International Journal of Obesity (1999) 23, 603±611
ß 1999 Stockton Press All rights reserved 0307±0565/99 $12.00
http://www.stockton-press.co.uk/ijo
Weight loss increases and fat loss decreases allcause mortality rate: results from two
independent cohort studies
DB Allison1*, R Zannolli1,2, MS Faith1, M Heo1, A Pietrobelli1,3, TB VanItallie1, FX Pi-Sunyer1 and SB Heyms®eld1
1
Obesity Research Center, St Luke's=Roosevelt Hospital, Columbia University College of Physicians and Surgeons, New York, NY, USA;
Department of Pediatrics, Policlinico LeScotte, University of Siena, Italy; and 3Department of Pediatrics, Scienti®c Institute H San
Raffaele, University of Milan, Milan, Italy
2
OBJECTIVE: In epidemiological studies, weight loss is usually associated with increased mortality rate. Contrarily,
among obese people, weight loss reduces other risk factors for disease and death. We hypothesised that this paradox
could exist because weight is used as an implicit adiposity index. No study has considered the independent effects of
weight loss and fat loss on mortality rate. We studied mortality rate as a function of weight loss and fat loss.
DESIGN: Analysis of `time to death' in two prospective population-based cohort studies, the Tecumseh Community
Health Study (1890 subjects; 321 deaths within 16 y of follow-up) and the Framingham Heart Study (2731 subjects; 507
deaths within 8 y of follow-up), in which weight and fat (via skinfolds) loss were assessable.
RESULTS: In both studies, regardless of the statistical approach, weight loss was associated with an increased, and
fat loss with a decreased, mortality rate (P < 0.05). Each standard deviation (s.d.) of weight loss (4.6 kg in Tecumseh,
6.7 kg in Framingham) was estimated to increase the hazard rate by 29% (95% con®dence interval CI), (14%, 47%,
respectively) and 39% (95% CI, 25%, 54% respectively), in the two samples. Contrarily, each s.d. of fat loss (10.0 mm in
Tecumseh, 4.8 mm in Framingham) was estimated to reduce the hazard rate 15% (95% CI, 4%, 25%) and 17% (95% CI,
8%, 25%) in Tecumseh and Framingham, respectively. Generalisability of these results to severely (that is, body mass
index BMI) 34) obese individuals is unclear.
CONCLUSIONS: Among individuals that are not severely obese, weight loss is associated with increased mortality
rate and fat loss with decreased mortality rate.
Keywords: body mass index; fat loss; weight loss; longevity; mortality; obesity
Introduction
There is an apparent paradox in the obesity ®eld.
Obesity is associated with numerous morbidities1 and
even small weight losses can be associated with shortterm reductions in risk factors for disease.2,3 Obesity
is also associated with an increased all-cause mortality
rate.4 In contrast, the majority of studies show that
weight loss is associated with an increased mortality
rate.5
This paradox may be attributable to: a) not restricting studies of weight loss to obese people; b) not
controlling for smoking status; c) not distinguishing
among different methods and causes of weight loss;
and d) not controlling for pre-existing disease.6,7 Only
two epidemiological studies have addressed these
concerns.8,9 Williamson et al 8 found that, among
never-smoking obese women who were apparently
free from wasting diseases and apparently lost
weight intentionally, weight loss was associated with
*Correspondence: Dr David B. Allison, Obesity Research Center,
St Luke's=Roosevelt Hospital, Columbia University College of
Physicians and Surgeons; 1090 Amsterdam Avenue, 14th ¯oor,
New York, NY 10025, USA..
E-mail: [email protected]
a decreased mortality rate. However, several factors
mitigate against immediate acceptance of the idea that
intentional weight loss, among obese people, decreases
the mortality rate. For example, there was no doseresponse relationship between amount of weight lost
and reduction of mortality in this study. Moreover, the
bene®cial effects of weight loss on mortality rate were
almost exclusive to women with co-morbidities.8 Also
noteworthy is a more recent study by Williamson et
al 9 presenting similar analyses for men. They found no
effect of apparently `intentional' weight loss on the allcause mortality rate.
An additional complication when interpreting previous data is the fact that studies used weight and
weight-based indices (for example, body mass index
(BMI), kg=m2) as implicit indicators of body fatness.
It is true that, especially after controlling for height,
weight is highly correlated with body fat. However,
weight is also highly correlated with the amount of
lean mass individuals have,10,11 making interpretation
of relations between weight and outcomes, such as
mortality, dif®cult. Indeed, Allison et al 12 combined
theoretical models and actual data to show that the Ushaped relationship frequently observed between BMI
and mortality could be a function of increasing fat
mass having a linear and increasing effect on mortality rate, but increasing lean mass having a linear and
Fat loss and mortality
DB Allison et al
604
decreasing effect on mortality rate. This is consistent
with other research implicating body composition,
rather than body weight per se, in determining
health risk. For example, Segal et al 13 compared
obese men to a group of football players, wrestlers
and power lifters who were matched for age and BMI,
but had a very low amount of body fat. Among these
subjects, the obese individuals had signi®cantly
greater elevations in cardiovascular disease (CVD)
risk factors (diastolic blood pressure (DBP), fasting
plasma insulin, lower high density lipoprotein (HDL)cholesterol and greater low density lipoprotein (LDL)cholesterol) than did the BMI-matched but low body
fat group.
This thinking about differential effects of fat and
lean mass on health from static models of body
composition can also be applied to dynamic models.
