Total body fat levels, as well as regional fat distribution, are associated with risk for a variety of diseases ^{(4)}. An upper body or visceral fat distribution pattern is associated with an increased risk for cardiovascular disease and diabetes independent of the effects of overall obesity ^{(7)}. As a result, it is important to accurately assess both total body fat and fat distribution, particularly in individuals with an upper body fat pattern.

Numerous techniques are available for the estimation of body fat and body fat distribution. Technologically advanced methodologies such as computed tomography, dual x-ray absorptiometry, and magnetic resonance imaging have improved the precision and accuracy of both body composition and body fat distribution assessment in the laboratory setting ^{(21)}. However, these methodologies are not available to many practitioners and require extensive time, expense, and expertise. In contrast, common field methods such as circumferences, skinfolds, and bioelectrical impedance are portable, relatively inexpensive, noninvasive, and require minimal amounts of time and training. As such, field methods have greater potential for use by hospitals, corporations, universities, and sports and fitness facilities ^{(6)}. Unfortunately, these practical methods for assessing body composition tend to be associated with greater error ^{(18)}.

A critical aspect regarding the accuracy of field methods lies in the use of appropriate regression equations. Numerous equations have been developed using some combination of anthropometric measures predicted against a criterion measure such as hydrostatic weighing ^{(12)}. These equations may be either population-specific or generalizable. Population-specific equations are selected based on the specific age, gender, ethnicity, and physical activity level of the individual ^{(10)}. In contrast, generalizable equations for assessing body composition are designed to apply to large, fairly homogeneous groups of men or women of varying ages and body composition ^{(5,13,14,23,26)}. Many generalizable equations use a variety of sites for skinfold and/or circumference measurements and assume a constant distribution of fat among individuals. However, the increasing prevalence of people classified as overweight in the United States ^{(15)} may increase the heterogeneity between people and limit the assumptions of fat distribution when using body fat estimation methods. As a result, it could be expected that the use of generalizable field equations on "uniquely shaped" individuals who do not share a standard or typical body fat distribution pattern would introduce greater error in the assessment of total body fat ^{(25)}.

Results from a pilot investigation conducted at our laboratory ^{(20)} suggested that the commonly used Jackson, Pollock and Ward ^{(13)} generalized skinfold equation may not be an appropriate field estimate of percent fat in women with a distinctly upper body fat distribution pattern. Thirty healthy, normal weight, premenopausal women were matched for body mass index (BMI) and characterized by waist to hip circumference ratio (WHR) into two distinct body fat distribution groups: lower body (LB; n = 16; WHR ≤ 0.75) and upper body (UB; n = 14; WHR ≥ 0.85). Percent fat was not different between groups based on hydrodensitometry (LB = 24.1 ± 1.8; UB = 23.7 ± 1.6). However, differences in mean percent fat by skinfolds approached significance (p = 0.09) between groups (LB = 24.1 ± 1.8%; UB = 29.1 ± 2.3%). These preliminary results led us to question the appropriateness of using generalized equations as estimates of percent fat in women with primarily UB fat distribution. Therefore, the purpose of this study was to determine whether various anthropometric and bioelectrical impedance regression equations accurately estimate body fat in women with either upper body (UB) or lower body (LB) fat distribution patterns as determined by waist-to-hip ratio.

#### METHODS

**Subjects.** Subjects were selected from over 100 premenopausal female volunteers between the ages of 24 and 53 recruited from the local community. All subjects completed a personal health and medical history questionnaire that served as a screening tool and provided basic descriptive data. All volunteers were screened for menstrual abnormalities, medication use (specifically drugs affecting water retention and metabolism), recent weight loss (>4 kg in previous month), and pregnancy. Once suitable subjects were identified, they were informed of the procedures and risks of the study and signed a written informed consent in accordance with the policies and procedures of the University Human Subjects Institutional Review Board.

