By continuing to browse this site you agree to us using cookies as described in About Cookies
Notice: Wiley Online Library will be unavailable on Saturday 7th Oct from 03.00 EDT / 08:00 BST / 12:30 IST / 15.00 SGT to 08.00 EDT / 13.00 BST / 17:30 IST / 20.00 SGT and Sunday 8th Oct from 03.00 EDT / 08:00 BST / 12:30 IST / 15.00 SGT to 06.00 EDT / 11.00 BST / 15:30 IST / 18.00 SGT for essential maintenance. Apologies for the inconvenience.
Department of Physiology, Ina Gittings Building #93, University of Arizona, Tucson, AZ 85721. E-mail: email@example.com
Objective: To examine the relationship between percentage of total body fat (%Fat) and body mass index (BMI) in early postmenopausal women and to evaluate the validity of the BMI standards for obesity established by the NIH.
Research Methods and Procedures: Three hundred seventeen healthy, sedentary, postmenopausal women (ages, 40 to 66 years; BMI, 18 to 35 kg/m2; 3 to 10 years postmenopausal) participated in the study. Height, weight, BMI, and %Fat, as assessed by DXA, were measured. Receiver operating characteristic analysis was performed to evaluate the ability of BMI to discriminate obesity from non-obesity using 38%Fat as the criterion value.
Results: A moderately high relationship was observed between BMI and %Fat (r = 0.81; y = 1.41x + 2.65) with a SE of estimate of 3.9%. Eighty-one percent of other studies examined fell within 1 SE of estimate as derived from our study. Receiver operating characteristic analysis showed that BMI is a good diagnostic test for obesity. The cutoff for BMI corresponding to the criterion value of 38%Fat that maximized the sum of the sensitivity and specificity was 24.9 kg/m2. The true-positive (sensitivity) and false-positive (1 − specificity) rates were 84.4% and 14.6%, respectively. The area under the curve estimate for BMI was 0.914.
Discussion: There is a strong association between %Fat and BMI in postmenopausal women. Current NIH BMI-based classifications for obesity may be misleading based on currently proposed %Fat standards. BMI >25 kg/m2 rather than BMI >30 kg/m2 may be superior for diagnosing obesity in postmenopausal women.
The NIH have established cut-off points using body mass index (BMI) for overweight and obesity at 25 kg/m2 and 30 kg/m2, respectively (1). According to preliminary data from the 1999 National Health and Nutrition Examination Survey (NHANES 1999), the prevalence of overweight (BMI 25.0 to 29.9 kg/m2) in adults ages 20 to 74 years increased by ∼2% between the NHANES II and NHANES 1999 surveys, whereas obesity (BMI ≥ 30.0 kg/m2) nearly doubled from ∼15% to 27% during the same period (2). In women ages 50 to 59 (typically the early menopausal years), the percentage considered obese (BMI ≥ 30.0 kg/m2) has increased by 47.3% between 1991 and 1998 (3). Overweight and obesity are known risk factors for insulin resistance, glucose intolerance, diabetes mellitus, hypertension, dyslipidemia, sleep apnea, arthritis, hyperuricemia, gall bladder disease, and certain types of cancer (4). This increase is an alarming statistic and emphasizes the importance of accurately assessing body composition, because it is expected that a significant extent of the risk of excess weight is caused by the contribution of excess fat to body mass. Therefore, it is important to determine the relationship between BMI and percentage of total body fat (%Fat).
DXA and underwater weighing (UWW) are accurate methods to estimate body fat. However, these methods are costly, difficult to access, and often impractical. Other, less expensive, but more accessible methods, such as bioelectrical impedance analysis (BIA) and skinfold thickness (SKF) have their own limitations. SKF measurements are dependent on the skill of the examiner and may vary widely when measured by different examiners (5,6). BIA may not be accurate in severely obese individuals and is not useful for tracking short-term changes in body fat brought about by diet or exercise (7,8). A weight-for-height measure such as BMI is a simple inexpensive method of determining overall fatness, although it is not a direct measure of body fat. Consequently, the use of BMI to predict body fatness would offer significant advantages, and in fact, several studies suggest that BMI is an acceptable predictor of body fatness (9,10,11,12,13). However, the use of BMI to predict %Fat and diagnose obesity has also been shown to have several limitations. Previous studies have shown that ethnicity, age, and sex may significantly influence the relation between %Fat and BMI (14,15,16,17). This suggests that %Fat and BMI may be more closely related within specific populations, and hence clearer definitions of obesity should be established for those groups. A recent report from a Centers for Disease Control and Prevention workshop recognized the potential differences between populations but reiterated the importance of BMI in population-based health promotion (18). It proposed that BMI can be an effective method on three levels: self-monitoring by the individual, identifying subpopulations where chronic disease is prevalent, and allowing population trends to be tracked and characterized. However, BMI's effectiveness is dependent on cut-off points that apply to all populations, something that currently recommended cut-off points may not do (18).
