease and/or cancer at
Am J Epidemiol
Vol. 147, No. 8, 1998
baseline, and survived the first 4 years of follow-up.
To ensure sufficient power, we restricted analyses to
race/sex groups with at least 100 fatal events, i.e.,
white men and women. The quadratic relation between
BMI and mortality persisted for both never smokers
and ever smokers; however, the BMI min increased
modestly from 23.5 to 23.7 (95 percent confidence
interval 22.0-25.6) for never-smoker white men and
decreased from 24.8 to 24.7 (95 percent confidence
interval 22.9-26.9) for never-smoker white women.
To obtain an estimate of the impact of the risk of
death associated with variation in BMI from the estimated BMIj^j,,, we calculated the proportions of the
sample within intervals associated with 10 percent and
20 percent increases in risk (table 6). A weighted
average of the group-specific estimates suggested that
at least half of the sample experienced no more than a
10 percent increase in risk, while three quarters of the
sample had an increase of no more than 20 percent.
Based on the confidence limits around the endpoints of
the risk intervals, the proportions of the cohort contained within the risk interval would be considerably
more extreme.
DISCUSSION
In this analysis of a representative sample of the US
population, we found consistent evidence of a non-
Downloaded from http://aje.oxfordjournals.org/ by guest on October 6, 2014
Women
Black
White
744
Durazo-Arvizu et a).
TABLE 3. Ten-yaar age-adjusted mortality rate and relative risk of death, by sex, rat» , and quintila of
body mass index: NHANES 1 Epfdemlologlc Follow-up Study, 1971-1987*
Qulntlleof
body mass Indexf
Hack man (26.7)$
£21.4
21.5-23.9
24.0-26.1
26.2-29.2
£29.3
Black women (27.2)
£22.2
22.2-25.5
25.6-28.7
28.8-33.0
£33.1
White women (25.7)
£20.9
21.0-23.0
23.1-25.4
25.5-29.1
£29.2
84
79
54
52
59
64
62
52
61
69
324
279
293
251
316
164
173
185
254
290
No. of
subjects
Relative risk
of death
Mortally rate
per 1,000
subjects
96%
confidence
Interval
1.59
1.33
1.00
1.04
1.18
576
482
361
375
425
512-648
413-547
290-^130
247
.71
.24
.00
.28
.39
323
235
189
242
263
269-371
187-273
148-234
192-288
213-307
921
918
969
894
942
1.24
1.05
1.06
1.00
1.17
354
300
303
285
334
324-376
275-325
276-324
256-304
305-^355
1,322
1.35
1.17
1.00
1.08
1.25
189
163
139
151
169-211
141-179
123-157
134-166
154-186
145
147
150
148
147
248
249
248
251
1,336
1,310
1,320
1,330
175
298-442
353-^187
• The age-adjusted rates wore obtained by the direct method, with 10-year strata of the entire sex-/racespecific cohort used as reference groups,
t Weight (kg)/height« (mi),
i Numbers in parentheses, body mass index of minimum mortality.
monotonic, U-shaped relation between BMI and mortality risk. A mathematical model in which the distribution of BMI values was made normal by taking its
inverse predicted the mortality experience with precision, and allowed estimation of the parameters of
interest using the full range of data. The point of
minimum mortality, estimated by this model, was on
average 0.4 BMI units below the group-specific means
(range, —1.1 to +1.5); in every instance, the 95 percent confidence interval of the BMI min included the
group mean. In addition, 70 percent of the population
was included in the range of BMI values which conferred no more than a 20 percent increase in all-cause
mortality risk.
