Effects of dietary supplements on lower-limb power and jump performance in physically active individuals and athletes: a systematic review and Bayesian network meta-analysis

Abstract

Objective

To evaluate the relative effects of dietary supplements, the certainty of evidence, and the robustness of findings within separate networks for peak power (PP), relative peak power (RPP), vertical jump (VJ), and countermovement jump (CMJ).

Methods

PubMed, Web of Science, Embase, and the Cochrane Library were systematically searched, with the final search conducted on September 8, 2026. Eligible studies were randomised parallel-group and crossover controlled trials published in English through August 31, 2026, involving healthy individuals with physical activity or training backgrounds. Four separate Bayesian random-effects networks were constructed. Effects were assessed using mean differences (MDs) in change scores, 95% credible intervals (CrIs), and prediction intervals. Risk of bias and certainty of evidence were assessed using RoB 2 and CINeMA, respectively. Subgroup analyses, network meta-regression, and sensitivity analyses were also conducted. This review was retrospectively registered in PROSPERO (CRD420261283835).

Results

A total of 119 studies were included. The PP, RPP, VJ, and CMJ networks comprised 47, 33, 31, and 44 studies, respectively, with some studies contributing to more than one outcome. Compared with placebo, β-hydroxy-β-methylbutyrate (HMB) yielded MDs of 64.70 W (95% CrI: 19.15 to 110.50) for PP and 0.84 W·kg⁻¹ (95% CrI: 0.42 to 1.28) for RPP. Favourable effects were also observed for β-alanine on PP and for creatine and vitamin D on RPP. Iron supplementation yielded an MD of 6.81 cm (95% CrI: 0.84 to 12.83) for VJ, supported by only one direct study. For CMJ, betaine yielded an MD of 3.72 cm (95% CrI: 1.24 to 6.18); the 95% CrIs for palmitoylethanolamide, caffeine combined with Rhodiola rosea, creatine, multi-ingredient supplements, and caffeine also excluded zero. The HMB findings for RPP remained relatively stable across multiple sensitivity analyses, whereas some other favourable comparisons had prediction intervals that included zero or showed attenuated findings in sensitivity analyses. Exploratory associations with training level and sport were observed for RPP, although interpretation was limited by sparse evidence and imbalanced covariate distributions. Of 1,636 pairwise comparisons, 17 were rated as moderate, 55 as low, and 1,564 as very low certainty. The HMB-versus-placebo comparisons for both PP and RPP were rated as moderate certainty.

Conclusions

Some supplements showed favourable effects within their respective networks, with the HMB findings for RPP remaining relatively stable. However, direct evidence was limited, and the certainty of evidence was very low for most comparisons. Findings should be interpreted in the context of the populations and supplementation regimens studied. Differences between networks cannot establish that outcome type modifies supplementation effects, and rankings alone should not be used to guide individualised supplementation regimens.

1. Introduction

Lower-limb power and jump performance are commonly used indicators of physical performance and training adaptations and are closely related to movements such as jumping, acceleration, and rapid changes of direction. Dietary supplements are used to support improvements in these outcomes, but their effects depend on the ingredients, supplementation regimens, and contexts of use. The International Olympic Committee consensus statement recognises evidence supporting certain supplements for specific aspects of exercise performance, while emphasising that their use should be considered in relation to the sport, nutritional status, and training demands [1]. Evaluating supplementation effects therefore requires clearly defined populations, exercise tasks, and outcomes, together with consideration of effect magnitude and uncertainty.

Previous systematic reviews and network meta-analyses have compared dietary supplements from several perspectives. Abreu et al. included 18 randomised placebo-controlled trials involving 307 high-level soccer players and qualitatively synthesised the effects of supplementation on jumping, sprinting, power, and strength performance [2]. Subsequent network meta-analyses extended the scope of these comparisons. Luo et al. evaluated the effects of different supplements on power and sport-specific performance in elite combat-sport athletes [3]. Deng et al. examined supplementation combined with physical conditioning in relation to strength, jumping, sprinting, and muscle mass, but pooled different jump tests into a single outcome [4]. Du et al. analysed vertical jump, peak power, and mean power separately in volleyball players [5]. Another network meta-analysis in soccer players covered outcomes including running distance, sprinting, jumping, and ratings of perceived exertion [6].

These studies provide a basis for comparing supplements, although their populations, training contexts, and approaches to outcome classification differ. Reviews restricted to a particular sport provide a more specific context for application, whereas reviews of supplementation combined with physical conditioning primarily reflect effects under concomitant training conditions. Further evidence synthesis requires clarification of the relative effects of different supplements on individual power and jump outcomes, as well as the extent to which these estimates depend on population characteristics, intervention regimens, and analytical assumptions.

The comparability of participants and intervention conditions also affects the interpretation of network meta-analysis. The participant-classification framework proposed by McKay et al. emphasises distinctions based on training and competitive level; physically active individuals and high-level athletes should not be regarded as a homogeneous population [7]. In network meta-analysis, the validity of indirect comparisons depends on the transitivity assumption, which requires reasonable comparability in the distribution of potential effect modifiers across direct comparisons. Imbalances in these factors may bias indirect estimates [8]. For dietary supplements, this assessment should consider dose, intervention duration, and concomitant training, in addition to sex, sport, and training level. A shared supplement name does not imply that all regimens have equivalent conditions of use, and estimates obtained by pooling regimens within a node still require interpretation in relation to the contributing studies.

Against this background, this study constructed separate Bayesian random-effects networks for PP, RPP, VJ, and CMJ in healthy individuals with different physical activity and training backgrounds. The primary objective was to estimate the effects of dietary supplements relative to placebo and other interventions within each outcome network and to assess the reliability of these comparisons using credible intervals, prediction intervals, and CINeMA assessments of certainty of evidence [9]. The secondary objectives were to explore associations of sex, sport, training level, intervention duration, and concomitant training with differences in effects through subgroup analyses and network meta-regression, where permitted by the available data and model identifiability, and to assess robustness through sensitivity analyses addressing jump-test type, supplement dose, correlation assumptions, and risk of bias. Conclusions were restricted to comparisons within each outcome network. Differences in effects or rankings between networks were considered descriptive and were not used to infer interactions between supplements and exercise tasks.

2. Methods

2.1. Protocol, reporting guidelines, and registration

This systematic review and Bayesian network meta-analysis was reported in accordance with the PRISMA 2020 statement [10] and the PRISMA extension for network meta-analyses (PRISMA-NMA) [11]. The completed PRISMA-NMA checklist is provided in Supplementary Material 1.