Indeed, there is evidence to suggest that in certain subgroups of a population, BMI may be more predictive of body fatness than in other subgroups. Wellens et al. showed a significantly higher correlation between BMI and %Fat in the upper BMI tertile than in lower tertiles in white women ages 20 to 45 years (19). Receiver operating characteristic (ROC) analysis suggested a BMI of 23 kg/m2 in white women might provide better diagnostic screening cut-offs for obesity (19). Whether the cut-off applies to postmenopausal women is not known. More recently, a large-scale study of elderly Italian women showed that BMI is an acceptable surrogate measure of body fatness for population studies, but it was deemed inadequate to replace a direct measure of body fatness at the individual level (20). Because postmenopausal women are likely to have increased fat mass and decreased fat-free mass (21), the relationship between %Fat and BMI may be different for this population. The change in this relationship might suggest that lower levels of BMI may be superior for predicting obesity. Therefore, the aim of this paper is to address the relationship between %Fat and BMI in a population of early postmenopausal women and moreover, consider the extent to which this relationship agrees with data from other studies. Also, given the aforementioned changes in fat and fat-free mass in postmenopausal women, the study uses ROC analysis to determine if there are more appropriate values for diagnosing obesity than those currently recommended by the NIH.
Research Methods and Procedures
Three hundred seventeen healthy, sedentary, postmenopausal women participating in the Bone Estrogen Strength Training study were included in the present study (22). The University of Arizona's Human Subjects Institutional Review Board approved the study and all participants gave written informed consent before participation in the study. The women were 3 to 10 years postmenopausal (40 to 66 years old). Menopause was defined as 12 consecutive months without a menstrual period. The sample included surgical as well as nonsurgical menopause. At entry into the study, they had a self-reported BMI >5th percentile and <95th percentile, were nonsmokers, and were either using hormone replacement therapy (HRT) for 1 to 6 years or had not used HRT in the preceding year. Subjects did not take any other medications known to affect bone metabolism and body composition and were not exercising or dieting before entering the study.
Participants’ %Fat was measured by DXA using a total-body scanner (model DPX-L; Lunar Radiation Corp., Madison, WI). DXA uses a constant potential X-ray source of 78 kVp and a rare-earth k-edge filter to achieve a congruent beam of stable dual-energy radiation with effective energies of 40 and 70 KeV. The scanner was calibrated daily against a standard calibration block supplied by the manufacturer. Subject position for the total body was standardized and identical to those described by Mazess et al. (23). All total body scans were completed in medium scan mode to ensure appropriate image resolution. All subjects were scanned twice within ∼2 weeks to improve accuracy by reducing potential day-to-day variability. The mean of the two scans was used in all statistical analysis. Initial scan analysis, including the placement of baselines distinguishing bone and soft tissue, edge detection, and regional demarcations, was performed by computer algorithms (version 1.3y; Lunar Corp.). Subsequently, the technician visually inspected all scans and adjustments were made as necessary. The %Fat coefficient of variation for repeated measurements was 2.8%.
Standing height (HT) and weight (WT) measurements were completed with subjects wearing light-weight clothing and no shoes. HT was measured to the nearest 0.1 cm using a Schorr measuring board (Schorr Products, Olney, MD). Subjects were asked to stand with their arms hanging freely at their sides, ankles or knees touching, and their weight distributed evenly on both feet. The mid-axillary plane was aligned perpendicular to the floor with the scapulae and buttocks in contact with the vertical board if possible, or whichever part touched the board first. The head was positioned in the Frankfurt Horizontal plane, and HT was measured during a maximal inhalation. WT was measured on a calibrated digital scale (model 770; SECA, Hamburg, Germany), accurate to 0.1 kg, while subjects stood with their weight evenly distributed and their arms hanging freely at their sides. The average of two measurements for both HT and WT were used as the criterion measurements. BMI was calculated as WT (kilograms) divided by HT squared (meters squared).