That is, greater clarity on the effects of weight
changes on health risks and mortality may be produced by separately considering the differential
effects of weight change per se vs changes in the
degree of fatness. Van Itallie and Yang14 suggested
that the degree of health bene®t achieved from weight
loss is likely to be dependent on the degree to which
the weight is lost as fat and lean mass is preserved.
Given the above, we hypothesised that, conditional
on fat loss, weight loss may be associated with
increased mortality rate because it is accompanied
by an undesirable loss of lean mass. Conversely, we
hypothesised that, conditional on weight loss, fat loss
will be associated with decreased mortality rate. We
tested these hypotheses, using data from two prospective longitudinal population-based cohort studies: 1)
the Tecumseh Community Health Study; and 2) the
Framingham Heart Study.
Methods
The Tecumseh sample
The Tecumseh Community Health Study is a prospective epidemiological study of residents from
Table 1
Tecumseh, Southeast Michigan. Data collection
began in 1957. Information was obtained from baseline medical history interviews, medical examinations, clinical measurements, laboratory work and
electrocardiograms. Mortality follow-up data for the
8637 white respondents to the original data collection,
was provided by the Tecumseh Mortality Follow-Up
Study.15 In data analyzed herein, `Mortality of cohort
members was monitored through local newspapers
and from individual questioning of Tecumseh residents. Recontact with all eligible members of the
cohort was attempted in 1977 ± 1979 with 99 percent
of success rate. Death certi®cates were obtained for all
deceased members of the cohort'.16 Time 1 corresponds to the baseline measurement. Time 2 was 4 y
later. Follow-up data were available for 16 y after
Time 2. During these 16 y, 321 deaths occurred (187
men; 134 women). Details on excluded and included
subjects are given in Table 1.
Subjects. Details of 1890 individuals (887 men and
1003 women) who had complete weight and skinfold
data at Time 1 and Time 2 and complete mortality
data are presented in Table 2.
Variables. Weight-change was calculated as weight
(kg) at Time 2 minus weight (kg) at Time 1 (positive
values therefore denote weight gains and negative
values weight loss). Two indicators of (subcutaneous)
fatness were available at both Time 1 and Time 2;
subscapular and triceps skinfold measurements. For
details on the measurements of these skinfolds in the
Tecumseh Community Health Study, see Ref. 17. Fatloss was calculated as skinfold thickness (triceps plus
subscapular) at Time 2 minus skinfold thickness
(triceps plus subscapular) at Time 1. An additional
analysis (data not shown) which converted the two
different skinfolds to a z-score, prior to summing, as
recommended by Wainer,18 produced essentially identical results. `Time to death' was the time to death
from any cause after Time 2.
Subjects excluded from included in the Tecumseh Community Health Study and Framingham Heart Study
Initial cohort
Respondents remaining after eliminating those with incomplete mortality follow-up
Respondents remaining after excluding those with missing BMI at Time 1
Respondents remaining after excluding those with missing weight loss informationa
Respondents remaining after excluding those with missing fat loss information
Respondents remaining after eliminating children (that is, age 18 y and under) at Time 1
Deaths among respondents included
Tecumseh study (n)
(M=F)
Framingham study (n)
(M=F)
8637
(4237=4400)
8601
(4220=4381)
8034
(3957=4077)
6027
(2937=3090)
3346
(1369=1707)
1890
(887=1003)
321
(187=134)
5209
(2336=2873)
5209
(2336=2873)
4405
(1941=2464)
3129
(1327=1802)
2731
(1160=1571)
2731
(1160=1571)
507
(290=217)
BMI ˆ body mass index.
Many subjects with missing BMI data at Time 2 are those that died between Time 1 and Time 2.
a
Fat loss and mortality
DB Allison et al
Table 2
605
Characteristics of the included subjects from the Tecumseh Community Health Study
Time 1
Age (y)
Smokers=Non-Smokers (n)b
BMI (kg=m2)
Height (cm)
Weight (kg)
Triceps skinfold (mm)
Subscapular skinfold (mm)
Time 2c
Weight (kg)
Triceps skinfold (mm)
Subscapular skinfold (mm)
Change scores
Weight-change (kg)
Fat-change (mm)
All
(n ˆ1890)
Men
(n ˆ 887)
Women
(n ˆ1003)
40.3 14.1 (18.0, 91.0)a
1159=731
25.1 4.3 (15.0, 49.2)
167.0 9.2 (138.0, 194.0)
70.3 13.8 (39.0, 139.0)
17.1 7.3 (3.0, 47.0)
19.9 8.6 (4.0, 58)
40.6 13.5 (18.0, 89.0)
706=181
25.5 3.5 (16.3, 43.9)
174.1 6.6 (152.0, 194.0)
77.3 11.8 (42.0, 139.0)
13.01 5.6 (3.0, 38.0)
16.8 9.8 (5.0, 58.0)
39.9 14.6 (18.0, 91.0)
453=550
24.8 4.8 (15.1, 49.2)
160.7 6.1 (138.0, 180.0)
64.2 12.5 (39.0, 126.0)
20.7 6.8 (4.0, 47.0)
18.6 9.3 (4.0, 58.0)
71.0 14.1 (37.0, 140.0)
21.0 9.5 (3.0, 61.0)
18.9 9.9 (4.0, 68.0)
78.2 11.9 (44.0, 140.0)
15.6 6.2 (3.0, 50.0)
17.8 8.0 (4.0, 61.0)
64.5 12.9 (37.0, 132.0)
25.8 9.3 (4.0, 61.0)
19.9 11.3 (4.0, 68.0)
0.66 4.6 (ÿ22.0, 19.0)
5.1 10.0 (ÿ39.0, 53.0)
0.96 4.3 (ÿ21.0, 13.0)
6.4 11.0 (ÿ25.0, 53.0)
0.39 4.9 (ÿ22.0, 19.0)
3.7 8.4 (ÿ39.0, 32.0)
BMI ˆ body mass index.
a
All data are means s.d. (range).
b
Ever=never.
c
Time 2 is four years later than Time 1.
