Prediction of fat mass from anthropometry at ages 7 to 9 years in Samoans: a cross-sectional study in the Ola Tuputupua’e cohort

Abstract

BACKGROUND/OBJECTIVE:

With increasing obesity prevalence in children globally, accurate and practical methods for quantifying body fat are critical for effective monitoring and prevention, particularly in high-risk settings. No population is at higher risk of obesity than Pacific Islanders, including children living in the independent nation of Samoa. We developed and validated sex-specific prediction models for fat mass in Samoan children.

SUBJECTS/METHODS:

Dual X-ray absorptiometry (DXA) assessments of fat mass and weight, height, circumferences, and skinfolds were obtained from 356 children aged 7–9 years old in the Ola Tuputupua’e “Growing Up” study. Sex-specific models were developed from a randomly selected model development sample (n = 118 females, n = 120 males) using generalized linear regressions. In a validation sample (n = 59 females; n = 59 males), Lin’s concordance and Bland-Altman limits-of-agreement (LoA) of DXA-derived and predicted fat mass from this study and other published models were examined to assess precision and accuracy.

RESULTS:

Models to predict fat mass in kilograms were: e^[(−0.0034355 * Age8 – 0.0059041 * Age9 + 1.660441 * ln (Weight (kg)) −0.0087281 * Height (cm) + 0.1393258 * ln[Suprailiac (mm)] − 2.661793)] for females and ê[−0.0409724 * Age8 − 0.0549923 * Age9 + 336.8575 * [Weight (kg)]⁻² − 22.34261 * ln (Weight (kg)) [Weight (kg)]⁻¹ + 0.0108696 * Abdominal (cm) + 6.811015 * Subscapular (mm)⁻² − 8.642559 * ln (Subscapular (mm)) Subscapular (mm)⁻² − 1.663095 * Tricep (mm)⁻¹ + 3.849035]for males, where Age8 = Age9 = 0 for children at age 7 years, Age8 = 1 and Age9 = 0 at 8 years, Age8 = 0 and Age9 = 1 at 9 years. Models showed high predictive ability, with substantial concordance (ρ_C > 0.96), and agreement between DXA-derived and model-predicted fat mass (LoA female = −0.235, 95% CI:−2.924–2.453; male = −0.202, 95% CI:−1.977–1.572). Only one of four existing models, developed in a non-Samoan sample, accurately predicted fat mass among Samoan children.

CONCLUSIONS:

We developed models that predicted fat mass in Samoans aged 7–9 years old with greater precision and accuracy than the majority of existing models that were tested. Monitoring adiposity in children with these models may inform future obesity prevention and interventions.

INTRODUCTION

Obesity and noncommunicable diseases account for over 75% of premature deaths in the Pacific and are emerging at earlier ages [1, 2]. Once excess body fat is established, children remain at higher obesity risk leading to increased adult risk [3]. Monitoring body fat is critical to detect and prevent obesity. Advanced techniques for body composition assessment exist (dual X-ray absorptiometry (DXA), computed tomography, and three-dimensional imaging [4, 5]), although they are rarely feasible in low-and-middle-income countries (LMICs). Complementing these techniques, valid prediction equations using anthropometric measures are accessible and inexpensive alternatives for indirect body fat estimation.

While Pacific Islander children were included in a few studies to develop prediction equations [6, 7], the only equation to specifically account for “Pacific Island” ancestry (to our knowledge) was published almost 20 years ago in Aotearoa New Zealand [8]. Given the increases in childhood body mass index (BMI) observed globally [2, 9], those earlier Pacific-specific equations, as well as, equations from settings with lower obesity prevalence [10–13], may not accurately predict body fat among children in contemporary LMICs characterized by high adiposity.

