Prediction of Body Fat by Anthropometry in Older Chinese People


Department of Medicine and Therapeutics, Prince of Wales Hospital, The Chinese University of Hong Kong, Shatin, Hong Kong. E-mail:


Objective: To derive regression equations for fat percentage by using simple anthropometric measurements applicable in normal and immobile (cannot stand or walk) older people.

Research Methods and Procedures: The study population comprised 352 females and 261 males, apparently well and community-dwelling, aged 69 to 82 years. Fifty-one females and 27 males were recruited for external validation. Body weight, standing height, arm span, triceps and biceps skinfold thicknesses (SFTs), and midarm circumference were measured. The reference method of total body fat percentage was dual-energy X-ray densitometry. Predictive equations for fat percentages were derived by stepwise multiple linear regression on anthropometric indices and gender.

Results: Upper-limb SFTs, body mass index, and gender yielded the more predictive equation. The SEE was 4.1% weight. There was a significant trend of underestimation in overweight subjects, especially in females. The equation using SFTs and midarm circumference was less reliable but more applicable to older immobile people and those with significant kyphoscoliosis.

Conclusions: The combination of body mass index and upper-limb SFTs gives reliable prediction of fat percentages in older Chinese people, except in the obese.


Body mass index (BMI) and skinfold thickness (SFT) have been used to estimate body fat by regression formulas either separately (1) or in combination (2). However, there are age-related and racial differences in the application of these equations (3). Therefore, there is a need to derive predictive equations for body fat with simple anthropometric measurements that can be used in older Chinese people.

We have reported previously that upper-limb SFTs provided reasonable estimates of fat percentages as estimated by DXA in older Chinese people (4). However, there was a tendency to underestimate fat in overweight people, and the 95% confidence interval (CI) of error of estimation was close to ±10% body weight.

Therefore, we derived more reliable equations for fat percentages by combining upper-limb SFTs with other commonly used anthropometric measurements. Special consideration was made for their applicability in disabled older people. BMI was chosen because of its independence from body fat distribution and its high correlation with body fat (1). In kyphotic people, however, the measurement of height is unreliable; arm span has been suggested to be a good substitute for height in older people (5). Therefore, BMI that was calculated by arm span was included in the regression analysis. In immobile (cannot stand or walk) people, the measurement of weight is burdensome and is rendered unreliable by fluid retention. Midarm circumference (MAC) has a high correlation with body fat in older people (4); therefore, it was also considered.

Research and Methods and Procedures

The study population was recruited from a public housing estate in Hong Kong and comprised 352 females and 261 males, aged 69 to 82 years, who were apparently healthy. Forty subjects of each gender were randomly selected for internal validation of regression equations. The remaining 312 female and 221 male subjects formed the determination group from which regression equations were derived. Fifty-one females and 27 males who were medically stable, self-ambulatory medical outpatients, aged 65 to 86 years, were recruited separately for external validation. The most common diagnoses were hypertension (N = 34) and musculoskeletal problems (N = 18). Those with significant fluid retention, type 2 diabetes, or malignancy were excluded. A simple questionnaire to inquire about level of physical activity and smoking habits was administered to all study subjects except the external validation subjects. This study was approved by the Clinical Research Ethical Committee of the Chinese University of Hong Kong.

Body weight was measured to the nearest 0.1 kg by a beam balance; subjects were wearing light indoor clothes. Standing height without shoes was measured to the nearest 0.5 cm by a height measure attached to a beam balance. Arm span was measured to the nearest 0.1 cm with the subject's back against a wall and arms extended at right angles to the body with palms facing forward (5). BMI (BMI = body weight/standing height2 [kg/m2]) and BMI calculated by arm span (BMI[arm span] = body weight/arm span2 [kg/m2]) were then calculated.

Triceps skinfold thickness (TSF) and biceps skinfold thickness (BSF) of the left arm were measured to the nearest 0.1 mm with a pair of calipers (Holtain, Crymych, UK), in duplicate. MAC was measured to the nearest 0.1 cm at the same level as TSF with a soft tape measure. The average of two readings was taken. The waist-to-hip ratio was measured in external validation subjects only.

All subjects had body fat percentages (Fat%DXA) estimated by DXA (Hologic QDR-2000 bone densitometer and Enhanced Array Whole Body software, version 5.67A; Hologic, Bedford, MA). All subjects wore a standard light gown.