The review protocol was initially developed in June 2025. Title and abstract screening began on July 1, 2025, followed by full-text screening in September 2025. Risk-of-bias assessment and data extraction were conducted in November and December 2025, respectively. The protocol was submitted to PROSPERO on January 12, 2026; the first statistical analyses began in March 2026, and registration was completed on June 28, 2026 (CRD420261283835). Because literature screening and data extraction preceded submission for registration, this review was retrospectively registered. Changes documented in the available manuscript versions and the methods ultimately implemented are summarised in Supplementary Material 2, together with their rationale. The available records do not establish whether each change was prespecified in the original registered protocol.

2.2. Information sources

PubMed, the Web of Science Core Collection, Embase, and the Cochrane Library were systematically searched. The initial search was conducted in July 2025, with a supplementary search in September 2025. All four databases were searched again in full on September 8, 2026, covering publications through August 31, 2026. Eligibility was restricted to peer-reviewed original studies published in English.

2.3. Search strategy

The search strategy was informed by previous relevant reviews and organised around three concepts: dietary supplements, study populations, and target outcomes. Terms were expanded to include synonyms, singular and plural forms, and spelling variants. Terms within each concept were combined using the Boolean operator “OR,” and the concepts were combined using “AND.” The PubMed search strategy was as follows:(((athlet*[tiab] OR sport*[tiab] OR player*[tiab] OR exercis*[tiab] OR trained[tiab] OR training[tiab] OR fitness[tiab] OR “physically active”[tiab] OR “physical activity”[tiab] OR “physical activities”[tiab] OR “recreationally active”[tiab] OR “recreational athletes”[tiab] OR “recreational athlete”[tiab] OR “recreational exercisers”[tiab] OR “recreational exerciser”[tiab] OR “recreationally trained”[tiab] OR “resistance trained”[tiab] OR “strength trained”[tiab] OR “endurance trained”[tiab] OR “aerobically trained”[tiab] OR “well trained”[tiab] OR “highly trained”[tiab] OR elite[tiab] OR collegiate[tiab] OR “college students”[tiab] OR “college student”[tiab] OR “university students”[tiab] OR “university student”[tiab] OR “healthy volunteers”[tiab] OR “healthy volunteer”[tiab] OR “healthy adults”[tiab] OR “healthy adult”[tiab] OR “healthy men”[tiab] OR “healthy women”[tiab] OR “healthy males”[tiab] OR “healthy females”[tiab] OR “young adults”[tiab] OR soccer[tiab] OR football[tiab] OR basketball[tiab] OR volleyball[tiab] OR handball[tiab] OR rugby[tiab] OR hockey[tiab] OR tennis[tiab] OR badminton[tiab] OR swimm*[tiab] OR cyclist*[tiab] OR cycling[tiab] OR runner*[tiab] OR running[tiab] OR sprinter*[tiab] OR rower*[tiab] OR rowing[tiab] OR wrestl*[tiab] OR boxer*[tiab] OR boxing[tiab] OR judo[tiab] OR judoka*[tiab] OR karate[tiab] OR taekwondo[tiab] OR gymnast*[tiab] OR weightlift*[tiab] OR powerlift*[tiab] OR bodybuild*[tiab] OR CrossFit[tiab] OR skater*[tiab] OR skating[tiab] OR skiing[tiab] OR skier*[tiab])) AND ((”dietary supplement*”[tiab] OR creatine[tiab] OR “beta-alanine”[tiab] OR carnosine[tiab] OR “beta-hydroxy-beta-methylbutyrate”[tiab] OR arginine[tiab] OR “arginine alpha-ketoglutarate”[tiab] OR citrulline[tiab] OR glutamine[tiab] OR tyrosine[tiab] OR taurine[tiab] OR carnitine[tiab] OR “acetyl-L-carnitine”[tiab] OR “branched-chain amino acid*”[tiab] OR leucine[tiab] OR isoleucine[tiab] OR valine[tiab] OR caffeine[tiab] OR “dietary nitrate*”[tiab] OR beetroot[tiab] OR betaine[tiab] OR “sodium bicarbonate”[tiab] OR “sodium citrate”[tiab] OR protein[tiab] OR “whey protein”[tiab] OR casein[tiab] OR “phosphatidic acid”[tiab] OR phosphatidylserine[tiab] OR “omega-3 fatty acid*”[tiab] OR “vitamin C”[tiab] OR “vitamin D”[tiab] OR “vitamin E”[tiab] OR magnesium[tiab] OR zinc[tiab] OR selenium[tiab] OR iron[tiab] OR antioxidant*[tiab] OR “alpha-lipoic acid”[tiab] OR pyruvate[tiab] OR “gamma-oryzanol”[tiab] OR “Rhodiola rosea”[tiab] OR “Withania somnifera”[tiab] OR spirulina[tiab] OR probiotic*[tiab] OR “Bacillus subtilis”[tiab] OR sialyllactose[tiab] OR ginseng[tiab] OR Cordyceps[tiab] OR curcumin[tiab] OR quercetin[tiab] OR resveratrol[tiab] OR polyphenol*[tiab] OR anthocyanin*[tiab] OR “tart cherry”[tiab] OR blackcurrant[tiab] OR pomegranate[tiab] OR carbohydrate*[tiab] OR maltodextrin[tiab] OR “ketone ester*”[tiab] OR “ketone salt*”[tiab]))) AND ((Wingate[tiab] OR WAnT[tiab] OR “peak power”[tiab] OR “relative peak power”[tiab] OR “absolute peak power”[tiab] OR “maximal power”[tiab] OR “maximum power”[tiab] OR “anaerobic power”[tiab] OR “anaerobic capacity”[tiab] OR “anaerobic performance”[tiab] OR “sprint cycling”[tiab] OR “cycle sprint”[tiab] OR “cycling sprint”[tiab] OR “power output”[tiab] OR “power performance”[tiab] OR “muscle power”[tiab] OR “muscular power”[tiab] OR “lower limb power”[tiab] OR “lower extremity power”[tiab] OR “leg power”[tiab] OR “explosive power”[tiab] OR “explosive strength”[tiab] OR “explosive performance”[tiab] OR explosiveness[tiab] OR jump*[tiab] OR “vertical jump”[tiab] OR “vertical jumping”[tiab] OR “jump height”[tiab] OR countermovement[tiab] OR “counter movement”[tiab] OR CMJ[tiab] OR CMJs[tiab] OR “squat jump”[tiab] OR “squat jumps”[tiab] OR SJ[tiab] OR Sargent[tiab] OR Abalakov[tiab] OR “athletic performance”[tiab] OR “sports performance”[tiab] OR “sport performance”[tiab] OR “physical performance”[tiab] OR “neuromuscular performance”[tiab] OR “muscle strength”[tiab] OR “muscular strength”[tiab]))

The complete search strategies as executed in each database, including field specifications and restrictions, are provided in Supplementary Material 3.