Samoa experiences the highest obesity prevalence globally [2] and will benefit from the development of predictive equations for monitoring fat mass in children. BMI is likely to poorly discriminate between fat and fat-free mass in Samoans, given its overestimation of body fat mass among Pacific Islanders resident elsewhere [8, 14]. In addition, at 3–7 years old, Samoans had a higher body fat percentage (BF%) than other ethnic groups, but a lower trunk-to-peripheral-fat ratio [15], suggesting that a different combination of body circumferences and skinfold thicknesses may be needed to account for their fat distribution.

Here, we developed and validated sex-specific models to predict fat mass in Samoan children aged 7–9 years using the Ola Tuputupua’e “Growing Up” cohort, and compared their predictive performance to existing models.

MATERIALS/SUBJECTS AND METHODS

Setting and study design

Samoa is a middle-income country with a population of >198,410 (92.6% of whom identify as Samoan) and has a median of 10 years of formal education [16–19]. Data were from a subset of the Ola Tuputupua’e cohort, which included 438 Samoan children aged 6–11 years old in 2019–2020. The cohort was purposively recruited in either 2015 (age 2–4 years) or 2017/18 (age 4–8 years) on ‘Upolu [20], the most populated Samoan island [17].

Yale (Protocol #2000020519) and Brown University Institutional Review Boards (IAA #18–41 959) and the Health Research Committee of the Samoa Ministry of Health approved study procedures. Parents provided written informed consent and children aged 7 years and older gave written assent.

Sample

The analysis sample included 356 children aged 7–9 years with complete anthropometric data and a full-body DXA scan in 2019–2020 (Supplementary Fig. 1). All children identified as having Samoan ancestry based on primary caregiver report of four Samoan grandparents. We chose to restrict the sample to include only 7–9 years old to avoid the likely onset of puberty and associated rapid growth and changes in body proportions [21–23].

Outcome assessment

Fat mass was estimated from a full-body DXA scan (Lunar iDXA, Encore 17, GE Healthcare Medicine, Madison, WI, USA) [24]. Children wore standard clothing (t-shirt, shorts, and shoeless). Scans were completed in pediatric mode by one of three trained DXA operators using standard procedures [15]. Daily quality control and assurance scans of a manufacturer-supplied phantom spine were performed to ensure scan reliability.

Anthropometric assessments

Anthropometric measurements were administered in standard clothing, recorded in duplicate, and averaged for analysis. If the first two measures differed by >0.1 kilograms (kg) for weight, >0.5 cm for height or circumference measurements, or >1 mm for skinfold thicknesses, then a third measurement was collected, and the closest two measures averaged. Height was measured to the nearest 0.1 cm with a portable stadiometer (Seca, Chino, CA, USA) and weight to the nearest 0.1 kg with a digital scale (Tanita Corporation of America, Arlington Heights, IL, USA). Averages of height and weight were used to calculate BMI. WHO child growth references were used to calculate BMI-for-age Z-scores [25] (used to describe the sample, but not in prediction models). Skinfold thicknesses were measured to the nearest 1 mm using Lange calipers (Beta Technology Inc., Houston, TX, USA). Body circumferences were measured to the nearest 0.1 cm with flexible measuring tapes (Seca, Chino, CA, USA) [26, 27].

Statistical analyses

We first described and compared sample characteristics by child age group and sex using generalized linear regressions. Child age was calculated by subtracting the reported date of birth from the measurement date and categorized into 1-year age groups because in this setting, recall of exact birthdate is less reliable than age in whole years. We presented means, standard deviations, and range of characteristics to improve comparisons with other studies [6, 8, 12]. We performed a sensitivity analysis to estimate the least square means and corresponding 95% confidence intervals of each characteristic by sex using age-adjusted generalized linear regressions. We plotted anthropometric characteristics versus fat mass and examined bivariate associations using Spearman correlations. Given our objective to develop sex-specific models and meaningful sample differences by sex, we sex-stratified the sample (n = 177 females, n = 179 males).

