Body fat estimation in the Mexican women’s national soccer team: a cross-sectional agreement study between a derived Mexican Football Federation anthropometric equation and dual-energy X-ray absorptiometry Authors: Carlos Abraham Herrera-Amante, Israel Ríos-Limas, Wiliam Carvajal-Veitía, Rodrigo Yáñez-Sepúlveda, Citlalli Daniela Morón-López, Karla Jessica Mendoza-Sánchez, Jaime Andrés Figueroa-Conde, César Octavio Ramos-García, José Francisco López-Gil Published: 2026 ABSTRACT AbstractBackgroundPrecise assessment of body composition is fundamental for monitoring training adaptations, guiding nutritional strategies, and informing performance-related decision-making in elite female soccer. Dual-energy X-ray absorptiometry (DXA) is considered a reference method; however, its limited accessibility restricts routine use in high-performance settings. Existing anthropometric equations often lack population-specific validation, particularly for elite Mexican female soccer players. This study aimed to derive and internally validate an anthropometric equation for estimating body fat percentage (%BF) and to evaluate its agreement with DXA.MethodsTwenty-seven elite players from the Mexican Women’s National Team underwent DXA assessment and International Society for the Advancement of Kinanthropometry (ISAK)-standardized anthropometry during a training camp. Sample-size adequacy was assessed a posteriori. A multiple linear regression model was derived using the sum of eight skinfolds (Σ8SF) and waist girth (WG) as predictors. Regression assumptions were verified through residual diagnostics, and multicollinearity was assessed using variance inflation factors. Predictive performance was evaluated using mean absolute error (MAE), root mean square error (RMSE), and standard error of estimate (SEE). Agreement with DXA was assessed using the intraclass correlation coefficient (ICC) and Bland–Altman analysis. Internal validation was performed using leave-one-out cross-validation (LOOCV). The six historical anthropometric equations were additionally refitted to the present sample to enable a like-for-like comparison of predictor structures.ResultsThe Mexican Footbal Federation equation (%BF = 0.12 × Σ8SF +0.39 × WG − 17.0) demonstrated moderate agreement with DXA (ICC = 0.71; mean bias: −0.05%; 95% LOA: −4.61% to +4.70%). The model explained 54% of the variance in DXA %BF (R² = 0.54; R²-adjusted = 0.50) and showed acceptable predictive performance (LOOCV Q² = 0.50; RMSE = 2.33%; MAE = 1.86%; SEE = 2.47%). After refitting each of the six historical anthropometric equations to the present sample, the Mexican Footbal Federation predictor combination retained the lowest RMSE (2.33% vs. 2.55–2.93%) and the highest ICC (0.71 vs. 0.44–0.63), supporting that the model’s advantage stems from its predictor structure rather than calibration on the test data.ConclusionsThe Mexican Football Federation equation provides a population-specific, field-applicable tool for estimating %BF in elite Mexican female soccer players. External validation in larger and more diverse cohorts is required before its widespread application. Introduction Body composition is a key determinant of performance in elite soccer, with body fat (BF) playing a particularly relevant role. Excess adiposity may impair mechanical efficiency, acceleration capacity, and the ability to sustain repeated high-intensity efforts during match play [1,2]. Recent evidence in professional female soccer players has also demonstrated associations between anthropometric profile, body composition, and physical performance, underscoring the importance of accurate assessment in this population [3]. As women’s soccer continues to grow globally, interest in understanding the structural and morphological characteristics that influence both health and performance has increased. Dual-energy X-ray absorptiometry (DXA) is widely recognised as the reference method for quantifying BF [4]. However, its routine use in sports environments is limited by cost, accessibility, and technical requirements. These constraints have led to the widespread adoption of anthropometric equations as practical field-based alternatives [5]. Despite their utility, such equations require rigorous validation against reference methods before being applied to specific populations. A recent systematic review identified 40 equations developed using multicomponent models for athletes, yet most have not been validated in women and exhibit marked population, sport, and morphological specificity [6]. The derivation of population-specific predictive equations from simple anthropometric or performance variables is a recurring theme across applied sports science, with comparable methodological approaches recently illustrated for predicting one-repetition-maximum strength from body mass in male athletes [7]; the present work extends that tradition to body-composition estimation in elite female soccer. Empirical evidence reinforces the need for contextual validation. In collegiate American football players, traditional anthropometric equations underestimated BF compared with DXA, prompting the development of population-specific models that were not generalisable to other contexts [8]. Similarly, equations generated in heterogeneous samples of women produced systematic errors when applied to female handball players [9]. In soccer, studies have also shown considerable variability in the agreement between anthropometric equations and DXA across countries and populations. For example, in Mexico, the equations proposed by Oliver, Ball, Civar, and Wilmore showed the greatest accuracy for estimating fat mass in professional soccer players [10], whereas in Colombia the Durnin and Womersley equation and a five-skinfold model demonstrated the best agreement [11]. Among international players, only the equation by Deurenberg et al. produced estimates comparable to DXA, while others showed relevant biases [12]. In female soccer players, some equations may estimate fat mass adequately at the group level but show important individual-level biases, limiting their utility for player monitoring [4]. This issue is particularly relevant in women’s soccer due to the physiological demands of the game and the considerable interindividual variability in body characteristics, even at professional levels. For instance, during the 2019 FIFA Women’s World Cup, players exhibited wide ranges in age, stature, and body mass, with goalkeepers