The Framingham sample
The Framingham Heart Study is a longitudinal, prospective cohort study initiated in 1948 to study 5209
residents in Framingham, Massachusets. Participants
have undergone biennial examinations since the
study's inception.19 These included measurements of
height, weight and various risk factors. At the 5th and
12th biennial examinations, measurements of skinfolds were also made. Examination procedures and
mortality follow-up information have been described
in detail elsewhere.20 Time 1 corresponds to the
measurement taken between 1954 and 1956. Time 2
was 14 y later. Follow-up data were available for eight
years after Time 2. During these eight years, 507
deaths occurred (290 men; 217 women). Details on
subjects excluded and included are given in Table 1.
Subjects. Details of 2731 individuals (1160 men and
1571 women) who had complete weight and skinfold
Table 3
data at Time 1 and Time 2 and complete mortality
data, are presented in Table 3.
Variables. Weight-change was calculated as weight
(kg) at Time 2 minus weight (kg) at Time 1. In the
Framingham Heart study there was only one indicator
of fatness, that is, the subscapular skinfold. For details
on the measurement of this skinfold in the Framingham Heart Study, see Cupples and D'Agostino.20 Fatchange was calculated as subscapular skinfold thickness at Time 2 minus the subscapular skinfold thickness at Time 1. Preliminary inspection of the data in
categorical analyses, as recommended by Zhao and
Kolonel,21 indicated that the linearity of the association between fat change and the mortality rate could
be improved by taking a monotonic transformation of
fat change. Thus, fat change was transformed by
taking its natural logarithm (after adding a constant
to insure all positive numbers) and then scaled to have
Characteristics of the included subjects from the Framingham Heart Study.
Time 1
Age (y)
Smokers=Non-smokers (n)b
BMI (kg=m2)
Height (cm)
Weight (kg)
Subscapular skinfold (mm)
Time 2c
Weight (kg)
Subscapular skinfold (mm)
Change scores
Weight-change (kg)
Fat-change (mm)
BMI ˆ body mass index.
a
Data are means s.d. (range).
b
Current=never ‡ former.
c
Time 2 is 14 y later than Time 1.
All
(n ˆ 2731)
Men
(n ˆ1160)
Women
(n ˆ1571)
50.0 7.9 (37.0, 70.0)a
1468=1263
25.7 3.9 (15.5, 45.7)
164.8 9.2 (138.4, 193.7)
69.9 13.0 (36.7, 125.1)
14.7 4.5 (3.0, 33.0)
49.8 7.8 (37.0, 70.0)
812=348
26.2 3.3 (15.5, 36.9)
172.3 7.0 (149.9, 193.7)
77.7 11.3 (40.8, 122.5)
14.3 4.1 (5.0, 33.0)
50.2 7.9 (37.0, 70.0)
657=914
25.3 4.3 (16.3, 45.7)
159.2 6.2 (138.4, 179.1)
64.1 11.1 (36.7, 125.2)
14.9 4.8 (3.0, 32.0)
70.3 13.2 (32.2, 128.4)
16.2 6.1 (3.0, 42.0)
77.7 11.8 (34.5, 125.2)
16.6 6.1 (4.0, 42.0)
64.9 11.6 (32.2, 128.4)
15.9 6.1 (3.0, 41.0)
0.46 6.7 (ÿ35.8, 27.2)
1.6 4.8 (ÿ15.0, 22.0)
0.03 6.6 (ÿ26.3, 22.7)
2.4 4.8 (ÿ14.0, 20.0)
0.82 6.8 (ÿ35.8, 27.2)
1.0 4.8 (ÿ15.0, 22.0)
Fat loss and mortality
DB Allison et al
606
its original mean and s.d. for ease of interpretation
`Time to death' was the time to death from any cause
after Time 2.
Statistical analysis
For both studies, we conducted a `primary' analysis
and a set of secondary=sensitivity analyses. The primary analysis consisted of Cox regression in which
time to death was used as the dependent variable,
weight change and fat change were used as independent variables, and height at Time 1, age, gender, and
smoking status were used as covariates. In addition,
we repeated this primary analysis after ®rst residualising fat change for weight change to reduce collinearity
as suggested by a reviewer. Results were not meaningfully different.
