Professional Documents
Culture Documents
1186 Am J Clin Nutr 2003;77:1186–91. Printed in USA. © 2003 American Society for Clinical Nutrition
SKINFOLD-THICKNESS PREDICTION EQUATIONS 1187
We developed a new skinfold-thickness equation (%BFnew) to oxide (99.9% purity; Cambridge Isotope Laboratories, Woburn,
predict %BF by using the 4C model as the reference in younger MA) in a solution of 75 mL deionized water. A second saliva sam-
and middle-aged adults (18–55 y). We evaluated the performance ple was collected exactly 2 h later. The 2 samples were centrifuged
of %BFnew compared with %BF estimates derived from 2 com- and the supernate was collected, sealed, frozen, and transported
monly used skinfold-thickness equations, namely those of Durnin to the Kettering-Scott Magnetic Resonance Laboratory (Ketter-
and Wormersley (%BFDW) and Jackson and Pollock (%BF JP). ing, OH) for analysis. Nuclear magnetic resonance spectrometry
Because technology that allows us to use a 4C model for valida- was used to determine the TBW from each sample. These values
tion studies is now accessible, we decided it was important to were then divided by 1.04 to account for the estimated 4% H+ non-
examine the possibility of providing clinicians and practitioners aqueous exchange (21).
with improved skinfold-thickness prediction equations. The hydrodensitometry method was used to measure BD
according to the method described by Siri (7). This method deter-
mines BD by using standardized hydrostatic weighing with cor-
SUBJECTS AND METHODS rection for residual volume. Residual volume was measured twice
on land to the nearest 0.1 L by nitrogen washout, using a com-
TABLE 1 TABLE 2
Descriptive characteristics of the validation and cross-validation groups1 Regression coefficients, partial explanation of variance (partial R2), and
significance of independent variables in the new percentage body fat
Value
(%BFnew) equations (validation group)
Validation group
Variable Coefficient Partial R2 P
Men (n = 274)
Age (y) 36.2 ± 11.0 (18.0–55.4) Men (n = 273)
Height (cm) 179.3 ± 6.6 (158.5–193.6) Intercept 20.94878 0.0091
Weight (kg) 79.0 ± 11.6 (45.2–109.0) Age 0.1166 0.5609 < 0.0001
BMI (kg/m2) 24.5 ± 3.2 (15.6–34.6) Stature 0.1166 0.0277 0.0082
Women (n = 230) Sum4 0.42696 0.0192 < 0.0001
Age (y) 35.1 ± 10.3 (18.0–55.6) Sum42 0.00159 0.0101 0.0002
Height (cm) 166.0 ± 5.6 (154.3–187.0) Model R2 0.6122 < 0.0001
Weight (kg) 63.5 ± 11.6 (43.3–108.6) Women (n = 229)
BMI (kg/m2) 23.0 ± 3.7 (16.5–40.6) Intercept 22.18945 0.0063
%BFnew equations
Croswell, Crymych, United Kingdom) and a metal or cloth (for The physical characteristics of subjects in the validation
upper arm only) tape measure. and cross-validation groups are shown in Table 1. In both
groups, men and women differed significantly in height,
Statistical analyses weight, and BMI, but age was not significantly different
The physical characteristics of the validation and cross-validation between the sexes. Men in the validation group were heavier
groups are presented as means with SDs and ranges. Unpaired t tests than were men in the cross-validation group. Physical char-
were used to determine significant differences between mean val- acteristics did not differ significantly between women in the
ues within each sex between groups (validation and cross-validation validation and cross-validation groups.
groups) and between the sexes within groups. Sex-specific step- Regression analyses in the validation group (Table 2) revealed
wise regression analyses were used to develop the %BFnew equa- that the sum of the triceps, subscapular, suprailiac, and midthigh
tions in the validation group. Variables included in the initial skinfold thicknesses (sum4) explained 56% (partial R2 = 0.56)
analyses included age; height (cm); weight (kg); BMI; the sum of of the variance of %BF in men and 65% (partial R2 = 0.64) of
the triceps, suprailiac, and midthigh skinfold thicknesses (sum3); the variance of %BF in women. In men, age, height, and sum42
and the sum of the triceps, subscapular, suprailiac, and midthigh were also significant contributors to the model, but to a lesser
skinfold thicknesses (sum4). Squares of sum3 (sum32) and sum4 degree, with no remaining variable’s partial R2 being > 0.027. In
(sum42) were also included to test for a curvilinear association women, in addition to the same set of variables that were significant
between skinfold thicknesses and %BF. in the model for men, BMI was also a significant predictor of %BF,
In the cross-validation group, the accuracy of the %BF new, although its effect size was relatively small (partial R2 = 0.024).