2.4. Eligibility criteria

Eligibility criteria were defined according to the PICOS framework, with reference to the Cochrane Handbook for Systematic Reviews of Interventions [12].

Population (P): Healthy individuals with a background in physical activity or exercise training were eligible. The target population encompassed Tiers 1–5 of the McKay participant-classification framework: recreationally active individuals, trained/developmental participants, highly trained/national-level athletes, elite/international-level athletes, and world-class athletes [7]. Two reviewers independently classified training level using the training frequency, years of training, competitive standard, level of competition, and sports participation reported in the original articles. When the population met the eligibility criteria but the information was insufficient to assign a specific training level, training level was recorded as “unknown,” and its influence was examined in the relevant sensitivity analyses.

Intervention (I): Eligible interventions were single-ingredient or multi-ingredient dietary supplements intended to improve lower-limb power or jump performance. Supplement identification and classification were informed by the International Olympic Committee consensus statement on dietary supplements [1] and the Australian Institute of Sport (AIS) Sports Supplement Framework [13].

Comparator (C): Eligible comparators included placebo, a no-supplement control, and other dietary supplements meeting the intervention definition used in this review. Placebo (PLA) and no-supplement control (BI) were modelled as separate network nodes to preserve the distinction between these control conditions. Trials directly comparing eligible supplements were also included when they met the remaining eligibility criteria.

Outcomes (O): Studies were required to report at least one of peak power (PP), relative peak power (RPP), vertical jump height (VJ), or countermovement jump height (CMJ).

Study design (S): Randomised parallel-group controlled trials and randomised crossover controlled trials were eligible.

Animal studies, reviews, conference abstracts, trial registration records, nonrandomized studies, and studies involving populations or interventions that did not meet these criteria were excluded. Information from duplicate publications or companion reports of the same study was reconciled to avoid counting the same study more than once.

2.5. Outcome definitions and measurement classification

Peak power (PP) was defined as the highest absolute power attained during a lower-limb cycle Wingate anaerobic test and was expressed in W. Relative peak power (RPP) was defined as PP normalised to body mass and was expressed in W·kg⁻¹ [14]. The source test for data included in the PP and RPP networks was verified study by study, and eligibility was restricted to the corresponding measures obtained from cycle Wingate tests.

Jump height was standardised to cm. The VJ and CMJ networks followed the outcome labels retained in the analytical dataset. VJ was an operational grouping rather than a single standardised movement protocol and included squat jumps, Sargent and Abalakov tests, and records described as CMJ in the original reports. Test names and movement procedures were subsequently checked, and restricted sensitivity analyses were conducted within the existing VJ network according to the verified test classification. These analyses did not reassign observations between networks. Records with confirmed jump height but unclear movement procedures were retained as having an unspecified protocol and examined in restricted sensitivity analyses.

Each individual test result contributed to only one network, although the movement procedures represented in the two networks were not entirely distinct. When a study reported different tests separately, the corresponding results could contribute to their respective networks. Arm-swing technique, measurement equipment, and the method used to summarise test results were recorded separately to inform the interpretation of differences between testing protocols [15,16].

2.6. Record management and study selection

All retrieved records were imported into EndNote. Duplicates were initially identified automatically, followed by manual review of records with incomplete bibliographic information or minor discrepancies using the DOI, PMID, title, first author, and publication year. Two reviewers (RCS and MXS) independently screened titles and abstracts against the prespecified PICOS criteria and assessed the full texts of potentially eligible reports. Disagreements were first resolved through discussion; when consensus could not be reached, a third reviewer (XLX) adjudicated. Multiple reports from the same trial were linked and treated as a single study to avoid double counting. One principal reason for exclusion was recorded for each report excluded after full-text assessment. The numbers of records screened and the reasons for exclusion at each stage are presented in the PRISMA flow diagram.

2.7. Data extraction

Two reviewers (RCS and MXS) independently extracted data using a standardised worksheet that had been developed and pilot-tested in advance. Disagreements were resolved by checking the original reports and through discussion, with adjudication by a third reviewer (XLX) when necessary. Extracted variables included author, publication year, country, study design, sample size, age, sex, sport, training history, McKay training tier, supplement ingredients and doses, intervention duration, concomitant training, and the means and standard deviations (SDs) of the relevant outcomes for each intervention arm.

Concomitant training was defined as general or sport-specific physical conditioning organised and delivered by a coach or strength and conditioning professional, with explicit objectives, a planned structure, a defined training period, and supervision. When a study reported multiple assessment time points for the same outcome, only data from the longest follow-up were retained in the primary analysis to avoid treating repeated measurements from the same participants as independent observations. When data were missing or reported in a format unsuitable for direct quantitative synthesis, the corresponding author was contacted for the required information. If suitable data remained unavailable, only the relevant outcome from that study was excluded from quantitative synthesis. Data extraction and conversion followed the Cochrane Handbook for Systematic Reviews of Interventions [12]. Variable definitions and coding rules are provided in Supplementary Material 4.

2.8. Conversion and synthesis of continuous outcomes

The primary analysis used mean differences (MDs) in change scores on the original measurement scales as the effect measure. Conversion of continuous outcomes and pooling of independent groups followed the Cochrane Handbook for Systematic Reviews of Interventions [12]. The mean change in each trial arm was calculated using Equation (1):

$${\overline{\text{X}}}_{\text{change}}={\overline{\text{X}}}_{\text{post}}-{\overline{\text{X}}}_{\text{pre}}$$ (1)

here, pre and post denote measurements before and after the intervention, respectively. When the original study directly reported the SD of the change score, that value was preferred. When it was not reported but pre- and post-intervention SDs were available, the change-score SD was estimated using Equation (2):

$${\text{SD}}_{\text{change}}=\sqrt{\left({\text{SD}}_{\text{pre}}^2+{\text{SD}}_{\text{post}}^2-2r{\text{SD}}_{\text{pre}}{\text{SD}}_{\text{post}}\right)}$$ (2)

here, r denotes the correlation between pre- and post-intervention measurements in the same participants. When this correlation was unavailable, the primary analysis assumed $r\text{=}0.5$ , and sensitivity analyses using alternative values assessed the influence of this assumption.