In the secondary=sensitivity analyses, we evaluated
whether the results remained generally consistent
when we: 1) treated weight change and fat change
as categorical rather than continuous variables, by
dividing the distribution along quintile lines; 2) used
logistic rather than Cox regression, to avoid the
assumption of proportional hazards; 3) controlled for
baseline values of weight and fat; 4) allowed for
interactions between baseline BMI (kg=m2) (or a
categorical dummy code for obesity status) and
weight change and fat change; 5) allowed for interactions between gender and weight change and fat
change; 6) incorporated information about change in
smoking status between Time 1 and Time 2; 7)
restricted the analysis to never smokers (Tecumseh
sample) or to never plus former smokers (Framingham
sample) at Time 2; and 8) to evaluate whether the
effects on weight change and fat change differed by
age, we added interaction terms between weight
change and fat change and age into the model, and
tested these terms for signi®cance. We did not exclude
subjects who died during the ®rst few years of followup, for reasons described elsewhere.22 Because our
thesis is that fat change and weight change each
confound control the other's association with mortality, we also examined the coef®cients for weight
change and fat change in separate reduced models,
to evaluate the effects of failing to control for one,
while estimating the effect of the other.
The two data sets were not pooled for the following
reasons. The data sets, though remarkably similar in
many ways, contained a number of important differences that we believe could have affected the outcome
and therefore militated against raw data pooling.
These differences included: 1) the lengths of followup available after the second measurements; 2) the
intervals between the ®rst and second measurements;
3) the average age at baseline; 4) the Tecumseh data
allowed categorization of subjects into never vs ever
smokers, whereas Framingham only allowed categorization into current vs former plus never smokers, and
5) the number of skinfold measures available.
Statistical analyses were conducted using the SPSS
statistical package.23 Figures were produced by the SPLUS software.24
Results
The Tecumseh sample
Primary analysis. Using Cox proportional hazards
regression,25 the log of the hazard ratio was regressed
on fat change and weight change while controlling for
age, gender, smoking status (ever smoker vs never
smoker) and height. The coef®cients for this estimated
model and the associated inferential statistics are
displayed in Table 4. Of primary note is that both
the coef®cients for weight change (P ˆ 0.0001) and
fat change (P ˆ 0.0112) were statistically signi®cant.
As hypothesized, the coef®cient for weight change
was negative in sign (b
bà ˆ 0.0559; 95% con®dence
intervals (CI) ˆ ÿ0.0839, ÿ0.0279) and the coef®cient for fat change was positive in sign
(b
bà ˆ ÿ0.0163; 95% CI ˆ ÿ0.0038, ÿ0.0288). This
indicates that weight loss is associated with increased
mortality rate and fat loss with decreased mortality
rate. Expressed in standard deviation units (s.d.), each
4.6 kg (that is, 1 s.d.) of weight loss resulted in a 29%
(95% CI ˆ 14%, 47%) increase in the hazard for
Table 4 Results of the primary analysis of the Tecumseh Community Health Study and the Framingham Heart Study. Time of Death
(Log of the hazard ratio) regressed (via Cox proportional hazards regression) on weight change and fat change, while controlling for
à for weight-change (negative) and fatage, gender, smoking status and height. Note the opposite signs of regression coef®cients (b
b)
change (positive)
Tecumseh
Age
Height
Male
Smokers=Non smokersa,b
Weight-change
Fat-change
a
bª
s.e.
P
(exp b)
bª
bª
s.e.
P
(exp b)
bª
0.0966
ÿ0.0026
0.6561
0.3436
ÿ0.0559
0.0163
0.0045
0.0093
0.1764
0.1295
0.0143
0.0064
< 0.0001
0.7769
0.0002
0.0080
0.0001
0.0112
1.1015
0.9974
1.9237
1.4100
0.9456
1.0164
0.0929
ÿ0.0058
ÿ0.6365
0.3623
ÿ0.0495
0.0393
0.0061
0.0069
0.1289
0.0984
0.0080
0.0107
< 0.0001
0.4010
< 0.0001
0.0002
< 0.0001
0.0002
1.0973
0.9942
0.5292
1.4366
0.9517
1.0401
Ever=never for the Tecumseh Community Health Study.
Current=never ‡ former for the Framingham Heart Study.
b
Framingham
Fat loss and mortality
DB Allison et al
mortality whereas a 1 s.d. (10.0 mm) increase in fat
loss resulted in a 15% (95% CI ˆ 4%, 25%) decrease
in the hazard for mortality.
Secondary analyses. Each secondary analysis con®rmed the general pattern observed in the primary
analysis. That is, whether we treated weight loss and
fat loss as categorical rather than continuous variables,
used logistic rather than Cox regression, controlled for
baseline values of weight and fat, restricted the
analysis to never smokers, or incorporated information about smoking status at Time 2 to account for
changes in smoking status, weight loss was associated
with an increased mortality rate and fat loss was
associated with a decreased mortality rate. The stability of the estimates from analysis to analysis was
quite high with the hazard ratios never changing more
than 10% (see Table 5). Moreover, in every case, with
two exceptions, the coef®cients for the effects of
weight loss and fat loss were statistically signi®cant
(P < 0.05). One of these exceptions was observed
from the analysis restricted to `never' smokers, in
which only 731 subjects were available in the analysis. In this analysis, given the small sample size, it is
not surprising that not all effects were statistically
signi®cant. However, the coef®cients for both weight
loss and fat loss remained in their predicted direction
(Weight change bbà ˆ ÿ0.0578; P ˆ 0.0080; Fat change
bà ˆ 0.0132; P ˆ 0.1508).
The second exception occurred in the categorical
analysis. Categorisation of weight and fat losses, for
example, by quintiles, is known to decrease statistical
power.26 Nevertheless, some (though not all) of the
higher categories of weight loss had statistically signi®cant increases in mortality rate, relative to the
lowest category of weight loss, and similarly, some
of the categories of fat loss had statistically signi®cantly higher mortality rates than the lower categories
of fat loss. Moreover, there was a clear trend toward
increasing mortality rate with higher categories of
weight loss and decreasing mortality rate with
higher categories of fat loss. Coef®cients from
the categorical models are displayed graphically in
Figure 1.