%BFDW, and %BFJP equations was tested by calculating correla- Interestingly, sum3, sum32, circumferential measures, and weight
tion coefficients. In addition, Bland-Altman plots (23) were used were not significant predictors of %BF in either model. The final
to determine bias and limits of agreement between equations. Bias equations, which were used in subsequent analyses, are as follows
was defined as (predicted %BF %BF4C) plotted on the y axis for men and women:
and the mean of %BF4C and predicted %BF plotted on the x axis.
Limits of agreement were defined as upper and lower 95% CIs and For men: %BFnew = 20.94878 + (age 0.1166)
were determined by the mean differences ± 2 SD. To determine (height 0.11666) + (sum4 0.42696)
each equation’s precision, root mean square error (RMSE) values (sum42 0.00159) (3)
were obtained by using univariate analysis. In all the analyses,
For women: %BFnew = 22.18945 + (age 0.06368)
%BF4C was used as the method of reference. As described previ-
ously, chest skinfold data were not available for men, so analyses + (BMI 0.60404) (height 0.14520)
with Jackson and Pollock equations are for women only. Alpha + (sum4 0.30919) (sum42 0.00099562) (4)
was set at P ≤ 0.05. We used a Bonferroni adjustment for multi- where height is in cm and sum4 is the sum of the triceps, sub-
ple comparisons of %BF methods and set at P < 0.025 for men scapular, suprailiac, and midthigh skinfold thicknesses.
SKINFOLD-THICKNESS PREDICTION EQUATIONS 1189
TABLE 3
Percentage body fat (%BF) estimates and measurements in the
cross-validation group1
Value
Men (n = 86)
%BFnew 22.7 ± 7.1 (11.1–36.6)
%BF4C 22.8 ± 9.5 (4.6–44.3)
%BFJP —
%BFDW 20.0 ± 7.02 (7.6–34.8)
Women (n = 91)
%BFnew 32.6 ± 5.2 (21.7–43.2)
%BF4C 32.8 ± 7.2 (14.1–46.2)
%BFJP 26.2 ± 5.92 (13.6–39.8)
%BFDW 31.0 ± 5.52 (20.4–43.9)
FIGURE 1. Bland-Altman plot showing the limits of agreement between FIGURE 3. Bland-Altman plot showing the limits of agreement between
percentage body fat calculated with the 4-compartment-model equation percentage body fat calculated with the 4-compartment-model equation
(%BF4C; the reference equation) and percentage body fat calculated with the (%BF4C; the reference equation) and percentage body fat calculated with the
new equation developed in the current study (%BFnew) in men. Durnin and Wormersley equation (%BFDW) in men.
1190 PETERSON ET AL
their means in women are shown in Figure 5. Bias with the Jack- groups similar to the one presented here. Bland-Altman plots pro-
son and Pollock equation is seen, with a mean difference of vided additional information regarding the underestimation of
6.6% (P < 0.05). Limits of agreement follow the same tendency %BF by the Durnin and Wormersley and Jackson and Pollock
of underestimation, with upper and lower confidence limits at equations. With the Durnin and Wormersley equations, the lower
3.3% and 16.5%, respectively. Relations are evident between the limits of agreement (mean differences 2 SD) were 12.9% and
mean of the measured and predicted values and the individual dif- 11.4% in men and women, respectively. These lower limits of
ferences in %BF in both sexes with all equations. In other words, agreement are considerably larger than the upper limit values, at
the relations seen in all equations are a function of %BF, with all 7.2% and 7.9% in men and women, respectively. On the basis of
equations having a tendency to produce larger underestimations these data, a man or woman is more likely to have their %BF
as %BF increased. underestimated using the Durnin and Wormersley equations. This
To address issues of precision, RMSEs of predicted %BF in the tendency to have the limits of agreement shifted toward underes-
cross-validation groups were investigated. %BFnew had a RMSE timations is even more pronounced in women using the Jackson
for men of 4.6% when predicting %BF, and the corresponding and Pollock equation, with upper and lower limits of agreement
value for %BFDW was 4.9%. In women, %BFnew equations had a of 3.3% and 16.5%, respectively. This indicates that a woman
RMSE of 5.0%. The corresponding values for %BFDW and %BFJP could have her %BF underestimated by as much as 16.5%. These
in women were both 4.9%. Overall, RMSEs were quite similar underestimations are most likely a result of systematic overesti-
among all the prediction equations. mations of BD. For example, the mean BD value with the Durnin
and Wormersley equation was higher, at 1.032 g/mL, than the
value of 1.028 g/mL that was measured by hydrodensitometry.