Surprisingly, despite the obvious statistical advantages, no prior studies have attempted to model the
relation between BMI and mortality using individuallevel data. Not only are the data used in the most
informative manner, arbitrary grouping is avoided,
statistical comparisons between populations become
possible, and the role of confounders can be more
effectively assessed. As a consequence, other aspects
of the epidemiology of obesity can be examined. As
noted above, the point estimate of BMI^,, varied
across groups, and the confidence intervals for the
most extreme comparison—black men versus white
women—just barely overlapped. Some investigators
have advanced a strong hypothesis regarding the role
of smoking and prevalent disease as either confounders or effect modifiers (8, 44). Despite its appeal and
widespread currency, very few empirical tests of this
hypothesis exist. Our data demonstrate that the impact
of these factors on the BMI-mortality relation is very
limited in the general population.
Given our interest in the role of relative weight in
the general population, this cohort provides a nearly
ideal data set, because it is a representative population
sample and the results are therefore generalizable. The
observation period for this cohort, which extended
through the late 1980s, increases its relevance. The
outcome of this analysis is broadly consistent with the
current US Department of Agriculture recommendations defining healthy weight as a BMI of 25 or less,
with an acceptable interval 6 BMI units wide (45).
Despite this agreement, our results are unusual in the
extent of the upturn in risk at lower BMI values;
Am J Epidemiol
Vol. 147, No. 8, 1998
Downloaded from http://aje.oxfordjournals.org/ by guest on October 6, 2014
White man (25.7)
£22.4
22.5-24.6
24.7-26.5
26.6-28.8
£28.9
No. of
deaths
Mortality and Body Mass Index
745
(B)
(A)
0.6
0.6-1
0.5
0.5-
0.4-
0.4-
0.3-
0.3-
02
02-
0.1 J
0.1-
CD
cd
"§
<-21.4
215-23.9
24.0-26.1
26.2-29.2
29.3+
<-22.2
22.3-25.5
25.6-28.7
28.8-33.0
33.1 +
25.5-29.1
29.2+
(D)
(C)
0.4-.
0.4
0.3-
0.3-
0.2-
0.2-
0.1 J
0.1 J
CD
ca
O
<-22.4
22.5-24.6
24.7-26.5
26.6-28.8
28.9+
<-20.9
21.0-23.0
23.1-25.4
Body Mass Index
(Quintiles)
FIGURE 2. Age-adjusted 10-year mortality rates by quintile of body mass Index (weight (kgj/height2 (m2)) for four race/sex groups, NHANES
I Epidemlologic Follow-up Study (1971-1987). (A), black men; (B), black women; (C), white men; (D), white women.
likewise, the optimal BMI found here would be considered high in the current debate, since the lower
bound of current recommendations extends to 19 (45).
The age range of these participants was broader than
that in many prior studies, which could potentially
have influenced the outcome (17,46), and the effect of
statistical adjustment for age is apparent in the contrast
between figures 1 and 2. Despite these caveats, since
recommendations are formulated for the entire population, it would be even more unreasonable to restrict
the data being analyzed to a specific age range.
Two previous reports from the NHANES I Epidemiologic Follow-up Study have addressed the issue of
mortality risk and relative weight among the elderly
(47, 48). Quantitatively similar outcomes were observed, and the U-shaped curve was apparent in all
subgroups, although no attempt was made to model
these relations. The increase in risk among lean individuals was particularly prominent among blacks;
however, these results were based on a relatively small
number of events and must be interpreted with appropriate caution. The higher BMI,,,;,, in blacks does sugAm J Epidemiol
Vol. 147, No. 8, 1998
gest, however, that while the shape of the relation may
be consistent across populations, the position of the
nadir could vary with the mean.
The value of the statistical approach proposed here
depends on several assumptions. First, it must be accepted that the underlying pattern of mortality being
modeled is an accurate description of the experience of
relevant human populations. Second, this pattern must
not have been distorted to any significant degree by
confounding factors, and therefore must reflect direct
cause-and-effect relationships. The first assumption
can be satisfied, at the level of phenomenology, by the
demonstration that the U-shaped relation is the "true"
or characteristic finding. The second assumption, on
the other hand, raises complex questions about etiologic processes which can only be answered by invoking an inference. Below, we briefly address the extent
to which each of these assumptions can be satisfied.