Interactions' between gender and weight loss
(P ˆ 0.6549) and fat loss (P ˆ 0.2430) were not signi®cant. Interactions between age and weight loss
(P ˆ 0.9040), and age and fat loss (P ˆ 0.5980),
were also not signi®cant. When interaction terms
between baseline BMI and weight loss and fat loss
were added to the model, only the interaction term for
Figure 1 Tecumseh Community Health Study sample, secondary analysis. Estimated hazard ratios (to the last category) of
individuals in the corresponding categories of weight change
(WC) and fat change (FC), and their standard errors (s.e.). The
numbers in parenthesis represent the minimum and maximum
values for the corresponding quintile-de®ned categories. For
both WC (kg) and FC (change in skinfolds thickness, mm), the
positive and negative values represent gains and losses, respectively. There is a clear trend toward a decreasing mortality rate
with higher categories of WC (d) and an increasing mortality rate
with higher categories of FC (j).
Table 5 Secondary analyses of the Tecumseh Community Health Study and the Framingham Heart Study. Evaluation of consistency
of results in response to variations of analytic technique
Weight change
Statistical model
Tecumseh
Basicb
Basic ‡ weight1c, fat1d
Basic for never smokers only
Basic ‡ smoking status at Time 2
Basic with logistic regression2
Framingham
Basic
Basic ‡ Weight1c, Fat1d
Basic for never and former smokers only
Basic ‡ smoking status at Time 2
Basic with logistic regression
Fat change
HR a
P
HR a
P
0.78
0.78
0.77
0.78
0.75
0.0001
0.0007
0.0080
0.0001
0.0010
1.18
1.18
1.14
1.18
1.25
0.0112
0.0311
0.1508
0.0099
0.0086
0.72
0.72
0.79
0.74
0.68
< 0.0001
< 0.0001
0.0039
< 0.0001
< 0.0001
1.20
1.20
1.19
1.20
1.23
0.0002
0.0004
0.0228
0.0002
0.0008
HR ˆ hazaro ratio.
hazard ratio per 1.0 standard deviation increase in weight change or fat change.
In the basic model, the effect of weight loss and fat loss have been adjusted for the following covariates measured at Time 1: age,
gender, height and smoking status.
c
Weight1 is weight (kg) at Time 1.
d
Fat1 is tricep and subscapular skinfolds measurement (mm) at Time 1.
e
In the logistic model, the coef®cient labelled HR is actually an odds ratio (OR).
a
b
607
Fat loss and mortality
DB Allison et al
608
weight loss and BMI was signi®cant (P ˆ 0.0022).
The sign of the weight loss by BMI interaction
coef®cient was negative indicating that, as baseline
BMI increased, the apparently deleterious effect of
weight loss decreased. That is, weight loss appeared
more benign among heavier than among lighter
people. It was observed from the Cox model, including interactions of BMI and weight loss and fat loss,
that weight loss was estimated to be bene®cial (that is,
reduce mortality rate) for individuals with initial BMI
34. However, the BMI by fat loss interaction term
was not signi®cant (P ˆ 0.6342), suggesting that the
bene®cial effects of losing fat were not only limited to
those people who had high BMIs at the beginning of
the study. This was consistent with an analysis treating obesity status as a categorical variable (that is,
BMI < 28 ˆ 0; BMI 28 ˆ 1). This analysis showed
a signi®cant interaction between obesity status and
weight loss (P ˆ 0.0079), but not fat loss
(P ˆ 0.3854), and indicated that weight loss did not
appear deleterious among obese individuals.
Finally, in analyses examining the effects of weight
change and fat change separately that is, not controlling for the other, the absolute values of the coef®cients were substantially reduced (29% for weight
change and 69% for fat change). This indicates that
consistent with our thesis failing to adjust total weight
change for fat change and vice versa causes a bias in
the estimation of effects.
The Framingham sample
Primary analysis. Again the log of the hazard ratio
was regressed (via Cox proportional hazards regression) on fat change and weight change while controlling for age, gender, smoking status (current smoker
or never smoker plus former) and height. The coef®cients for this estimated model and associated inferential statistics are shown in Table 4. Of primary note,
is that both the coef®cients for weight change and fat
change (P < 0.0001, P ˆ 0.0002, respectively) were
statistically signi®cant. Moreover, as hypothesized,
the coef®cient for weight change was negative in
sign (b
bà ˆ ÿ0.0495; 95% CI: ÿ0.0652, ÿ0.0338) and
the coef®cient for fat change was positive in sign
(b
bà ˆ 0.0393; 95% CI: ÿ0.0183, ÿ0.0603). Thus,
weight loss was associated with increased mortality
rate and fat loss with decreased mortality rate.
Expressed in s.d., each 6.7 kg (that is, 1 s.d.) of
weight loss, would result in a 39% (95% CI, 25%,
54%) increase in the hazard ratio for mortality,
whereas a 1 s.d. (4.8 mm) increase in fat loss would
result in a 17% (95% CI, 8%, 25%) decrease in the
hazard ratio for mortality.
Secondary analyses. As in the Tecumseh data analysis and as highlighted in Table 5, each of the
secondary analyses con®rmed the general pattern
Figure 2 Framingham Heart Study sample, secondary analysis.