When only the standard error (SE) of a trial-arm mean was reported, it was converted to an SD using Equation (3), where n is the sample size of the corresponding trial arm:

$$\text{SD}=\text{SE}\times \sqrt{\text{n}}$$ (3)

When only the lower and upper bounds of a 95% confidence interval for a mean were reported, the SE was estimated using Equation (4), and the SD was then calculated using Equation (3):

$$\mathit{SE}=\frac{U-L}{2c_{0.975}}$$ (4)

here, L and U are the lower and upper bounds of the 95% confidence interval, respectively, and $c_{0.975}$ is the critical value of the relevant distribution. The method used to construct the interval in the original study was followed where available. When the sample size was known and a t distribution was used or the distribution was unspecified, the t critical value with n−1 degrees of freedom was used. A value of 1.96 was used when a normal-based method was explicitly reported or the conditions for a large-sample approximation were met. These conversions were restricted to SEs or confidence intervals for the mean of a single trial arm.

The MD in change scores between the intervention and comparator groups was calculated using Equation (5):

$$\text{MD}={\overline{\text{X}}}_{\text{change},\text{I}}-{\overline{\text{X}}}_{\text{change},\text{C}}$$ (5)

here, I and C denote the intervention and comparator groups, respectively. Positive values indicate a change favoring the intervention, whereas negative values favour the comparator. PP and RPP were retained in W and W·kg⁻¹, respectively, and VJ and CMJ were expressed in cm. None of the four outcomes was converted to a standardised mean difference.

For parallel-group trials with multiple independent dose groups assigned to the same supplement node, change-score data were combined within the study for the primary analysis. The pooled mean was calculated using Equation (6), where N is the sum of the sample sizes across the independent dose groups:

$${\overline{\text{X}}}_{\text{combined}}=\left[{\text{Σ}}_{\text{i}=1}^{\text{m}}{\text{n}}_{\text{i}}{\overline{\text{X}}}_{\text{i}}\right]/\text{N}$$ (6)

The pooled SD was calculated using Equation (7). Here, m is the number of dose groups combined; $n_i$ , ${\overline{X}}_i$ , and ${\mathit{SD}}_i$ are the sample size, mean change, and SD of the change score in group i, respectively; and ${\overline{X}}_{\mathit{combined}}$ is the pooled mean. The total sum of squares includes both within-group variability and between-group variability arising from differences among group means:

$${\text{SS}}_{\text{total}}={\text{Σ}}_{\text{i}=1}^{\text{m}}({\text{n}}_{\text{i}}-1){\text{SD}}_{\text{i}}^2+{\text{Σ}}_{\text{i}=1}^{\text{m}}{\text{n}}_{\text{i}}{\left({\overline{\text{X}}}_{\text{i}}-{\overline{\text{X}}}_{\text{combined}}\right)}^2$$

$${\text{SD}}_{\text{combined}}=\sqrt{[{\text{SS}}_{\text{total}}/(\text{N}-1)]}$$ (7)

Crossover trials were treated as repeated measurements in the same participants. When different doses or administration conditions were combined, condition-specific mean changes were averaged with equal weights, and the pooled variance was calculated from the condition-specific variances and between-condition covariances. The sample size reflected the actual number of participants. The primary analysis assumed $q\text{=}0.5$ for the correlation between change scores under different intervention conditions, with alternative values examined in sensitivity analyses. This parameter was treated separately from the pre–post correlation r in Equation (2).

For multi-arm trials comparing different interventions, all distinct intervention arms were retained and estimated jointly in the network model. Within-study covariance structures accounted for correlations arising from shared comparators and crossover designs.

2.9. Handling of crossover trials

Randomised crossover trials were treated as repeated measurements in the same participants under different intervention conditions, with between-condition correlations retained in the sampling covariance matrix [12]. For change scores under the intervention and comparator conditions, the SD of the paired difference was expressed using Equation (8):

$${\text{SD}}_{\text{paired}}=\sqrt{{\text{SD}}_{\text{Δ},I}^2+{\text{SD}}_{\text{Δ},C}^2-2q{\text{SD}}_{\text{Δ},I}{\text{SD}}_{\text{Δ},C}}$$ (8)

here, ${\text{SD}}_{\text{Δ},I}$ and ${\text{SD}}_{\text{Δ},C}$ are the SDs of change scores under the two conditions, and q is the correlation between those change scores in the same participants. The SE of the paired difference was calculated using Equation (9), where n is the number of participants who completed the corresponding paired measurements and were included in the analysis:

$${\text{SE}}_{\text{paired}}=\frac{{\text{SD}}_{\text{paired}}}{\sqrt{n}}$$ (9)

The primary model estimated effects jointly using the variances and covariances of the conditions; Equations (8) and (9) describe the sampling uncertainty of the corresponding paired contrast. When usable correlation information was unavailable, q was set to 0.5 and examined in sensitivity analyses separately from the pre–post correlation r. Participants contributing to multiple conditions were not counted repeatedly. Period effects, washout periods, and carryover effects were considered in the risk-of-bias assessment using information from the original reports. A correlation assumption does not itself remove carryover effects.

2.10. Definition of network nodes

Separate evidence networks were constructed for PP, RPP, VJ, and CMJ. Single-ingredient supplements were classified into nodes by ingredient. Regimens containing two or more active ingredients were modelled as complete combination nodes, without estimating separate component effects or assuming that the combination effect euqalled the sum of its component effects. Different doses of the same ingredient were assigned to a single node in the primary analysis, with within-study data combined according to study design. Differences in dose and administration were examined in sensitivity analyses. Heterogeneity among formulations classified as multi-ingredient supplements (MIPS) was further examined by excluding this node or splitting it into study-specific nodes.

Placebo (PLA) and no-supplement control (BI) were retained as separate nodes, with PLA as the primary reference. All distinct intervention arms in multi-arm trials were retained, including direct comparisons between supplements. Node abbreviations, ingredients, doses, and contributing studies are provided in the Supplementary Materials. Nodes supported by a single study were retained in the estimation, with evidence sparsity considered when interpreting effects and assessing certainty of evidence.