Statistical analyses were conducted using the SPSS/PC (SPSS for Windows, V.9.0; SPSS Inc., Chicago, IL). Means, SDs, and Pearson's product moment correlations were calculated. Independent sample Student's t tests were used to compare age, years past menopause, height, weight, BMI, and %Fat between women who used HRT and women who did not use HRT. Linear regression analyses were performed to evaluate the relationship between %Fat by DXA and BMI. The slopes of the regression lines for the two groups were not different (p > 0.05). Therefore, data from the two groups were combined for further analysis. Significance was inferred at p < 0.05. All results were expressed as means and SDs.
ROC analysis is a way of evaluating the accuracy of a diagnostic test by summarizing the potential of the test to discriminate between the absence and presence of a health condition (24,25). In this study, ROC analysis examined the ability of BMI to discriminate obesity from non-obesity using ≥38%Fat as the criterion value for obesity, based on the recent recommendation for older women by Lohman et al. (26,27). The slightly higher %Fat value of 38% (30% to 35% is usually recommended for obesity in women) allows for a small increase in fatness with age without a concomitant increase in risk factors (26,27). By this delineation, obese women (≥38% fat) who are correctly classified as obese by BMI represent true-positive cases, whereas obese subjects classified as non-obese represent false-negative cases. Non-obese women correctly classified as non-obese represent true-negative cases, whereas non-obese subjects classified as obese represent false-positive cases. These assessments of BMI's screening performance were repeated using slightly lower and higher %Fat cut-off values (35% and 40%) to address alternative definitions of obesity.
The sensitivity of BMI refers to the probability that BMI will classify a subject as obese when the subject is truly obese. The specificity is the probability that BMI will classify a subject as non-obese when the subject is truly non-obese (true-negative). The true-positive rate (sensitivity) was plotted against the false-positive rate (1 − specificity) across a range of BMI values. This relationship gives an estimate of BMI that yields the minimal number of false-negative and false-positive cases (e.g., the BMI best able to discriminate between obese and non-obese subjects). This estimate is selected by determining the BMI value that maximizes the sum of sensitivity and specificity. The sum of sensitivity and specificity gives the overall performance of the cut-off point, with higher sums indicating superior performance (28).
Another index provided by ROC analysis is the area under the curve (AUC), which reflects the overall accuracy of BMI in predicting obesity. The closer the AUC is to 1.0, the greater the chance that any given obese subject will have a higher BMI than that of any given non-obese subject. Conversely, if BMI is unable to distinguish between obese and non-obese (i.e., where there is no difference between the two distributions), the AUC will be close to 0.5, and the ROC curve will approach a diagonal line.
A literature review was performed based on a comprehensive MEDLINE search and a search for articles cross-referenced in related published reports. Only studies that presented both BMI and %Fat by DXA, hydrodensitometry, or a four-component model (4CM) in healthy postmenopausal-aged women (range of mean ages, 42 to 75 years) were included. In cases where only mean values of HT and WT were presented, a BMI score was calculated using these values. Twenty-six studies were then examined in two ways to determine the extent to which our sample could be generalized. First, the mean values of BMI and %Fat were plotted against our regression line to determine whether their results supported the relationship between BMI and %Fat that we found. This was accomplished by determining whether their mean values fell within 1 SE of estimate (SEE) from the regression line derived in our study. Second, the mean values of BMI and %Fat for each of the studies were plotted, and a regression line was created. This regression line was then plotted against the regression line of the present study. In the event that a single study examined multiple populations (i.e., different age, ethnicity, or gender), only those fitting our criteria were used. In cases where multiple populations within a given study fit our criteria, each population was treated as an independent sample for comparison. Thus, a single study may have been used more than once in comparing it with our study. In cases where both UWW and DXA %Fat were presented, only the DXA measure was included in our analysis.
