Nonlinear association between body mass index and physical fitness in primary school children: a single-school cross-sectional study using generalized additive models Authors: Junhui Zhu, Mingling Zhao, Meiling Zheng, Ling Wang, Qiuyue Li Published: 2026 ABSTRACT AbstractBackgroundBody mass index (BMI) is associated with physical fitness in children, but the functional form of this relationship is debated. Using routinely collected fitness-monitoring data from a single primary school, this study examined whether BMI is nonlinearly associated with physical fitness in primary school children, treating the analysis as a methodological, single-school demonstration rather than a population-representative survey.MethodsWe analysed records from 1,572 students (808 boys, 764 girls; Grades 1–6) nested within 37 classes in one school in Tianfu New Area, Sichuan, China. BMI was calculated from measured height and weight. Four fitness indicators available for the whole sample—vital capacity, 50-m sprint, sit-and-reach, and 1-min rope skipping—were analysed as the primary outcomes; a composite index (mean of sex- and grade-standardised z-scores, with the 50-m sprint reverse-coded so that higher scores denote better fitness) was retained as a secondary descriptive summary. Quadratic regression and generalized additive models (GAMs; penalised cubic B-splines) adjusted for sex and grade were used to model BMI. Primary overall and sex-stratified models used cluster-robust standard errors at the class level; grade-stratified exploratory models, each based on only a few classes, were assessed with model-based smooth tests. Nonlinearity was quantified with the effective degrees of freedom (EDF) of the BMI smooth and a Wald test, and exploratory subgroup analyses were corrected for multiple comparisons (Benjamini–Hochberg false discovery rate, FDR).ResultsThe class-level intraclass correlation was non-negligible for BMI (0.124) and the 50-m sprint (0.085), confirming the need to account for clustering. After cluster-robust adjustment, BMI showed a statistically supported nonlinear association with the 50-m sprint (EDF = 4.4, P < 0.001), 1-min rope skipping (EDF = 5.1, P < 0.001), and the composite index (EDF = 5.7, P < 0.001), but not with vital capacity (EDF = 1.0, P = 0.44). For the 50-m sprint the best predicted performance was well localised at a BMI of about 16.2 kg/m²; for 1-min rope skipping the optimum was less precisely located, and for the composite index the cluster bootstrap did not identify a stable interior optimum (it spread the maximum across ≈ 15.7–21.2 kg/m², with a median of ≈ 19.7 kg/m² well above the full-sample point estimate). Predicted performance declined at both lower and higher BMI. The quadratic turning point (~ 18.5 kg/m²) was higher than the GAM peak and moved toward the GAM estimate when sparse BMI tails were trimmed, indicating sensitivity of the parabola to extreme values. In exploratory, FDR-corrected subgroup analyses, nonlinearity remained statistically supported in girls and in Grades 1 and 3 but not in boys or other grades; these subgroup findings are hypothesis-generating only.ConclusionsIn this single-school sample, the BMI–fitness association was nonlinear for speed- and coordination-related tasks; for the 50-m sprint the best predicted performance was well localised at an intermediate BMI (about 16.2 kg/m²), whereas for the composite index no stable single optimum could be identified. Because the data come from one school and lack body-composition and physical-activity measures, the estimated optima are sample-specific descriptive features and should not be read as intervention targets. The study illustrates how routine school fitness data and GAMs can describe BMI–fitness patterns beyond categorical comparisons, and motivates multi-school, longitudinal confirmation.Keywords: Children, Body mass index, Physical fitness, Generalized additive models, Nonlinear association, Cluster-robust inference, School health Introduction The physical health of children and adolescents has become a shared concern across educational quality, public-health governance, and the Healthy China agenda [13]. Although national monitoring suggests that the overall fitness profile of Chinese students has improved, persistent problems remain, including excess weight, inadequate endurance, and uneven development across physical-fitness domains [23]. Long-term evidence further indicates that BMI levels among Chinese children and adolescents have continued to rise over recent decades, with overweight and obesity becoming increasingly prevalent in both national and regional samples [8, 10]. At the same time, undernutrition and excess weight now coexist within the same school-aged population, making the health profile of children and adolescents substantially more complex than a simple binary classification of “normal” versus “abnormal” weight [22]. Against this background, school-based health promotion should move beyond body-size categories alone and pay closer attention to the ways in which anthropometric status is linked to functional physical performance. A substantial body of research has shown that BMI is associated with physical fitness, but the form of this association is not consistently linear [15]. Studies in Chinese children and adolescents have repeatedly reported poorer speed, endurance, flexibility, and overall fitness among overweight or obese groups, while some evidence also suggests disadvantage