2.11. Risk of bias assessment

RCS and MXS used the Cochrane RoB 2 tool to assess outcome-specific results for the effect of assignment to intervention (intention-to-treat, ITT) [17]. The corresponding official tools were used for parallel-group and randomised crossover trials, covering all 119 included studies. Study and outcome records were linked using consistent identifiers. Disagreements were resolved through discussion, with adjudication by XLX when necessary.

The assessment covered five domains: the randomisation process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result. Period and carryover effects were also considered for crossover trials. Domain-level and overall judgements were categorised as low risk, some concerns, or high risk. Initial recommendations were generated using the tool’s decision algorithms, followed by final judgements based on the original reports, with any changes and their reasons recorded. Overall risk was not determined by simply counting domains with concerns. Analytical assumptions, such as imputed correlations, were examined through sensitivity analyses and did not, by themselves, lead to a high-risk judgement.

2.12. Bayesian network meta-analysis

All statistical analyses were conducted in R 4.6.1 under Windows 11. Each outcome was analysed using a Bayesian random-effects consistency model with a normal likelihood and an identity link, with the change-score MD on the original measurement scale as the effect measure [18]. Models included study-specific baselines and relative treatment effects. Within-study covariance matrices accounted for sampling correlations arising from shared comparators and crossover designs. A common between-study heterogeneity parameter was assumed within each outcome network. For random contrasts relative to a common comparator within the same multi-arm study, the variance was ${\tau }^2$ and the covariance was ${\tau }^2$ /2.

Prior distributions were specified using the outcome-specific reference scale S. The values of S were fixed at 149.707578 W for PP, 1.55 W·kg⁻¹ for RPP, 6.393971 cm for VJ, and 5.1 cm for CMJ. The same scale was retained across models for a given outcome. Study baselines were assigned normal priors with a mean of 0 and an SD of 10S, and relative treatment effects were assigned normal priors with a mean of 0 and an SD of 2S. The between-study SD $\tau$ was assigned a half-normal prior with a scale of 0.5S. Fixed-effect consistency models were used only for model-fit comparisons and did not replace the primary random-effects models on the basis of these comparisons. Model fit was evaluated using conditional residual deviance, posterior predictive checks, and the deviance information criterion (DIC) [19]. DIC was compared only for the same outcome, dataset, and computational definition.

The primary models were fitted in JAGS through rjags 4-17 using four independent chains. Each chain underwent 3,000 adaptation iterations and 6,000 warm-up iterations, followed by 12,000 retained draws with a thinning interval of 1. Sampling quality was assessed using trace plots, the rank-normalised convergence diagnostic $\hat{R}$ , bulk and tail effective sample sizes (ESS), and Monte Carlo standard errors (MCSE) [20]. The criteria were $\hat{R}\text{<}1.01$ , both bulk and tail ESS ≥ 400, and an MCSE-to-posterior-SD ratio ≤ 0.05. When these criteria were not met, sampling was extended or an equivalent posterior computation method was used.

For some regression and supplementary analyses, the same Gaussian likelihood, priors, and covariance structure were used, with conditionally Gaussian parameters integrated out analytically, followed by numerical integration over $\tau$ and conditional posterior sampling. Integration results were checked by grid refinement and comparison with JAGS results. The computation method and diagnostics for each model are retained in the Supplementary Materials. Initial sampling runs that did not converge were not used for reporting results. Primary effects, primary heterogeneity summaries, and final evidence-profile intervals were taken from the MCMC analyses; paired supplementary model comparisons retained their numerical-integration estimates.

Treatment effects were reported as posterior means with 95% credible intervals (CrIs), together with the posterior probability of an effect favoring the intervention. The primary random-effects models also provided 95% prediction intervals (PIs) to describe the distribution of potential true effects in future exchangeable studies. All nodes were compared pairwise, and ranking probabilities, cumulative ranking curves, and the surface under the cumulative ranking curve (SUCRA) were reported [21]. Rankings were interpreted alongside effect magnitude, interval width, and certainty of evidence. Differences in effects or rankings across outcome networks were considered descriptive and were not used to infer interactions between supplements and exercise tasks.

2.13. Heterogeneity, inconsistency, and transitivity

Between-study heterogeneity was primarily described using the posterior distributions and 95% CrIs of $\tau$ , ${\tau }^2$ , and the standardised quantity $\frac{\tau }{S}$ . A descriptive network-level $I^2$ was calculated using a typical sampling variance. Supplementary frequentist analyses reported the decomposition of Q, $I^2$ , and random-effects results for direct comparisons estimated by restricted maximum likelihood. These measures were interpreted in relation to the outcome scale, study precision, and prediction intervals; network homogeneity was not judged using a single $I^2$ threshold.

Global inconsistency was assessed by comparing residual deviance and DIC between the random-effects consistency model and a design-specific random-effects model, supplemented by a design-by-treatment interaction test accounting for within-study covariance [22]. Total Q was decomposed into within-design and between-design components to distinguish heterogeneity within the same comparison design from inconsistency across designs. Local assessment used Bayesian node splitting to compare direct evidence with independent indirect evidence [23]. The inconsistency factor was defined using Equation (10):

$$\omega =d_{\mathrm{direct}}-d_{\mathrm{indirect}}$$ (10)

To construct indirect evidence, all studies containing both treatments in the comparison were removed. Node splitting was performed only when an identifiable indirect pathway remained in the reduced network. The direct and indirect components shared a common heterogeneity parameter, and the 95% CrI and two-sided posterior tail probability for $\omega$ were reported. Triangular and chordless quadrilateral loops were also identified, accounting for correlations between contrasts from multi-arm studies. Loop inconsistency factors, 95% confidence intervals, and tests adjusted using the Holm procedure within each outcome were reported. Loops formed entirely by a single multi-arm study were not treated as independent inconsistency tests. Non-estimability was not interpreted as consistency, and a lack of detected statistical differences did not establish complete agreement between direct and indirect evidence.

Transitivity was assessed by examining the distribution of potential effect modifiers across direct comparisons [8], including sex, sport, training level, duration, concomitant training, ingredient combinations, dose, comparator type, and jump-test protocol. Observed imbalances were further examined using subgroup analyses, meta-regression, and sensitivity analyses. When data were insufficient, assessment remained descriptive, and the implications for individual comparisons were considered in the CINeMA indirectness domain. Statistical consistency tests did not replace assessment of transitivity.

