Contents - Previous - Next

Body composition of Third World populations and BMI

Measured body composition

Information is quite limited on body composition in groups of those of non-European origin measured by the more direct methods, such as densitometry and hydrometry. A search of the literature identified 12 papers with 29 samples of 945 men and 7 papers with 12 samples of 394 women. One sample of women was rejected because of very high lung residual volumes in the determination of body density. Two-thirds of the measurements were based on densitometry and the remainder on hydrometry. For men, the correlation between % fat and BMI was r = 0.72 (P < 0.01) but this was due largely to the effect of two outlying groups of urban Mexicans (Valencia et al., 1992). Omitting these reduced the correlation to 0.26 (n.s.). In women, the correlation was r= 0.53 (P < 0.05). There is clearly insufficient information to comment on population differences in body composition in relation to BMI from these data.

Estimated data

In the absence of much direct body composition data, it was decided to investigate the relationship between estimated body composition and BMI using group mean data.

The database arid analyses

A literature search was made for samples of 10 or more adults in rural areas of Third World countries with data on height, weight, sex, and the skinfold thicknesses used in the equations of Durnin & Womersley (1974), namely biceps, triceps, subscapular and supra iliac. The groups sought were non-European, non-urban, except that some student and service personnel were included. Ages were 20 years and over, although samples with age ranges 18-30 years, for example, were included. The women were non-pregnant as far as could be ascertained. The mean data of 173 samples of men in 70 papers, representing the results of over 13 600 individuals and 112 samples of women of over 9100 individuals, were entered into the database. A full list of references is available from the author.

Body density was estimated from the equations of Durnin & Womersley (1974) and converted to % body weight as fat using the equation of Siri (1961). Estimates were made using 4, 3 or 2 skinfolds depending on the information available. However, estimates were not made based on the biceps plus one other. Unweighted % fat, FM (kg), and FFM (kg) were regressed on BMI by standard techniques.

To investigate population differences in the relationship of BMI and body composition the 173 samples of men and 112 samples of women were categorized as (i) 'African' - from south of the Sahara, (ii) 'Indo-Mediterranean' Indian sub-continent, Middle East and the Lapps, but not other Europeans, (iii) 'Asian origin' - mainly North, Central and South

Americans but with 9 samples of Chinese and Japanese, and (iv) 'Pacific' peoples, excluding Hawaiians who were regarded as too 'urbanized'. % fat, FM and FFM were regressed on BMI and the separate regression lines tested for significant differences in regression coefficients and intercepts by analysis of covariance (Snedecor & Cochrane, 1967, pp .432-436). Homogeneity of residual variance was assessed by Bartlett's test (Snedecor & Cochrane, 1967, pp. 296-298).

Several objections can be raised to these procedures. The estimation equations for estimating body fat could be population specific, perhaps because of ethnic differences in external fat patterning. This is almost certainly true but the bias in estimating the group mean may be less than that for individual values. Ethnic differences in fat distribution may contribute to a further problem. % fat estimated by using 2,3 or 4 skinfolds in groups of say Africans can be quite different. Perhaps the greatest error is introduced by age estimations. One-third of the samples of men had no description of age and others had wide age ranges, e.g. 20-50 years. For these, the equation for 17-72-year-olds was used. If these individuals were 40-49 years of age, the mean % fat would be underestimated by 3-4%. Further, using mean height and mean weight to calculate the sample BMI may be biased if the distribution of weight or height is skewed and a meta-analysis such as this on unweighted data is of a weak design. However, it represents a more rigorous approach to the description of population differences in body composition in relation to BMI than a casual narrative account of the extensive indirect data on this topic.

The results

The scatter plots of % fat, FM, FFM and body weight on BMI are shown in Figs 4 and 5. The relationship between % fat and BMI appears linear in men and in women. At any BMI, women have a higher % fat than men. In contrast, for FM and BMI, the relationship appears discontinuous with breaks and a steepening of the slope around FM 10 kg in men and 15 kg in women and BMI 25 kg/m². These may represent a shift from normal variation to conditions of overweight and obesity. From the Forbes equation the relationship would be expected to be smooth and continuous. Conversely, FFM and BMI have a heteroscedastic distribution (i.e. a different spread) and the relationship appears to have a curvilinear element. Attempts to linearize the relationships by single or double log transformation, square rooting, introducing a quadratic term or treating it as a rectangular hyperbola increased the proportion of the explained variance significantly in some cases but never by >1%.