at the lower end of the weight spectrum [7, 19]. More recent analyses from Chinese school populations have further shown that BMI and adiposity may be related to multiple dimensions of fitness, including sprinting, coordination, flexibility, and composite fitness performance [16]. Similar patterns have been observed in regional samples from Xinjiang and Shandong, where physical fitness varied substantially across BMI strata [2, 20]. Research on Tibetan youth and older Chinese adolescents has added further nuance by showing that the BMI–fitness relationship may differ by developmental stage, sex, altitude context, and fitness component [11, 21]. In Chinese high-school students, this association has even been described as parabolic, suggesting that a moderate BMI may be more favorable than either a low or a high BMI for health-related physical fitness [14]. Similar concerns have also been reported internationally. Evidence from Brazilian children and adolescents showed that linear models alone were insufficient to describe the BMI–fitness relationship because meaningful nonlinear variation remained across the BMI continuum [12]. A comparable pattern was reported in Polish adolescents, where several fitness outcomes followed distinctly nonlinear trajectories across BMI levels [9]. More recently, research on primary-school students in ethnic-minority areas of Sichuan again underscored the need to examine heterogeneity across the distribution rather than assuming a single average effect [18]. Taken together, these findings suggest that conventional group comparisons or ordinary linear regression can detect broad between-group differences but may be less suitable for describing potential thresholds, plateaus, or turning points across the full BMI range. From a methodological perspective, the generalized additive model (GAM) provides a useful framework for examining such patterns [6]. By incorporating smooth functions into a regression structure, a GAM can model continuous exposure–response relationships without imposing a rigid parametric form in advance [5, 17]. This flexibility has made GAMs valuable in health research when the effects of continuous predictors are unlikely to be strictly linear [1, 3]. For school-based fitness data, which typically involve continuous anthropometric indicators, developmental gradients, and sex heterogeneity, GAMs offer a practical way to describe potentially nonlinear associations more flexibly than simple grouping or ordinary linear models. Against this background, the present study used routinely collected physical-fitness monitoring data from a single primary school in Tianfu New Area, Sichuan, to address three questions. First, is BMI nonlinearly associated with each of the major individual fitness indicators in primary-school children? Second, is a comparable pattern observed for a composite summary of fitness? Third, does the association vary across sex and grade groups? We emphasise at the outset that, because the data come from one school, this study is intended as a methodological, single-school demonstration of how continuous-exposure modelling can describe BMI–fitness relationships in routine school data, rather than as a population-representative survey or a basis for setting intervention thresholds. By modelling BMI as a continuous exposure rather than relying only on BMI categories, and by explicitly accounting for the clustering of students within classes, the study aimed to provide a more refined and statistically appropriate description of the BMI–fitness relationship in this sample. Materials and methods Data source and study participants This was a school-based cross-sectional study using annual physical-fitness monitoring records from primary-school students in Grades 1–6 in one school in Tianfu New Area, Sichuan, China. The original dataset included grade, class, sex, date of birth, height, weight, vital capacity, 50-m sprint, sit-and-reach, 50 m × 8 shuttle run, 1-min sit-ups, and 1-min rope skipping. After data verification, a total of 1,572 valid records were retained for analysis (808 boys and 764 girls), nested within 37 classes. By grade, the sample comprised 297 students in Grade 1, 323 in Grade 2, 271 in Grade 3, 271 in Grade 4, 245 in Grade 5, and 165 in Grade 6. Because all records originate from a single school, findings are explicitly treated as sample-specific throughout. Data cleaning and variable construction Height, weight, vital capacity, 50-m sprint, sit-and-reach, 1-min sit-ups, and 1-min rope skipping were converted into numeric format. For the 50 m × 8 shuttle run, times originally recorded in minutes-and-seconds notation were converted into seconds. Examination of the core variables did not identify values that clearly fell outside physiologically plausible ranges; height ranged from approximately 118.5 to 167.0 cm, weight from 18.7 to 81.0 kg, and BMI from 12.5 to 29.4 kg/m². These distributions were consistent with the expected range for primary-school children, so no additional exclusion of extreme values was performed. BMI was calculated using the standard formula BMI = weight (kg) / height (m)². Although the raw dataset included dates of birth, no unified date of physical-fitness assessment was available; converting date of birth into exact age under these circumstances could have introduced systematic error. Accordingly, grade fixed effects, rather than exact chronological age, were used to control for developmental