International Journal of Obesity (2000) 24, 1145±1152
ß 2000 Macmillan Publishers Ltd All rights reserved 0307±0565/00 $15.00
www.nature.com/ijo
Indirect estimates of body composition are
useful for groups but unreliable in individuals
LS Piers1*, MJ Soares2, SL Frandsen3 and K O'Dea1
1
Centre for Population Health and Nutrition, Monash Institute of Public Health, Monash Medical Centre, 246 Clayton Road, Clayton,
VIC 3168, Australia; 2Department of Nutrition, Dietetics and Food Science, School of Public Health, Curtin University of Technology,
GPO Box U 1987, Perth, WA 6845, Australia; and 3The School of Health Sciences, Deakin University, 225 Burwood Highway, Burwood,
VIC 3125, Australia
OBJECTIVE: To assess the usefulness of the body mass index (BMI) in identifying individuals classi®ed as overweight
or obese based on estimates of body fat percentage (BF%) obtained by the deuterium dilution (BF%DD) method. In
addition, to assess the accuracy of bioelectrical impedance analysis (BIA) and skinfold thickness (SFT) measurements
in the estimation of body composition of Australians at the individual and group level.
DESIGN: Cross-sectional study.
SUBJECTS: One hundred and seventeen healthy Australian volunteers of European descent, comprising of 51 males
and 66 females, ranging in age from 19 to 77 y.
MEASUREMENTS: BMI was calculated from body weight and height. Fat-free mass (FFM) was estimated from
measures of total body water (TBW) using deuterium dilution (FFMDD), SFT using the equations of Durnin and
Womersley (Br J Nutr 1974; 32: 77 ± 97) (FFMSFT), and BIA using the equations of Lukaski et al (J Appl Physiol 1986; 60:
1327 ± 1332) (FFMLU), Segal et al (Am J Clin Nutr 1988; 47: 7 ± 14) (FFMSe) and Heitmann (Eur J Clin Nutr 1990; 44: 831 ±
837) (FFMHe). Estimates of fat mass (FM) were calculated as the difference between body weight and FFM, while BF%
was calculated by expressing FM as a percentage of body weight.
RESULTS: BMI had poor sensitivity and positive predictive value in identifying individuals as being overweight=obese
as classi®ed by BF%DD. Furthermore, estimates of FFM (and hence FM) from BIA or SFT could not be used
interchangeably with DD, without the risk of considerable error at the individual level. At the group level errors
were relatively smaller, though statistically signi®cant. While FFMSFT could be corrected by the addition of the bias
(1.2 kg in males and 0.8 kg in females), no simple correction was possible with BIA estimates of FFM for any of the
equations used. However, an accurate prediction of FFMDD was possible from the combination of FFMHe, biceps SFT
and mid-arm circumference in both males and females. The bias of this prediction was small ( < 0.15 kg), statistically
non-signi®cant in both sexes, and unrelated to the mean FFM obtained by the two methods. The revision of
Heitmann's estimate of FFM using anthropometric variables described in this study had the best sensitivity (79%),
speci®city (96%) and positive predictive value (92%) in identifying overweight=obese individuals in comparison to the
other equations tested.
CONCLUSION: BMI was a poor surrogate for body fatness in both males and females. The currently recommended
equations for the prediction of body composition from SFT and BIA provided inaccurate estimates of FFM both at the
individual and group level as compared to estimates from DD. However, Heitmann's equations, when combined with
measures of the biceps SFT and mid-arm circumference, provided better estimates of FFM both at the individual and
group level.
International Journal of Obesity (2000) 24, 1145±1152
Keywords: body composition; body mass index; deuterium dilution; bioelectrical impedance analysis; skinfold
thickness; methodology
Introduction
The currently recommended classi®cation of weight
status is the body mass index (BMI). According to this
classi®cation individuals with a BMI of 18.5 ±
*Correspondence: LS Piers, Centre for Population Health and
Nutrition, Monash Institute of Public Health, Monash Medical
Centre, 246 Clayton Road, Clayton, Victoria 3168, Australia.
E-mail: [email protected]
Received 5 October 1999; revised 28 January 2000; accepted
24 May 2000
24.9 kg=m2 are considered have a `normal weight',
those with a BMI of 25 ± 29.9 kg=m2 are classi®ed as
being overweight (or pre-obese), and those with a
BMI of 30 kg=m2 and above are considered to be
obese.1 BMI has been chosen as a surrogate measure
of body fatness based on its simplicity and because of
its association with mortality.1 However, based on this
classi®cation, a heavy muscular individual may be
labelled `obese', as BMI is simply calculated by
dividing body weight (kg) by height2 (m2), and does
not incorporate any measure of body composition.
There is also evidence that the relationship between
BMI and body fat percentage (BF%) may be different
in different ethnic groups,2 ± 6 implying that a BMI-
Accuracy of indirect estimates of body composition
LS Piers et al
1146
based classi®cation of weight status would necessarily
be population or possibly even cohort-speci®c. In
effect, this would make direct comparisons of
weight status between different populations groups
in epidemiological studies impossible to interpret.
Hence, better measures of weight status are required
to distinguish between body weight associated with
muscle and that associated with fat. A classi®cation of
weight status based on body fat percentage (BF%)
would overcome this problem and would allow direct
comparison of weight status between different
populations and=or ethnic groups.