Estimated hazard ratios (to the last category) of individuals in the
corresponding categories of weight change (WC) and fat change
(FC), and their standard errors (s.e.). The numbers in parenthesis
represent the minimum and maximum values for the corresponding quintile-de®ned categories. For both WC (kg) and FC
(change in skinfold thickness, mm), the positive and negative
values represent gains and losses, respectively. There is a clear
trend toward a decreasing mortality rate with higher categories
of weight change (d) and an increasing mortality rate with
higher categories of fat change (u).
observed in the primary analysis. Speci®cally, the
same secondary analyses that were applied to the
Tecumseh data produced remarkably similar results
with respect to the Framingham data. For example,
there was, again, a clear trend toward increasing
mortality rate with greater weight loss and decreasing
mortality rate with greater fat loss. These coef®cients
from the categorical models are displayed graphically
in Figure 2.
However, unlike in Tecumseh, when interaction
terms between baseline BMI and weight change
(P ˆ 0.1397) and fat change (P ˆ 0.5973) were
added to the model, none were signi®cant (although
the direction of effect was the same as that observed in
Tecumseh), nor were those between a dichotomous
indication of obesity and weight change (P ˆ 0.3272)
and fat change (P ˆ 0.6852). Interactions between age
and weight change (P ˆ 0.6980), and age and fat
change (P ˆ 0.4035), were also not signi®cant. Thus,
in contrast to the Tecumseh study, the apparently
deleterious effects of weight loss and the bene®cial
effects of fat loss were independent of baseline BMI.
Interaction terms between gender and weight change
(P ˆ 0.3582) and fat change (P ˆ 0.0889) were also
not signi®cant.
Finally, consistent with the Tecumseh results in
analyses examining the effects of weight change and
fat change separately (that is, not controlling for the
others), the absolute values of the coef®cients were
substantially reduced (31% for weight change and
99% for fat change), indicating the strong bias resulting from considering weight change without considering fat change and vice versa.
Fat loss and mortality
DB Allison et al
Discussion
Using two population-based, prospective cohort studies, the Tecumseh Community Health Study and the
Framingham Heart Study, we analysed time to death
as a function of both weight loss (that is, loss of
kilograms) and fat loss (that is, reduction in skinfold
thickness), while controlling for some possible confounding factors: age, gender, smoking status and
height. Results were consistent with the stated hypothesis that weight loss increases whereas fat loss
decreases the all-cause mortality rate. The deleterious
effects of weight loss and the bene®cial effects of fat
loss were independent of baseline BMI in the Framingham study. In the Tecumseh study where weight
loss was observed to be more benign among heavier
than among lighter people, the bene®cial effects of
weight loss did not appear until a BMI 34 (that is,
in severely obese people).
It is noteworthy that results were con®rmed in two
completely independent cohorts. Although these two
cohort studies had much in common (for example,
similar sample sizes and measurements) there were
also important differences (for example, the interval
between the two weight and fat measurements and the
length of follow-up). Such replication strengthens
con®dence that the ®ndings were not likely to be
due to chance or some `quirk' of an individual data
set, and suggests a robust phenomenon.
In interpreting these data, it is important to realize
that weight change and fat change are positively
correlated. That is, in general, when individuals lose
weight, they lose fat and vice versa. In the studies
analyzed herein, the correlations between weight loss
and fat loss were between 0.50 ± 0.60. By analysing
weight change and fat change simultaneously in the
statistical model, we are able to look at their independent effects. Because weight loss is essentially equal
to loss of body fat mass plus loss of lean body mass,
when weight loss and fat loss are analysed simultaneously in the same statistical model, weight loss
essentially becomes an indicator of lean body mass
loss. Therefore, it may be reasonable to interpret these
results as indicating that fat mass loss is bene®cial
whereas lean body mass loss is deleterious and that
the extent to which weight loss is bene®cial or
deleterious will depend on the composition of that
weight loss. It is also noteworthy that after controlling
for the linear effects of weight loss and fat loss, there
were no signi®cant quadratic effects (data not shown),
indicating that there is no statistically signi®cant
evidence for a non-monotonic association between
our indicators of fat loss and weight loss and mortality
rate. This implies that where we have written `weight
loss is associated with increased mortality rate and fat
loss with decreased mortality rate', the reader could as
appropriately read, `weight gain is associated with
decreased mortality rate and fat gain is associated with
increased mortality rate.' Nevertheless, inspection of
Figure 1 and Figure 2 suggests that, although not
statistically signi®cant, there may be some departures
from linearity, such that the effects of weight gain and
fat gain may not simply be the opposite of the effects
of weight loss and fat loss. Future research in larger
samples may be necessary to systematically determine
the extent to which the results observed are due to fat
loss and weight loss per se vs fat gain and weight gain,
or both.
The idea that lean body mass loss is deleterious is
consistent with clinical research on diseased subjects.
For example, Tellado et al 27 studied mortality rate as
a function of several biomedical markers and weight
loss in a sample of 73 patients with sepsis and
malnutrition. To measure nutritional status, they
used the ratio of exchangeable sodium to exchangeable potassium (the Nae=Ke ratio), that is the extracellular mass as a function of body cell mass.28
Results indicated that a high Nae=Ke ratio was a
powerful predictor of mortality rate in this sample.
Comparable results were reported by Kotler et al,29
who studied survival time among AIDS patients. They
found that the lowest total potassium values (indicative of low lean body cell mass) were associated with
the earliest death. This was observed even though low
body cell mass was associated with a wide range of fat
mass values.