stage. Treatment of missing test items The four indicators used as primary outcomes—vital capacity, 50-m sprint, sit-and-reach, and 1-min rope skipping—were recorded for every student (0% missing). For the remaining two items, missingness was entirely structural and determined by grade-specific testing arrangements under the National Student Physical Health Standards: 1-min sit-ups are administered in Grades 3–6 and the 50 m × 8 shuttle run in Grades 5–6. Within the eligible grades, completeness was 100% (sit-ups: 952/952 students in Grades 3–6; shuttle run: 410/410 students in Grades 5–6), and no student in an eligible grade had a missing value for the corresponding item. Because the indicator of “missingness” is therefore deterministic in grade—which is itself adjusted for in every model—there is no residual missing-at-random information to recover, and statistical imputation would add modelling assumptions without addressing any genuine information loss. We therefore did not impute, and analysed sit-ups and the shuttle run only within their eligible grades as supplementary outcomes. As an additional check, Supplementary Table S1 compares students who were and were not eligible for each grade-restricted item on sex, BMI, and the four whole-sample indicators; the groups differed mainly in ways expected from their differing grade (developmental) composition and were balanced on sex, reinforcing that this is structural, grade-driven missingness already adjusted for in every model. The composite index was built only from the four fully observed indicators so that it is defined identically for every student. Primary outcomes and the composite index Following reviewer guidance, the four individual indicators were treated as the primary outcomes, because each captures a distinct physiological capacity—respiratory function (vital capacity), speed (50-m sprint), lower-limb/trunk flexibility (sit-and-reach), and lower-limb power and coordination (1-min rope skipping)—and may relate to BMI through different mechanisms. To provide a single, interpretable secondary summary, a composite index was also constructed. Because absolute performance differs markedly across grade and sex owing to growth and maturation, each of the four indicators was first standardised to a z-score within every sex-by-grade cell, and the composite index was defined as the arithmetic mean of the four z-scores. Aggregation was therefore performed on a common, dimensionless within-cell scale on which the four indicators carry equal weight, rather than on incommensurable raw units. Direction of coding and verification For the 50-m sprint a shorter time denotes better performance, so its z-score was reverse-coded (multiplied by -1) before aggregation; the three remaining indicators, for which higher values denote better performance, retained their original orientation. Consequently, higher composite values denote better overall fitness relative to same-sex, same-grade peers. We verified this orientation empirically: the composite index correlated positively with 1-min rope skipping (r = 0.59), sit-and-reach (r = 0.51) and vital capacity (r = 0.14) and negatively with 50-m sprint time (r = -0.48), exactly as intended, and it correlated strongly and positively with the school-computed total fitness score (r = 0.79). This confirms that the reverse-coding was applied correctly and that the composite is oriented so that larger values represent better fitness. Statistical analysis Continuous variables were summarised as mean ± standard deviation and compared between boys and girls using independent-samples t tests (Welch). To quantify the degree to which students resembled their classmates, we estimated the class-level intraclass correlation coefficient (ICC) for BMI and for each outcome using a one-way random-effects decomposition. Nonlinear modelling For each outcome we fitted two complementary models adjusted for sex and grade. First, a quadratic regression entered BMI as linear and quadratic terms, Y = β0 + β1·BMI + β2·BMI² + β3·Sex + β4·Grade + ε, providing an interpretable parametric summary and a turning point at -β1/(2β2). Second, a generalized additive model represented BMI with a penalised cubic B-spline (P-spline): a cubic B-spline basis defined by 10 equally spaced knots spanning the observed BMI range (giving 13 basis functions), penalised by a second-order difference penalty on adjacent coefficients [4], with sex and grade entered as parametric terms. The smoothing parameter was selected by generalized cross-validation. The flexibility actually used by each smooth is reported as its effective degrees of freedom (EDF); an EDF of 1 indicates an essentially linear fit, whereas larger values indicate increasing curvature. Accounting for class-level clustering Because students were nested within 37 classes, all inferential quantities (regression coefficients, confidence intervals, and tests of nonlinearity) for the primary overall and sex-stratified models were computed with cluster-robust (sandwich) standard errors clustered at the class level. For these GAMs, the statistical significance of nonlinearity was assessed with a cluster-robust Wald χ² test comparing the full smooth with its linear projection, referenced to the dimension of the penalised B-spline basis (13 basis functions). Because this nominal basis dimension exceeds the effective degrees of freedom actually used by each penalised smooth (range 1.0–7.1; Table 3), it provides a