presenting higher BF compared with outfield players [13]. Despite these advances, evidence in elite female soccer players remains limited, and to date no anthropometric equations have been validated against DXA in Mexican female players. This gap restricts the precision of body composition assessments in this population, especially considering the rapid growth of Liga MX Femenil and the absence of population-specific tools. Accordingly, the aim of the present study was to generate and internally validate an anthropometric equation for estimating body fat percentage (%BF) in elite Mexican female soccer players and to evaluate its agreement with DXA. Methods Study design A cross-sectional study was conducted to derive and internally validate a multivariable predictive model for estimating body fat percentage (%BF). All participants attended the assessment centre on a single occasion for data collection. The study was designed and reported in accordance with the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guidelines [14]. Setting All assessments were conducted between January 3 and 5, 2026, at the High-Performance Centre (CAR, in Spanish) of the Mexican Football Federation (FMF, in Spanish) in Mexico City. The CAR provides controlled laboratory conditions and specialised equipment for body composition assessment. The study adhered to the ethical principles of the Declaration of Helsinki [15] and the Regulations of the General Health Law for Health Research [16]. Written informed consent was obtained from all participants in accordance with Articles 20–22 of the aforementioned regulations. The research protocol was approved by the Research Ethics Committee of the Metropolitan Autonomous University, Xochimilco Unit (code: DCBS.CD.071.2) and registered at ClinicalTrials.gov (identifier: NCT06941064). Participants Twenty-seven female soccer players from the Mexican Women’s National Team (senior team) participated in the study. All players were active members of the national team at the time of assessment and had prior international competition experience. Participants were consecutively included based on their availability during the national team training camp. This sampling strategy constitutes a convenience sample; however, it reflects the total accessible population of this highly homogeneous cohort of elite athletes. Assessments Participants were instructed to arrive wearing close-fitting Lycra clothing (shorts and a sports shirt with the back and arms exposed) to facilitate identification of anatomical landmarks. All metal objects (e.g. rings, bracelets, piercings) were removed prior to testing. Assessments were conducted during the national team training camp in the following sequence:Hydration statusAnthropometric measurementsDXA body composition analysis Hydration status Hydration status Anthropometric measurements Anthropometric measurements DXA body composition analysis DXA body composition analysis Hydration status Hydration status was assessed using salivary osmolality measured with a portable osmometer (MX3 Hydration Testing System®, MX3 Diagnostics, Austin, TX, USA). Following manufacturer recommendations [17], participants refrained from consuming food or liquids for at least 5 minutes prior to saliva collection. Participants first swallowed residual saliva and then produced a fresh sample. With the mouth open and the tongue exposed, the analyser was positioned at approximately 75° to facilitate collection. A single measurement was obtained per participant. As an inclusion criterion, salivary osmolality values were required to fall between 65 and 100 mOsm/L. All 27 screened players met the hydration criterion. No participants were excluded on hydration grounds. Descriptive osmolality values are reported in Section 3.1. Anthropometry Anthropometric measurements were performed by an International Society for the Advancement of Kinanthropometry (ISAK) Level 2–certified anthropometrist following the standards of the ISAK [18].Body mass (kg): digital scale, accuracy 50 g (SECA® 874, Hamburg, Germany)Stretch stature (cm): portable stadiometer, accuracy 1 mm (SECA® 217, Hamburg, Germany)Waist girth (cm): flexible, inelastic steel tape, accuracy 1 mm (Rosscraft®, Surrey, Canada)Skinfolds (mm): Harpenden caliper, accuracy 1 mm (Holtain Ltd., Crymych, UK) Body mass (kg): digital scale, accuracy 50 g (SECA® 874, Hamburg, Germany) Body mass (kg): digital scale, accuracy 50 g (SECA® 874, Hamburg, Germany) Stretch stature (cm): portable stadiometer, accuracy 1 mm (SECA® 217, Hamburg, Germany) Stretch stature (cm): portable stadiometer, accuracy 1 mm (SECA® 217, Hamburg, Germany) Waist girth (cm): flexible, inelastic steel tape, accuracy 1 mm (Rosscraft®, Surrey, Canada) Waist girth (cm): flexible, inelastic steel tape, accuracy 1 mm (Rosscraft®, Surrey, Canada) Skinfolds (mm): Harpenden caliper, accuracy 1 mm (Holtain Ltd., Crymych, UK) Skinfolds (mm): Harpenden caliper, accuracy 1 mm (Holtain Ltd., Crymych, UK) All instruments were calibrated prior to testing. Each anthropometric variable was measured at least twice. A third measurement was taken when the difference between the first two exceeded 5% for skinfolds or 1% for other variables. When two measurements were obtained, the mean was used; when three were obtained, the median was used, following ISAK guidelines. The sum of eight skinfolds (Σ8SF) was calculated, including the triceps, subscapular, biceps, iliac crest, supraspinale, abdominal, thigh, and calf sites. Body fat percentage (%BF) was also estimated using classical anthropometric equations, presented here in chronological order: Sloan [19], the 4‑site formula commonly known as the Faulkner equation [20], Durnin and Womersley [21], the 6‑site equation of Yuhasz [22], and Jackson and Pollock [23]. Although widely referred to as the “Faulkner equation”, historical analyses have shown that this 4‑site formula was not developed by Faulkner but corresponds to an unpublished derivation by Yuhasz, later disseminated under Faulkner’s name [24]. Dual-energy X-ray absorptiometry (DXA) Body composition was assessed using a Lunar iDXA system (General Electric®, Madison, WI, USA). Daily quality control procedures were performed according to manufacturer recommendations [25[25]][25]. All procedures followed standardised protocols and current best practices for DXA assessment [25–28]. The Lunar iDXA system demonstrates excellent in vivo precision, with coefficients of variation of