Birth weight, conventionally used as a measure of fetal growth, is a composite of many components including, bone, internal organs, muscle, fat and fluids. These may be determined by different regulatory mechanisms, both environmental and genetic. It has been suggested that fat-mass reflects the intra-uterine environment, whereas fat-free mass is more likely to be altered by genetic factors [1, 2].

A major determinant of body composition at birth is gender. Males tend to be longer and heavier than females, but females have more subcutaneous fat [1, 3–5]. Other factors known to influence body composition include parity and maternal BMI which have stronger associations with measures of fat at birth than measures of skeletal size [1, 2], and maternal [2, 6] and paternal [1, 7, 8] height which are more associated with skeletal size. The effects of smoking are less clear with some studies reporting a reduction only in skeletal measures as a result of maternal smoking [7, 9–12], whereas others find a general decrease in all aspects of growth [13–15].

Identifying factors related to fetal growth has become an important part of the study of the developmental origins of adult disease. Relationships between low birth weight and later ill health such as diabetes, cardiovascular disease, stroke, and obesity, are well established [16–19]. Not only is birth weight seen to be important, but also how thin, or how short, the baby is [20–22]. Considering components of birth weight separately, and investigating their determinants, may give more insight into these associations.

There are various methods of obtaining measures of body composition in utero and at birth. Studies have used DEXA scans [7, 23, 24], ultransonography [25–27], and total body electrical conductivity [9, 28]. These methods require expensive equipment, specialist training, and the more accurate methods are likely to be difficult to do in large samples. A simpler method for assessing body composition at birth is for midwives to take detailed anthropometry using standardized techniques, which can provide information on different aspects of growth. These measurements can be analysed individually [2, 5, 13, 14, 29]. However, often these measurements tend to be highly intercorrelated leading to problems of multicollinearity in regression analysis. Careful decisions, therefore, need to be made as to which measurements are the most appropriate and convey most information describing the aspect of growth of interest. A method able to combine them to provide more general estimates of size and body composition (as DEXA and ultrasonography do) would have considerable advantages. Mathematical equations to obtain estimates of body composition have been used previously, but these equations are not internally derived [1, 30].

An alternative way of summarizing body size and composition using anthropometric measurements could be to use principal components analysis (PCA) [31], a method previously used to describe different aspects of birth size [10, 23, 32, 33]. This statistical approach can be used to investigate the underlying structure of a dataset, by reducing the data to meaningful components. This method enables a large number of correlated variables to be summarized in terms of a relatively small number of uncorrelated principal components. The extent to which the principal components capture the variation in the original variables can be quantified in terms of the proportion of variance explained. The components can aid interpretation and represent meaningful constructs that parsimoniously describe the multivariate data.

We used principal components analysis to reduce multivariate measurements at birth to two more meaningful independent components representing different aspects of birth size: skeletal size and fat. We examined their relationships with established determinants of birth weight.

Principal components analysis produced two clear components allowing recognition of discrete aspects of birth size when using the rotated analysis. The rotated components, which can be considered to represent "skeletal size" and "fat", had different associations with factors known to alter birth weight.

There are various methods of obtaining measures of fat and skeletal size in utero and at birth. Direct measures of body composition can be obtained using DEXA scans [7], ultransonography [25–27], and total body electrical conductivity [9, 28]. Although DEXA and ultrasonography would be the gold standard for measuring body composition, these methods are likely to be expensive and impractical in a large epidemiological study. Anthropometric measurements provide an alternative method where measures such as head circumference, length, and limb length can represent skeletal size, and measures of skinfold thicknesses can represent fat. Mathematical equations for combining these measurements to obtain estimates of body composition have previously been defined and can be applied retrospectively to the data [35, 36]. Principal components analysis is another way of summarizing body size and composition using these anthropometric measurements and enables easier analysis of the data by reducing the problem of multicollinearity. It is a well established statistical method that can easily be implemented and allows the components to be internally derived from the data.

It has been suggested that genetic factors are more likely to alter skeletal size whereas the intrauterine environment has more of an effect on fat mass[1]. We were able to gain some insight into potential genetic determinants by also producing components for the mother and father, which had not been carried out in other studies looking at PCA on birth size. Maternal and paternal skeletal size components were found to have significant associations with the skeletal size components in boys and girls (Tables 6 and 7), in keeping with the notion that genetic factors are more likely to affect skeletal growth. This finding is consistent with associations found between parental height and birth length seen in other studies [2, 6–8, 10].