Bioelectrical impedance analysis (BIA) is a relatively simple to use and objective method to estimate
body composition. It has been proposed as a suitable
®eld method for use in large-scale epidemiological
studies and for use in clinical situations.7 It is also
widely used in the health and ®tness industry to
estimate body composition. The validity of BIA
estimates of TBW and FFM have been widely published.8 ± 11 In these studies TBW is measured using
isotope dilution, or FFM is measured using hydrodensitometry, or a multi-compartment model of body
composition is derived using a combination of criterion methods in a group of volunteers. Resistance,
reactance, age, gender and anthropometric measurements are then made at the same time and related to
the estimates of either TBW or FFM. BIA estimates of
body composition are, therefore, a statistical association.12 While many investigators agree that BIA
estimates of body composition are valid at the group
level, very few if any suggest that they are valid at the
level of the individual. However, BIA is used in the
clinical setting to identify individuals with high body
fat, and is used in epidemiological studies to classify
individuals into groups with differing amounts of
body fat. It is important to establish the accuracy of
the method to estimate body composition at the level
of the individual.
Recently, both skinfold thickness (SFT) measurements with use of Durnin and Womersley's equations,13 and BIA with the use of the equations of
Segal et al10 have been recommended for use in the
Australian population of Anglo-Celtic ancestry to
estimate BF%.14 However, their usefulness in individuals was not addressed. We therefore assessed the
accuracy of both SFT, using Durnin and Womersley's
equations, and BIA, using three sets of published
equations, to estimate FFM both at the level of the
individual and also the group, against estimates from
TBW measurements using deuterium dilution (DD).
In all 117 healthy Australian volunteers of European
descent ranging in age from 19 to 77 y were studied.
The BIA equations used to estimate FFM were those
of Lukaski et al, derived from hydro-densitometric
estimates of FFM,8 and Segal et al, also derived from
hydro-densitometric estimates of FFM,10 and Heitmann's BIA equations derived using a multi-compartment model from measurements of TBW and
potassium.11
International Journal of Obesity
Subjects and methods
Subjects
Adult male and female subjects, across a wide age
range, were recruited by advertisement in the local
media and through personal approach. All were resident in Melbourne, Australia. Inclusion criteria were
as follows: (1) absence of clinical signs or symptoms
of chronic disease; (2) Weight stability ( 2 kg for
preceding 12 months); (3) not on any medication that
could affect body composition. All subjects gave
written informed consent to participate in the study.
The Deakin University Ethics Committee approved
the experimental protocol, and all measurements were
made at the clinical rooms of the Toorak campus of
Deakin University.
Anthropometry and skinfold thickness measurements
Standing height was measured using a stadiometer
®xed to the wall and recorded to the nearest 0.1 cm.
Body weight was measured between 7 and 10 am after
an overnight fast and immediately after voiding, with
subjects wearing light indoor clothing and no shoes,
on a beam balance and recorded to the nearest 100 g.
Mid-arm, waist and hip circumferences were measured as described by Callaway et al.15 Skinfold
thicknesses (SFT) at four sites (biceps, triceps, subscapular and supra-iliac) were measured on the right
side of the body and recorded to the nearest 0.2 mm.16
Each skinfold was measured in triplicate and the mean
of the three measurements used for further analyses.
The sum of the four skinfold thicknesses were then
used in the sex and age speci®c equations of Durnin
and Womersley for the prediction of body density.13
Body fat percentage (BF%) was calculated from body
density. Fat mass (FM) was then calculated and FFM
was estimated from the difference of body weight and
FM.
Deuterium dilution
Total body water (TBW) was measured by deuterium
oxide (2H2O) dilution (DD) in all 117 subjects after an
overnight fast as previously described.17 Fat-free mass
from DD (FFMDD) was calculated from TBW assuming a hydration of 0.732.18 The value of 0.732
assigned to FFM hydration to derive fat-free mass in
adults can also be employed in body composition
studies involving the elderly.19 FM was calculated
from the difference of body weight and FFMDD and
expressed as a percentage of body weight. We have
previously demonstrated that FFM estimated from
TBW measurements is similar to FFM estimated
from dual energy X-ray absorptiometry (DEXA) in
the same group of subjects.17
Accuracy of indirect estimates of body composition
LS Piers et al
Bioelectrical impedance analysis
Resistance and reactance measurements were made
using a single frequency four-terminal impedance
plethysmograph (RJL Systems, model 101, Detroit,
MI, USA). All measurements were made early in the
morning between 7 and 10 am after an overnight fast.
Disposable electrodes (Nikotabs-E, Medical Equipment Services, Australia) cut in half were used.
Each half was positioned in the mid-line of the
dorsal surfaces of the hands and feet, proximal to
the metacarpal ± phalangeal and metatarsal ± phalangeal joints, between the distal prominence of the
radius and ulna and between the medial and lateral
malleoli at the ankle respectively. Speci®cally, the
proximal edge of one electrode was in line with the
proximal edge of the ulnar tubercle at the wrist and
the proximal edge of the other was in line with the
medial malleolus of the ankle.20 Each subject lay still
and supine with arms by their sides and all four limbs
abducted. An excitation current of 800 mA at 50 kHz
was introduced at the distal electrodes of the hand and
foot, and the voltage drop was detected by the
proximal electrodes.20 The lowest resistance (Rs) and
reactance (Xcs) values assuming a series model were
recorded. The Rs and Xcs were then used in three
separate equations to determine FFM: the general
equation of Lukaski et al;8 the gender-speci®c equations, not requiring a priori designation of `lean' or
`obese' using anthropometry, of Segal et al;10 and the
gender-speci®c equations of Heitmann11 as set out
below:
Lukaski et al8
FFMLU …kg† ˆ 0:756 height2 =Rs ‡ 0:110
weight ‡ 0:107 X cs ÿ 5:463
Segal et al10
FFMSe …kg†< ˆ height2 0:00132 ÿ Rs 0:04394
‡ weight 0:30520 ÿ age 0:16760
‡ 22:66827
Classi®cation of subjects by body fat percentage
1147
All subjects were classi®ed into normal weight or
overweight=obese, based on estimates of BF%
obtained by DD. Hence, male subjects with a
BF% 25 and females with a BF% < 30 were considered normal weight, while males with a BF% > 25
and females with a BF% > 30 were considered overweight=obese. This classi®cation formed the reference
classi®cation, and allowed for the calculation of
sensitivity, speci®city and positive predictive value
of classi®cations when based independently on either
BMI, SFT or BIA.