Subjects were then classified into one of two groups by WHR: lower body (LB) or upper body (UB) fat distribution. WHR categorization values were set as: LB, WHR ≤ 0.73; and UB, WHR ≥ 0.80. These categories represented groups that were at least one standard deviation apart in WHR values. Once grouped, each subject was individually matched for age and percent fat (based on hydrostatic weighing). A total of 36 women were finally appropriately matched (UB, *N* = 18; LB, *N* = 18) and served as subjects for this study.

**Anthropometric measurements.** Stature was measured to the nearest eighth of an inch using a standard stadiometer. Body mass was assessed to the nearest quarter of a pound with subjects wearing only a swimsuit. Values were converted to metric units to determine BMI. Circumference measurements were assessed at nine sites using a Gulick tape measure. Sites measured included the neck, arm, forearm, wrist, waist, hips, proximal thigh, mid thigh, and calf. All sites except for the hip were measured over bare skin with subjects wearing underwear. Measurements were recorded to the nearest 0.10 cm, and a minimum of two measures was obtained from each site assuring a difference of no more than 0.20 cm between trials. An average of the two closest measurements for each site was calculated. Neck, arm, forearm, wrist, thigh, and calf measurements were determined with the subject standing in accordance with the guidelines of Callaway et al. ^{(3)}. Waist (at the narrowest point between the bottom of the rib cage and the umbilicus) and hip locations (at the greatest extension of the buttocks) were marked on the skin with subjects standing for subsequent supine measurement. Two trained graduate student technicians obtained measurements. Before the start of data collection, an interrater reliability check was conducted, and reliability between the two technicians was established at r = 0.99 for each of the circumference sites.

Skinfolds were assessed at 11 locations (triceps, biceps, chest, horizontal and vertical midaxillary, subscapular, horizontal and vertical abdominal, suprailiac, thigh, and calf) and were measured on the right side of the body using a single calibrated Lange (Cambridge Scientific Industries, Cambridge, MD) caliper. Readings were taken 2-4 s after releasing the caliper. All skinfolds were obtained at standardized sites according to Harrison et al. ^{(9)}. In addition, the midaxillary, suprailiac, and abdomen measures were taken in accordance with the recommendations of Jackson et al. ^{(13)} (i.e., vertical fold for midaxillary and abdominal and at the anterior axillary line for the suprailiac). A minimum of two skinfolds was recorded at each location with a difference of no greater than 2 mm allowed between acceptable measures. An average of the two closest measurements was calculated for each site. Skinfolds were obtained by two experienced, trained graduate student technicians demonstrating an interrater reliability range of r = 0.866 (thigh) to 1.0 (triceps) for each individual skinfold site and an overall intraclass correlation coefficient of 0.99 for the sum of all sites.

The appropriate anthropometric data were entered into five different equations to determine body density using the generalizable equations of Jackson et al. ^{(13)} [JPW] (7 and 3-site), Tran and Weltman ^{(23)} [TW], Durnin and Womersley ^{(5)} [DW] and Vogel et al. ^{(26)} [V] (see Appendix A). Total body fat was calculated from body density using the equation of Siri ^{(19)}.

**Hydrostatic weighing.** The simplified oxygen dilution technique of Wilmore et al. ^{(27)} was used to measure residual lung volume using a Beckman (Palo Alto, CA) O_{2} analyzer (Model OM-11) and a Beckman CO_{2} analyzer (Model LB-2) with flow control (Model R-1). Before testing, O_{2} and CO_{2} analyzers were calibrated to register the correct percentage of calibrated gas (certified to the nearest 0.01%) with flow control adjusted to 500 mL·min^{−1}. Measurements of residual lung volume were repeated three times on land in the seated position assumed during the submersion phase of underwater weighing. A minimum of two values within 100 mL of each other was averaged and recorded as the representative residual volume.