Statistical Analysis

Linear regression equations for Fat%DXA were derived by using stepwise multiple regression analysis on the determination group. The following variables were initially considered: BMI, BMI(arm span), log (TSF + BSF), and MAC. Gender was included as a dummy variable (male = 1, female = 2). The validity of the predictive regression equations was examined in the internal validation groups and the external validation groups as suggested by Bland and Altman (6). Error of estimation was defined as fat percentage as calculated by the following equation: (Fat%Equation) − (Fat%DXA). Bias was the mean of error of estimation; 95% CI of error were calculated by bias ± 1.96 × SD of error. Correlation between error and mean of Fat%DXA and Fat%Equation was ascertained by linear regression. For group comparison, Student's t tests were used for parametric data; Mann–Whitney tests were used for nonparametric data.


The characteristics of all of the study subjects are shown inTable 1. They were comparable with local norms (7), except that fewer male subjects reported regular exercise than the local average of men of >70 years of age (56% vs. 69%) (7). The regression equations derived from the determination group subjects are shown in Table 2. Equation 1 was the most predictive equation when all variables were considered in stepwise multiple linear regression analysis. Equation 2 was derived after BMI was excluded, for the benefit of kyphotic people whose height is unreliable. Equation 3 was derived after excluding BMI and BMI(arm span) for the benefit of immobile older people.

Table 1.  Characteristics of the study subjects
  • *

    p < 0.01

  • p < 0.05, significantly different from females (Mann–Whitney test for triceps and biceps skinfolds, unpaired t test for age and other anthropometric variables, andχ2 test for lifestyle variables).

  • More than one hour per day.

 Mean (SD)RangeMean (SD)Range
Age (years)76.7 (3.1)69–8273.8 (2.9)*69–79
BMI (kg/m2)23.4 (3.9)12.5–34.622.7 (3.6)13.8–36.9
Height (cm)147.0 (5.7)125–162.5161.0 (6.0)*144–188.5
Arm span (cm)153.3 (6.3)133–173167.4 (6.9)*146–185
Tricep SFT (mm)17.2 (6.8)2.4–35.310.6 (4.5)*1.8–25.9
Bicep SFT (mm)7.2 (3.4)2.1–22.25.7 (3.3)*1.0–24.5
MAC (cm)26.3 (3.5)12.1–36.527.2 (3.2)17–38
Fat% by DXA34.7 (8.5)10.0–51.924.8 (7.3)*7.1–42.2
Regular exercise2236314656
Outdoor walking24670127*49
Current smoker31973*28
Total352 261 
Table 2.  Regression equations to estimate body fat percentages from anthropometry
EquationR2CoefficientSESEE (% weight)
  • *

    1 for men, 2 for women.

Equation 10.81  4.1
Sex* 6.1370.413 
BMI 1.1200.069 
Log (TSF+ BSF) 17.3081.403 
Intercept −27.1491.225 
Equation 20.81  4.1
Sex 6.3480.425 
BMI (arm span) 1.1250.073 
Log (TSF+ BSF) 17.281.451 
Intercept −25.5721.212 
Equation 30.76  4.6
Sex 6.5720.552 
Log (TSF+ BSF) 23.6531.566 
MAC 0.7740.085 
Intercept −30.7021.786 

The anthropometric and lifestyle characteristics of the internal validation subjects were comparable with those of the whole study groups. The external validation subjects were also comparable with regard to age, BMI, and fat percentages, except that the male subjects were older (76.3 ± 4.1 years). The waist-to-hip ratios were 0.85 ± 0.06 and 0.91 ± 0.07 in males and females, respectively. The bias and 95% CI of error in the internal and external validation groups when Equations 1, 2, and 3 were applied are shown in Table 3. The errors of estimation (Fat%Equation minus Fat%DXA) were plotted against mean fat percentages (mean of Fat%Equation and Fat%DXA). The plot for Equation 1 is shown inFigure 1. There was a trend toward underestimation of fat percentages with increasing fat percentages in both genders with all three equations, but it was more marked in female external validation subjects, particularly for Equations 2 and 3. The plot of errors of estimation against mean fat percentages for Equations 2 and 3 had similar patterns, but those for Equation 2 had a wider scatter at both extremes of fatness.