approximately 0.82% for total fat mass and 0.86% for %BF under test–retest conditions with repositioning [29]. No internal test–retest reliability assessment was performed in the present study; consequently, a technician-specific minimum detectable change (MDC) cannot be derived from these data. Statistical analysis Data normality was assessed using the Shapiro–Wilk test. Due to non-normal distribution, descriptive statistics are presented as median and interquartile range (IQR). Associations between anthropometric variables and DXA %BF were evaluated using Spearman’s rank correlation coefficient (ρ), interpreted according to Hopkins [30]: trivial (0.00–0.09), small (0.10–0.39), moderate (0.40–0.69), strong (0.70–0.89), or very strong (≥ 0.90). Although raw variables were non-normal, regression assumptions were verified through residual diagnostics. A multiple linear regression model was derived, with DXA %BF as the dependent variable. Predictor variables were selected based on physiological relevance and exploratory associations. Several candidate models were examined; the final model was selected based on interpretability, compliance with regression assumptions, and predictive performance. The sum of eight skinfolds (Σ8SF) and waist girth (WG) were retained as predictors, providing the best balance between model fit and parsimony. No automated variable selection procedures were used. Regression assumptions were evaluated through:LinearityNormality of residuals (Shapiro–Wilk)Homoscedasticity (Breusch–Pagan)Independence (Durbin–Watson)Multicollinearity (VIF < 5) Linearity Linearity Normality of residuals (Shapiro–Wilk) Normality of residuals (Shapiro–Wilk) Homoscedasticity (Breusch–Pagan) Homoscedasticity (Breusch–Pagan) Independence (Durbin–Watson) Independence (Durbin–Watson) Multicollinearity (VIF < 5) Multicollinearity (VIF < 5) Model performance was assessed using R, R², R²-adjusted, ANOVA, MAE, and RMSE. Diagnostic plots (residuals vs. fitted, Q–Q plots) and Cook’s distance were used to identify influential observations. Internal validation was performed using leave-one-out cross-validation (LOOCV) to obtain Q². Agreement between the FMF equation and DXA was evaluated using the intraclass correlation coefficient (ICC 3,1) with absolute agreement [31]. Bland–Altman analysis was used to assess systematic and proportional bias. Limits of agreement (LOA) were calculated as mean difference ± 1.96 SD [32]. Sample-size adequacy for model derivation was further evaluated a posteriori, to assess the adequacy of the achieved convenience sample, not to determine recruitment targets, using the closed-form criterion proposed by Riley et al. for two-predictor multivariable prediction models with a continuous outcome [33]. For the observed coefficient of determination of the FMF model (R² = 0.54) and a target uniform shrinkage of ≥ 0.90, the minimum required sample size approximated 22 participants, a threshold satisfied by the realised sample (n = 27). This criterion supplements rather than replaces the conventional observation-to-variable heuristic and acknowledges that estimates remain modest in absolute terms; this caveat is reiterated in the limitations section. In addition to applying the six previously published anthropometric equations in their original form, each equation was refitted to the present sample preserving its original predictor structure (Sloan: iliac crest and triceps skinfolds; Faulkner-style four-site formula: sum of four skinfolds; Durnin & Womersley: log-transformed four-site sum plus age; Yuhasz: sum of six skinfolds; Jackson & Pollock three-site: sum of three skinfolds, its square, and age). Coefficients were re-estimated by ordinary least squares so that error metrics derived from refitted models can be compared with the FMF equation under identical estimation conditions. The Jackson & Pollock seven-site model could not be refitted because the chest and axilla skinfolds are not part of the ISAK eight-site protocol used here. As a complementary sensitivity analysis, non-linear and machine-learning approaches were also explored [34]. The FMF model was benchmarked against nine alternative algorithmic classes, including regularised linear models (Ridge, Lasso, ElasticNet, and Bayesian Ridge), an instance-based method (k-nearest neighbours), a kernel-based method (support vector regression with a radial basis function kernel), ensemble tree methods (random forest with maximum depth restricted to three and gradient boosting with maximum depth restricted to two), and a small multilayer perceptron with a single hidden layer of four units. Models were evaluated using two predictor sets: the FMF set (Σ8SF + waist girth) and the full set of 13 candidate anthropometric variables. Out-of-fold predictions were generated using 5 × 5 nested cross-validation, with hyperparameter tuning performed exclusively within the inner loop to prevent data leakage, while the outer loop provided unbiased estimates of predictive performance. RMSE, MAE, and the cross-validated coefficient of determination (Q²) were calculated from the out-of-fold predictions, and 95% confidence intervals were estimated using a 1,000-replicate bias-corrected and accelerated (BCa) bootstrap procedure. Permutation importance (200 repeats) was computed for the best-performing regularised models to characterise predictor contributions. Analyses were implemented using scikit-learn 1.4 and statsmodels 0.14 in Python 3.12. Statistical significance was set at α < 0.05. Results Twenty-seven female soccer players from the Mexican Women’s National Team were included in the analysis. Descriptive characteristics of the sample are presented in Table 1. Table 1.Descriptive characteristics of the sample.VariableMedian (IQR)Age (years)26.2 (22.8–29.3)Stretch stature (cm)167.0 (163.5–171.1)Body mass (kg)62.8 (59.0–67.3)Body fat (%) by DXA21.0 (18.2–23.1)Triceps skinfold (mm)10.2 (9.0–11.8)Subscapular skinfold (mm)8.6 (7.3–9.8)Biceps skinfold (mm)4.2 (3.4–5.2)Iliac crest skinfold (mm)12.4 (10.2–15.1)Supraspinale skinfold (mm)8.4 (5.9–9.5)Abdominal skinfold (mm)12.2 (8.9–13.9)Thigh skinfold (mm)13.6 (11.0–16.5)Calf skinfold (mm)7.0 (5.0–8.0)Waist girth (cm)71.0 (69.3–73.8) Σ8SF (mm)77.4 (69.7–85.8)Note: Values are presented as median and interquartile range (IQR). Descriptive characteristics of the sample. Note: Values are presented as median and interquartile range (IQR). Descriptive characteristics Median age was 26.2 years (IQR: 