Few investigators doubt the increase in health risk
associated with obesity. Although controversy does
exist over the occurrence of excess mortality among
persons who are lean, the majority of population-based
Downloaded from http://aje.oxfordjournals.org/ by guest on October 6, 2014
i_
746
Durazo-Arvizu et al.
TABLE 4. Estimated body mass Index (BMI) of minimum mortality based on logistic regression
analysis, by sex and race: NHANES 1 Epidemlologlc Follow-up Study, 1971-1987
Adjustment
factors)
Mean
BMI*
No. oJ
subjects
No. of
deaths
M*Lnt
95%
corfktenca
interval
27.2(1.20)$
26.7(1.43)
27.7(1.79)
27.1 (1.17)
24.8-29.5
23.9-29.5
24.2-31.2
24.8-29.4
26.9(1.08)
27.4 (2.44)
26.5(1.15)
26.8(1.06)
24.8-29.0
22.6-32.2
24.3-28.8
24.7-28.9
25.0
25.4
24.2
24.8
(0.55)
(0.83)
(0.69)
(0.53)
23.9-26.1
23.7-27.0
22.8-25.5
23.8-25.9
24.7 (0.54)
23.8(0.91)
24.8 (0.65)
24.3 (0.54)
23.7-25.8
22.0-25.5
23.5-26.1
23.3-25.4
Black men
Age
Ever smokers
Never smokers
Age and smoking
25.6
25.2
25.9
737
493
244
Age
Ever smokers
Never smokers
Age and smoking
27.9
26.6
28.6
1,243
537
706
328
208
120
Black women
308
108
200
White men
25.7
25.2
26.1
4,644
3,395
1,305
1,463
1,072
391
White women
Age
Ever smokers
Never smokers
Age and smoking
25.3
24.1
25.8
6,618
2,840
3,778
1,066
414
652
* Weight (kg)/heighti (m»).
t BMI of minimum mortality.
t Numbers in parentheses, standard error (estimates based on delta method).
studies demonstrate some degree of upturn at the
lower end of the weight-for-height distribution (3-7,
12-18). If we accept this observation as being characteristic of unselected population samples, the interpretation of the causal process remains much in doubt.
Two basic hypotheses have been put forward. Some
investigators believe that observations made in whole
populations are fatally flawed by the impact of confounding factors, primarily smoking and illnessrelated weight loss. In the face of this potential limitation, it is argued that analyses of highly selected,
homogeneous subgroups are more revealing of the true
cause-and-effect relationship (10, 15). An alternative
point of view suggests that representative population
samples provide more reliable findings, since the
quantitative impact of confounding is likely to be
complex and heterogeneous, both between populations
and across the range of BMIs within populations, and
the relation between BMI and mortality may therefore
be biased among subgroups in unpredictable ways.
Finally, a series of long-term studies have yielded data
suggesting that excessive leanness may actually be a
cause of illness, most notably lung cancer (49-52). If
this is true, elimination of the susceptible members of
the lean subgroup will obviously bias the outcome.
Unfortunately, in the absence of complete knowledge
of confounding factors, hypotheses related to the
shape of the "true" BMI-mortality relation cannot be
tested directly.
Do other means of testing these hypotheses exist? It
has been suggested that observational studies of
weight change provide unique, and perhaps better,
information about the risk of overweight. An overview
of these studies supported the view that weight change
values close to the population mean are associated
with the best outcome and that persons who fail to gain
weight have increased risk (46). In the face of this
additional evidence, we conclude that there is no basis
for rejecting the hypothesis that mortality risk increases at both extremes of weight.
Because a nonmonotonic model best describes the
relation between BMI and mortality in the sample
studied here, before and after control for confounders,
the model which was developed has several advantages. The full range of data can be used in a single
analysis without the need to define arbitrary quantiles.