Then, we used random sampling in each sex-stratified sample to select a model development sample and validation sample in a 2:1 ratio (using SAS 9.4 proc surveyselect). This approach and the sample sizes were comparable to other studies [12, 28]. We verified homogeneity of each sample by comparing age distribution, sex, and anthropometric characteristics using Kruskal-Wallis tests (all p > 0.05, data not shown).

Since the fat mass was right-skewed, the natural logarithm of fat mass was used as the model outcome and for presentation, was back-transformed to absolute fat mass. We developed sex-specific models in the following steps:

1. Started with child age group, weight, and height

2. Identified the best fitting fractional polynomial function term for each variable using STATA commands fp and mfp [29, 30]

3. Added the anthropometric characteristic with the highest Spearman correlation with DXA-derived fat mass

4. Identified the best-fitting fractional polynomial function term for this characteristic (repeating Step 2)

5. Kept the characteristic if the model fit improved with a lower Bayesian information criterion (BIC) and/or adjusted R² value

6. Repeated Steps 3–5 until no remaining variable produced a lower BIC

7. Tested for interactions between child age groups with the remaining characteristics

8. Kept the interaction term if the model had the lowest BIC

We cross-validated the developed models for precision and accuracy within their respective validation sample (n = 59 females; n = 59 males) using Lin’s concordance correlations (ρ_c) [31, 32]. Concordance plots showed changes in ρ_C as a function of the tightness of the data about its reduced major axis (data precision) and the nearness of the data’s reduced major axis to the line of perfect concordance (data accuracy) [31]. To complement ρ_c, Bland and Altman’s limits-of-agreement (LoA) [33] and Wilcoxon signed-rank tests were used. Prediction models were considered to cross-validate if data revealed substantial concordance (0.95 < ρ_C ≤ 0.99, using cutoffs proposed by McBride [34]), no mean departure from agreement (95% CI of LoA includes 0), and no meaningful differences between predicted and observed values (p < 0.05) [35].

We further assessed model fit and performance using 1) plots of model residuals versus predicted fat mass and each model predictor to examine homoscedasticity, 2) R² to describe the proportion of fat mass variance explained by the model, and change in R² to describe the additional proportion of variance in fat mass explained by predictors added to the model, and 3) root mean square error (RMSE) to assess the average difference between predicted and observed values (a measure of accuracy). The value of RMSE estimated here, using STATA command regress, is equivalent to the standard error of estimate [36]. We conducted sensitivity analyses to compare the study models using 1-year age groups and age in years.

Lastly, we compared our models to existing models by Slaughter [10], Goran [11], Dezenberg [12], and Hudda [13] because the anthropometric parameters used in each equation matched those collected in our study (Supplementary Table 1). Using the validation sample, we compared observed fat mass based on DXA with the predicted values calculated from the sex-specific prediction models in this study and existing models. We compared model fit and performance using BIC, ρ_c, LoA, Wilcoxon signed-rank tests, R², adjusted R², and RMSE.

A two-sided α of 0.05 was specified. Analyses were conducted using SAS version 9.4. (SAS Institute, Cary, NC, USA) and STATA 16.0 (STATA Corp LLC, College Station, TX, USA).

RESULTS

Sample characteristics

Samoans had high adiposity at age 7–9 years, (28.21 ± 5.98% total body fat in females and 24.18 ± 6.60% in males, Table 1) and after adjustment for age in years (Supplementary Table 2). Fat mass was strongly correlated with all anthropometric characteristics, including body weight (female ρ = 0.936 and male ρ = 0.891), BMI (female ρ = 0.880, male ρ = 0.863), circumferences (ρ range:0.715–0.881) and skinfold thicknesses (ρ range:0.628–0.789, Fig. 1).

Table 1.

Sample anthropometric characteristics and body composition by age group and sex (n = 356)^a.