deliberately conservative reference distribution for the test; the improvement in fit relative to the linear model was additionally summarised by the change in Akaike information criterion (ΔAIC). To probe the stability of the estimated turning points and peaks, we used a nonparametric cluster bootstrap that resampled whole classes with replacement (300 replicates) and recomputed each estimate, yielding 95% percentile intervals. Peak-performance BMI values were extracted from the fitted curves within the central region of the BMI distribution (5th–95th percentiles) to avoid artefacts driven by sparse tails, and were treated as descriptive, sample-specific features rather than validated thresholds. Table 3Generalized additive model results for BMI and each outcome (penalised cubic B-spline; cluster-robust nonlinearity test)Outcome n EDF s(BMI)Wald χ² (df)ΔAICNonlinearity PPeak BMI50-m sprint (s)15724.3853.5 (13)21.68< 0.00116.171-min rope skipping (reps)15725.0636.0 (13)13.81< 0.00116.17Sit-and-reach (cm)15727.0825.9 (13)10.130.01720.93Vital capacity (mL)15721.0013.2 (13)-0.000.435—Composite index (z)15725.7439.0 (13)23.12< 0.001not localised §EDF = effective degrees of freedom of the BMI smooth (1 ≈ linear). Wald χ² is the cluster-robust test of the smooth term against its linear projection (referenced to the penalised B-spline basis dimension, 13; a conservative reference df relative to the smaller EDF). ΔAIC = AIC(linear) − AIC(GAM); larger favours the spline. Peak BMI is the within-sample location of best predicted performance (5th–95th percentile range) and is shown only where nonlinearity was supported; for the 50-m sprint it denotes the minimum-time BMI. § For the composite index the smooth is nonlinear (P < 0.001) but its maximum is not a stable feature: the cluster bootstrap (Table 5) spreads the optimum across ≈ 15.7–21.2 kg/m² (median ≈ 19.7), so no single peak BMI is reported (see Sect. 3.3) Generalized additive model results for BMI and each outcome (penalised cubic B-spline; cluster-robust nonlinearity test) EDF = effective degrees of freedom of the BMI smooth (1 ≈ linear). Wald χ² is the cluster-robust test of the smooth term against its linear projection (referenced to the penalised B-spline basis dimension, 13; a conservative reference df relative to the smaller EDF). ΔAIC = AIC(linear) − AIC(GAM); larger favours the spline. Peak BMI is the within-sample location of best predicted performance (5th–95th percentile range) and is shown only where nonlinearity was supported; for the 50-m sprint it denotes the minimum-time BMI. § For the composite index the smooth is nonlinear (P < 0.001) but its maximum is not a stable feature: the cluster bootstrap (Table 5) spreads the optimum across ≈ 15.7–21.2 kg/m² (median ≈ 19.7), so no single peak BMI is reported (see Sect. 3.3) Exploratory subgroup analyses and multiplicity Sex- and grade-stratified GAMs were fitted as exploratory analyses. To limit false-positive findings arising from testing eight subgroups, the nonlinearity P values from these models were adjusted for multiple comparisons using the Benjamini–Hochberg false-discovery-rate (FDR) procedure; both raw and FDR-adjusted P values are reported. Because each grade comprises only 4–8 classes, cluster-robust testing is unstable within single grades; sex strata (which span all 37 classes) were therefore tested with cluster-robust statistics, whereas grade strata were tested with a model-based F test for the smooth, and all subgroup results are interpreted only as hypothesis-generating signals. Software and reproducibility Two-sided P < 0.05 defined statistical significance. All analyses were performed in Python 3.12 using NumPy 2.4 and SciPy 1.17; the B-spline basis was constructed with SciPy’s spline routines, and the penalised GAM, cluster-robust covariance, cluster bootstrap, and FDR correction were implemented in a self-contained scripted pipeline. The key modelling specifications, including the spline basis, number of knots, penalty structure, smoothing-parameter selection method, clustering correction, bootstrap procedure, and multiplicity correction, are reported above to support reproducibility. As a robustness check, the penalised-spline GAMs were independently re-fitted with a separate re-implementation that used a different B-spline basis construction and an independent grid search for the smoothing parameter. This reproduced the substantive findings—an essentially linear association for vital capacity (EDF ≈ 1), a well-localised mid-range optimum near 16 kg/m² for the 50-m sprint, and a nonlinearly shaped but not reliably localised optimum for the composite index—confirming that the conclusions are not artefacts of a single analysis pipeline. Results Sample characteristics and class-level clustering Table 1 presents the basic characteristics of the sample. Boys had a higher BMI than girls (17.53 vs. 16.73 kg/m², P < 0.001) and greater vital capacity, whereas girls performed better in sit-and-reach and 1-min rope skipping. Median BMI increased across grades, with a more pronounced rise in Grades 5–6 (Fig. 1). Table 1Basic characteristics of the study participants (mean ± SD)VariableTotal (n = 1572)Boys (n = 808)Girls (n = 764)P valueBMI (kg/m²)17.14 ± 2.6417.53 ± 2.9216.73 ± 2.24< 0.001Height (cm)133.85 ± 11.87133.97 ± 11.33133.71 ± 12.420.666Weight (kg)31.35 ± 9.3132.11 ± 9.6630.55 ± 8.86< 0.001Vital capacity (mL)1951.24 ± 523.352137.58 ± 522.371754.16 ± 446.85< 0.00150-m sprint (s)10.52 ± 1.1710.42 ± 1.1710.64 ± 1.16< 0.001Sit-and-reach (cm)9.27 ± 5.606.85 ± 4.8711.82 ± 5.16< 0.0011-min rope skipping (reps)122.68 ± 31.57120.42 ± 32.13125.07 ± 30.800.003Independent-samples (Welch) t tests compared boys and girls. Grade distribution: Grade 1, 297; Grade 2, 323; Grade 3, 271; Grade 4, 271; Grade 5, 245; Grade 6, 165 Basic characteristics of the study participants (mean ± SD) Independent-samples (Welch) t tests compared boys and girls. Grade distribution: Grade 1, 297; Grade 2, 323; Grade 3, 271; Grade 4, 271; Grade 5, 245; Grade 6, 165 Distribution of BMI by sex and grade. Boxes show the interquartile range and median; whiskers show the non-outlier range Because students were nested within 37 classes, we quantified class-level clustering before inference. The intraclass correlation was non-negligible for BMI (ICC = 0.124) and for the 50-m sprint (ICC = 0.085 for the standardized score), moderate for sit-and-reach (0.070), and small for rope skipping (0.010), vital capacity (0.003) and the composite index (0.017). Although several ICCs are modest, the clustering of the exposure (BMI) is substantial; ignoring it would understate uncertainty. Therefore, the primary overall and sex-stratified models used class-level cluster-robust standard errors; grade-stratified models were treated as exploratory and assessed with model-based smooth tests because each grade contained only a small number of classes. BMI and the individual fitness indicators (primary analyses) After adjustment for sex and grade and with cluster-robust standard errors, the quadratic models supported curvature for each of the three performance-based individual indicators (Table 2). The 50-m sprint showed a U-shaped association (positive quadratic term, P = 0.002), whereas 1-min rope skipping and sit-and-reach showed inverted-U patterns (negative quadratic terms, P = 0.004 and P = 0.001). The quadratic term for vital capacity was far from significant (P = 0.947), indicating an essentially monotonic, near-linear association in this sample. Importantly, the curvature for the performance indicators persisted after accounting for clustering, although the cluster-robust P values were larger than their model-based counterparts, confirming that ignoring clustering would have overstated precision. Table 2Quadratic regression of BMI on the individual fitness indicators and the composite index, adjusted for sex and grade (cluster-robust 95% CIs and P values)OutcomeBMI β (95% CI) P BMI² β (95% CI) P Turning pointAdj R²50-m sprint (s) †-0.240 (-0.418, -0.061)0.0080.0077 (0.0028, 0.0126)0.00215.600.4471-min rope skipping (reps)8.538 (2.193, 14.883)0.008-0.2476 (-0.4178, -0.0773)0.00417.240.233Sit-and-reach (cm)1.658 (0.622, 2.694)0.002-0.0444 (-0.0711, -0.0177)0.00118.670.264Vital capacity (mL)11.787 (-25.430, 49.004)0.535-0.0338 (-1.0226, 0.9550)0.947—0.891Composite index (z) ‡0.222 (0.088, 0.356)0.001-0.0060 (-0.0096, -0.0025)< 0.00118.450.007† The 50-m sprint is modelled on completion time, so a positive BMI² with a U-shape means time is shortest (performance best) near the turning point ‡ Composite index shown last as a secondary summary. Turning point shown only where the quadratic term was significant Quadratic regression of BMI on the individual fitness indicators and the composite index, adjusted for sex and grade (cluster-robust 95% CIs and P values) † The 50-m sprint is modelled on completion time, so a positive BMI² with a U-shape means time is shortest (performance best) near the turning point ‡ Composite index shown last as a secondary summary. Turning point shown only where the quadratic term was significant Generalized additive models gave a consistent but more flexible picture (Table 3; Fig. 2). The effective degrees of freedom (EDF) quantify the curvature actually supported by the data: the 50-m sprint (EDF = 4.4), 1-min rope skipping (EDF = 5.1) and the composite index (EDF = 5.7) all showed clearly nonlinear smooths (cluster-robust P ≤ 0.001), whereas vital capacity collapsed to an essentially straight line (EDF = 1.0, P = 0.435). Sit-and-reach showed weaker and more irregular curvature (EDF = 7.1, P = 0.017) and is interpreted cautiously. For the 50-m sprint, predicted performance was best at a well-localised mid-range BMI of about 16.2 kg/m² and declined toward both extremes; for 1-min rope skipping the smooth was similarly nonlinear, but the location of its optimum was less precisely estimated (Sect. 3.4, Table 5). The 95% confidence band widened markedly at the sparse upper tail of the BMI distribution (Fig. 2), so curve behaviour beyond roughly the 95th percentile should not be over-interpreted. GAM smooth curves for BMI and the four individual fitness indicators, adjusted for sex and grade. A 50-m sprint; B 1-min rope skipping; C sit-and-reach; D vital capacity. Shaded bands are cluster-robust 95% CIs; grey margins mark the sparse BMI tails (<5th / >95th percentile), where CIs widen and curves are least reliable. For the 50-m sprint, the y-axis is completion time (lower = better) Table 5Class-level cluster-bootstrap intervals (300 replicates) for the turning point and the GAM peakOutcomeQuadratic TP: median (95% CI)GAM peak: median (95% CI)Composite index18.50 (17.00, 19.41)19.68 (15.71, 21.16)50-m sprint15.68 (12.61, 16.98)16.26 (15.76, 16.89)1-min rope skipping17.23 (13.46, 18.39)16.74 (15.14, 20.87)Wide intervals (e.g. the composite index) indicate an optimum that is not reliably localised; narrow