Statistics
Data were analysed using the SPSS for Windows
(Version 6.1.3, SPSS Inc., USA) statistical software
package. All data are presented as mean s.d., unless
otherwise stated. Normality of the distribution of
variables was assessed by the one sample Kolmogorov ± Smirnov Goodness of Fit test. Differences
between gender groups were determined using
unpaired t-tests while differences in estimates of
body composition variables within individuals were
determined using paired t-tests. FFM obtained by
different techniques were correlated using Pearson's
correlation coef®cients and compared by the method
suggested by Bland and Altman.21 The Bland and
Altman method requires the absolute difference
between the two measures in question when made in
the same individual to be small and statistically
unrelated to the mean of the two measures.21 The
upper and lower limits of agreement (LOA) between
the estimates of FFM by the two methods were
computed by adding and subtracting 1.96s.d.,
to the mean difference between the two estimates, respectively. Sensitivity, speci®city and positive
predictive value were calculated as described by
Beaglehole et al.22
FFMSe …kg†, ˆ height2 0:00108 ÿ Rs 0:02090
‡ weight 0:23199 ÿ age 0:06777
‡ 14:59453
(These equations do not require a priori designation
of `lean' or `obese' using anthropometry)
Heitmann11
FFMHe …kg†< ˆ 0:279 height2 =Rs ‡ 0:245 weight
‡ 0:231 height ÿ 0:077 age
ÿ 14:94
FFMHe …kg†, ˆ 0:279 height2 =Rs ‡ 0:181
weight ‡ 0:231 height ÿ 0:077
age ÿ 14:94
Note: height (cm), weight (kg) and age (y).
Results
Sixty-six women and 51 men were studied, ranging in
age from 19 to 77 y, with BMI ranging from 17 to
40 kg=m2. Subject characteristics are given in Table 1.
All four skinfold thicknesses were measured in 48
males and 63 females. In the remaining three males
and three females either the subscapular or the suprailiac skinfold were dif®cult to measure, hence the data
on these subjects were excluded from analyses associated with SFT. In addition, mid-arm (MAC) was
not measured in one of the female subjects. Females
had signi®cantly lower body weight (P < 0.0005),
height (P < 0.0005), BMI (P ˆ 0.01), mid-arm
(P < 0.0005), waist circumference (P < 0.0005),
waist to hip circumference ratio (P < 0.0005), and
TBW (P < 0.0005), but higher biceps and triceps
International Journal of Obesity
Accuracy of indirect estimates of body composition
LS Piers et al
1148
relationship (P < 0.0005) between BMI and BF%DD
(r ˆ 0.62, 95% CI 0.42, 0.76), BF%LU (r ˆ 0.72, 95%
CI 0.55, 0.83), BF%Se (r ˆ 0.77, 95% CI 0.63, 0.86),
BF%He (r ˆ 0.90, 95% CI 0.83, 0.94) and BF%SF
(n ˆ 48, r ˆ 0.66, 95% CI 0.46, 0.80). However, in
both males and females, there was considerable overlap of the 95% CI of BF%DD between the different
BMI groups (Table 2).
Table 1 Subject characteristics
Males (n ˆ 51)
Females (n ˆ 66)
Variable
Mean
s.d.
Mean
s.d.
Age (y)
Weight (kg)
Height (cm)
BMI (kg=m2)
Mid-arm circumferencea
Skinfold thicknessb(mm)
Biceps
Triceps
Subscapular
Supra-iliac
Waist (cm)
Hip (cm)
Waist to hip ratio
TBW (kg)
Resistance (O)
Reactance (O)
36
77.7
177.1
24.8
32.1
18
11.8
6.9
3.6
3.0
36
62.1*
165.0*
22.9*
28.5*
18
11.3
6.9
4.2
3.7
5.2
11.3
14.9
18.4
84.1
98.1
0.85
45.3
481
50
2.5
5.4
6.5
9.0
11.2
6.2
0.07
5.4
52
8
8.6*
18.9*
15.4
15.9
71.1*
95.9
0.74*
31.8*
608.*
64.*
4.3
5.5
6.7
5.9
10.8
8.3
0.07
4.2
67
10.2
Fat-free mass
Comparison of FFMDD with FFMSF. FFMDD was
signi®cantly different from FFMSF in males
(P < 0.021), and females (P ˆ 0.013; Tables 3 and
4). The difference in FFM, estimated by the two
methods, was not signi®cantly correlated to the
mean of the two estimates in males (P ˆ 0.976) or
females (P ˆ 0.156; Table 4).
* Signi®cantly different on an unpaired t-test (P < 0.05).
n ˆ 65 in females. bn ˆ 48 in males and n ˆ 63 in females.
a
SFT (P < 0.0005), resistance (P < 0.0005) and reactance (P ˆ 0.014) values when compared to the males
on an unpaired t-test.