The implications of this study, if con®rmed in
additional studies, are clear and profound. For example, weight loss might only be advisable under conditions promoting a suf®cient proportion of the lost
weight as fat. When prospective cohort studies are
conducted using direct measurements of total fat
mass, via more sophisticated body composition measurements, it will then be possible to calculate directly
the proportion of weight loss as fat necessary to
achieve a net reduction in mortality rate. Given that
the present study measured relative fatness, rather
than fat mass per se, we can only say that a higher
proportion of weight lost as fat is desirable. We
cannot specify the `minimal' desirable proportion.
Lean body mass is positively related to the body fat
over a wide range of body weights.30 From this
relationship, Forbes30 predicted that induced fat loss
would be accompanied by a loss of lean body mass
inversely related to the initial body fat content. That
is, given a ®xed level of caloric restriction, individuals
with a relatively greater amount of total body fat were
expected to lose relatively smaller proportions of their
weight as lean mass.30 Thus, clinical prescriptions for
weight loss may need to consider patients' body
composition at baseline and hence their projected
lean mass loss for a ®xed level of weight loss.
The samples studied herein consist of adults of
European ancestry living in the United States under,
presumably, reasonable conditions of health and nutrition. Moreover, although there was certainly a large
number of obese individuals, where obesity is de®ned
as, for example, a BMI > 28, there were few severely
obese individuals. Thus, it is not clear that these
609
Fat loss and mortality
DB Allison et al
610
results are generalisable to children, individuals of
other ethnic backgrounds, individuals living under
other circumstances, or severely obese or severely
underweight individuals. To some extent, this is
supported by a signi®cant interaction term between
baseline BMI and weight loss in the Tecumseh study,
which suggested that among severely obese individuals weight loss might be bene®cial regardless of its
composition. Clearly more research is needed in this
area.
One avenue for future research is to replicate these
®ndings in a study in which body composition is
measured by more sophisticated methods. At present,
we are aware of no such studies. However, in the third
National Health and Nutrition Examination Survey
(NHANES III study),31 body composition measurements were taken on all individuals, via bioimpedance
analysis. Moreover, in the currently planned
NHANES IV study, it is anticipated that body fat
will be measured by dual energy X-ray absorptiometry (DEXA). Hopefully, follow-up studies of these
two cohorts would be conducted allowing replication
of the present ®ndings. Moreover, by comparing the
effects of fat mass loss and weight loss in a common
metric (that is, kilograms), one can determine the
proportion of weight that must be lost as fat to achieve
a net reduction in mortality risk.
In this regard, it must also be pointed out that the
measures of fatness used herein consisted of triceps
and subscapular skinfolds. It is conceivable, although
perhaps unlikely, that an individual could lose fat
from these two regions of the body, but simultaneously gain an equal or greater amount of fat in
other parts of the body such that, despite a reduction
in skinfold thickness, the individual may not have
reduced their fatness overall. This can only be rigorously evaluated in future research in which direct
estimates of total fat mass are obtained.
A more speculative direction for future research
concerns the development of treatments that preferentially promote fat loss relative to lean loss. Presently,
most treatments for obesity are geared toward producing reduced caloric intake and, to some extent,
increased caloric expenditure. This results in weight
loss, but such methods usually do not preferentially
promote fat loss, per se. In contrast, in some animal
studies, certain newer compounds under investigation
(for example, ciliary neurotrophic factor and leptin)32
appear to preferentially produce fat loss. It therefore
may be possible that investigators will be able to
develop therapeutics aimed speci®cally at fat loss per
se. A second approach might be to put more emphasis
on exercise, which tends to preserve lean mass.33
Finally, additional research might evaluate the
current obesity treatments in terms of their effects
on body composition. Virtually all studies of obesity
treatment measure weight loss, but only a minority
measure fat loss. Moreover, there has been relatively
little exploration of the effects of various currently
existing treatments, conditions of treatment, and rates
of weight loss, in terms of their differential effects on
degree of body fat loss. Such research might help us
select the treatments and conditions that would be
expected to produce the greatest bene®ts for obese
patients.
Acknowledgements
This research was funded by NIH grants
R29DK47256, R01DK51716, TD32DK37352, and
P30DK26687.
This paper uses data supplied by the Inter-University Consortium on Political and Social Research
(ICPSR) from the Tecumseh Heart Study and the
National Heart, Lung and Blood Institute, NIH,
DHHS from the Framingham Heart Study. The
views expressed in this paper are those of the authors
and do not necessarily re¯ect the views of ICPSR, the
Tecumseh Study, the National Heart, Lung and Blood
Institute or the Framingham Study.
References
1 Pi-Sunyer FX. Health implications of obesity. Am J Clin Nutr
1991; 53: (Suppl), S1595 ± S1603.
2 Goldstein DJ. Bene®cial effects of modest weight loss. Int J
Obes 1992; 16: 397 ± 415.
3 Pi-Sunyer FX. A review of long-term studies evaluating the
ef®cacy of weight loss in ameliorating disorders associated
with obesity. Clin Ther 1996; 18: 1006 ± 1035.
4 Troiano RP, Frongillo EA, Sobal J, Levitsky DA. The relationship between body weight and mortality: a quantitative analysis of combined information from existing studies. Int J
Obesity 1996; 20: 63 ± 75.
5 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 ± 743.