Hydrostatic weighing (UWW) was accomplished in a 640-gallon capacity water tank containing a rectangular chair of PVC pipe suspended from a cable that was attached to a load cell connected to a computer. The weight of the chair was assessed before each underwater weighing test. Each subject was instructed to immerse herself completely under water, expire all air from the lungs, and remain as motionless as possible. Measurements of underwater weighing consisted of 5-10 trials per subject. Within each trial, 2-4 readings were obtained from the computer. Mean underwater weight was calculated from the highest values with the least variability from each trial. This mean weight was used to determine body density according to the equation of Buskirk ^{(2)}. Percent fat was estimated using the formula of Siri ^{(19)}.

**Bioelectrical impedance.** Total body resistance was assessed using the Valhalla Medical Products 1990-B Bioresistance Body Composition Analyzer (San Diego, CA). Subjects came to the laboratory a minimum of 4 h post-prandially. All were informed to avoid strenuous exercise, alcohol, and caffeine for 12 h before the test and to consume a minimum of 64 fluid ounces of water in the 24-h period preceding the test. During testing, subjects were instructed to lie still in a supine position with arms and legs comfortably apart on a padded table with a nonconductive surface. Normal room temperature (34-36°C) was maintained. The skin was cleaned of oil and debris using alcohol pads. Electrodes were placed on the dorsal surface of the right arm and foot. Proximal electrodes were placed to bisect the lateral and medial malleolus of the wrist and ankle. Distal electrodes were placed at the base of the second metacarpal and metatarsal phalangeal joints. A single analysis was run on each subject before skinfold and hydrodensitometry testing. The same measured resistance data were entered into three generalizable equations to estimate total body water and calculate fat free mass and percent body fat based on the equations of Lohman ^{(17)} [L]; Gray et al. ^{(8)} [G], and Van Loan and Mayclin ^{(24)} [VLM] (see Appendix A).

**Statistical analysis.** Statistical analyses were performed using the SPSS v. 6.1 software for Windows. Descriptive statistics were calculated for subject characteristics. Mean differences in %BF between UWW and the five anthropometric and the three BIA equations for each group were computed with a 2 × 5 ANOVA with repeated measures (group by method) for anthropometric data and a 2 × 3 ANOVA with repeated measures for the BIA data. When significant group by method interactions were indicated, a Dunnett's *t post hoc* analysis was used to treat UWW as a control with which all other groups were compared. Linear regression and Pearson product moment correlation defined the relationship between the prediction equations and UWW. An alpha level of 0.05 was used to indicate statistical significance.

#### RESULTS

The power for all significant effects was calculated at or above 0.81 to detect a large effect (ω^{2} = 0.15) between groups. Results are expressed as means ± SD. Subject characteristics are summarized by group in Table 1. Subjects were individually matched for age and percent fat from hydrostatic weighing thus there were no significant differences between groups. The groups showed no statistically significant differences in BMI; however, it should be noted that the upper body fat group had a greater variability in body mass and were heavier and slightly taller than the lower body fat group.

Figure 1 shows that compared to UWW, three of five anthropometric equations significantly overestimated (*P* < 0.05) %BF in UB women. Table 2 displays the mean ± SD for the %BF estimations and the body density data. The analyses indicated no significant differences between any of the anthropometric estimates and UWW in the LB group. There was a consistent underestimation of body density resulting in an overestimation of %BF (range = 3-8%) in the UB group in the JPW 7-site, DW, and TW equations as compared with UWW. The JPW 3-site equation had a higher amount of variability in the measurement and indicated no statistically significant (p = 0.09) difference compared with UWW. Only the Vogel et al. ^{(26)} anthropometric equation predicted %BF values that were similar to UWW in the UB group. The V equation calculates %BF directly and thus does not estimate %BF from the body density.

Figure 1-Comparison ... Image Tools |
Table 2 Image Tools |

Figure 2 indicates percent body fat derived from the three BIA equations (L, G, and VLM) established for Caucasian female subjects compared with UWW. The G equation significantly overestimated %BF in UB as compared with LB by about 4%. The equation of VLM significantly over-estimated %BF for both LB and UB by about 7%, whereas the L equation underestimated %BF prediction as compared with UWW in the LB group. Table 2 displays the mean ± SD for the BIA %BF estimations.