intervals (e.g. the 50-m sprint) indicate a well-identified optimum Class-level cluster-bootstrap intervals (300 replicates) for the turning point and the GAM peak Wide intervals (e.g. the composite index) indicate an optimum that is not reliably localised; narrow intervals (e.g. the 50-m sprint) indicate a well-identified optimum Composite index (secondary summary) As a single secondary summary, the composite index reproduced the qualitative pattern seen for the performance indicators (Tables 2 and 3; Fig. 3A): a significant quadratic term (P = < 0.001) and a nonlinear GAM smooth (EDF = 5.7, ΔAIC = 23.1, cluster-robust P < 0.001), with lower predicted relative fitness at both the low and high ends of the BMI range. Unlike the well-localised optimum of the 50-m sprint, however, the composite curve did not have a reliably identified interior optimum. Although its full-sample maximum fell near a BMI of 16.26 kg/m², the cluster bootstrap spread the location across about 15.7 to 21.2 kg/m² (Sect. 3.4, Table 5) and yielded a bootstrap median (≈ 19.7 kg/m²) far above the full-sample point estimate, with many resampled maxima lying near the upper boundary of the central BMI range; the single-point maximum is therefore not a stable feature of the curve. We accordingly do not report a composite peak BMI (Table 3) and describe the composite only as nonlinear in shape—with lower predicted relative fitness at both the low and high ends of the BMI range, but no dependable interior optimum. As expected from its within-cell standardisation, BMI explained only a small share of variance in the composite (adjusted R²=0.007); the association is thus statistically detectable but modest in magnitude, and is best read as a description of curve shape rather than of explanatory importance. Composite fitness index. A Overall GAM smooth for the composite index (nonlinear in shape, but with no reliably localised interior optimum: the cluster bootstrap spread the maximum across ≈ 15.7–21.2 kg/m², median ≈ 19.7 kg/m², Table 5); B sex-stratified smooths showing supported nonlinearity in girls but not boys. Shaded bands are 95% CIs; grey margins mark sparse BMI tails Sensitivity of turning points and peaks The quadratic turning point for the composite index (18.45 kg/m²) was higher than its full-sample GAM maximum (16.26 kg/m²). This difference is expected: a parabola is symmetric and is pulled toward the sparse high-BMI tail, whereas the spline can follow a flatter optimum. Two analyses confirm this interpretation. First, trimming the BMI tails moved the quadratic turning point steadily toward the GAM estimate, from 18.45 kg/m² (full range) to 17.20 kg/m² when the most extreme 10% of BMI values were removed (Table 4), demonstrating the parabola’s sensitivity to extreme observations. Second, a class-level cluster bootstrap showed that the location of the composite maximum is itself imprecisely estimated (95% interval 15.7–21.2 kg/m²), with many resampled maxima lying near the upper boundary of the central BMI range (bootstrap median ≈ 19.7 kg/m²), so the composite optimum is not reliably localised, whereas the 50-m sprint peak was well identified (15.8–16.9 kg/m²) (Table 5). Together these results indicate that only the 50-m sprint optimum is well identified; for the composite index the location of best fitness is not reliably estimated and should not be read as a precise value or threshold. Table 4Sensitivity of the composite-index quadratic turning point to trimming of the BMI tailsBMI range used n BMI interval (kg/m²)Quadratic turning pointFull range157212.5–29.418.45Trim 2.5/97.5%149213.9–24.018.13Trim 5/95%141414.2–22.717.89Trim 10/90%125714.5–20.617.20As progressively more extreme BMI values are removed, the quadratic turning point moves toward the composite full-sample GAM maximum (~ 16.3 kg/m²), illustrating the parabola’s sensitivity to sparse tails Sensitivity of the composite-index quadratic turning point to trimming of the BMI tails As progressively more extreme BMI values are removed, the quadratic turning point moves toward the composite full-sample GAM maximum (~ 16.3 kg/m²), illustrating the parabola’s sensitivity to sparse tails Exploratory sex- and grade-stratified analyses In exploratory analyses corrected for multiple comparisons (Table 6; Fig. 3B), nonlinearity remained statistically supported in girls (FDR P = 0.006) but not in boys (FDR P = 0.534), whose smooth was essentially linear (EDF = 1.4). Among grades, only Grade 1 (FDR P < 0.001) and Grade 3 (FDR P < 0.001) retained significant curvature after FDR adjustment; the apparent curves in other grades did not survive correction. Because each grade contains few classes and some small-stratum smooths are unstable (e.g. the high EDF in Grade 3), these subgroup results are presented only as hypothesis-generating signals and the corresponding peak BMI values are not interpreted as grade- or sex-specific thresholds. Table 6Exploratory sex- and grade-stratified GAMs for the composite index, with Benjamini–Hochberg FDR correctionStratum n EDFΔAICRaw PFDR PPeak BMIBoys8081.430.050.4670.534—Girls7644.1617.970.0020.00616.66Grade 12973.5713.82< 0.001< 0.00116.27Grade 23232.911.530.0660.119—Grade 32717.4417.92< 0.001< 0.00120.14Grade 42714.930.270.0890.119—Grade 52452.231.220.0780.119—Grade 61651.00-0.001.0001.000—Sex strata were tested with cluster-robust statistics (they span all 37 classes); grade strata, each containing only 4–8 classes, were tested with a model-based