In addition, the parameters associated with the nadir of
the curve, and its confidence interval, can be specified.
The estimation of a "range of normal values" is further
aided by the use of this model. Incremental increases
Am J Epidemiol Vol. 147, No. 8, 1998
Downloaded from http://aje.oxfordjournals.org/ by guest on October 6, 2014
Age
Ever smokers
Never smokers
Age and smoking
Mortality and Body Mass Index
747
TABLE 5. Estimated body mass index (BMI) of minimum mortality based on logistic regression
analysis, excluding persons who died during the first 4 years of follow-up, by sex and race: NHANES I
Epidemiologic Follow-up Study, 1971-1987
Adjustment
factors)
No. of
subjects
Mean
BMI*
No. of
deaths
95%
BMI^f
confidence
Interval
26.4 (0.89)$
26.3(1.17)
26.5(1.23)
26.4 (0.88)
24.7-28.2
24.0-28.6
24.1-29.0
24.7-28.1
26.2 (0.99)
27.7 (2.63)
25.4(1.10)
26.2 (0.98)
24.3-28.2
22.5-32.8
23.3-27.6
24.2-28.1
24.8 (0.61)
25.5(1.01)
23.5 (0.82)
24.6 (0.59)
23.6-26.0
23.5-27.5
21.9-25.2
23.5-25.8
24.9
24.1
24.8
24.5
23.7-26.0
22.3-25.9
23.3-26.2
23.3-25.6
Black man
Age
Ever smokers
Never smokers
Age and smoking
246
159
87
655
444
211
25.8
25.7
25.9
Black woman
Age
Ever smokers
Never smokers
Age and smoking
28.0
27.1
28.6
1,177
509
668
25.8
25.6
26.2
4,285
3,074
1,211
242
80
162
White men
1,104
807
297
White women
Age
Ever smokers
Never smokers
Age and smoking
6,429
2,766
3,663
25.2
24.4
25.8
877
340
537
(0.58)
(0.92)
(0.73)
(0.58)
• Weight (kg)Aieight» (m»).
t BMI of minimum mortality.
i Numbers in parentheses, standard error (estimates based on delta method).
in mortality risk can be examined, and the relation
with the underlying distribution of BMIs in the population can be determined. Across the great majority of
attained values, the excess risk was well under 20
percent in comparison with the minimum, suggesting
that most individuals do not experience a substantial
increase in mortality risk from modest overweight or
underweight. The proposed model also makes it possible to compare population subgroups in a systematic
fashion. In partial confirmation of the generalizability
of the shape of the BMI-mortality relation, a similar
model fitted the experience of each group.
This analysis has potential limitations as a basis for
inferring cause and effect, most of which are inherent
in all such studies, as described above. First, data on
morbid events were not presented here, since this was
not a focus of this paper, although morbidity is clearly
an important consequence of obesity. Second, the pop-
TABLE 6. Range of body mass Index (BMI) values associated with an Increase In mortality risk of no
more than 10% or 20%, by sex and race: NHANES I Epidemiologic Follow-up Study, 1971-1987*
Increase Inriskerf £10%
Black men
Black women
White men
White women
Weighted average§
hcrease In rbk of £20%
BM It range
%ol
population?
BMI range
25.0-29.7
23.3-31.5
22.0-28.4
21.3-28.4
34
48
62
53
23.3-32.5
22.2-33.8
21.1-30.0
2O.3-30.4
55
%0l
population^
56
63
77
69
70
* These values were based on logistic regression analysis, adjusting for age and smoking status at baseline.
t Weight (kgyneighti (m»).
X Percentage of individuals whose BMI fell within the specified range.
§ Weighted by the number of subjects per group.
Am J Epidemiol
Vol. 147, No. 8, 1998
Downloaded from http://aje.oxfordjournals.org/ by guest on October 6, 2014
Age
Ever smokers
Never smokers
Age and smoking
748
Durazo-Arvizu et al.