22.8–29.3), with a median stretch stature of 167.0 cm (163.5–171.1) and body mass of 62.8 kg (59.0–67.3). DXA %BF showed a median of 21.0% (18.2–23.1). Skinfold and girth measurements are summarised in Table 1, with Σ8SF showing a median of 77.4 mm (69.7–85.8). Salivary osmolality (MX3 portable osmometer) showed a median of 82 mOsm/L (IQR 74–90 mOsm/L; range 67–98 mOsm/L); all 27 screened players satisfied the 65–100 mOsm/L hydration criterion and were therefore retained in the analytic sample. Associations between anthropometric variables and DXA %BF Spearman correlation analysis identified several significant associations between anthropometric variables and %BF assessed by DXA (Figure 1). The sum of eight skinfolds (Σ8SF) showed the strongest association with DXA %BF among all variables examined, with a moderate positive correlation (ρ = 0.63, 95% BCa CI [0.35–0.80], p < 0.001). Among individual skinfolds, the calf (ρ = 0.55, 95% BCA CI [0.21–0.79], p = 0.003), subscapular (ρ = 0.52, 95% BCa CI [0.13–0.80], p = 0.006), iliac crest (ρ = 0.49, 95% BCa CI [0.12–0.72], p = 0.010), and supraspinale (ρ = 0.45, 95% BCa CI [0.08–0.73], p = 0.017) presented the most notable correlation coefficients among the variables analysed, while waist girth also demonstrated a moderate positive correlation (ρ = 0.44, 95% BCa CI [0.01–0.75], p = 0.021). According to the predefined thresholds, all observed correlations were within the moderate range. Spearman correlation coefficients (ρ) between anthropometric variables and body fat percentage assessed by DXA (n = 27). Data normality was assessed using the Shapiro–Wilk test; due to non-normal distribution (p < 0.05), non-parametric Spearman rank correlations were performed. All presented correlations are statistically significant (p < 0.05). Derivation of the FMF anthropometric equation Table 2 presents the regression coefficients and diagnostic statistics for the FMF equation predicting %BF. Both Σ8SF and waist girth contributed significantly to the model (p < 0.05). The intercept was not statistically significant (β₀ = −17.0, p = 0.085). Model diagnostics indicated good overall performance, with a standard error of estimate (SEE) of 2.47% BF. Variance inflation factors showed no evidence of multicollinearity (VIF = 1.86), and residual independence was supported by the Durbin–Watson statistic (1.98). The intercept (β₀ = −17.0; 95% CI −36.43 to +2.51; p = 0.085) was retained despite its non-significant p-value because forcing the regression through the origin worsened model fit (RMSE 2.33% vs. 2.48% with the intercept removed; AIC 128.32 vs. 129.73) and because the FMF equation is intended for use within the observed predictor range (Σ8SF 49–143 mm; waist girth 65–80 cm), in which a regression line through the origin would be biologically implausible (a hypothetical individual with zero subcutaneous fat and zero waist girth would not exhibit zero DXA %BF). The non-significance of β₀ is therefore a consequence of the restricted predictor range observed in elite athletes rather than evidence that the intercept can be dropped; uncertainty associated with the intercept is reiterated as a limitation in Section 5. Table 2.Regression coefficients and model diagnostics of the FMF anthropometric equation for estimating %BF.FMF equationANOVASEEVIF (maximum)Durbin–WatsonParameterEstimate SE t-value p-value95% CI Lower95% CI UpperF p %BF = 0.12 × (Σ8SF) +0.39 × (WG) − 17.08.72<0.0012.471.861.98 β₀ (Intercept)−17.09.433−1.7980.085−36.432.51 β₁ (Σ8SF)0.120.03853.1540.0040.0420.201 β₂ (Waist girth)0.390.14042.8090.010.1050.684Note: SE = standard error; CI = confidence interval. %BF = estimated body fat percentage. Σ8SF = sum of eight skinfolds (mm), including triceps, subscapular, biceps, iliac crest, supraspinale, abdominal, thigh, and calf. WG = waist girth (cm). Coefficients were derived using a multiple linear regression model (ordinary least squares) to derive the FMF anthropometric equation. Caution: the 95% CI for the intercept (β₀) spans from −36.43 to +2.51, a range of nearly 39 percentage points. This exceptionally wide interval reflects the limited sample size and restricted predictor range of this elite cohort, and indicates that the equation may produce increasingly uncertain estimates if applied outside the derivation sample’s predictor bounds (Σ8SF 49–143 mm; waist girth 65–80 cm). The model should not be extrapolated beyond these ranges. Regression coefficients and model diagnostics of the FMF anthropometric equation for estimating %BF. Note: SE = standard error; CI = confidence interval. %BF = estimated body fat percentage. Σ8SF = sum of eight skinfolds (mm), including triceps, subscapular, biceps, iliac crest, supraspinale, abdominal, thigh, and calf. WG = waist girth (cm). Coefficients were derived using a multiple linear regression model (ordinary least squares) to derive the FMF anthropometric equation. Caution: the 95% CI for the intercept (β₀) spans from −36.43 to +2.51, a range of nearly 39 percentage points. This exceptionally wide interval reflects the limited sample size and restricted predictor range of this elite cohort, and indicates that the equation may produce increasingly uncertain estimates if applied outside the derivation sample’s predictor bounds (Σ8SF 49–143 mm; waist girth 65–80 cm). The model should not be extrapolated beyond these ranges. Residual diagnostics Residual analyses confirmed that the assumptions of multiple linear regression were satisfied. The Shapiro–Wilk test indicated normally distributed residuals (W = 0.97, p = 0.63). The Breusch–Pagan test showed no evidence of heteroscedasticity (χ² = 1.84, p = 0.39). The Durbin–Watson statistic suggested no meaningful autocorrelation in the residual sequence. Comparative predictive performance and agreement of anthropometric equations Figure 2 illustrates the Bland–Altman plot for the FMF model, while Table 3A summarises its performance in comparison to six previously published anthropometric equations. The FMF model demonstrated the highest agreement with DXA, yielding the highest ICC (0.71) and the lowest mean bias (−0.05% BF), with relatively narrow limits of agreement (−4.61% to 4.70%). Because no internal DXA test–retest reliability assessment was performed, these limits of agreement reflect the combined uncertainty of the reference method and the FMF equation. Linear regression between method differences and DXA values indicated a statistically significant proportional bias (β = −0.46, p < 0.001), although the magnitude was small. The FMF model also showed the strongest association with DXA (%BF), with the highest adjusted coefficient of determination (R²-adjusted = 0.50) and the lowest prediction errors (MAE = 1.86%; RMSE = 2.33%). LOOCV confirmed internal consistency (Q² = 0.50), and no influential observations were detected (maximum Cook’s D = 0.21). In contrast, the six previously published equations demonstrated lower agreement and higher error estimates. All exhibited negative LOOCV Q² values (range: −1.13 to −15.32). The Sloan and Yuhasz equations showed the largest systematic biases (−9.37% and −8.42% BF, respectively) and the highest prediction errors (RMSE = 13.89% and 11.29%). To enable a like-for-like comparison of predictor structures, each historical equation was additionally refitted to the present sample by re-estimating its coefficients via ordinary least squares while preserving its original predictor set (Table 3B, Panel B). After refitting, error metrics decreased substantially for every historical equation (RMSE range 2.55–2.93%; MAE range 1.88–2.38%) and ICC values improved (range 0.44–0.63), confirming that a large fraction of the apparent superiority of the FMF model in Panel A was attributable to calibration. Even after refitting, however, the FMF predictor combination (Σ8SF + waist girth) retained the lowest RMSE (2.33%), the highest ICC (0.71), and the highest cross-validated Q² (0.43), supporting the conclusion that the model’s advantage lies in its predictor structure rather than in calibration on the test sample. The Jackson & Pollock seven-site model could not be refitted because the chest and axilla skinfolds are not part of the ISAK eight-site protocol used here and were therefore unavailable in the present dataset. Relationship between DXA-measured body fat percentage (%BF) and values predicted by the FMF anthropometric equation. (A) Bland–Altman plot illustrating the agreement between methods; the central dashed line represents the mean bias (−0.05%), while the outer dashed lines indicate the 95% limits of agreement (−4.61% to 4.70%). (B) Linear association between DXA-measured %BF and FMF-predicted %BF. R, multiple correlation coefficient; R 2, coefficient of determination; R 2a, R 2adjusted; RMSE, root mean square error; ICC, intraclass correlation coefficient. Table 3A.Predictive performance, agreement, and error metrics of anthropometric equations for estimating body fat percentage (%BF).Performance indicatorFMF ModelDurninJP7JP3FaulknerYuhaszSloanICC0.710.330.180.170.070.040.01Bias (%BF)−0.05−3.12−4.21−3.86−6.11−8.42−9.37Lower LOA (95%)−4.61−9.87−11.3−10.5−14.8−18.2−21.4Upper LOA (95%)4.703.632.873.122.571.342.66 R 0.740.560.610.410.540.630.43 R²0.540.320.370.170.30.40.18 R²-adjusted0.50.290.330.140.260.360.15 LOOCV Q²0.5−1.13−2.23−1.89−5.43−9.78−15.32 MAE (%)1.864.65.75.27.910.612.8 RMSE (%)2.335.026.185.848.7211.2913.89Cook’s D Number of influential cases (D >1)0011223Maximum Cook’s D 0.210.280.340.310.440.480.52Note: ICC = intraclass correlation coefficient; LOA = limits of agreement (Bland–Altman method); R = Pearson correlation coefficient; R² = coefficient of determination; Q² = predictive coefficient of determination; LOOCV = leave-one-out cross-validation; MAE = mean absolute error; RMSE = root mean square error. Cook’s D indicates influential observations. JP3 and JP7 refer to the Jackson and Pollock anthropometric equations based on three- and seven-skinfold models, respectively. %BF = body fat percentage. For previously published equations, predictions were generated using the original fixed formulas without model refitting. Accordingly, R, R², R²-adjusted, and error metrics reflect the association and agreement between predicted %BF and DXA values within the present sample, rather than re-estimated model parameters. Q² values for these equations should be interpreted as relative predictive performance indices, and Cook’s D values represent the influence of observations on the residuals (predicted − DXA) rather than on re-fitted regression models. Predictive performance, agreement, and error metrics of anthropometric equations for estimating body fat percentage (%BF). Note: ICC = intraclass correlation coefficient; LOA = limits of agreement (Bland–Altman method); R = Pearson correlation coefficient; R² = coefficient of determination; Q² = predictive coefficient of determination; LOOCV = leave-one-out cross-validation; MAE = mean absolute error; RMSE = root mean square error. Cook’s D indicates influential observations. JP3 and JP7 refer to the Jackson and Pollock anthropometric equations based on three- and seven-skinfold models, respectively. %BF = body fat percentage. For previously published equations, predictions were generated using the original fixed formulas without model refitting. Accordingly, R, R², R²-adjusted, and error metrics reflect the association and agreement between predicted %BF and DXA values within the present sample, rather than re-estimated model parameters. Q² values for these equations should be interpreted as relative predictive performance indices, and Cook’s D values represent the influence of observations on the residuals (predicted − DXA) rather than on re-fitted regression models. Table 3B.Panel B. Predictive performance and agreement of refitted historical anthropometric equations and the FMF equation against DXA-measured body fat percentage in the present sample.Equation (refitted) R²ICCBias (%BF)Lower LOAUpper LOARMSE (%) Q² (LOOCV)Durnin & Womersley (logΣ4SF + age)0.450.630.00−4.96 +4.962.550.29Sloan (iliac crest + triceps)0.270.440.00−5.70 +5.702.930.13Yuhasz (Σ6SF)0.370.550.00−5.29 +5.292.720.30Faulkner-style (Σ4SF)0.300.470.00−5.63 +5.632.890.17Jackson & Pollock 3-site (Σ3SF + Σ3SF² + age)0.270.440.00−5.69 +5.692.930.00FMF (Σ8SF + waist girth) [reference]0.540.71−0.05−4.61 +4.702.330.43Note: Each historical equation was refitted to the present sample (n = 27) by ordinary least-squares re-estimation of its coefficients while preserving its original predictor structure. R² = coefficient of determination in the present sample; ICC = intraclass correlation coefficient (3,1, absolute agreement) between refitted prediction and DXA; LOA = 95% Bland–Altman limits of agreement; RMSE = in-sample root mean square error; Q² = leave-one-out cross-validated coefficient of determination. By construction, refitted models have bias = 0 in-sample. The Jackson & Pollock seven-site model