BMI and body fat percentage
In females (n ˆ 66) BMI was signi®cantly correlated
(P < 0.0005) to BF%DD (r ˆ 0.75, 95% con®dence
interval (95% CI) 0.62, 0.84), BF%LU (r ˆ 0.80,
95% CI 0.69, 0.87), BF%Se (r ˆ 0.90, 95% CI 0.84,
0.94) or BF%He (r ˆ 0.92, 95% CI 0.87, 0.95) and
BF%SF (n ˆ 63, r ˆ 0.68, 95% CI 0.52, 0.79) Likewise, in males (n ˆ 51) there was also a signi®cant
Comparison of FFMDD with FFMLU. FFMDD was
signi®cantly different from FFMLU in males
(P < 0.0005) and females (P ˆ 0.001) Tables 3 and
4). The difference in FFM, estimated by the two
methods, was signi®cantly correlated to the mean of
the two estimates in males (P ˆ 0.032) and females
(P ˆ 0.036; Table 4).
Comparison of FFMDD with FFMSe. FFMDD was
signi®cantly different from FFMSe in males
(P ˆ 0.027), but not females (P ˆ 0.83; Tables 3 and
4). The difference in FFM, estimated by the two
methods, was signi®cantly correlated to the mean of
the two estimates in females (P ˆ 0.001) but not males
(P ˆ 0.86; Table 4).
Table 2 Body mass index (BMI) and percentage body fat in each gender based on recommended BMI cut offsa
Males
Females
BMI < 18.5 18.5 BMI < 25 25 BMI < 30
No. of subjects
BMI (kg=m2)
Body Fat (%)b
0
Ð
Ð
Ð
31
22.4 1.5
16.4 6.7
(13.9 ± 18.8)
14
27.1 1.5
21.6 8.5
(16.7 ± 26.5)
BMI 30
BMI < 18.5
18.5 BMI < 25 25 BMI < 30
6
4
31.9 1.1
17.7 0.4
31.7 5.2
24.6 4.2
(26.2 ± 37.2) (17.9 ± 31.3)
51
21.7 1.7
27.1 6.1
(25.4 ± 28.8)
7
27.3 1.6
38.0 6.3
(32.1 ± 43.8)
Values are mean s.d. Figures in parentheses are 95% con®dence intervals.
From WHO.1
b
From deuterium dilution.
a
Table 3 Fat-free mass (FFM) from deuterium dilution, skinfold thickness and bioelectrical impedance
analysis
FFM (kg) estimated from:
Deuterium dilution
SFTa and Durnin and Womersley's equations13
BIA using Lukaski et al 's equation8
BIA using Segal et al equations10
BIA using Heitmann's equations11
BIA using the modi®ed Heitmann's equations
Males
Mean s.d. (n ˆ 51)
Females
Mean s.d. (n ˆ 66)
61.9 7.3
60.3 7.2*
59.3 6.5*
60.7 7.2*
60.6 5.7*
61.8 6.7
43.5 5.8
42.4 5.4*
42.5 5.3*
43.3 4.9*
44.3 4.6*
43.4 5.8
SFT ˆ skinfold thickness; BIA ˆ Bioelectrical impedance analysis.
n ˆ 48 in males and n ˆ 63 in females.
*Signi®cantly different from corresponding estimate based on DD (P 0.05) on a paired t-test.
a
International Journal of Obesity
BMI 30
4
35.4 4.4
46.5 4.9
(38.7 ± 54.3)
Accuracy of indirect estimates of body composition
LS Piers et al
Table 4 Mean difference ( s.d.) and limits of agreement (LOA) in estimates in fat free mass measured by deuterium dilution (FFMDD)
vs estimates from skinfold thickness and bioelectrical impedance
Comparison
of FFMDD vs FFM from:
Durnin and Womersley
Lukaski et al
8
Segal et al 10
Heitmann11
Modi®ed Heitmann
(this study)
13
Group
No. of
subjects
Mean
difference (kg)
s.d.
Male
Female
Male
Female
Male
Female
Male
Female
Male
Female
48
63
51
66
51
66
51
66
51
65
1.19
0.81
2.61
0.87
1.24
0.06
1.27
7 0.94
0.13
0.11
3.46
2.52
2.69
2.04
3.87
2.29
3.72
2.34
3.13
2.07
1149
LOAc(kg)
P-value
Correlation
coef®cienta
P-valueb
Upper
Lower
0.021
0.013
< 0.0005
0.001
0.027
0.83
0.018
0.002
0.76
0.67
0.004
0.18
0.30
0.26
0.03
0.41
0.45
0.51
0.22
0.19
0.98
0.16
0.032
0.036
0.86
0.001
0.001
< 0.0005
0.13
0.14
8.0
5.8
7.9
4.9
8.8
4.5
8.6
3.6
6.3
4.2
7 5.6
7 4.1
7 2.7
7 3.1
7 6.3
7 4.4
7 6.0
7 5.5
7 6.0
7 3.9
a
c
Correlation coef®cient and bsigni®cance of the difference between the two methods vs mean of the two methods.
LOA ˆ Limits of agreement ˆ mean difference s.d.1.96.
r2 ˆ 0:82; standard error ˆ 3:15 kg; ANOVA;
Comparison of FFMDD with FFMHe . FFMDD was
signi®cantly different from FFMHe in males
(P ˆ 0.018) and females (P ˆ 0.002; Tables 3 and 4).