ACKNOWLEDGMENTS
This work was supported in part by Cooperative Agreement U83/CCU512480 from the Centers for Disease Control and Prevention and grant DK 52329 from the National
Institutes of Health.
REFERENCES
1. van Itallie TB. Health implications of overweight and obesity
in the United States. Ann Intern Med 1985;103:983-8.
2. Lipton RB, Liao Y, Cao G, et al. Determinants of incident
non-insulin-dependent diabetes mellitus among blacks and
whites in a national sample: The NHANES I Epidemiologic
Follow-up Study. Am J Epidemiol 1993;138:826-39.
3. Schroll M. A longitudinal epidemiological survey of relative
weight at age 25, 50 and 60 in the Glostrup population of men
and women born in 1914. Dan Med Bull 1981;28:106-16.
4. Tuomilehto J, Salonen JT, Marti B, et al. Body weight and risk
of myocardial infarction and death in the adult population of
eastern Finland. Br Med J (Clin Res Ed) 1987;295:623-7.
5. Vandenbroucke JP, Mauritz BJ, de Bruin A, et al. Weight,
smoking, and mortality. JAMA 1984;252:2859-60.
6. Stevens J, Keil JE, Rust PF, et al. Body mass index and body
girths as predictors of mortality in black and white men. Am J
Epidemiol 1992; 135:1137-46.
7. Association of Life Insurance Medical Directors of America
and Society of Actuaries. Build Study, 1979. Philadelphia,
PA: Recording and Statistical Corporation, 1980.
8. Garrison RJ, Feinleib M, Castelli WP, et al. Cigarette smoking
as a confounder of the relationship between relative weight
and long-term mortality: The Framingham Heart Study.
JAMA 1983;249:2199-203.
9. Lindsted K, Tonstad S, Kuzma JW. Body mass index and
patterns of mortality among Seventh-day Adventdst men. Int J
Obes 1991;15:397-406.
10. Lee IM, Manson JE, Hennekens CH, et al. Body weight and
mortality: a 27-year follow-up of middle-aged men. JAMA
1993;270:2823-8.
11. Wilcosky T, Hyde J, Anderson JJ, et al. Obesity and mortality
in the Lipid Research Clinics Program Follow-up Study.
J Clin Epidemiol 1990;43:743-52.
12. Folsom AR, Kaye SA, Sellers TA, et al. Body fat distribution
and 5-year risk of death in older women. JAMA 1993;269:
483-7.
13. Comstock GW, Kendrick MA, Livesay VT. Subcutaneous
fatness and mortality. Am J Epidemiol 1966;83:548-63.
14. Lew EA, Garfinkel L. Variations in mortality by weight
among 750,000 men and women. J Chronic Dis 1979;32:
563-76.
15. Manson JE, Willett WC, Stampfer MJ, et al. Body weight and
mortality among women. N Engl J Med 1995;333:677-85.
16. Waaler HT. Height, weight and mortality: the Norwegian
experience. Acta Med Scand Suppl 1984;679:1-56.
17. Harris T, Cook EF, Garrison R, et al. Body mass index and
mortality among nonsmoking older persons: The Framingham
Heart Study. JAMA 1988;259:1520-4.
18. Andres R. Mortality and obesity: the rationale for age-specific
height-weight tables. In: Andres R, Bierman EL, Hazzard WR,
eds. Principles of geriatric medicine. New York, NY:
McGraw-Hill Book Company, 1985:311-18.
19. Heymsfield SB, Darby PC, Muhlheim LS, et al. The calorie:
myth, measurement, and reality. Am J Clin Nutr 1995;
62(suppl):1034S-41S.
20. Lands WE. Alcohol and energy intake. Am J Clin Nutr 1995;
62(suppl):1101S-1106S.
21. Colditz GA, Giovannucci E, Rimm EB, et al. Alcohol intake
in relation to diet and obesity in women and men. Am J Clin
Nutr 1991;54:49-55.