2.14. Subgroup analyses and network meta-regression

Subgroup analyses and univariable network meta-regression were conducted separately within each of the four outcome networks for five factors: sex, sport, McKay training level, intervention duration, and concomitant training. Results were not interpreted as independent associations mutually adjusted for all five factors. Duration was categorised as ≤7, 8–28, 29–56, and >56 days, and a continuous-duration model was also fitted. Studies with unknown sex, training level, sport, or concomitant training status were excluded only from the corresponding covariate analysis. They were neither assigned to a known category nor removed from the primary network on that basis.

Categorical variables were indicator-coded. The preferred reference categories were male participants, training Tier 1, power athletes, no concomitant training, and >56 days. When the preferred category was unavailable, a category supported by the data was used and documented. Continuous duration was first transformed using ${\log }_2$ and then centred at the median duration of independent studies contributing to the relevant treatment class, as shown in Equation (11). Here, $T_i$ denotes intervention duration in days for study i, and ${\tilde{T}}_c$ is the median duration for class c:

$$z_i={\log }_2(T_i),x_{i,c}=z_i-{\log }_2\left({\tilde{T}}_c\right)$$ (11)

Each treatment node retained its own baseline relative effect. Within a model, supplement nodes shared treatment-class interaction coefficients, BI had separate class interaction coefficients when identifiable, and the PLA interaction term was fixed at 0. The model comparing treatment b with treatment a can be written as Equation (12):

$$y_{i,\mathit{ab}}=(d_b-d_a)+(h_{i,b}-h_{i,a})+u_{i,\mathit{ab}}+{\epsilon }_{i,\mathit{ab}}$$

$$h_{i,t}={\beta }_{g(t)}^T{x}_{i,g(t)}$$ (12)

here, $d_t$ is the baseline effect of treatment t relative to PLA; g(t) denotes the treatment class; $h_{i,t}$ is the covariate interaction term for that class; and $u_{i,\mathit{ab}}$ and ${\epsilon }_{i,\mathit{ab}}$ are the study random effect and sampling error, respectively. For categorical variables, $\beta$ is a coefficient vector; for continuous duration, $\beta$ represents the change in MD per doubling of intervention duration. Interaction coefficients were assigned normal priors with a mean of 0 and an SD of 2S. Different population strata within a study had separate baselines but shared study-level treatment random effects. The common supplement interaction term cancels in the direct difference between two supplements. This specification does not establish separately for each ingredient that the same effect-modification relationship applies. The common class-level interaction was used as a parsimonious specification because the available data provided limited support for treatment-specific interactions.

Before model fitting, the design-matrix rank and covariate support were checked. Coefficients subject to complete confounding or insufficient information were classified as non-estimable. Subgroup forest plots primarily presented conditional effects for the relevant categories from the same interaction model, distinguishing direct evidence within a subgroup, indirect pathways within a subgroup, and extrapolation relying on assumptions across subgroups. When an interaction was not identifiable, only effects supported by observations were displayed; estimates were not supplied for unobserved categories. Interaction coefficients, 95% CrIs, and residual heterogeneity were reported, with comparisons against a model without interactions fitted to the same covariate dataset. Shaded regions in continuous-regression plots represented 95% CrIs for mean relative effects. Study-level regression analyses were not adjusted for multiple comparisons and were interpreted as exploratory. Differences in statistical significance within subgroups or visual trends were not treated as evidence of between-subgroup differences, and no individual-level causal inference was made [24].

2.15. Small-study effects and sensitivity analyses

Comparison-adjusted funnel plots were used to assess small-study effects [25]. The primary plots were centred on the random-effects pooled estimate for each direct comparison under common heterogeneity, with supplementary plots centred on network estimates. Supplement-versus-control contrasts were oriented so that positive values favoured the supplement. Comparisons between supplements were oriented according to a fixed ordering of treatment names, without using observed rankings to determine direction. All available direct contrasts were retained in the plots, with 90%, 95%, and 99% contours.

Egger-type generalised least squares regression included comparison-specific fixed intercepts and an SE term. The covariance matrix incorporated both sampling correlations and between-study heterogeneity, with inference based on study-clustered CR1 robust standard errors [26]. Formal tests were restricted to supplement-versus-control comparisons supported by at least two independent studies per direct comparison. The model was required to include at least 10 independent studies, a full-rank design matrix, and at least three residual degrees of freedom. Tests across the four outcomes within the same analysis series were adjusted using the Holm procedure. Additional analyses were restricted to PLA-controlled trials, parallel-group trials, or studies meeting risk-of-bias criteria, retained only the most precise contrast per study, or sequentially omitted individual studies. When too few studies were available, assessment was descriptive. Funnel-plot asymmetry and test results were used to assess small-study effects and were not taken alone as proof of publication bias, the direction of missing evidence, or the absence of bias.

All sensitivity analyses re-estimated the random-effects networks on the original measurement scales. First, the pre–post correlation r and the between-condition crossover correlation q each took values of 0.1, 0.3, 0.5, 0.7, and 0.9, forming 5 × 5 combinations that were compared with the primary-analysis setting of 0.5/0.5. Only variances dependent on the relevant assumptions were recalculated; change-score SDs directly reported in the original studies were unchanged. Further analyses were restricted to randomised parallel-group trials, excluded high-risk results, retained only low-risk results, or retained only results rated as low risk by both the tool recommendation and the assessor’s judgement. The influence of change-score SD sources and conversion methods was also examined.

Sensitivity analyses of study composition and node handling included sequential omission by original study ID, simultaneously removing all strata and related records from that study [27]; restriction to studies with sample sizes at or above the median study sample size for the relevant outcome; separate examination of 1-day, >1-day, and >28-day regimens and studies involving participants at training Tiers ≥2 or ≥3; restriction to comparisons involving concomitant training or matched training conditions; removal of BI arms, removal of studies containing BI, or merging of BI with PLA; removal of MIPS or combination-ingredient nodes; splitting of MIPS into study-specific nodes; and separation of major ingredients into single-day and multiple-day regimen nodes where supported by the data. Restricted analyses estimated only the network component connected to PLA, with any lost nodes and comparisons recorded.

To examine VJ measurement heterogeneity, analyses based on verifiable test names and movement protocols retained specific jump types, excluded unknown or specified types, or substituted another reported test from the same study. Dose analyses were conducted within individual supplements and restricted to records with verifiable ingredients, numerical doses, units, and regimens. Results were compared after dose restrictions, splitting into dose-specific nodes, and sequential omission of regimens. No common continuous dose relationship was constructed across different ingredients or non-convertible units. Where administration methods could be distinguished, data from each method separately replaced the pooled estimate in the analysis, while preserving correlations among measurements from the same participants. Administration categories were not inferred when they could not be distinguished.