6 Lee IM, Paffenbarger RS Jr. Is weight loss hazardous? Nutr
Rev 1996; 54: (Suppl), S116 ± S124.
7 Williamson DF. Lingering questions about intentional weight
loss. Nutrition 1996; 12: 819 ± 820.
8 Williamson DF, Pamuk E, Thun M, Flanders D, Tim Byers,
Heath C. Prospective study of intentional weight loss and
mortality in never-smoking overweight US white women aged
40 ± 64 years. Am J Epidemiol 1995; 141: 1128 ± 1141.
9 Williamson DF, Pamuk E, Thun M, Flanders D, Heath C,
Byers T. Prospective study of intentional weight loss and
mortality in overweight men aged 40 ± 64 years. Obes Res
1997; 5 (Suppl 1): S94.
10 Baumgartner RN, Heyms®eld SB, Roche AF. Human body
composition and the epidemiology of chronic disease. Obes
Res 1995; 3: 73 ± 95.
11 Heyms®eld SB, Allison DB, Heshka S, Pierson RN Jr.
Assessment of human body composition. In: Allison DB
(ed). Handbook of Assessment Methods for Eating Behaviors
and Weight-Related Behaviors. Sage: Thousand Oaks, 1995,
pp 515 ± 560.
12 Allison DB, Faith MS, Heo M, Kotler DP. Hypothesis concerning the U-shaped relation between body mass index and
mortality. Am J Epidemiol 1997; 146: 339 ± 349.
13 Segal JC, Verschuren WMM, VanLeer EM, Kromhout D.
Overweight, underweight and mortality. J Clin Inves 1987; 80:
1050 ± 1055.
14 Van Itallie TB, Yang M-U. Diets and weight loss. New Engl J
Med 1977; 297: 1158 ± 1161.
15 Tecumseh Management Committee. Tecumseh mortality
follow-up study, 1978 [Computer ®le]. Tecumseh Management
Committee: Ann Arbor, 1992. Inter-university Consortium for
Political and Social Research [distributor], 1993.
Fat loss and mortality
DB Allison et al
16 Butler WJ, Ostrander LD Jr, Carman WJ, Lamphiear DE.
Mortality from coronary heart disease in the Tecumseh study.
Long-term effect of diabetes mellitus, glucose tolerance and
other risk factors. Am J Epidemiol 1985; 121: 541 ± 547.
17 Hawthorne V, et al. Tecumseh Community Health Study,
1959 ± 1969 [Computer ®le]. University of Michigam, Institute
of Gerontology: Ann Arbor, 1989. Inter-University Consortium for Political and Social Research [distributor], 1989.
18 Wainer H. Estimating coef®cients in linear models: It don't
make no nevermind. Psychol Bull 1976; 83: 213 ± 217.
19 Dawber TR. The Framingham Study. Harvard University
Press. Cambridge, 1980.
20 Cupples LA, D'Agostino RB. Some risk factors related to the
annual incidence of cardiovascular disease and death using
pooled repeated biennals measurements: Framingham Heart
Study, 30 year follow-up. Sect 34. In: Kannel WB, Wolf PA,
Garrison RJ (eds). The Framingham Study: an epidemiological
investigation of cardiovascular disease. DHHS: Washington
1986.
21 Zhao LP, Kolonel LN. Ef®ciency loss from categorizing
quantitative exposures into qualitative exposures in case-control studies. Am J Epidemiol 1992; 136: 464 ± 474.
22 Allison DB, Heo M, Flanders DW, Faith MS, Williamson DF.
Examination of early mortality exclusion as an approach to
control for confounding by occult disease in epidemiological
studies of mortality risk factors. Am J Epidemiol 1997; 146:
672 ± 680.
23 SPSS. SPSS v. 7.5. SPSS Inc: Chicago, 1997.
24 S-PLUS v. 3.3. MathSoft Inc: Seattle, 1995.
25 Cox DR, Oakes D. Analysis of survival data. Chapman &
Hall: London, 1984.
26 Selvin S. Statistical analysis of epidemiological data. Oxford
University Press: New York, 1996.
27 Tellado JM, Garcia-Sabrido JL, Hanley JA, Shizgal HM,
Christou NV. Predicting mortality based on body composition
analysis. Ann Surg 1989; 209: 81 ± 87.
28 Forse RA, Shizgal HM. The Nae=Ke ratio: a predictor of
malnutrition. Surg Forum 1980; 31: 89 ± 92.
29 Kotler DP, Wang J, Pierson RN. Body composition studies in
patients with the acquired immunode®ciency syndrome. Am J
Clin Nutr 1985; 42: 1255 ± 1265.
30 Forbes GB. Human body composition: growth, aging, nutrition, and activity. Springer-Verlag: New York, 1987.
31 National Center of Health Statistics. Plan and operation of the
Third National Health and Nutrition Examination, 1988 ± 94.
Vital Health Stat 1994; 1.
32 Gloaguen I, Costa P, Demartis A, Lazzaro D, Di Marco A,
Graziani R, Paonessa G, Chen F, Rosenblun CI, Van der Ploeg
LH, Cortese R, Cliberto G, Laufer R. Ciliary neurotrophic
factor corrects obesity and diabetes associated with leptin
de®ciency and resistance. Proc Natl Acad Sci USA 1997; 94:
6456 ± 6461.
33 Blair SN, Cooper KH. Dose of exercise and health bene®t.
Arch Intern Med 1997; 157: 486.
611