22. Williamson DF, Forman MR, Binkin NJ, et al. Alcohol and
body weight in United States adults. Am J Public Health
1987;77:1324-30.
23. Comoni-Huntley J, Barbano HE, Brody JA, et al. National
Health and Nutrition Examination I—Epidemiologic
Follow-up Survey. Public Health Rep 1983;98:245-51.
24. Hillar H. Plan and operation of the Health and Nutrition
Examination Survey, United States, 1971-1973. (Vital and
Health Statistics, Series 1, nos. 10a, 10b). Hyattsville, MD:
National Center for Health Statistics, 1973. (DHEW publication no. (HSM) 73-1310).
25. Engel A, Maurer K, Collins E. Plan and operation of the
HANES I Augmentation Survey of adults 25-74 years: United
States, 1974-1975. (Vital and Health Statistics, Series 1, no.
Am J Epidemiol
Vol. 147, No. 8, 1998
Downloaded from http://aje.oxfordjournals.org/ by guest on October 6, 2014
ulation of inference was restricted to blacks and whites
living in the United States. Third, a mathematical
model can distort the underlying relations if the fit is
not precise. We attempted to guard against this potential bias in the analytic process by using both quantiles
and model-based approaches, although goodness of fit
was formally tested. As noted above, the outcomes in
terms of the shape of the relation and the apparent
minimum were consistent with both methods. Finally,
the transformation used for these data may not be
appropriate for all populations.
No disagreement exists about the health risk associated with obesity, and the findings for the middle of
the distribution are likewise consistent. Even studies
which emphasize a positive monotonic relation between BMI and risk do not demonstrate a significant
difference between the lowest quantiles and those at
the mean (10, 15). Given the high prevalence of obesity in the United States and the acknowledged risk, it
would seem appropriate to provide the public with
information regarding this well-established relation
and to reserve the discussion of benefit or harm associated with leanness for scientific debate.
In summary, we used a mathematical function to
model the U-shaped relation between BMI and mortality observed in a random sample of the US population. In each of the four race/sex groups, the BMI min
was close to the sample mean, and values associated
with no more than a 20 percent increase in risk included almost three quarters of the population. Analyses restricted to never smokers yielded similar results. We placed emphasis on the entire population, in
analogy with the "intention-to-treat" convention of
randomized trials. Stratification was carried out only
for subgroups in which a strong rationale existed for a
confounding effect, and the outcome was interpreted
as complementary to findings in the whole sample,
particularly since the general relation was similar in all
groups. Discrepancies with prior studies may reflect
the choice of the population sample and the definition
of subgroups.
Mortality and Body Mass Index
26.
27.
28.
29.
31.
32.
33.
34.
35.
36.
37.
Am J Epidemiol
Vol. 147, No. 8, 1998
38. Kay R, Little S. Transformations of the explanatory variables
in the logistic regression model for binary data. Biometrika
1987;74:495-501.
39. Guerrero VM, Johnson RA. Use of the Box-Cox transformation with binary response models. Biometrika 1982;69:
309-14.
40. Tokey JW. Exploratory data analysis. Reading, MA: AddisonWesley Publishing Company, 1977.
41. D'Agostino RB, Balanger A, D'Agostino RB Jr. A suggestion
for using powerful and informative tests of normality. Am Stat
1990;44:316-21.
42. Noreen EW. Computer intensive methods for testing
hypotheses: an introduction. New York, NY: John Wiley and
Sons, Inc, 1989.
43. Miller RG. Survival analysis. New York, NY: John Wiley and
Sons, Inc, 1981.
44. Manson JE, Stampfer MJ, Hennekens CH, et al. Body weight
and longevity: a reassessment JAMA 1987;257:353-8.