Factors associated with the fat component were largely those related to the intrauterine environment (Table 7) with parity and maternal glucose associated with fat in both boys and girls, and the strongest predictor of fat in the girls being maternal fat component (Table 7). Similar relationships of parity with fat measures have recently been reported [1, 2] and it has been suggested this is due to the mothers' vascular system becoming better adapted to the transfer of energy to the fetus after the first pregnancy [2]. Our findings are also in agreement with other studies that have found maternal glucose [27, 28] and maternal BMI [2, 37] to be good predictors of measures of fat and soft tissue in babies at birth.

Maternal smoking appeared to alter only the skeletal size component (Fig 3b), suggesting the intra-uterine environment does play a role in determining skeletal size. This is consistent with previous research [7, 9, 10], particularly the studies by Lindley et al[11, 12] who found reduced birth length, and head circumference, but increased ponderal index in babies born to smoking mothers, suggesting fat deposition is maintained despite deficits in other aspects of fetal growth. Other studies have found maternal smoking to be associated with a reduction in all measures of size, including fat and skinfold thicknesses [11, 13, 14], although the relationship with fat reported by Zaren et al[13] was weak, and in the study by Cliver et al[14], skinfolds were only affected in babies born to heavy smokers. In multiple regression analysis (Table 7) smoking did show an independent association with the fat component in the boys, but not the girls, which may account for some of the differences in the literature. This sexual dimorphism has been seen before [15] and it was suggested that it may be due to hormonal differences.

These findings are all consistent with previous research. The disparity between the associations of gestation and the components of birth size, however, has not been described previously. In our data the strongest independent predictor of the skeletal size component in both boys and girls was gestation, but it showed no association with the fat component. It is important to note these results are only based on term babies, so gestation in this study refers to the limited period of 37–42 weeks. Our data suggest that after 37 weeks gestation the increase in birth weight is as a result of further skeletal growth, rather than fat deposition (Fig 4a and Table 7). In contrast, other studies have found increases in skinfold thickness [3, 5, 38] and fat mass [25] with gestational age at term. In our data, the subscapular skinfold thickness had a weak but significant correlation with gestation (r = 0.096, p = 0.011), but the tricep skinfold thickness did not (r = 0.067, p = 0.076). These differences may be partly due to what the components represent. A recent study by Guihard-Costa et al. [29] found that although skinfold thicknesses increased with gestation, their ratio with body weight between 33 and 42 weeks gestation significantly decreased, and this may reflect what is seen with our data. This finding warrants further investigation as our cross-sectional data are not conclusive.

Other studies have used principal components analysis to produce components describing birth size [10, 32, 33], although they differ slightly from the components produced from our data. Denham et al[32] used principal components analysis to describe differences in neonatal size between black and white babies of low socioeconomic status. The components they obtained were similar to the unrotated components produced in our data (Table 3), representing general body size, and contrasts between skeletal and fat measures (referred to as "body composition"). Hindmarsh et al[10] examined size and shape at birth and reported 4 components in total, although the third and fourth components together explained only an additional 15% of the variation in birth measures. The first two components again were similar to our unrotated components. A key difference in our analysis, was the use of varimax rotation on the components (Table 4). The advantage of this method over the unrotated analysis, is that we obtained a clear distinction between measures of skeletal size and fat which are more meaningful in terms of explaining body size. Components representing skeletal size and fat were produced in the study by Evans et al[33] investigating the effect of frequent prenatal ultrasound examinations on birthweight. However, their components were obtained by a slightly different method: two separate analyses were carried out, one entering in only skeletal measures, and one entering in only fat measures, rather than all measures being entered into one analysis. Similarly, Koo et al[23] derived two components representing fat: one from PCA on neonatal body circumferences, and one from PCA on skinfold thickness measurements. The advantage of our method is that no clinical assumptions need to be made about the data prior to analysis. The components are derived purely from the data.

It is clear that this analysis will need to be repeated in another population, particularly as the difference in the effect of gestation on skeletal size and fat is not widely reported in the literature. Further research into how representative the components are of skeletal size and fat compared with more complex measures of body composition would also be of interest. However, other associations found with the components largely confirm what has been found in other studies looking at determinants of body composition, suggesting principal components analysis is a valid method for analysing birth size.

In conclusion, we have provided further evidence demonstrating principal components analysis is a useful method for analysing anthropometric data in babies at birth. The addition of varimax rotation enabled more distinct, clinically relevant, components to be produced than those seen in previous studies. We have clearly shown differences in the determinants of skeletal size and fat components at birth, which support findings from previous research, as well as highlighting possible new areas for further study. This approach may be a useful tool in future investigations into the developmental origins of adult disease.