Table 3.  Bias and 95% CIs of error when regression equations were applied to validation subjects*
 Validation subjects
 Females (N = 40)Males (N = 40)Females (N = 51)Males (N = 27)
  • *

    Presented as bias (95% CI of error), % weight;Error, fat percentage calculated by equation minus Fat%DXA; Bias, mean of error; 95% CI of error, bias ± 1.96 × SD of error.

  • **

    Significant difference between Fat%DXA and fat percentage calculated by equation, p < 0.01, paired t test.

  • Skinfolds, log (TSF + BSF).

  • Significant correlation between error and fat percentage on regression, p < 0.05.

  • §

    p < 0.001.

Equation 1: BMI and skinfolds−0.6 (−8.7, 7.4)−1.8 (−10.4, 6.9)−2.1 (−11.0, 6.8)** §−0.1 (−7.6, 7.5)
Equation 2: BMI (arm span) and skinfolds0.4 (−8.0, 8.9)−0.8 (−9.9, 8.4)−2.1 (−12.1, 7.9)** §0.2 (−6.9, 7.3)
Equation 3: Skinfolds and MAC−0.3 (−8.1, 7.6)§−1.4 (−10.3, 7.5)−2.0 (−11.7, 7.7)** §−0.2 (−7.7, 7.3)
Figure 1.

Scatter plot of error of estimation against mean Fat% in internal and external validation subjects for predictive Equation 1, using BMI and upper-limb SFTs. (A) Internal validation subjects. (B) External validation subjects. Dashed lines signify the upper and lower 95% CI of the error of estimations. Significant correlations between error of estimation and mean Fat% were found in male (A) and female (B).


DXA is recognized as a reliable and noninvasive means of measuring body composition. It has been validated in younger, local Chinese women (8). Although it can differ from the recognized gold standard of the four-compartment model by up to 10% weight in older adults (9), its readings can be converted to that of the latter by predictive equations (10).

The Chinese study subjects had significantly lower BMIs, TSFs, and MACs than their American counterparts (11), but they were comparable with those reported in local community surveys (12). Moreover, their levels of physical activity and smoking, which may have important influences on body composition in older people (13), were also similar to population norms (7). The predictive equations derived in this study should therefore be generally applicable to older Chinese people.

BMI and upper-limb anthropometry were selected for study primarily because they can be easily measured without the need to undress. We previously reported predictive equations of fat percentages from upper-limb SFTs alone in each gender, with explained variances of 0.70 and 0.51 in female and male subjects, respectively (4). The addition of BMI and gender into the equation increased the explained variance to 0.81 and reduced the SE of estimation (SEE) from ∼5% to 4.1%.

Upper-limb anthropometry is liable to underestimate fat percentages in older people because of the centripetal fat distribution associated with aging (14). The addition of BMI that is independent of fat distribution into the predictive equations did not appear to significantly reduce the trend of underestimation, suggesting that BMI shares similar tendencies for underestimation with overweight. This could be explained by the loss of lean mass with aging and lack of exercise, thus increasing the relative proportion of fat even without weight increase. Because the study subjects were mostly physically active, the accuracy of the equation may be reduced in those who are physically inactive.

The trend of underestimation appeared to be more marked in female outpatient subjects. A possible explanation was that many of these subjects had hypertension that is associated with centripetal fat distribution and internal abdominal fat deposition. There was a significant correlation between waist-to-hip ratio and error of estimation with all three equations (correlation coefficients ranged from −0.39 to −0.48). The inclusion of waist-to-hip ratio or waist circumference may therefore improve the reliability of the predictive equation.

Although Equation 2 using BMI(arm span) had higher explained variance than Equation 3 using MAC, the validity of the two equations was similar on the Bland and Altman (6) model of error analysis. Because the measurement of body weight is problematic, Equation 3 is preferred for immobile people and those with kyphoscoliosis. When compared with the equation using SFTs only (4), Equation 3 had a narrower 95% CI of errors and higher explained variance. Because immobile people have accelerated loss of lean and bone mass, further validation is warranted.

We conclude that the combination of BMI with upper-limb SFTs gave a more reliable estimation of body fat than upper-limb SFTs alone in older Chinese individuals. The equation using MAC and upper-limb SFTs can be used in immobile older people and in those with significant kyphoscoliosis.


No outside funding/support was provided for this study.