In men, the 10 samples with the lowest % fat were mainly Africans, Masai (Day, Bailey & Robinson, 1979), Hazda (Hiernaux and Boedhi Hartono, 1980) and Kalahari Bushmen (Elsner, 1963). The two highest were Samoans (Bindon & Baker, 1985). The small number of groups above 20% fat were mainly older groups. In women, the highest groups were mainly Samoans (Bindon & Baker, 1985), Tokelau (Prior et al., 1977), a middle-aged Indian group (Sanjeev et al., 1991) with 37% fat at a BMI of 24 and two groups of Israeli Kurds and Yemenites who fell some way to the left of the distribution (Davies et al., 1973). Their high FM resulted in their having some of the lowest FFM.

The correlation coefficients between BMI and anthropometric and body composition variables are shown in Fig. 6. BMI was highly correlated with weight and poorly correlated with height in men and women. The correlations of % fat and BMI are similar to those found for groups of individuals in the West. BMI was more highly correlated with FM than with % fat and was well correlated with FFM too. This confirmed that BMI is more a measure of size (kg) than of composition (%) and as much a measure of lean as of fat (Norgan & Ferro-Luzzi, 1982). Correlations are also given for the 123 and 77 samples where mean triceps and subscapular skinfolds were available separately. Here, the sum or log sum of the skinfolds correlated almost as well as % fat, suggesting that the transformation of skinfolds to composition is unnecessary. The subgroup behaved similarly to the total group in terms of the correlation with other variables.

It is worth comparing body weight and BMI as estimators of body composition. The correlations of weight with % fat and FM are as high as those for BMI and body composition. In the case of FFM, correlations with weight are higher, r = 0.83 for men and 0.85 for women, than with BMI, r = 0.64 and 0.76, respectively.

Effects of age

The correlation coefficients in groups of subjects 19-29 years were much lower than those of the total group or of 30-45-year-olds (Fig. 7). In the case of BMI and height, the as 20-24-year-olds behaved in the same way as 25-29-year-olds.

Considering samples where age was well described (a mean given in the report and/or a range <10 years), correlations between % fat or FM and BMI were noticeably lower than those described above. For 106 samples of men, the r for BMI and % fat was 0.42 (vs 0.68) and for FM was 0.56 (vs 0.78). For BMI and height, r was -0.24 (vs -0.07), although when age was added to the regression the coefficient was not significant (t = 1.61, n.s.). For 61 samples of women, the BMI relationships were % fat r = 0.67 (vs 0.82), FM = 0.83 (vs 0.92) and FFM = 0.60 (vs 0.76).

Sex differences

The regression coefficients of % fat with BMI in men and women, i.e. 1.45 and 1.58 respectively, were not significantly different (F = 0.65, d.f. = 1,283, n.s.) but the intercept terms were (F = 567, d.f. = 1,284, P < 0.001). For FM and FFM, both regression coefficients and intercepts were significantly different between the sexes but only just so in the case of FFM on BMI (F = 5.5, P < 0.05). The regression coefficients and intercepts and the analysis of covariance statistics are shown in Table 1.

Ethnic differences

The % fat and BMI relationships of the four groups of samples are shown in Figs 8 and 9. The linear relationship between % fat and BMI persists although in the samples of 'Asian origin' the correlation coefficients are only 0.45 (P < 0.05 for women, <0.01 for men). The correlation coefficients between BMI and body composition are shown in Fig. 10. This also shows that in each group BMI is more highly correlated with FM than with % fat, and remains highly correlated with FFM. What is again noticeable is that BMI is correlated with height in several groups, particularly in women and 'Pacific' men.

Table 1. The statistics of the regressions of body composition variables on body mass index in men and women of different ethnic groups