The difference in FFM, estimated by the two methods,
was signi®cantly correlated to the mean of the two
estimates in males (P ˆ 0.001) and females
(P < 0.0005; Table 4).
d:f : 3; 44; F ˆ 70:0; significance P < 0:00005:
FFMDD in females ˆ FFMHe 1:083 ‡ MAC
0:391 ÿ biceps SFT 0:368
ÿ 12:504
r2 ˆ 0:88; standard error ˆ 2:07 kg; ANOVA;
d:f : 3; 58; F ˆ 143:4; Significance P < 0:00005:
Regression analysis
FFMDD was regressed in a step-wise manner on
FFMHe, age, biceps, triceps, subscapular and suprailiac SFT, mid-arm, waist and hip circumferences, and
waist to hip circumference ratio in males and females
separately. In both males and females, FFMHe, biceps
SFT and MAC were found to best predict FFMDD as
given below:
The estimate of FFM thus obtained ie by modifying
the estimate of FFM obtained by using the equation of
Heitmann (Mod-FFMHe), was not signi®cantly different from FFMDD and the bias (FFMDD 7 ModFFMHe) was small and unrelated to the mean FFM
obtained by the two methods (Table 4). The estimates
of FFM obtained by the equations of Lukaski et al 8
and Segal et al 10 were not improved by the addition of
anthropometric variables since in both instances, bias
was signi®cantly related to the mean of the two
estimates.
FFMDD in males ˆ FFMHe 0:980 ‡ MAC
1:019 ÿ biceps SFT 1:009
ÿ 24:709
Table 5 The sensitivity, speci®city and positive predictive value of body mass index (BMI), skinfold thickness (SFT) and bioelectrical
impedance analysis (BIA) in classifying individuals as overweight or obese against reference estimates of body fat percentage based
on deuterium dilution
Classi®cation based on:
Body Mass Indexd
BF% from SFTe
BF% from BIA using Heitmann's equationse
BF% from BIA using modi®ed Heitmann equationse
Sensitivitya(%)
Speci®cityb (%)
Positive predictive valuec(%)
47.7
73.7
61.4
79.1
86.3
84.9
93.2
95.9
67.7
71.8
84.4
91.9
a
Sensitivity …%† ˆ
No: of subjects classified overweight or obese using BMI …or BF%SFT or BF%BIA †
100
No: of subjects classified overweight or obese using BF%DD
b
Specificity …%† ˆ
No: of subjects classified normal weight using BMI …or BF%SFT or BFBIA †
100
No: of subjects classified normal weight using BF%DD
c
Positive predictive value …%† ˆ
No: of subjects classified overweight or obese using BF%DD
100
No: of subjects classified overweight or obese using BMI …or BF%SFT or BF%BIA †
d
e
Normal weight: BMI < 25; overweight=obese: BMI 25.
Overweight=obese ˆ BF% > 25 in males and BF% > 30 in females.
International Journal of Obesity
Accuracy of indirect estimates of body composition
LS Piers et al
1150
The sensitivity, speci®city and positive predictive
value of BMI, SFT and BIA to identify those classi®ed as overweight or obese by DD (BF%DD > 25 in
males and BF%DD > 30 in females) is presented in
Table 5. BMI had poor sensitivity, good speci®city
and an average positive predictive value. Both SFT,
using Durnin and Womersley's equations,13 and BIA,
using Heitmann's equation,11 had greater sensitivity
and positive predictive value but similar speci®city
when compared to BMI. Modi®cation of Heitmann's
equations resulted in an improvement in sensitivity,
speci®city and positive predictive value (Table 5).
Discussion
This study demonstrates that despite the signi®cant
correlation between BMI and BF%DD, BMI only
explained, on average, 50% (between 18 and 85%)
of the variance in BF%DD. There was also a considerable overlap in the 95% CI of BF%DD among the
various BMI groups, in both males and females (Table
2), which indicates that the use of BMI alone cannot
be used to discriminate between individuals of differing BF%DD. Hence, other factors need to be considered when an estimate of body fat is required, in
addition to height and weight. These results are
similar to an earlier study in female adolescent volunteers and adult patients.23 The use of BMI as a
surrogate for body fatness is also hard to justify at
the level of the individual given its poor sensitivity
and positive predictive value (Table 5). The poor
sensitivity suggests that if one relies on BMI alone,
only about 50% of overweight=obese individuals
would be identi®ed. Furthermore, the positive predictive value indicates that only about two-thirds of those
identi®ed as overweight=obese would be truly overweight=obese. However, the higher speci®city of BMI
implies that it is better at identifying those individuals
who are not overweight=obese. Malina and Katzmarzyk,24 in a study on adolescents, also concluded that
BMI had a low sensitivity but high speci®city to
identify those at risk of obesity or those that were
obese.