	Total	Age group (years)			P b	Sex		P b
		7	8	9		Female	Male
n (%)		155 (43.54)	116 (32.58)	85 (23.88)		177 (49.72)	179 (50.28)
Weight (kg)	30.56 ±8.98 (18.00–82.30)	26.07 ± 4.88 (18.00–51.10)	32.37 ± 8.93 (21.45–70.50)	36.29 ± 10.63 (22.30–82.30)	<0.001	30.21 ± 8.26 (18.25–68.50)	30.91 ± 9.65 (18.00–82.30)	0.460
Height (cm)	129.75 ± 8.07 (109.50–155.50)	124.29 ± 5.40 (109.50–135.80)	131.20 ± 6.58 (116.10–150.35)	137.74 ± 6.36 (121.95–155.50)	<0.001	129.60 ± 8.31 (113.15–150.30)	129.90 ± 7.85 (109.50–155.50)	0.730
BMI (kg/m2)	17.86 ± 3.33 (13.52– − 40.25)	16.76 ± 2.15 (13.52–27.71)	18.55 ± 3.47 (13.81–33.98)	18.90 ± 4.22 (14.28–40.25)	<0.001	17.72 ± 3.03 (14.11–32.99)	17.99 ± 3.59 (13.52–40.25)	0.436
BMI Z-score (SD)c	0.82 ± 1.19 (−1.70–6.72)	0.57 ± 1.02 (−1.70–4.65)	1.08 ± 1.26 (−1.65–5.73)	0.92 ± 1.30 (−1.20–6.72)	0.002	0.71 ± 1.04 (−1.20–4.91)	0.93 ± 1.32 (−1.70–6.72)	0.080
Circumferences (cm)
Abdominal	62.50 ± 9.04 (49.45–114.05)	58.99 ± 5.99 (49.45–91.20)	64.53 ± 9.07 (51.45–102.10)	66.13 ± 11.17 (50.75–114.05)	<0.001	62.41 ± 8.71 (50.75–102.10)	62.59 ± 9.37 (49.45–114.05)	0.849
Waist	60.50 ± 7.27 (48.40–95.45)	57.75 ± 5.17 (48.40–82.40)	61.98 ± 7.45 (50.25–95.45)	63.47 ± 8.53 (49.50–94.30)	<0.001	59.77 ± 6.80 (49.50–91.40)	61.21 ± 7.66 (48.40–95.45)	0.062
Hip	67.85 ± 8.86 (50.15–117.20)	63.66 ± 6.32 (50.15–90.20)	70.02 ± 8.51 (54.70–105.20)	72.54 ± 9.89 (59.85–117.20)	<0.001	68.14 ± 8.02 (50.15–105.20)	67.57 ± 9.63 (52.05–117.20)	0.546
Mid-upper arm	20.19 ± 3.37 (9.35–37.70)	18.98 ± 2.32 (14.65–28.50)	20.79 ± 3.57 (9.35–34.05)	21.57 ± 3.92 (15.85–37.70)	<0.001	20.19 ± 3.23 (15.55–34.05)	20.18 ± 3.50 (9.35–37.70)	0.975
Mid-calf	25.50 ± 3.44 (13.50–40.00)	24.06 ± 2.51 (18.35–35.20)	26.17 ± 3.52 (13.50–36.20)	27.22 ± 3.76 (20.40–40.00)	<0.001	25.58 ± 3.31 (19.25–37.25)	25.43 ± 3.57 (13.50–40.00)	0.675
Skinfolds (mm)
Bicep	7.82 ± 3.44 (4.00–23.00)	6.91 ± 2.53 (3.00–17.15)	8.49 ± 3.53 (3.75–0.30)	8.57 ± 4.30 (3.20–22.00)	<0.001	8.07 ± 3.35 (3.20–20.30)	7.57 ± 3.52 (3.00–22.00)	0.175
Tricep	9.73 ± 3.68 (4.00–23.00)	8.74 ± 2.74 (4.00–19.05)	10.57 ± 4.00 (4.30–22.20)	10.39 ± 4.27 (4.20–23.00)	<0.001	10.20 ± 3.61 (4.00–23.00)	9.26 ± 3.69 (4.00–22.20)	0.015
Subscapular	8.43 ± 4.27 (2.15–28.00)	7.07 ± 2.64 (2.50–18.50)	9.40 ± 4.57 (2.15–22.50)	9.58 ± 5.44 (4.20–23.00)	<0.001	8.92 ± 4.17 (3.00–27.00)	7.94 ± 4.32 (2.15–28.00)	0.029
Suprailiac	7.72 ± 4.61 (2.70–29.85)	6.48 ± 3.24 (2.70–21.00)	8.67 ± 4.99 (3.00–29.85)	8.69 ± 5.60 (3.20–27.30)	<0.001	8.43 ± 4.69 (2.75–29.85)	7.02 ± 4.44 (2.70–24.40)	0.004
DXA-derived body composition
Fat mass (kg)	8.45 ± 5.00 (2.97–41.23)	6.56 ± 2.66 (2.97–21.93)	9.41 ± 5.31 (4.07–36.77)	10.58 ± 6.45 (4.47–41.23)	<0.001	8.91 ± 4.59 (4.04–36.77)	8.00 ± 5.35 (2.97–41.23)	0.086
Fat-free mass (kg)	22.14 ± 4.47 (13.79–46.45)	19.57 ± 2.76 (13.79–28.92)	22.96 ± 4.18 (16.40–37.70)	25.71 ± 4.54 (16.62–46.45)	<0.001	21.29 ± 4.07 (14.37–35.48)	22.98 ± 4.69 (13.79–46.45)	<0.001
Lean mass (kg)	21.07 ± 4.27 (13.01–44.46)	18.63 ± 2.64 (13.01–27.71)	21.85 ± 4.00 (15.47–36.18)	24.45 ± 4.35 (15.77–44.46)	<0.001	20.25 ± 3.88 (13.70–33.80)	21.88 ± 4.48 (13.01–44.46)	<0.001
Total body fat (%)	26.19 ± 6.61 (15.75–54.06)	24.49 ± 5.16 (15.74–43.55)	27.60 ± 7.02 (17.28–54.06)	27.34 ± 7.66 (15.79–50.64)	<0.001	28.21 ± 5.98 (18.05–54.06)	24.18 ± 6.60 (15.75–50.64)	<0.001