	Men				Women
	Slope	Intercept	R2	SEE	Slope	Intercept	R2	SEE
% Fat
All samples	1.45	-17.8	0.46	4.1	1.58	-8.6	0.68	4.2
Between sexes F =	0.7	567.4***
African	1.18	-12.4	0.23	3.0	1.32	4.5	0.41	3.0
Asian	0.71	-3.7	0.21	2.8	0.90	5.4	0.19	4.1
Indo-Med.	2.12	-27.8	0.52	4.0	2.46	-21.8	0.67	4.3
Pacific	2.14	-34.2	0.73	4.0	1.71	-12.3	0.89	3.2
Between groups F =	9 3 ***	77.9***			3 70	41.8***
Fat mass
All samples	1.44	-22.8	0.76	3.1	1.76	-24.4	0.92	2.8
Between sexes F =	7.7***	217.2***
African	0.99	-13.5	0.39	1.7	1.11	-11.3	0.58	1.7
Asian	0.70	-8.4	0.38	1.8	1.04	-10.0	0.39	2.9
Indo-Med.	1.81	-27.9	0.67	2.5	2.08	-28.6	0.86	2.1
Pacific	2.16	-39.8	0.87	2.7	1.96	-29.0	0.95	2.2
Between groups F =	20.6***	78.1***			8.6***	***
Fat-free mass
All samples	1.22	24.3	0.64	3.9	0.90	18.0	0.76	3.0
Between sexes F =	5.5*	919.9***
African	1.47	21.3	0.17	4.6	0.71	24.1	0.12	3.5
Asian	1.76	12.5	0.53	3.3	0.90	18.1	0.4	2.4
Indo-Med.	1.13	24.5	0.36	3.6	0.58	22.0	0.27	2.4
Pacific	1.40	19.1	0.70	3.0	1.01	15.3	0.89	1.8
Between groups F =	0.7	31.3***			1.20	41.8***

*P < 0.05; ***P < 0.001.

The regression coefficients, intercept terms, R2 and the standard error of the estimate (SEE) of the regression equations of body composition on BMI are given in Table 1. A significant difference in intercepts was found in all analyses performed. However, the regression coefficients for % fat in the four groups of women (F =3.7, n.s.) and for FFM in men (F=0.7, n.s.) or women (F = 1.2, n.s.) were not significantly different. In all other cases, the regression coefficients of the different groups differed significantly.

The lines for % fat on BMI of all the samples of men and for the different ethnic groups are plotted in Fig. 11 to show their range and overlap. Also plotted are the lines for the Italian shipyard workers (ITC) and the equations of Womersley and Durnin (1977) over the range of data here. Firstly, the regression coefficients for all samples, for Womersley and Durnin's subjects and for ITC men are quite similar and are unlikely to be significantly different. This has been found to be true for these men and ITC (F = 0.7, d.f. = 1,309, n.s.). But it can be expected that the intercepts would be and indeed are significant for men and for ITC (F = 46.7, d.f. = 1,310, P < 0.001). The relevance of differences in intercept when the data of the independent variable do not approach zero is debatable.

Considering the different ethnic groups in Fig. 12, the equations for women would give similar estimates of % fat in all groups except Indo -Mediterraneans over the range of BMI commonly found. This is less clear cut for the men. Pacific and Indo-Mediterraneans have steeper slopes than Africans and 'Asians'. 'Asians', who make up most of the world's population, have, in general, low correlations between BMI and body composition. This seems worthy of further investigation. Whether the low correlations and those in Africans are a consequence of the limited range of the data is not clear. In spite of low R2, the SEEs for African and 'Asian' groups are lower than those of other groups.

For the estimates of FM and FFM, the statistics in Table 1 suggest that the equations of the different groups would produce different estimates of FM for a given BMI but similar estimates of FFM, even though intercepts were significantly different. Comparing the European shipyard workers to these men, there was a difference in regression coefficients (F = 25.7, P < 0.001) but not in intercepts (F = 0.9, n.s.) for FFM on BMI, but for FM on BMI both the regression coefficients (F = 5.4, P < 0.05) and intercepts (F = 61.3, P < 0.001) were significantly different. It is likely that a similar situation obtains for the data of Womersley & Durnin (1977).

Multiple regression of body composition on BMI, sex (men = 0, women = 1) and ethnic group (1-4) showed that sex and ethnic group contributed significantly to the regression (except for the ethnic group in the FM regression) but the SEEs were not lower than those obtained when treating the sexes and groups separately. When considering the sexes separately, the contribution of ethnic group was significant only in the regression of FFM but again SEE were only 0.2 kg lower. There is little to be gained from these aggregative approaches which are in any case contra-indicated by analysis of covariance.

In the 106 samples of men where age was well established, age made a significant contribution to the regression of % fat on BMI (t statistic of the coefficient = 2.04, P < 0.05) and FFM (t=-2.76, P<0.01) but not FM (t= 1.81, n.s.). This significance was not maintained in the 61 samples of women with good mean ages. Age made no significant contribution to the regressions of the components of body composition with BMI.

Contents - Previous - Next