Table 1 presents the means, SDs, and ranges of physical characteristics by HRT status and all subjects combined. Our sample was 87.7% white (n = 278), 9.5% Hispanic (n = 30), 1.6% Asian American (n = 5), 0.9% African American (n = 3), and 0.3% Native American (n = 1). Forty-nine percent of the subjects were using some form of HRT. Age and years past menopause were significantly different (p < 0.05) between women using HRT and women not using HRT, but no significant differences were observed when comparing height, weight, BMI, and %Fat. Consequently, the results that follow are based on all subjects combined and not by HRT status.
Table 1. Descriptive statistics of postmenopausal women by HRT status and independent of HRT status
HRT (n = 155)
No HRT (n = 162)
HRT + No HRT (n = 317)
Mean ± SD
Mean ± SD
Mean ± SD
p < 0.05.
p < 0.01.
HRT, hormone replacement therapy; BMI, body mass index; FFM, fat-free mass.
Based on BMI classifications suggested by the NIH (1), 1.6% were underweight, 45.1% were within the healthy weight range, 38.5% were overweight, and 14.8% were obese. Linear regression analysis showed that %Fat and BMI were highly correlated (r = 0.81) with a SEE of 3.9%. Figure 1 illustrates the relationship and shows the resulting regression line [%Fat = 1.41(BMI) + 2.65]. The association between BMI and %Fat was not altered by adjusting for years past menopause or years on HRT. Using this equation to estimate the %Fat values associated with the NIH BMI classifications, the following was found: BMI of 19 kg/m2 equals 29.4%Fat, BMI of 25 kg/m2 equals 37.9%Fat, BMI of 30 kg/m2 equals 45.0%Fat, and BMI of 35 kg/m2 equals 52.0%Fat.
Gallagher et al. (29) have proposed the use of the inverse of BMI to predict body fatness because a curvilinear relationship may exist between BMI and %Fat. Linear regression using the inverse of BMI to predict %Fat in our sample resulted in r = 0.83, SEE = 3.7%, and %Fat = −951.7(BMI) + 76.8. Using this equation to predict %Fat values associated with the BMI classifications, the following was found: BMI of 19 kg/m2 equals 26.7%Fat, BMI of 25 kg/m2 equals 38.7%Fat, BMI of 30 kg/m2 equals 45.1%Fat, and BMI of 35 kg/m2 equals 49.6%Fat.
Twenty-six studies were examined that resulted in 36 samples. Their BMI and %Fat values were plotted against our regression line (Table 2; Figure 2). Eighty-one percent of the studies selected from the literature review fell within 1 SEE from the regression equation in our study.
Table 2. Mean body mass index and mean total body fat percentage in postmenopausal age populations
Not specified in this paper.
Calculated from height and weight group means or using a prediction equation.
Calculated from individual subject data presented in paper.
A comparison of the BMI cut-off point for classifying obesity as suggested by the NIH and that proposed by this study is presented in Table 3. ROC analysis was used to examine the ability of BMI to discriminate obesity from non-obesity using ≥38%Fat as the criterion value for obesity. The ROC curve depicted in Figure 3 shows the cut-off point for BMI that maximized the sum of the sensitivity and specificity for detecting obesity defined as 38%Fat was 24.9 kg/m2. The true-positive (sensitivity) and false-positive (1 − specificity) rates corresponding to the cut-off value were 84.4% and 14.6%, respectively, which means BMI will classify a subject as obese 84.4% of the time when the subject is truly obese but will misclassify them as obese 14.6% of the time. The AUC estimate was 0.914, meaning that a randomly selected obese subject will have a greater BMI than a randomly selected non-obese subject 91% of the time. The 95% confidence interval was 0.877 to 0.942. The lower limit of 0.877 is high, suggesting good diagnostic performance by BMI. Using the NIH cutoff for obesity of 30 kg/m2, the sensitivity was 25.6 and the specificity was 99.3.