Prior sensitivity analyses changed the half-normal scale for $\tau$ from 0.5S to 0.25S and 1S and also used a half-t prior with four degrees of freedom and a scale of 0.5S. The SD of the normal prior for treatment effects was additionally set to S and 4S. Scenarios were compared in terms of effect direction, 95% CrIs, posterior probabilities favoring the intervention, and heterogeneity. Where rankings were compared, ranks and SUCRA were recalculated within the set of nodes shared by the primary analysis and the corresponding scenario. Missing or non-estimable comparisons were not treated as zero effects. Robustness was judged from substantive changes in effect magnitude and uncertainty, with each type of analysis presented graphically, rather than solely from whether a statistical significance threshold was crossed.

2.16. Certainty of network evidence

All pairwise comparisons were assessed using the CINeMA framework, with emphasis on comparisons against PLA [9]. The official R package cinemar and the CINeMA website were used to cross-check the assessment implementation and website exports. Final effect estimates, 95% CrIs, and 95% PIs were synchronised with the primary MCMC analyses. Imprecision and heterogeneity were reassessed using these intervals; the other domain judgements and the stated overall downgrading procedure were retained. The six domains were within-study bias, reporting bias, indirectness, imprecision, heterogeneity, and incoherence. Overall certainty was classified as high, moderate, low, or very low.

The contribution matrix retained variance information from multi-arm and crossover designs. Within-study bias was assessed using the contribution-weighted average of the confirmed RoB 2 ratings. Indirectness was assigned the rating accounting for the largest share of contributions, with ties resolved in favour of the higher level of concern. Reporting bias was assessed using available registration tracking, measured outcomes that could not be synthesised, and small-study effects. Imprecision and heterogeneity were evaluated using a distribution-based reference range of ±0.2S, with S retained at the fixed outcome-specific scale. This range was not a validated minimal clinically important difference. Website rules were used to assess the position of the CrI relative to the effect regions and the additional uncertainty in the PI beyond the CrI. Incoherence was assessed using the website’s Separate Indirect from Direct Evidence (SIDE) and global design-by-treatment interaction rules, with these findings reported separately from Bayesian node-splitting results.

Some concerns and major concerns were taken as starting points for downgrading by one and two levels, respectively. For imprecision and heterogeneity, only concerns arising from distinct sources were combined; the resulting downgrade was then compared with that for incoherence, and the larger was retained. Independent concerns relating to within-study bias, reporting bias, and indirectness were counted separately, with a maximum overall downgrade of three levels. This approach to combining concerns represented assessor judgement rather than the website’s automatic scoring rule. The rationale for each domain, combined downgrading decisions, and final ratings were retained for each comparison and interpreted alongside the effects, intervals, and network structure.

3. Results

3.1. Study selection and characteristics of included studies

The updated search was completed on September 8, 2026, covering literature published before September 2026. The four databases yielded 36,338 records: 8,605 from PubMed, 11,514 from Embase, 11,027 from the Web of Science Core Collection, and 5,192 from the Cochrane Library. After removing 19,115 duplicate records within the updated search and 1,537 records overlapping with previous searches, 15,686 new unique records remained. Following screening, 609 reports underwent full-text assessment; 570 were excluded or linked to existing studies, leaving 39 additional independent studies. Together with the 80 previously included studies, the review included 119 studies. The selection process and report counts are presented in the flow diagram.

The included studies were published between 1992 and 2026 and comprised 106 randomised parallel-group trials and 13 randomised crossover trials. Intervention duration ranged from 1 to 210 days. Studies included participants of different sexes, sports, and training levels: 75 enroled only males, 17 only females, 17 mixed-sex populations, and three reported separate male and female strata; seven did not clearly report participant sex. Information was insufficient to assign a specific McKay tier in 28 studies. Their original descriptions of training background were retained, and training tier was treated as missing only in the corresponding tier-based analyses. Study characteristics, intervention ingredients, and specific doses are presented in the table of included study characteristics in Supplementary Material 4.

All 119 studies contributed to at least one outcome network [28–146]. After retaining the relevant assessment time points, 388 study-arm–outcome records remained. Following within-node pooling according to study design, the primary models used 356 records. The PP, RPP, VJ, and CMJ networks included 47, 33, 31, and 44 studies, with corresponding outcome-specific sample sizes of 1,464, 939, 1,002, and 1,226 participants, respectively. Some studies reported multiple outcomes; therefore, these study counts and sample sizes overlap and were not summed across networks Figure 1.

Figure 1.

PRISMA study selection flow diagram.

3.2. Risk of bias

RoB 2 assessments covered all 119 studies and comprised 156 result-level records, defined by outcome and comparisons requiring separate judgements. Of these, 140 records came from parallel-group trials and 16 from crossover trials. Overall risk was rated as low for 70 records (44.9%), some concerns for 79 (50.6%), and high for seven (4.5%). Among the 48 PP assessment records, 23 were rated as low risk and 25 as some concerns. The corresponding counts for low risk, some concerns, and high risk were 15, 15, and three among the 33 RPP records; 14, 16, and one among the 31 VJ records; and 18, 23, and three among the 44 CMJ records. Assessment records and independent studies represent different counting units.

Concerns were concentrated in the selection of the reported result and the randomisation process, with 71 and 44 records rated as some concerns in these domains, respectively. High-risk judgements primarily involved missing outcome data and deviations from intended interventions, affecting six and two records, respectively; some records had concerns in more than one domain. Domain-level judgements, supporting information from the original reports, and complete risk-of-bias plots for parallel-group and crossover trials are provided in Supplementary Material 5.

3.3. Evidence networks and model diagnostics

The PP, RPP, VJ, and CMJ networks included 38, 22, 27, and 27 nodes and 61, 28, 48, and 44 direct comparisons, respectively. The 47 PP studies contributed 49 analytical units, whereas the number of analytical units equalled the number of studies in the other three networks. Stratified units were not counted as additional independent studies. All four networks were connected to PLA; however, direct evidence was unevenly distributed, with 49/61, 22/28, 44/48, and 35/44 direct comparisons supported by only one study, respectively.