F test. FDR correction was applied across all eight subgroup tests. Peak BMI shown only where curvature survived FDR correction Exploratory sex- and grade-stratified GAMs for the composite index, with Benjamini–Hochberg FDR correction Sex strata were tested with cluster-robust statistics (they span all 37 classes); grade strata, each containing only 4–8 classes, were tested with a model-based F test. FDR correction was applied across all eight subgroup tests. Peak BMI shown only where curvature survived FDR correction Supplementary grade-restricted outcomes For the two grade-restricted items, higher BMI was associated with fewer sit-ups and longer shuttle-run times, but neither showed statistically supported nonlinearity (1-min sit-ups, Grades 3–6, n = 952: EDF = 1.14, P = 0.507; 50 m × 8 shuttle run, Grades 5–6, n = 410: EDF = 2.29, P = 0.123). These outcomes are therefore consistent with an approximately linear disadvantage at higher BMI rather than a clear curvilinear pattern, and are reported only as supplementary given their restricted grade coverage. Discussion In this single-school sample, BMI was nonlinearly associated with speed- and coordination-related fitness: predicted performance was best at an intermediate BMI that was well localised near 16.2 kg/m² for the 50-m sprint, less precisely located for 1-min rope skipping, and not reliably localised for the composite index (see Table 5) and declined at both lower and higher BMI, whereas vital capacity was essentially linear in BMI. The pattern was robust to class-level clustering and is broadly consistent with previous reports of nonlinear or non-monotonic BMI–fitness relationships in children and adolescents. Because the analysis was cross-sectional and confined to one school, these are sample-specific associations, not causal effects or generalisable thresholds. Why fitness may be lower at both low and high BMI The U- and inverted-U shapes are biologically plausible through distinct pathways at the two ends of the distribution. At the lower end, a low BMI in this age range often reflects limited fat-free (muscle) mass and smaller energy and strength reserves; because power-, speed- and jump-based tasks depend on absolute force production, children with very low body mass may generate less propulsive force and perform less well. At the upper end, a high BMI more often reflects greater fat mass, which adds non-propulsive load that must be accelerated and supported; in weight-bearing tasks such as sprinting and repeated jumping this mechanical load reduces movement economy and relative power-to-mass, lowering performance. That the 50-m sprint and rope skipping—both weight-bearing and power-dependent—showed the clearest curvature and the lowest peak BMI, while vital capacity (a body-size-scaled respiratory volume) was monotonic, is consistent with a mechanical-load interpretation. Crucially, however, BMI cannot distinguish fat mass from fat-free mass, so these mechanisms remain inferential; body composition was not measured here and the proposed pathways should be regarded as hypotheses for studies with direct adiposity and muscle-mass measures. Exploratory sex and grade differences Evidence for nonlinearity was statistically supported in girls but not in boys after multiplicity correction. This may partly reflect composition of the index—flexibility and coordination components may be more stable in girls at this age—or greater heterogeneity in boys’ activity patterns and maturation that dilutes the BMI signal. The grade-restricted findings (curvature surviving correction only in Grades 1 and 3) are similarly tentative: each grade contains few classes, several small-stratum smooths were unstable, and the corresponding peak locations varied widely. We therefore treat all subgroup results as hypothesis-generating only and do not propose sex- or grade-specific BMI targets. Value of modelling BMI as a continuous exposure Categorising BMI into underweight, normal, overweight and obese groups before comparing fitness is practical but can obscure within-category gradients and impose artificial cut-points. Modelling BMI continuously with a penalised spline described the shape of the association directly and showed where a linear assumption would have been too restrictive (clearly for the performance tasks, but not for vital capacity). At the same time, the spline analyses themselves show why the extracted peaks must be read cautiously: the composite-index curve does not yield a reliably localised interior optimum, and confidence bands widen substantially in the sparse BMI tails. The contribution of this approach is thus descriptive refinement, not the generation of automatic intervention rules. Implications and the importance of unmeasured factors Routine school fitness data, analysed with appropriate models, can describe BMI–fitness patterns beyond categorical comparisons and may help schools identify domains (here, speed and coordination) where body-mass load is most relevant. These descriptive signals should not, however, be translated into BMI targets. Physical fitness in children is shaped by many factors not captured by BMI or available in these records—body composition (fat versus fat-free mass), habitual physical activity and sedentary time, diet and energy availability, sleep, pubertal stage, and family socioeconomic context—any of which could confound or modify the BMI–fitness association. The small variance explained by BMI underscores