BMI body mass index, DXA dual X-ray absorptiometry.

^aMeans ± standard deviations presented with the minimum and maximum values of the range in parentheses.

^bBased on generalized linear regression models to assess for characteristic differences by either child age group or sex.

^cZ-scores for body mass index (BMI) was calculated based on age and sex using the 2007 World Health Organization child reference.

Fig. 1. Sex-specific associations of anthropometric characteristics with fat mass derived from dual-energy X-ray absorptiometry (DXA) in Samoan children at ages 7 to 9 years (N = 356).

Anthropometric characteristics of 177 females and 179 males are shown in gray and black dots, respectively. The associations between DXA-derived fat mass and A weight, B height, C BMI, D BMI Z-score based on the World Health Organization child growth references, body circumferences for E abdomen, F waist, G hip, H mid-upper arm, I mid-calf, and skinfold thicknesses for J biceps, K triceps, L subscapular, and M suprailiac were estimated using Spearman correlation coefficients for females (ρ_F) and males (ρ_M).

Sex-specific model development and performance

Derived from the model development sample (Supplementary Tables 3, 4), the final prediction equations to model fat mass (kg) were: e^[(−0.0034355 * Age8 − 0.0059041 * Age9 + 1.660441 * ln (Weight (kg))−0.0087281 * Height (cm) + 0.1393258 * ln[Suprailiac skinfold (mm)] − 2.661793)] for females and ê[−0.04097 24 * Age8 − 0.0549923 * Age9 + 336.8575 * [Weight (kg)]⁻²− 22.34261 * ln (Weight (kg)) [Weight (kg)]⁻¹ + 0.01086 96 * Abdominal circumference (cm) + 6.811015 * Subscapular skinfold (mm)⁻²− 8.642559 * ln (Subscapular skinfold (mm)) Subscapular skinfold (mm)⁻²− 1.663095 * Tricep skinfold (mm)⁻¹ + 3.849035] for males, Where Age8 = 0 and Age9 = 0 for the 7 year age group, Age8 =1 and Age9 = 0 for the 8 year age group, Age8 = 0 and Age9 = 1 for the 9 year age group.