DISCUSSION Although 0.004 g/mL may initially appear to be a small difference
We developed new skinfold-thickness prediction equations and in BD estimates, when using these 2 values to predict %BF, the
compared them with the equations of Durnin and Wormersley and differences in mean predicted %BF values are almost 2%. A rela-
Jackson and Pollock, and with a 4C model, which we used as the tively small under- or overestimation of BD can result in consid-
reference equation. To our knowledge, no previous study has used erable errors in predicted %BF values.
a 4C model to cross-validate several skinfold-thickness prediction %BFnew, %BFJP, and %BFDW all had similar estimates of preci-
equations. We chose the Durnin and Wormersley and Jackson and sion, as shown by the similar RMSE values. A prediction equa-
Pollock equations because of their popularity in the field and in tion’s RMSE is certainly important, however, it should not be con-
research settings (18, 24–31). Previous studies have explored the sidered the only criterion when choosing a proper equation.
predictive accuracy and precision of the Jackson and Pollock and Although %BF DW and %BF JP values were similar in terms of
Durnin and Wormersley equations in different populations (18, RMSE values and precision, we showed that accuracy was less
24–31). Our study is also unique in that we present data from than desirable in the Durnin and Wormersley and Jackson and Pol-
healthy, white men and women who had characteristics compara- lock equations in a cohort similar in age and physical character-
ble to those of the groups used by Durnin and Wormersley and istics to their derivation samples.
Jackson and Pollock. Providing precise and accurate body-composition information
The skinfold-thickness equations developed in our laboratory is of great importance. The comparisons provided in this paper
had no significant mean differences with the %BF4C model in men should aid in making a more informed decision, considering pre-
or women in the cross-validation group. In this cohort, %BFDW cision and accuracy, when choosing among popular equations for
and %BFJP means were significantly lower than the %BF4C mean, predicting %BF. In summary, we have shown that our newly
which may need to be considered when using the Durnin and developed generalized equations, which use the 4C model as a ref-
Wormersley and Jackson and Pollock equations to predict %BF in erence, can accurately predict %BF in healthy, white adults. Our
SKINFOLD-THICKNESS PREDICTION EQUATIONS 1191
newly developed skinfold-thickness equations are important for 2 fat estimations from a four-compartment model by using density,
reasons. First, the %BFnew equations were validated and cross- water, and bone mineral measurements. Am J Clin Nutr 1992;55:
validated with large samples of men and women spanning a wide 764–70.
age range (18–55 y). The validation sample used by Durnin and 12. Williams DP, Going SB, Lohman TG, Hewitt MJ, Haber AE. Estima-
Wormersley was similar to ours in size, but they did not cross- tion of body fat from skinfold thicknesses in middle-aged and older
validate their equations. Jackson and Pollock did cross-validate, men and women: a multiple component approach. Am J Hum Biol
1992;4:595–605.
but their cross-validation did not provide an analysis of their equa-
tions’ potential to over- or underestimate %BF. Second, the 13. Withers RT, LaForgia J, Pillans RK, et al. Comparisons of two-, three-,
and four-compartment models of body composition analysis in men
%BFnew equations were developed and validated by using a 4C
and women. J Appl Physiol 1998;85:238–45.
model as a reference. The 4C model has been shown to be more
14. Heyward VH, Stolarczyk LM. Applied body composition assessment.
accurate in measuring %BF than the 2C method, which was the Champaign, IL: Human Kinetics, 1996.
criterion method for Durnin and Wormersley and Jackson and Pol-
15. Clasey JL, Kanaley JA, Wideman L, et al. Validity of methods of body
lock. When validated against the 4C model, the Durnin and Worm- composition assessment in young and older men and women. J Appl
ersley equations underestimated %BF in men and women and the Physiol 1999;86:1728–38.