that it is only one modest correlate of fitness. Accordingly, the mid-range BMI associated with best performance here should be interpreted as a feature of this sample under adjustment for sex and grade only, and not as an optimal or recommended value. Limitations Several limitations should be emphasised. First, the data come from a single school in one district of Sichuan, China; the sample is not population-representative and the specific BMI values reported may not generalise to other regions, settings, or ethnic groups. Second, the design is cross-sectional, so no causal or temporal inference is possible. Third, important determinants and potential confounders—body composition, habitual physical activity, sedentary behaviour, diet, sleep, pubertal/maturational status, and socioeconomic factors—were unavailable, so the mechanisms underlying the associations could not be tested and residual confounding cannot be excluded. Fourth, because no unified assessment date was recorded, exact age could not be computed and grade was used as a proxy for developmental stage. Fifth, although we addressed class-level clustering with cluster-robust inference and a cluster bootstrap, we could not fit a full random-effects GAMM, and unmeasured school- or class-level factors may remain. Finally, the sex- and grade-stratified analyses are exploratory; despite FDR correction and conservative interpretation, they require confirmation in larger, multi-school samples. Conclusion Using routine fitness data from 1,572 primary-school students in a single school, and after explicitly accounting for class-level clustering, we observed that BMI was nonlinearly associated with speed- and coordination-related physical fitness. For the 50-m sprint the best predicted performance occurred around a relatively well-defined BMI of about 16.2 kg/m²; for 1-min rope skipping the optimum was less precisely located, and for the composite index no stable interior optimum could be identified. For all of these outcomes, predicted performance was nonetheless lower at both low and high BMI. Vital capacity, by contrast, was essentially linear in BMI. In exploratory, FDR-corrected subgroup analyses, the nonlinear pattern was statistically supported in girls and in Grades 1 and 3 but not elsewhere; these subgroup results are hypothesis-generating only. These findings suggest that routine school fitness data, analysed with continuous-exposure models, can describe BMI–fitness patterns more faithfully than categorical comparisons alone. Given the single-school, cross-sectional design and the absence of body-composition and physical-activity measures, the results should be read as sample-specific, descriptive associations—consistent with the view that both low and high BMI may accompany poorer fitness—rather than as causal evidence or as BMI thresholds for intervention. Confirmation in larger, multi-school, longitudinal studies with direct measures of body composition and activity is needed before any practical recommendation is made. Acknowledgements The authors sincerely thank the participating students and their parents/guardians for their cooperation in this study. We also acknowledge the support of the school staff in the organization of the physical fitness assessments and data collection. Abbreviations Akaike information criterion Body mass index Confidence interval Effective degrees of freedom False discovery rate Generalized additive model Intraclass correlation coefficient Standard deviation Authors’ contributions J.Zhu and M.Zhao contributed substantially to the conception and design of the study, data interpretation, methodological revision, and manuscript preparation. M.Zhao and M.Zheng collected and verified the data. M.Zhao performed the initial statistical analysis, and J.Zhu contributed to the re-analysis, methodological refinement, and revision of the response to reviewers. J.Zhu and M.Zhao drafted and revised the manuscript. L.W. contributed to data interpretation and critical revision of the manuscript. Q.L. supervised the study. All authors read and approved the final manuscript. Funding The authors declare that no specific funding was received for this research. Data availability The datasets used and/or analyzed during the current study are not publicly available because they contain identifiable or potentially sensitive information from minors and are subject to ethical and institutional restrictions. De-identified data may be available from the corresponding author on reasonable request and with permission from the relevant institution and ethics committee. Declarations Ethics approval and consent to participate This study involving human participants was reviewed and approved by the Ethics Committee of Yibin University (Approval No. YBULLXS26041601). All procedures were performed in accordance with the Declaration of Helsinki and relevant guidelines and regulations. Written informed consent was obtained from the participants’ legal guardians, and assent was obtained from the children where appropriate. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets used and/or analyzed during the current study are not publicly available because they contain identifiable or potentially sensitive information from minors and are subject to ethical and institutional restrictions. De-identified data may be available from the corresponding author on reasonable request and with permission from the relevant institution and ethics committee.