Models fitted the data well and homoscedasticity assumptions held (Supplementary Figs. 2–4). Each anthropometric characteristic alone explained between 52% and 91% of the variation in the outcome (Supplementary Table 5). When we compared models using child age in years and 1-year groups results were similar (Supplementary Tables 6, 7).

Cross-validation of sex-specific prediction models

In the validation sample, models showed high predictive performance for fat mass with minimal systematic deviation from the dashed line of perfect concordance (except for those >20 kg fat mass) (Fig. 2). Models had substantial agreement (female ρ_C = 0.961, 95% CI:0.947–0.975; male ρ_C = 0.973, 95%CI: 0.960–0.986, Table 2). There were no meaningful differences between predicted and observed fat mass (Wilcoxon signed-rank test p > 0.05).

Fig. 2. Concordance and 95% limits of agreement between the observed fat mass using dual-energy X-ray absorptiometry and model predicted fat mass in the validation sample set of Samoan children (n = 59 females, n = 59 males).

Model-predicted and observed fat mass of females and males are shown in gray and black, respectively. Concordance plots (A, B) show the tightness of the data about its reduced major axis and the nearness of the data’s reduced major axis to the dashed line of perfect concordance (with identify regression: y = x). Bland and Altman’s limits of agreement plots (C, D) show the regression line of the difference between predicted and observed values against the pairwise means, in comparison with the dashed line of perfect mean agreement (when y = 0).

Table 2.

Comparison of the Ola Tuputupua’e study model and previously published models from non-Samoan samples to predict fat mass among Samoan children at ages 7–9 years in the validation sample.

	Predicted, kg Mean ± SD (min–max)	P for differences with observeda	Concordanceb, ρC (95% CI) (r, Cb)	Mean of Predicted -Observed, kg (95% LoA)c
Female (n = 59)
Study Model	8.961 ± 4.468 (4.461–29.600)	0.667	0.961 (0.947–0.975) (0.982, 0.979)	−0.235 (−2.924, 2.453)
Slaughter et al. [10]	5.594 ± 2.594 (1.852–14.523)	<0.001	0.474 (0.353–0.595) (0.762, 0.622)	−3.602 (−10.933, 3.730)
Goran et al. [11]	8.649 ± 2.537 (5.178–15.256)	0.239	0.645 (0.564–0.726) (0.851, 0.758)	−0.547 (−7.522, 6.429)
Dezenberg et al. [12]	7.313 ± 3.450 (2.936–19.531)	<0.001	0.760 (0.697–0.841) (0.925, 0.831)	−1.883 (−7.018, 3.253)
Hudda et al. [13]	9.037 ± 5.032 (4.552–33.845)	0.378	0.984 (0.976–0.991) (0.988, 0.996)	−0.159 (−1.985, 1.668)
Male (n = 59)
Study Model	7.705 ± 3.786 (3.609–19.337)	0.213	0.973 (0.960–0.986) (0.978, 0.995)	−0.202 (−1.977, 1.572)
Slaughter et al. [10]	5.754 ± 3.284 (2.271–14.189)	<0.001	0.768 (0.687–0.849) (0.922, 0.833)	−2.153 (−5.458, 1.151)
Goran et al. [11]	8.306 ± 2.469 (5.158–14.783)	0.002	0.818 (0.770–0.867) (0.936, 0.874)	0.398 (−3.561, 4.357)
Dezenberg et al. [12]	6.312 ± 3.271 (1.603–14.633)	<0.001	0.839 (0.779–0.900) (0.943, 0.890)	−1.596 (−4.565, 1.373)
Hudda et al. [13]	8.108 ± 3.912 (3.273–21.375)	0.184	0.958 (0.937–0.979) (0.961, 0.997)	0.201 (−2.055, 2.456)

kg kilograms, min minimum value, max maximum value, SD standard deviation, ρ_C Lin’s concordance correlation coefficient, CI confidence interval, r Spearman coefficient correlation between predicted and observed fat mass, C_b bias correlation factor, LoA Bland and Altman’s limit of agreement