45. US Department of Agriculture, Agricultural Research Service,
Dietary Guidelines Advisory Committee. Report of the Dietary Guidelines Advisory Committee on the dietary guidelines for Americans, 1995, to the Secretary of Health and
Human Resources and the Secretary of Agriculture. Washington, DC: US Department of Agriculture, 1995.
46. Andres R, Muller DC, Sorkin JD. Long-term effects of change
in body weight on all-cause mortality: a review. Ann Intern
Med 1993;119:737-43.
47. Tayback M, Kumanyika S, Chee E. Body weight as a risk
factor in the elderly. Arch Intern Med 1990;150:1065-72.
48. Comoni-Huntley JC, Harris TB, Everett DF, et al. An overview of body weight of older persons, including the impact on
mortality: The National Health and Nutrition Examination
Survey I—Epidemiologic Follow-up Study. J Clin Epidemiol
1991;44:743-53.
49. Knekt P, Heliovaara M, Rissanen A, et al. Leanness and
lung-cancer risk. Int J Cancer 1991;49:208-13.
50. Kabat GC, Wynder EL. Body mass index and lung cancer risk.
Am J Epidemiol 1992;135:769-74.
51. Goodman MT, Wilkens LR. Relation of body size and the risk
of lung cancer. Nutr Cancer 1993;20:179-86.
52. Kark JD, Yaari S, Rasooly I, et al. Are lean smokers at
increased risk of lung cancer? The Israel Civil Servant Cancer
Study. Arch Intern Med 1995;155:24O9-16.
Downloaded from http://aje.oxfordjournals.org/ by guest on October 6, 2014
30.
14). Hyattsville, MD: National Center for Health Statistics,
1978. (DHEW publication no. (PHS) 78-1314).
National Center for Health Statistics. Plan and operation of the
NHANES I Epidemiologic Follow-up Study, 1987. (Vital and
Health Statistics, Series 1, no. 27). Hyattsville, MD: National
Center for Health Statistics, 1992. (DHEW publication no.
(PHS) 92-1303).
Abraham S, Johnson CL, Najjar MF. Weight by height and
age for adults 18-74 years, United States, 1971-1974: age and
sex distributions of weight by single inches of height for
adults, 18-74 years of age in the civilian, noninstitutional
population of the United States. (Vital and Health Statistics,
series 11, no. 208). Hyattsville, MD: National Center for
Health Statistics, 1978. (DHEW publication no. (PHS)
78-1656).
Cooper RS, Ford E. Comparability of risk factors for coronary
heart disease among blacks and whites in the NHANES-I
Epidemiologic Follow-up Study. Ann Epidemiol 1992;2:
637-45.
Liao Y, Cooper RS, McGec DL. Iron status and coronary heart
disease: negative findings from the NHANES I Epidemiologic
Follow-up Study. Am J Epidemiol 1994; 139:704-12.
World Health Organization. International classification of diseases. Manual of the international statistical classification of
diseases, injuries, and causes of death. Ninth Revision. Geneva, Switzerland: World Health Organization, 1977.
McLaughlin JK, Dietz MS, Mehl ES, et al. Reliability of
surrogate information on cigarette smoking by type of informant. Am J Epidemiol 1987;126:144-6.
Machlin SR, Kleinman JC, Madans JH. Validity of mortality
analysis based on retrospective smoking information. Stat
Med 1989;8:997-1009.
Hosmer DW Jr, Lemeshow S. Applied logistic regression.
New York, NY: John Wiley and Sons, Inc, 1989.
Cox DR. Regression models and life tables (with discussion).
J R Stat See B 1972;34:187-220.
Dobson AJ. An introduction to generalized linear models.
London, England: Chapman and Hall Ltd, 1990.
Nevill AM, Holder RL. Body mass index: a measure of fatness
or leanness? Br J Nutr 1995;73:507-16.
Cornfield J, Gordon T, Smith WW. Quantal response curves
for experimentally uncontrolled variables. Bull Inst Int Stat
1961;38:97-115.
749