Body composition changes with increasing age, and
for the same BMI young and older individuals may
have very different body composition.17,25,26 Indeed,
the observation that a relatively high BMI (27 ±
30 kg=m2 for men, 30 ± 35 kg=m2 for women) is associated with minimum hazard in persons older than
70 y has been made.27 This would suggest that a BMIbased classi®cation of weight status to identify those
at increased risk of chronic disease might not be
universally applicable across all age groups. In a
study of postmenopausal women, abnormalities of
lipid metabolism correlated with the amount of
upper-half-body fat, irrespective of age and BMI.28
This implies that body composition and fat distribution may be more useful than BMI in identifying those
International Journal of Obesity
with metabolic disorders that predispose to chronic
disease. In a study of female adolescent volunteers
and adult patients, a highly signi®cant relationship
was found between BMI and BF% as obtained by
DEXA absorptiometry.23 However, only 58% of the
variance in BF% in adolescents and 66% in adults
could be predicted by BMI, similar to values reported
in this study. They also observed that without any
change in BMI an adolescent's BF% could vary by as
much as 7 3% to ‡ 7%. In addition, for an individual
adult the same BMI could correspond to difference in
fat of 5%.23 Similarly, in other studies on young
subjects, con®dence limits on a BMI ± fatness association were wide, with individuals of similar BMI
showing large differences in total body FM and in
body fat percentage.29 ± 31
We also compared estimates of FFM from SFT and
the equations of Durnin and Womersley,13 and those
from BIA and the equations of Lukaski et al, 8 Segal
et al10 and Heitmann11 with estimates of FFM
obtained from DD. We used the equations of Segal
et al,10 which did not require a priori designation of
`lean' or `obese' using anthropometry, as the authors
themselves questioned the practical application of
their fatness speci®c equations.10 While the bias
between SFT-derived FFM and DD was signi®cant
in both gender groups, there was no relation between
bias and the mean of the two estimates (Table 4).
Hence, the addition of 1.2 kg in males and 0.8 kg in
females to the SFT estimate is recommended. All
equations used to predict FFM from BIA data,
except for the Segal et al10 equation in females, also
resulted in a small but statistically signi®cant bias in
each gender group. In contrast to the SFT data we
found a signi®cant relationship between the bias and
the mean FFM obtained by BIA and DD, for all
comparisons made (Table 4). Hence, a simple correction factor cannot be used for BIA derived estimates
of FFM. Results of other studies in Anglo-Celtic
Australians14,32 support these conclusions.
At the level of the individual, BIA estimates of
FFM could not be used with any degree of con®dence,
as the limits of agreement (LOA) of the BIA estimates
of FFM were wide (Table 4). For example, an estimate of FFM obtained by use of the Segal et al
equation for males,10 could be up to 8.8 kg higher or
6.3 kg lower than the estimate obtained by DD in the
same individual. For an 80 kg man with a FFM of
60 kg, this could theoretically represent an estimated
range of BF% of between 14.0% and 33%. Clearly,
such differences in estimates of body fat cannot be
tolerated at the level of the individual. This result is
disappointing, as BIA has been proposed as a simple
tool to estimate body fat in epidemiological studies.7
The LOA of BF% by BIA and SFT against DEXA in
two studies of Anglo-Celtic Australians were also
wide, indicating poor agreement between both SFT
and BIA with DEXA at the individual level.14,32 This
is also evident from an earlier study by Tagliabue
et al,33 who made BIA measurements on 38 healthy
Accuracy of indirect estimates of body composition
LS Piers et al
adults before and after a low calorie diet for 5 weeks.
A mean weight loss of 4.2 2.3 kg was observed.
According to BIA estimates, FFM decreased in 28
subjects and increased in 10. In four cases, the
estimated reduction in FFM was greater than the
weight loss! In another study on body composition
in groups of potentially malnourished patients, an
estimation of the level of agreement in the percentage
of lean tissue between DEXA absorptiometry and BIA
by the Bland and Altman method showed a bias
of 7 0.07% and LOA from 7 8.0% to 7.8%.34
Correlation between whole-body impedance measurements and various bio-conductor volumes, such as
TBW and FFM, are experimentally well established;
however, the reason for the success of the impedance
technique is much less clear.12 In practice BIA
machines introduce into the body a known current,
most often about 800 mA, usually at a frequency of
50 kHz. The current passes between two electrodes
and generates a voltage between different points in the
body. The electrodes are usually placed on the wrist
and ankle and the current passes through all conducting material between the source and detecting electrodes. The body is a volume conductor and the current is
carried by charged ions. Because the current passes
along the path of least resistance, the paths will differ
from person to person because of differences in body
size, shape, electrolytes, ¯uid distribution and measurement conditions. It will also vary in the same
person from time to time as these characteristics
change. Obese individuals with underlying cardiovascular and=or renal conditions are likely to have
alterations in extra-cellular water (ECW) to TBW
ratio, which would then lead to under or overestimates
of FFM based on measures of TBW. The subjects
participating in this study were healthy volunteers,
had no clinical evidence of oedema or dehydration,
and were not suffering from an illness or on medication that was likely to in¯uence the ratio of ECW:
TBW volume. Other factors that are known to in¯uence BIA estimates of body composition are electrode
con®guration,35 skin temperature,36 exercise-induced
dehydration,37 prior food consumption,38,39 and body
position.40 All these variables, however, were
controlled for in our study.
It is well established that the limbs make the largest
contribution to whole-body resistance.12 It is therefore, not surprising that step-wise regression revealed
that the biceps SFT and MAC together with FFMHe
provided a better estimate of FFMDD both in males
and females, resulting in a smaller bias and narrower
LOA (Table 4). While the LOA did improve, the
improvement was not large enough to improve estimates of FFM at the individual level. In contrast to
our study, Rutishauser et al,32 who used DEXA
estimates of body fat as their reference standard,
found that the comparability and precision of body
fat estimates derived from Heitmann's BIA equations11 could not be improved by adjusting for differences in umbilical circumference. This is not
surprising as the trunk accounts for < 10% of total
resistance.12 MAC was not reported in that study and
hence we cannot comment on this variable. However,
it is conceivable that inter-individual differences in
appendicular limb length to standing height may
contribute to the bias observed with various BIA
models. Unfortunately, appendicular limb length was
not measured in the current study, but future studies
could address this hypothesis.