Table 3 records the correlation coefficients for body density and fat free mass (FFM) data from the various equations with that from UWW. In addition, the standard error of the estimate (SEE) is recorded to allow a comparison with the published SEE of the original equation sample population. Figure 3 (a-h) displays the individual regression curves between UWW with the estimated %BF from each equation.

Table 3 Image Tools |
Figure 3-Relationshi... Image Tools |

#### DISCUSSION

The purpose of this study was to assess the accuracy of several generalizable anthropometric and bioelectrical impedance equations to estimate percent body fat in women with different patterns of regional adiposity. Generalizable prediction equations are commonly used by professionals in hospitals, fitness facilities, and wellness/health promotion programs as a means to assess body fat and monitor health risk of clients. To be appropriately used, these prediction equations need to apply to a population that is representative of the sample population upon which they were developed.

Compared with men, women have a greater spectrum of possibilities for body fat distribution patterning ^{(25)}. The diversity of body fat distribution patterning in women may complicate and perhaps limit the generalizability of the prediction of percent body fat in women by anthropometry. In addition, the prevalence of women who are classified as overweight (i.e., BMI ≥ 25.0) has increased approximately 9% since several of the original regression equations were generated ^{(15)}. These changes in population characteristics over the last 20 yr may serve to increase the variability and limit the generalizability of the original body fat prediction equations. Results from this study indicated that three of five anthropometric equations (DW, JPW, and TW but not JPW-3 or V) overestimated %BF in UB women as compared with hydrodensitometry. Only the equation of Vogel et al. ^{(26)} indicated mean percent fat values similar to UWW in the UB group. When the JPW 3-site equation was employed, it demonstrated a greater amount of variability in the measurement than the other techniques. This variability most likely attenuated any statistically significant difference between groups. The sites included in the JPW 3-site equation are the triceps, suprailiac, and thigh. Perhaps the greater variability in the JPW 3-site equation resulted from using a limited number of sites on individuals with distinct or "unusual" body fat distribution patterns. For example, according to Durnin and Wormersley ^{(5)} there are individuals who, "because of unusual fat distribution the likelihood of larger error may be reduced by using multiple skinfolds" (p. 96). Therefore, although the %BF predicted from the JPW 3-site equation was not significantly different from UWW in the UB group, it may be less reliable in this "unusual" group and thus inappropriate to be used for UB-shaped women. Of interest was that no significant differences in %BF were found between any of the anthropometric derived values and UWW in the matched LB group of women. Thus, according to our data, all of the anthropometric equations were able to predict %BF in the LB group appropriately.

The significant and high correlations of each equation with hydrostatic weighing indicated that each equation was able to predict body fat consistently. However, it was clear that for the UB group the values were uniformly higher than UWW for JPW7, TW, and DW. Theoretically, the sample populations selected to develop the original generalizable equations accounted for the heterogeneity of body shape and body fat distribution patterning in women. In each of the original equations, body density values had fairly low standard errors of estimate (range: 0.0079-0.0116 g·cc^{−1}) in generating the equations (Table 3). The current data indicated that for the entire population the SEE of the body density measures were within 0.006 g·cm^{−1} or less of the corresponding original published prediction equations. This indicates that our sample was an appropriate representation of the original sample. However, none of the original reports described WHR as a characteristic variable for the subjects. The prevalence of UB fat distribution is relatively low in healthy normal weight women. The mean WHR of women aged 20-50 is approximately 0.77 ± 0.06 ^{(16,26)}. Assuming WHR is a normally distributed variable within the population, most women (67%) would have a WHR of 0.80 or smaller and only 16% women would have a WHR of 0.83 or greater (i.e., one SD above the mean). Indeed, a WHR of 0.89 in the normal weight (nonobese) population of women would be quite rare (less than 3%) (i.e., two SD above the mean). It is quite likely that the sample populations used to generate the original equations consisted of very few UB-shaped women. It is also likely, then, that these equations are not appropriate for women with a UB fat distribution pattern. The equation that seemed to apply to both UB and LB groups in this study was the Vogel et al. ^{(26)} equation. Perhaps the V equation was not influenced by body fat location because the original sample population consisted of wide ethnic diversity of subjects (e.g., military personnel) representing women with a greater heterogeneity of body fat distribution. In addition, the V equation uses no skinfold measures and consists of circumference measures at sites that are very reliable to measure (i.e., forearm, neck, and wrist), which might account for its increased ability to predict %BF in women with distinctly different regional adiposity patterns.