^aDifferences between the predicted and observed values of fat mass were assessed using Wilcoxon signed-rank tests.

^bConcordance is the product of r and C_b that measures how far the best fit regression between the predicted and observed values deviates from the identity regression (the $4\stackrel{\circ }{5}$ line, intercept = 0, slope = 1).

^cBased on the Bland and Altman’s limits of agreement procedure (1986), with 0 indicating perfect mean agreement between the mean difference and pair-wise means for predicted and observed values

Comparison of prediction models

In the validation sample, our models showed better predictive performance, with the highest ρ_C (Table 2) 96% of the total variation in fat mass explained, (female R² = 0.964 and male R² = 0.956) and with smaller errors (female RMSE = 0.861 and male= 0.802 kg) compared to the majority of existing non-Samoan derived equations (Supplementary Table 8). While there was poor agreement (ρ_C ≤ 0.90) between DXA-derived fat mass and predicted fat mass from Slaughter et al., Dezenberg et al., and Goran et al. equations, substantial agreement (0.95 < ρ_C ≤ 0.99) was observed using Hudda et al.

DISCUSSION

This study developed sex-specific models to predict fat mass from anthropometry in a sample of Samoan children aged 7–9 years. Overall, the equations show high predictive ability, with substantial concordance and agreement between DXA-derived fat mass and model-predicted fat mass, and low individual error. The models predicted DXA-derived fat mass in Samoan children with greater precision and accuracy than the majority of other existing equations that were tested and may be used as a tool for monitoring body fat in this high-obesity-risk population.

Weight had the strongest positive correlation with fat mass and accounted for ~90% of the variation in fat mass of both females and males. The models consisted of different combinations of anthropometric predictors with weight, which may be explained by sex differences in regional accumulation and storage of body fat. At ages 7–9 years, females had a higher total BF% and trunk-to-peripheral fat ratio than males (data are not shown). This was similar to what we previously observed at ages 3–7 years in a representative subset of the cohort [15], which may be related to nutrition transitions that have been associated with high and rising prevalence of obesity [37, 38], and increasingly high body size at earlier ages in this setting [20].

To our knowledge, no study has developed and validated models to predict DXA-derived fat mass in children in any Pacific Island country. Modeling has often focused on the prediction of percentages and not absolute values of body fat. Total BF%, $\frac{fatmass}{weight}*100$, has been used to classify obesity and predict cardiometabolic health in children [39, 40], but includes fat mass as the numerator and a component of the denominator, and consequentially over-adjusts for weight meaning it is not an ideal model outcome [41]. Rush et al. developed a model to predict fat-free mass using deuterium dilution measurements and bioelectrical impedance analysis (BIA) in New Zealand European, Maori, and Pacific Island children aged 5–14 years old [8]. While a direct comparison of our study with Rush et al. was not possible given that we did not conduct BIA measures, future studies should compare the predictive performance of both models.

Based on our findings, our sex-specific models should be used in place of models developed using non-Samoan samples, which poorly predicted fat mass. Fat mass was underestimated by the Slaughter et al. model and overestimated by Goran et al., consistent with previous studies in middle-and-high-income settings [9, 28, 42, 43]. Apart from differences in sample size and age range, the poor performance may be explained by differences in population-level adiposity levels [9] and nutritional environments [37, 38]. On average, fat mass in the 1980–1990s samples of US children in Slaughter et al. and Goran et al. was nearly 2 kg lower than that of the 2019–2020 Samoan sample [10, 11]. Since these previously existing models were unlikely derived from samples with a similar body fat distribution to Samoans and may have differences in pubertal development, inaccurate estimations were expected.