Table 3. Performance comparison of BMI-based cutoffs for diagnosing obesity using 35%, 38%, and 40% body fat, NIH (1) vs. present study
BMI-based classification of overweight and obesity has been widely accepted by the research community but has not yet been broadly adopted by primary care practitioners (18). The weak relation of BMI to important health outcomes in a varied range of populations and the inability of BMI to accurately assess body composition for the individual are important limitations of the index that may explain the dissatisfaction among some clinicians. In fact, the use of BMI to determine obesity or as a surrogate for body composition has been shown to have several limitations (14,15,16,17,20,53). These studies demonstrate that adjusting a subject's body weight for height will not necessarily yield consistent levels of body fatness. Therefore, two individuals with identical BMI values may have considerably different %Fat levels, particularly if they vary in ethnicity, age, or sex. Although data on the effect of ethnicity have been mixed, Deurenberg et al. concluded in a meta-analysis that the relationship between %Fat and BMI does vary among ethnic groups (14). However, this study did not include Hispanics, and no studies to date have found that the relationship between %Fat and BMI in Hispanics is different from whites. It has also been shown that variations in age and sex have a significant impact on the relationship between BMI and adiposity (9,15). When comparing young and old subjects and men and women, older subjects and women will have a higher %Fat (for a given weight and height) despite similar BMIs (15). Consequently, it is likely that there are populations in which the BMI classifications proposed by NIH will be misleading.
This study provides a more homogeneous population, thus eliminating potential inequities associated with age, sex, or ethnicity. However, this is also a limitation of the study. The characteristics of this sample may not be fully representative of all postmenopausal women, particularly older postmenopausal women. The study is further limited because all subjects were paid volunteers and may not characterize a random sampling of all postmenopausal women. Nevertheless, the current sample size of 317 subjects is large enough to define within narrow limits the magnitude of the correlation between BMI and %Fat (r = 0.81). Furthermore, we were encouraged to find that other studies examining this population (Table 2) agree with our findings and for the most part (81%) fall within 1 SEE from our regression equation.
There are several reasons why some studies did not fit within 1 SEE from the regression equation in our study. First, there were studies that did not specify ethnicity (36,46), which could affect the relationship between %Fat and BMI (14). Second, several of the studies had a much smaller sample size than ours, reducing the precision of their measures (36,40,46,47). These samples were also older than the current study sample (36,40,46,47), and although we found that years past menopause did not alter the relationship between %Fat and BMI, samples with substantially older subjects may be affected. Finally, not all of these studies used DXA to determine %Fat in this population (36,40,41). Studies using two component models such as UWW to predict %Fat could potentially be inaccurate because of variations in the densities of the body constituents. Postmenopausal women are a population that have experienced changes in the density of their fat-free mass, in particular, reductions in bone mineral density (21,54,55). The use of DXA, as in this study, allows a three-component measure, which, in a population where bone is known to undergo dramatic changes, may remove inaccuracies caused by potential component variations. However, because the evidence suggests that the variability between the methods (UWW, DXA, and 4CM method) only ranges from 1% to 3% (35), all were considered for a more inclusive comparison.
Previous studies have created prediction formulas to estimate %Fat based on BMI (9,29,36). Formulas derived from age- and sex-specific regression equations have been shown to have an estimation error of about 4% in elderly men and women (9,36). The equation developed by Deurenberg et al. incorporates the confounding variables of age and sex, but it used densitometry and only included ∼60 women in our age range (9). Gallagher et al. (29) also developed an equation for predicting %Fat (4CM) from the inverse of BMI that incorporated both age and sex. The sample used to derive the prediction equation consisted of 671 white and African-American subjects (380 women and 225 men) and varied widely in age, but still yielded a high correlation (r = 0.90) and a SEE of 4.3%. The current study sample was more ethnically and age homogeneous, included only postmenopausal women (n = 317), and it resulted in a lower SEE (3.9%) for the prediction of %Fat from BMI. Accordingly, the results of our study are more directly applicable to the postmenopausal population.
In this population, it is apparent that BMI is as effective for predicting %Fat as BIA or SKF. The SEE (3.9%) was similar to those found in previous studies (9) and comparable with using SKF or BIA with potentially less measurer error and cost than either of these methods. This does not necessarily mean that BMI is an effective screening tool for diagnosing obesity. However, the AUC determined by the ROC analysis (0.91) does support BMI's effectiveness at diagnosing obesity. The AUC in this sample suggests that any randomly selected obese subject will have a higher BMI than a randomly selected non-obese subject 91% of the time. The 95% confidence interval (0.88 to 0.94) demonstrates that at even the lower limit (0.88), the diagnostic ability of BMI is good.