was not refitted because the chest and axilla skinfolds are not part of the ISAK eight-site protocol used in this study and were therefore unavailable. Panel B. Predictive performance and agreement of refitted historical anthropometric equations and the FMF equation against DXA-measured body fat percentage in the present sample. Note: Each historical equation was refitted to the present sample (n = 27) by ordinary least-squares re-estimation of its coefficients while preserving its original predictor structure. R² = coefficient of determination in the present sample; ICC = intraclass correlation coefficient (3,1, absolute agreement) between refitted prediction and DXA; LOA = 95% Bland–Altman limits of agreement; RMSE = in-sample root mean square error; Q² = leave-one-out cross-validated coefficient of determination. By construction, refitted models have bias = 0 in-sample. The Jackson & Pollock seven-site model was not refitted because the chest and axilla skinfolds are not part of the ISAK eight-site protocol used in this study and were therefore unavailable. Descriptive performance of the FMF model across %BF strata As a descriptive supplement, prediction errors of the FMF model were examined across two collapsed DXA %BF categories (Table 4): below the sample median (≤21%) and at or above the sample median (>21%). Given the small total sample (n = 27) and the resulting subgroup sizes (n = 11–16 per category), this analysis is strictly exploratory. No statistical comparisons were performed and RMSE values should be interpreted with caution given their high sampling variability at these group sizes. The four-category stratification originally considered yielded subgroups of n = 5–8, which is statistically insufficient to support meaningful RMSE computation; the two-category presentation is offered solely for descriptive context. Bias values were +0.4% BF in the lower stratum and −0.3% BF in the higher stratum (n = 11–16), with RMSE values of 1.9% and 2.6% BF respectively. These figures are presented for descriptive orientation only and should not be interpreted as subgroup performance benchmarks (n = 11–16). Given the small overall sample and the attendant statistical fragility of any subgroup metric, no further inference is drawn from Table 4. Table 4.Descriptive performance of the FMF model by DXA %BF stratum (collapsed two-category presentation; descriptive only).DXA %BF range n DXA %BF (mean ± SD)Predicted %BF (mean ± SD)Bias (%) RMSE(%)≤21% (below median)1117.5 ± 1.517.8 ± 1.7 +0.41.9>21% (above median)1623.2 ± 1.922.9 ± 2.1−0.32.6Note: DXA = dual-energy X-ray absorptiometry; %BF = body fat percentage; SD = standard deviation; RMSE = root mean square error. Bias was calculated as the mean difference between predicted %BF and DXA (%BF_predicted − %BF_DXA). The original four-category stratification (subgroups n = 5–8) was collapsed into two categories split at the sample median (21%) to reduce statistical fragility; the resulting subgroup sizes (n = 11 and n = 16) remain modest. All values are purely descriptive; no statistical comparisons were performed. Descriptive performance of the FMF model by DXA %BF stratum (collapsed two-category presentation; descriptive only). Note: DXA = dual-energy X-ray absorptiometry; %BF = body fat percentage; SD = standard deviation; RMSE = root mean square error. Bias was calculated as the mean difference between predicted %BF and DXA (%BF_predicted − %BF_DXA). The original four-category stratification (subgroups n = 5–8) was collapsed into two categories split at the sample median (21%) to reduce statistical fragility; the resulting subgroup sizes (n = 11 and n = 16) remain modest. All values are purely descriptive; no statistical comparisons were performed. Machine-learning benchmarking against the FMF equation Figure 3 summarises the predictive performance of the FMF OLS equation and nine alternative algorithmic classes under 5 × 5 nested cross-validation. With the FMF predictor set (Panel A), the cross-validated RMSE of the OLS reference was 2.67% BF (95% BCa CI 2.11–3.38). Regularised linear alternatives produced marginally lower point estimates (Ridge 2.61; ElasticNet 2.63; Bayesian Ridge 2.63; Lasso 2.69), but all confidence intervals overlapped extensively with the OLS reference (differences of −0.06 to +0.03 percentage points). Non-linear models systematically performed worse than OLS (k-NN 2.76; SVR-RBF 2.84; random forest 3.30; gradient boosting 3.38), and the multi-layer perceptron collapsed (RMSE ≈ 11.5), consistent with the well-known instability of neural networks at small sample sizes. With the full set of 13 candidate predictors (Panel B), OLS overfitted substantially (RMSE 5.29). Regularised linear models partially recovered (Lasso 3.63; Ridge 3.70; Bayesian Ridge 3.47), and the best non-linear model was k-NN (2.98); however, none of these models matched the cross-validated RMSE of the parsimonious two-predictor OLS-FMF model (2.67). Permutation importance for the regularised linear models identified waist girth, biceps and calf skinfolds, and the subscapular skinfold as the dominant predictors, consistent with the FMF predictor structure. Cross-validated predictive performance of the FMF OLS equation and nine alternative machine-learning models under 5 × 5 nested cross-validation. Bars show out-of-fold root mean square error (RMSE) and error bars represent 95% bias-corrected-and-accelerated confidence intervals (1,000 replicates). The dashed vertical line marks the RMSE of the FMF OLS reference. Both panels share an identical x-axis range (0 to 6.0% BF) to enable direct visual comparison of out-of-sample RMSE across predictor sets. Panel A: models trained on the FMF predictor set (Σ8SF + waist girth). Panel B: models trained on all 13 anthropometric predictors. A multi-layer perceptron with one hidden layer of four units was also evaluated but is not displayed because its RMSE (~11.5% with the FMF predictor set and ~16.0% with the 13-predictor set) is more than four times the next-worst model, reflecting the instability of neural networks at small sample sizes (n = 27). Discussion The primary aim of this study was to derive and internally validate an anthropometric equation for estimating body fat percentage (%BF) in elite Mexican female soccer players and to examine its agreement with DXA as the reference method. Among all equations evaluated, the FMF model demonstrated the most favourable alignment with DXA, showing moderate