In conclusion, our results suggest that BMI is a poor
surrogate for body fatness in both males and females.
There was a considerable scatter in BF%DD at any
given BMI. BMI also had a poor sensitivity and
positive predictive value when it was used to identify
subjects classi®ed as being overweight=obese by DD
estimates of body fatness (Table 5). Estimates of body
composition from SFT or BIA could not be used
interchangeably with those from DD, without the
risk of considerable error, at the level of the individual. However, at the level of the group, while errors
were relatively smaller, they were still statistically
signi®cant. The use of simple correction factors
(1.2 kg in males and 0.8 kg in females) improved the
accuracy of SFT estimates of FFMDD of both males
and females in this study. The combination of biceps
SFT, MAC measurements and FFMHe provided a
better estimate of FFMDD. The bias from the revised
equations of Heitmann were small ( < 0.15 kg), nonsigni®cant (P > 0.05) and unrelated to the mean of the
two methods. Importantly this revision improved the
LOA in both males and females and resulted in
increased sensitivity, speci®city and positive predictive value, when compared to the original (Table 5).
Future studies in population representative samples are
needed to extend the usefulness of such modi®cations
to the equations of Heitmann.11
1151
Acknowledgements
LSP was the recipient of a postdoctoral research
fellowship from Deakin University, Australia. He is
currently supported by a grant (981019) from the
National Health and Medical Research Council, Australia. MJS is the recipient of a Kellogg's Senior
Lectureship. We wish to thank Ms Leanne L McCormack, Mr Mark Lutschini, Dr Karen Walker, Ms
Connie Karschimkus, Ms Olga Strommer and Ms
Bronwyn Diffey for their assistance with the project,
Ms Gail Dyson for assistance with the NMR analyses,
and Dr Kevin G Rowley for useful comments on the
manuscript.
References
1 World Health Organisation (WHO). Obesity: preventing and
managing the global epidemic. WHO: Geneva, 1997.
2 Norgan NG. Population differences in body composition in
relation to the body mass index. Eur J Clin Nutr 1994;
48(Suppl 3): S10 ± S25.
3 Rush EC, Plank LD, Laulu MS, Robinson SM. Prediction of
percentage body fat from anthropometric measurements: comparison of New Zealand European and Polynesian young
women. Am J Clin Nutr 1997; 66: 2 ± 7.
International Journal of Obesity
Accuracy of indirect estimates of body composition
LS Piers et al
1152
4 Gurrici S, Hartriyanti Y, Hautvast JG, Deurenberg P. Relationship between body fat and body mass index: differences
between Indonesians and Dutch Caucasians. Eur J Clin Nutr
1998; 52: 779 ± 783.
5 Deurenberg P, Yap M, van Staveren WA. Body mass index
and percent body fat: a meta analysis among different
ethnic groups. Int J Obes Relat Metab Disord 1998; 22:
1164 ± 1171.
6 Gurrici S, Y Hartriyanti Y, Hautvast JGAJ, Deurenberg P.
Differences in the relationship between body fat and body
mass index between two different Indonesian ethnic groups:
the effect of body build. Eur J Clin Nutr 1999; 53:
468 ± 472.
7 Brodie D, Moscrip V, Hutcheon R. Body composition measurement: a review of hydrodensitometry, anthropometry, and
impedance methods. Nutrition 1998; 14: 296 ± 310.
8 Lukaski HC, Bolonchuk WW, Hall CB, Siders WA. Validation of tetrapolar bioelectrical impedance method to
assess human body composition. J Appl Physiol 1986; 60:
1327 ± 1332.
9 Khaled MA, Lukaski HC, Watkins CL. Determination of total
body water by deuterium NMR. Am J Clin Nutr 1987; 45:
1 ± 6.
10 Segal KR, Van Loan M, Fitzgerald PI, Hodgdon JA, Van
Itallie TB. Lean body mass estimation by bioelectrical impedance analysis: a four-site cross-validation study. Am J Clin
Nutr 1988; 47: 7 ± 14.
11 Heitmann BL. Evaluation of body fat estimated from body
mass index, skinfolds and impedance. A comparative study.
Eur J Clin Nutr 1990; 44: 831 ± 837.
12 Foster KR, Lukaski HC. Whole-body impedance Ð what does
it measure? Am J Clin Nutr 1996; 64(Suppl 3): 388S ± 396S.
13 Durnin JV, Womersley J. Body fat assessed from total body
density and its estimation from skinfold thickness: measurements on 481 men and women aged from 16 to 72 y. Br J Nutr
1974; 32: 77 ± 97.
14 Wattanapenpaiboon N, Lukito W, Strauss BJ, Hsu-Hage BH,
Wahlqvist ML, Stro-d DB. Agreement of skinfold measurement and bioelectrical impedance analysis (BIA) methods with
dual energy X-ray absorptiometry (DEXA) in estimating total
body fat in Anglo-Catic Australians. Int J Obes Relat Metab
Disord 1998; 22: 854 ± 860.
15 Callaway CW, Chumlea WC, Bouchard C, Himes JH, Lohman
TG, Martin AD, Mitchell CD, Mueller WH, Roche AF,
Seefeldt VD. Circumferences. In: Lohman TG, Roche AF,
Martorell R (eds). Anthropometric standardisation reference
manual. Human Kinetics: Champaign, IL, 1988, pp 39 ± 54.