Interestingly, the Hudda et al. equation, specifying children with “other” ethnic origin [13], accurately predicted DXA-derived fat mass in our study sample. The Hudda et al. model was the only one of the four existing models to include ethnicity as a predictor and had approximately ten times the number of children in their development sample, which may partly explain the good model fit and performance. Their sample was also pooled from four UK population-based studies of children with a wide range of ages (4–15 years) and reported ethnic origins (including n = 239 “other” predominantly mixed ethnicities), which may account for better applicability [13]. Although we derived fat mass with DXA rather than deuterium dilution as used by Hudda et al., our findings did align with another validation study finding [44]. Their equation worked very well for children across a wide range of ethnicities and countries, including Australia and Aotearoa New Zealand where many Pacific Islanders live [44]. The Hudda et al. equation could be useful in the Samoan setting when skinfold thickness measurements may be difficult to obtain.

Limitations and strengths

Given that the sample was not selected purposively to develop prediction equations, the models may not be generalizable to populations that do not share similar anthropometric characteristics. In addition, different relationships may be observed between our anthropometric predictors and fat mass in other settings. For children with >20 kg of fat mass, the study models provided underestimations. We were restricted to the data presently available in the cohort for this analysis and although the study sample was greater in size compared to Rush et al. [7], we hope to expand the narrow age range with future data collection. If additional data become available, external validity may be assessed. For random selection of model development and validation samples, alternative methods may be used in the future to partition the data (i.e., k-fold cross-validation). We focused on DXA-derived fat mass but future work may consider other outcomes including fat and fat-free mass indices which account for height. We further recognize that using anthropometry to predict fatness in children has potential issues related to inherent sex differences by age, pubertal development (which is not measured in this cohort, but may be ongoing in some of the older children), and inter-interviewer variability. Any use of prediction equations should be accompanied by uniform and refresher training in anthropometric measurement, particularly to minimize errors in skinfold thickness measurements [45].

Despite these limitations, this study has many strengths. We used Ola Tuputupua’e data to develop and validate the first prediction model of fat mass for children in Samoa, where obesity is highly prevalent and preventative interventions are needed. Pacific Islanders are one of the fastest-growing ethnic groups in the United States and this work may be generalizable to other settings with high obesity prevalence and significant economic and nutrition transitions. We measured fat mass using DXA, an accurate method for body composition measurement and used Lin’s concordance correlation coefficient as a measure of agreement for fat mass allowing comparison with other studies [32, 34]. Polynomial terms for anthropometric predictors not only allowed us to account for non-linear relationships with fat mass, but also to improve final model fit and performance. Moreover, this work aligns with obesity prevention efforts in Samoa and the Healthy Islands vision of Pacific Island leaders to reduce the noncommunicable disease burden [46].

Public health implications and future directions

The newly developed equations may be used in routine surveillance, particularly in the ongoing Samoa Ministry of Health School health program, as an obesity screening tool, and add value to obesity-related research studies. The Excel calculator provided in the Supplementary file would allow simple calculations of fat mass from the relevant predictor variables. Accurate and precise estimation of fat mass with anthropometry rather than DXA will save costs and time to complete the physical measurements, given that there is only one DXA scanner in Samoa.

To further inform public health and primary care practice in Samoa, future work requires a larger sample to assess the predictive ability of these models to classify obesity. In this population, it is unclear whether established thresholds for childhood obesity based on fat mass are appropriate for identifying later cardiometabolic risk. Continuation of the Ola Tuputupua’e cohort will allow us to address these questions. Monitoring and identification of high body fat in children should be integrated into public health programs to prevent obesity at the earliest opportunity and these efforts may be supported by utilizing prediction equations in the future.

DATA AVAILABILITY

Data are available upon reasonable request to the corresponding author.