agreement (ICC = 0.71), negligible mean bias (−0.05% BF), and lower prediction error (MAE 1.86%, RMSE 2.33%) compared with the six traditional equations, and these conclusions held after each historical equation was refitted to the present sample (Table 3B, Panel B), confirming that the FMF model’s advantage stems from its predictor structure rather than from in-sample calibration alone. Although these values represent a meaningful improvement over generic models, they must be interpreted within the context of the sample. The restricted %BF range (IQR 4.9 percentage points) mechanically attenuates correlation-based metrics, including ICC, which would likely be higher in a population with broader adiposity variability. The 95% limits of agreement (−4.61% to 4.70% BF) indicate that individual estimates may deviate from DXA by up to approximately 4.7 percentage points in either direction. Limits of agreement of this magnitude span approximately 9.3 percentage points, which is wider than the seasonal %BF changes typically reported in elite female soccer (often ≤2% BF). The FMF equation is therefore most appropriately positioned as a tool for squad profiling and between-group comparisons (e.g. across positional lines, age groups, or training phases) rather than for tracking small within-athlete changes; the latter application requires a method whose precision is smaller than the change of interest, which DXA or a comparable reference technique provides [4]. These findings are consistent with recent work in Latin American female athletes. In Colombian professional women’s soccer, Durnin & Womersley produced the smallest inter-method difference relative to DXA, yet still exhibited wide limits of agreement and substantial individual variability, limiting their diagnostic utility [35]. Similarly, in Medellín high-performance youth athletes, all commonly used equations showed broad limits of agreement despite moderate-to-good ICC values, leading the authors to conclude that such models lack precision for individual assessment [11]. These parallels reinforce the present interpretation: even when mean bias is small, anthropometric equations not developed within the target population tend to show limited individual-level accuracy. Traditional equations consistently showed lower agreement and larger prediction errors. Notably, the Durnin and Womersley [21] equation emerged as the best-performing conventional alternative despite not incorporating lower-limb skinfolds. The FMF equation reduced RMSE by 54% relative to Durnin & Womersley (from 5.02% to 2.33% BF). These findings reinforce the well-established principle that anthropometric equations exhibit sport-, sex-, and morphology-specific behaviour, and that models developed within the target population provide superior validity compared with those derived from heterogeneous or non-athletic samples [6,36,37]. Similar systematic errors have been documented when applying generalised equations to American football [8], handball [9], and male professional soccer players [10,38]. In female handball players, for example, independently developed equations significantly over- or underestimated DXA values, whereas sport-specific models showed markedly superior accuracy [9]. This pattern mirrors the present findings and underscores the need for population-specific approaches. The structure of the FMF equation is also biologically plausible. The inclusion of waist girth and the sum of eight skinfolds reflects the combined influence of regional adiposity and overall subcutaneous fat distribution on DXA %BF in elite female soccer players. This aligns with evidence showing that body composition in women’s soccer is dynamic [39,40] and associated with physical performance [2,3]. From a practical standpoint, improving the precision of field-based estimation is relevant for performance monitoring, nutritional planning, and health surveillance. Nevertheless, the FMF equation should be considered a candidate tool pending external validation. Agreement with DXA and validity of the FMF model The FMF equation demonstrated moderate agreement with DXA, with negligible mean bias and substantially lower error than traditional equations, an advantage that persisted after the historical equations were refitted to the present sample (Table 3B, Panel B). The limits of agreement indicate that while group-level descriptive metrics (mean and median %BF for squads or positional lines) are robust, individual-level precision is constrained to a range that is broader than the seasonal %BF changes typically of interest in elite female soccer. This pattern mirrors findings in Colombian women’s soccer, where even the best-performing classical equation of Durnin & Womersley showed wide limits of agreement and substantial dispersion around the bias line [35]. Likewise, in Medellín youth athletes, all equations exhibited broad limits of agreement despite acceptable ICC values [11], reinforcing that ICC alone cannot determine individual-level accuracy. The FMF model’s reduced RMSE relative to traditional equations underscores the value of population-specific derivation, particularly in elite female athletes whose morphological characteristics differ from those of general or mixed-sex samples. The machine-learning benchmarking analysis (Section 3.7, Figure 3) provides an empirical test of whether the FMF OLS equation can be improved by more flexible algorithmic approaches. The answer is that, in the present sample, it cannot. Under nested cross-validation, no model, including linear-regularised, kernel-based, tree-ensemble, or neural approaches, produced a cross-validated RMSE outside the 95% confidence interval of the FMF reference when restricted to the same two predictors. Likewise, none of the models trained on the full set of 13 candidate predictors matched the parsimonious two-predictor OLS-FMF model. We interpret this convergence as evidence that the FMF equation captures the linear signal available in this dataset and that, in samples of this size, the bias–variance trade-off favours a parsimonious linear model over more flexible alternatives. The exploration of non-linear and machine-learning approaches remains a worthwhile direction for future studies once larger and more heterogeneous samples become available. The analytic protocol implemented here, based on 5 × 5 nested cross-validation with BCa confidence intervals, is offered as a reproducible template for future work. Relevance of