16 Harrison GG, Buskirk ER, Lindsay Carter JE, Johnston FE,
Lohman TG, Pollock ML, Roche AF, Wilmore J. Skinfold
thicknesses and measurement technique. In: Lohman TG,
Roche AF, Martorell RI(eds) Anthropometric standardisation
reference manual. Human Kinetics: Champaign, IL, 1988, pp
55 ± 70.
17 Piers LS, Soares MJ, McCormack LM, O'Dea K. Is there
evidence for an age-related reduction in metabolic rate? J Appl
Physiol 1998; 85: 2196 ± 2204.
18 Pace N, Rathburn EN. Studies on body composition. III. The
body water and chemically combined nitrogen content in
relation to fat content. J Biol Chem 1945; 158: 685 ± 691.
19 Sergi G, Perini P, Bussolotto M, Zurlo F, Malvasi L, Carraro
R, Del Prato S, Giantin V, Enzi G. Body composition study in
the elderly: comparison between tritium dilution method
and dual photon absorptiometry. J Gerontol 1993; 48:
M244 ± 248.
20 Lukaski HC. Methods for the assessment of human body
composition: traditional and new. Am J Clin Nutr 1987; 46:
537 ± 556.
International Journal of Obesity
21 Bland JM, Altman DG. Statistical methods for assessing
agreement between two methods of clinical measurement.
Lancet 1986; 1(8476): 307 ± 310.
22 Beaglehole R, Bonita R, Kjellstrom T. Basic epidemiology.
World Health Organization: Geneva, 1993.
23 Hannan WJ, Wrate RM, Cowen SJ, Freeman CP. Body mass
index as an estimate of body fat. Int J Eat Disord 1995; 18:
91 ± 97.
24 Malina RM, Katzmarzyk PT. Validity of the body mass index
as an indicator of the risk and presence of overweight in
adolescents. Am J Clin Nutr 1999; 70: 131S ± 136S.
25 Heitmann BL. Body fat in the adult Danish population aged
35-65 y: an epidemiological study. Int J Obes 1991; 15: 535 ±
545.
26 Gallagher D, Visser M, SepuÂlveda D, Pierson RN, Harris T,
Heyms®eld SB. How useful is body mass index for comparison of body fatness across age, sex, and ethnic groups? Am J
Epidemiol 1996; 143: 228 ± 289.
27 Allison DB, Gallagher D, Heo M, Pi-Sunyer FX, Heyms®eld
SB. Body mass index and all-cause mortality among people
age 70 and over: the Longitudinal Study of Aging. Int J Obes
Relat Metab Disord 1997; 21: 424 ± 431.
28 Douchi T, Ijuin H, Nakamura S, Oki T, Yamamoto S,
Katanozaka M, Nagata Y. The relation between body-fat
distribution and lipid metabolism in postmenopausal women.
J Obstet Gynaecol Res 1996; 22: 353 ± 358.
29 Marshall JD, Hazlett CB, Spady DW, Conger PR, Quinney
HA. Validity of convenient indicators of obesity. Hum Biol
1991; 63: 137 ± 153.
30 Schaefer F, Georgi M, WuÈhl E, Scharer K. Body mass index
and percentage fat mass in healthy German schoolchildren
and adolescents. Int J Obes Relat Metab Disord 1998; 22:
461 ± 469.
31 Pietrobelli A, Faith MS, Allison DB, Gallagher D, Chiumello
G, Heyms®eld SB. Body mass index as a measure of adiposity
among children and adolescents: a validation study. J Pediatr
1998; 132: 204 ± 210.
32 Rutishauser IH, Pasco JA, Wheeler CE. The in¯uence of body
build on estimates of body composition from anthropometric
measurements in premenopausal women. Eur J Clin Nutr
1995; 49: 248 ± 255.
33 Tagliabue A, Cena H, Trentani C, Lanzola E, Silva S. How
reliable is bio-electrical impedance analysis for individual
patients? Int J Obes Relat Metab Disord 1992; 16:
649 ± 652.
34 Ghosh S, Meister D, Cowen S, Hannan WJ, Ferguson A. Body
composition at the bedside. Eur J Gastroenterol Hepatol 1997;
9: 783 ± 788.
35 Lukaski HC. Comparison of proximal and distal placements of
electrodes to assess human body composition by bioelectrical
impedance. Basic Life Sci 1993; 60: 39 ± 43.
36 Caton JR, Mole PA, Adams WC, Heustis DS. Body composition analysis by bioelectrical impedance: effect of skin temperature. Med Sci Sports Exerc 1988; 20: 489 ± 491.
37 Brodie DA, Eston RG, Coxon AY, Kreitzman SN, Stockdale
HR, Howard AN. Effect of changes of water and electrolytes
on the validity of conventional methods of measuring fat-free
mass. Ann Nutr Metab 1991; 35: 89 ± 97.
38 Gallagher M, Walker KZ, O'Dea K. The in¯uence of a
breakfast meal on the assessment of body composition using
bioelectrical impedance. Eur J Clin Nutr 1998; 52: 94 ± 97.
39 Fogelholm M, SievaÈnen H, Kukkonen-Harjula K, Oja P, Vuori
I. Effects of meal and its electrolytes on bioelectrical impedance. Basic Life Sci 1993; 60: 331 ± 332.
40 Pinilla JC, Webster B, Baetz M, Reeder B, Hattori S, Liu L.
Effect of body positions and splints in bioelectrical impedance
analysis. J Parenter Enteral Nutr 1992; 16: 408 ± 412.