With the regression equation from this study, BMI is disparate from %Fat for the overweight and obese classifications. According to the current classifications from NIH (1), a BMI of 25 to 30 kg/m2 indicates overweight and 30 to 35 kg/m2 indicates class I obesity. Our results indicate that the corresponding ranges for %Fat are 38% to 45% and 45% to 52% for overweight and class I obesity, respectively. The relationship for overweight and obesity is linear across the majority of %Fat and BMI values, but the relationship is not linear at lower BMIs. The corresponding %Fat for a BMI of 19 kg/m2 is 29.4% (or 26.7% using the inverse of BMI that relates more linearly to %Fat), an amount of fatness clearly not associated with underweight health-related risks. However, this is likely explained by the fact that our sample only had five subjects (1.6%) <19 kg/m2.
ROC analysis was performed to obtain a more analytical measure of the effectiveness of BMI for diagnosing obesity in postmenopausal women. In younger women (20 to 45 years), Wellens et al. (19) showed that BMI provided a sensitive measure for estimating adiposity based on a 33%Fat criterion measure but did not examine a postmenopausal population. In the absence of a clearly defined relationship between %Fat and morbidity and mortality in postmenopausal women, our cut-off of 38%Fat for obesity was based on the recommendation of Lohman et al. (26,27). With this criterion value, 24.9 kg/m2 was the most sensitive and most specific BMI for the present population of postmenopausal women, meaning that it is the optimal value to discriminate between obesity and non-obesity. Other commonly used %Fat cut-off points for obesity (35% and 40%) yielded a range of 24.0 to 25.3 kg/m2 when used as criterion values (Table 3). These values are consistent with the 25.5 kg/m2 determined by Sardinha and Teixeira (44) in a population of women over 50 years, employing similar methodology. It also closely corresponds to the cut-off point determined by Calle et al. (56) in a large-scale prospective study showing that women with a BMI higher than 25 kg/m2 had a significantly increased risk of death from cardiovascular disease.
According to our data, the NIH classification for obesity of 30 kg/m2 gives a false-positive rate of 0.7 (sensitivity = 99.3) and a true-positive rate (specificity) of 25.6 (Table 3). This indicates that although a BMI of 30 kg/m2 will likely not classify a non-obese postmenopausal woman as obese, it may inaccurately classify many obese women as non-obese. This is a potentially harmful misclassification because the prevalence of adult obesity has increased (2), and it is an important correlate of chronic disease (4). Given the greater health risks associated with obesity, it is perhaps more important to classify everyone who is truly obese as obese even at the risk of increasing the false-positive rate. Therefore, 25 kg/m2 may be a better obesity criterion value than 30 kg/m2 because it maximizes the sum of sensitivity and specificity, which seems more appropriate in the case of predicting obesity as an excess of body fatness and not merely weight.
Although we support the use of BMI criterion values for population studies, it is important to remember that the relationship between BMI and %Fat is not precise and creating absolute cut-off measures may be unreliable for any given individual (20). Nevertheless, BMI-based classifications are still important and necessary in the primary care and public health settings (18). BMI criterion values such as those proposed here are limited by the lack of clearly defined standards that establish a relationship between %Fat and risk factors associated with morbidity and mortality in postmenopausal women. Even so, our selection of ≥38%Fat is likely to be consistent with an increased health risk for postmenopausal women, and therefore, the corresponding BMI of >24.9 kg/m2 should be considered when classifying obesity in postmenopausal women. Given that the moderately high relationship between %Fat and BMI is representative of other studies of postmenopausal women and that there is a dramatic difference between our cut-off point (24.9 kg/m2) and that proposed by the NIH for the population as a whole (30.0 kg/m2), our results clearly support the need for population-specific BMI classifications. In conclusion, in the context of recommended health-related %Fat cut-offs, the results of our study suggest that a BMI of 30 kg/m2 may be too conservative. The currently accepted NIH cut-off for the category of overweight (25 kg/m2) could be a better criterion to diagnose obesity in postmenopausal women.
This study was supported by National Institutes of Health grant AR39559.