anthropometric predictors The predictors retained in the FMF equation, waist girth and Σ8SF, are consistent with known patterns of fat distribution in women. Waist girth reflects central adiposity, which contributes substantially to DXA %BF, while Σ8SF capture global subcutaneous fat. In elite female soccer players, adiposity distribution is influenced by sport-specific demands, hormonal milieu, and training history. Similar findings have been reported in female handball players, where skinfold sums and regional girths were among the strongest predictors of DXA-measured fat mass [9]. The failure of traditional equations to accurately estimate %BF in this population likely reflects their derivation from mixed or non-athletic samples with different fat distribution patterns. The FMF equation’s structure aligns with evidence that anthropometric predictors must reflect the morphological characteristics of the target population to achieve valid estimates. Practical implications for women’s football From an applied perspective, the FMF equation offers a more precise field-based alternative for estimating %BF when DXA is unavailable or impractical. This is particularly relevant in national-team environments, where logistical constraints may limit repeated laboratory assessments [41]. Similar conclusions have been drawn in Colombian women’s soccer, where the need for accessible, low-cost, and sport-specific tools has been emphasised [35]. While the limits of agreement restrict its use for detecting small individual changes, the negligible mean bias supports its utility for group-level monitoring, squad profiling, and longitudinal evaluation of training cycles. Importantly, body composition data should be interpreted within an athlete-centred framework, integrated with performance, health, and contextual indicators rather than used as isolated targets [42,43]. Future research should prioritise external validation across competitive levels and regions and explore whether incorporating additional low-burden variables enhances predictive accuracy without compromising feasibility. As highlighted by Serafini et al. [6], advancing population-specific anthropometric models remains a priority in applied sports science. Limitations and future directions Several limitations must be acknowledged. First, the derivation sample (n = 27) is small for model derivation. Although the achieved sample satisfied the Riley et al. criterion, larger samples are generally recommended to improve coefficient stability and support future external validation [33]. LOOCV supported internal consistency (Q² = 0.50), but cross-validation in small samples offers limited protection against optimistic bias. Second, the restricted %BF range (IQR 18.2–23.1%) likely attenuated correlation-based metrics, meaning agreement may be underestimated relative to more heterogeneous populations. Third, the intercept (β₀ = −17.0) did not reach statistical significance (p = 0.085), introducing uncertainty at predictor-range extremes. Fourth, comparison equations were benchmarked without refitting, which may disadvantage them relative to the FMF model. Fifth, the FMF equation depends on ISAK-standardised anthropometry performed by a certified assessor; performance may decline with less experienced evaluators, and caliper choice may influence measurements [44]. Finally, no internal DXA test–retest reliability assessment was conducted in this study; accordingly, a technician-specific minimum detectable change cannot be derived, and the limits of agreement reported in Section 3.5 cannot be attributed solely to the FMF equation, they reflect the combined uncertainty of both the reference method and the anthropometric model. These limitations echo those reported in Colombian and European studies, where sample size, population specificity, and methodological variability similarly constrained generalisability [9,11,35]. Three additional limitations merit explicit acknowledgement. First, the present sample was not stratified by playing position. Because the physiological and morphological demands of elite female soccer differ substantially between goalkeepers, defenders, midfielders, and forwards, and because position-specific biomarker and performance profiles have been documented in elite female players [45], position-stratified derivation would constitute a desirable next step once larger samples accumulate. Second, while the precision of the Lunar iDXA densitometer is well documented in the literature (in vivo CVs ≈ 0.82–0.86% under repositioning), no test–retest precision study was conducted in the present cohort, and an in-house minimum detectable change (MDC ≈ 2.77 × SDm) was therefore not derived. In its absence, the limits of agreement reported here cannot be partitioned into criterion-method error and FMF-equation error, and small deviations from DXA within the MDC envelope cannot be confidently attributed to the new equation. Third, although the intercept of the FMF model was retained for the structural reasons outlined in Section 3.3, its non-significance (p = 0.085) and wide 95% confidence interval (−36.43 to +2.51) introduce uncertainty at the extremes of the predictor range; the model should accordingly be used within the Σ8SF and waist-girth values represented in this derivation sample until external validation in broader populations is available. Conclusions The anthropometric FMF equation developed in this study demonstrated the highest concordance with DXA among all models evaluated in elite players of the Mexican Women’s National Team, substantially reducing prediction error and eliminating the systematic bias observed in traditional equations. Its predictor structure, Σ8SF and waist girth, captures the adiposity patterns characteristic of this population and provides a practical, population-specific alternative for estimating %BF when DXA access is limited. Nonetheless, the limits of agreement indicate that its primary utility lies in squad profiling and group-level comparisons rather than in detecting small individual changes over time. Overall, these findings highlight the importance of population-derived tools and position the FMF equation as an applicable resource in high-performance women’s soccer environments. Future research should externally validate this model in larger and more diverse cohorts to confirm its generalisability and explore non-linear extensions integrating anthropometric, functional, and contextual variables.