Body fat distribution is known to have a strong genetic component and will vary across ethnic groups ^{(1)}. The population of women in this study were predominately Anglo-American/Caucasian. However, because it is known that Native Americans primarily carry fat in the UB region ^{(22)} (mean WHR = 0.81 ± 0.07), we also chose to examine the accuracy of another anthropometric equation specifically developed for Native American women ^{(11)} to predict %BF in our subject population (see Appendix A). Interestingly, the Hicks et al. ^{(11)} skinfold equation also significantly overestimated %BF in the UB group compared with UWW. The results indicated an overestimation of %BF for UB of about 8% (i.e., 38.06 ± 11.9% compared with 30.83 ± 8.19% for UWW). Similar to the previous results, the Hicks et al. ^{(11)} equation was not significantly different in %BF compared with UWW in the LB group (30.06 ± 3.5 vs. 29.7 ± 8.0). Thus, although southwestern Native Americans who formed the population for the Hicks et al. ^{(11)} study and the UB group of women in this study may have had similar regional fat distribution patterns, this equation was not found to be effective for predicting %BF in UB women.

Regional fat patterning had little effect on %BF estimated by BIA equations. We hypothesized that the location of the body fat would influence the electrical conductivity differentially in UB compared with LB. This then would produce differences in %BF between groups. The results indicated no obvious trend for regional adiposity to alter the prediction of percent body fat using BIA. In general, the VLM equation overestimated %BF in both UB and LB compared with UWW. However, the L ^{(15)} equation tended to underestimate %BF in the LB group whereas the G equation ^{(8)} significantly overestimated %BF in the UB group similar to that previously reported for the anthropometric data. It is unclear what significance these differences between equations have in identifying regional adiposity using BIA. It may be that the primary assumption of the BIA technique (i.e., the body is one cylinder) is not appropriate for women with disparate body shapes. Continued research on the use of the BIA technique (perhaps with a segmental approach) with women with different body fat distribution patterning is clearly warranted.

The consistent overestimation of %BF in the UB group by the skinfold methods is an indication that subcutaneous skinfold thicknesses at various sites on women are influenced by regional adiposity determined from WHR. Indeed the data from the V equation (that uses no skinfolds) and the BIA results seem to substantiate that regional adiposity has little influence on estimates that are not derived from skinfold measures. Thus, the differences between groups must reside in the "pinching" of the skinfolds. The technique of locating and pinching the skinfolds was well practiced and standardized by our technicians. It was unlikely that a systematic bias or error would occur to consistently overestimate the skinfold values. Consequently, we believe that the "standardized" skinfold sites used in the DW, TW, and JPW equations are not appropriate for women with upper body fat distribution.

Because excess UB fat is associated with a greater likelihood of developing metabolic disorders and disease, those with UB fat are the clinically important group when evaluating health status. Thus, any practical body fat prediction method must be able to accurately distinguish this group. Results from this study indicate that anthropometric and some BIA equations are accurate for predicting %BF in LB fat "shaped" women but are not appropriate for women with primarily abdominal fat patterning. We recommend that until an appropriate equation is developed, care be taken in interpreting body fat data in upper body fat women (WHR >0.80) if anthropometric methods are used to assess percent body fat.