All four random-effects models met the prespecified sampling diagnostic criteria. The maximum R-hat values for PP, RPP, VJ, and CMJ were 1.007, 1.001, 1.009, and 1.002, respectively. The minimum bulk effective sample sizes were 622, 1,023, 433, and 2,552, respectively; all minimum tail effective sample sizes exceeded 400, and the maximum ratio of Monte Carlo standard error to posterior standard deviation did not exceed 0.050. Trace plots and parameter-level diagnostics are presented in the model diagnostic figures and Supplementary Material 6.

The DIC values for the four random-effects models were 183.14, 129.18, 123.07, and 162.04, respectively, which were 1.90, 1.27, 1.70, and 0.87 higher than those of the fixed-effect models fitted to the same datasets. Given these small differences in model fit, the prespecified random-effects models were retained for the primary effect estimates, rankings, and pairwise comparisons.

3.4. Primary network effects versus placebo

All effects are posterior mean differences in change scores (MDs), with positive values indicating greater improvement relative to PLA. Values in parentheses are 95% credible intervals (CrIs). Prediction intervals (PIs) are reported separately to distinguish uncertainty in the average relative effect from the expected range of true effects in a future exchangeable study (Figure 2).

Figure 2.

Network plots of eligible interventions for peak power (PP), relative peak power (RPP), vertical jump (VJ), and countermovement jump (CMJ).

3.4.1. Peak power

HMB and beta-alanine (BALA) showed MDs versus PLA of 64.70 W (95% CrI: 19.15 to 110.50) and 40.96 W (95% CrI: 5.10 to 76.78), respectively. These comparisons were supported by two and four direct studies, respectively. The 95% PIs were 14.57 to 114.91 W for HMB and −0.47 to 82.40 W for BALA; the latter included zero. The MD for BI versus PLA was −43.35 W (95% CrI: −77.34 to −9.12). The 95% CrIs for all other supplements versus PLA included zero.

3.4.2. Relative peak power

The MDs versus PLA were 0.84 W/kg (95% CrI: 0.42 to 1.28) for HMB, 0.66 W/kg (95% CrI: 0.09 to 1.21) for vitamin D (VD), and 0.36 W/kg (95% CrI: 0.03 to 0.66) for creatine (CR). These comparisons were supported by four, three, and ten direct studies, respectively. The corresponding 95% PIs were 0.16 to 1.54, −0.13 to 1.41, and −0.29 to 0.97 W/kg. The 95% CrIs for all other nodes included zero.

3.4.3. Vertical jump

Iron supplementation (FE) showed an MD versus PLA of 6.81 cm (95% CrI: 0.84 to 12.83), with a 95% PI of 0.58 to 13.10 cm. This node was supported by one direct study involving 11 participants in the iron-supplementation group. The MDs for probiotics plus spirulina (PS+SPI) and antioxidants plus vitamin E (AOX+VE) were 3.50 cm (95% CrI: −1.91 to 8.82) and 2.85 cm (95% CrI: −0.18 to 5.87), respectively. The 95% CrIs for these comparisons and all other nodes included zero.

3.4.4. Countermovement jump

Betaine (BET), palmitoylethanolamide (PEA), caffeine plus Rhodiola (CAF+GRE), creatine (CR), multi-ingredient supplements (MIPS), and caffeine (CAF) had 95% CrIs excluding zero versus PLA. Their MDs were 3.72 cm (95% CrI: 1.24 to 6.18), 2.87 cm (95% CrI: 0.41 to 5.29), 2.30 cm (95% CrI: 0.31 to 4.17), 2.26 cm (95% CrI: 0.46 to 3.94), 1.79 cm (95% CrI: 0.33 to 3.24), and 1.25 cm (95% CrI: 0.02 to 2.37), respectively. The corresponding numbers of direct studies were 2, 1, 1, 7, 3, and 6. The 95% PIs for BET and PEA were 0.90 to 6.53 cm and 0.04 to 5.65 cm, respectively; the PIs for the remaining comparisons included zero.

3.5. SUCRA rankings

In the complete networks including PLA and BI, the three highest SUCRA values for PP were observed for L-carnitine (LC, 0.888), KA (0.871), and AAKG (0.862); those for RPP were observed for leucine (LEU, 0.819), arachidonic acid (ARA, 0.784), and HMB (0.776); those for VJ were observed for FE (0.921), PS+SPI (0.744), and AOX+VE (0.738); and those for CMJ were observed for BET (0.857), PEA (0.769), and CAF+GRE (0.709). Ranking probabilities, cumulative ranking curves, and a heatmap of rankings recalculated within supplements are provided in Supplementary Material 7. SUCRA values from the complete networks and those recalculated after removing controls are presented separately.

Higher rankings were not always accompanied by precise effect estimates. In PP, LC, KA, and AAKG were each supported by only 1 study at the node level, and their 95% CrIs versus PLA all included 0; KA had no direct comparison with PLA. The mean rank of LC was 5.14, with a 95% rank interval of 1–34. In RPP, the 95% CrIs for LEU and ARA also included 0, whereas in VJ, the effect of FE remained supported by a single study despite its high SUCRA value. Rankings across the different networks describe only the relative positions within their respective bodies of evidence (Figure 3).

Figure 3.

Bayesian network meta-analysis estimates versus placebo for peak power (PP), relative peak power (RPP), vertical jump (VJ), and countermovement jump (CMJ). Points represent posterior means; coloured and pale lines indicate 95% credible and prediction intervals, respectively.

3.6. Pairwise comparisons

The four networks yielded 1,636 unique pairwise comparisons: 703, 231, 351, and 351 for PP, RPP, VJ, and CMJ, respectively. Of these, 642, 203, 303, and 307, respectively, had no corresponding direct comparison studies. The 95% CrIs excluded 0 for 28, 4, 6, and 10 comparisons, respectively, and included 0 for all remaining comparisons. Complete effect estimates and the numbers of direct studies are provided in Supplementary Material 8.

3.6.1. PP

Beyond comparisons with PLA, the MD for HMB versus dietary nitrate (DNO) was 69.55 W (95% CrI: 7.94–128.31), and that for BALA versus DNO was 45.81 W (95% CrI: 3.27–85.71); neither comparison had a corresponding direct study. Some larger differences involved BI or tyrosine (TYR). For example, the MD for HMB versus BI was 108.04 W (95% CrI: 50.77–165.00), and that for HMB versus TYR was 134.89 W (95% CrI: 45.28–224.57); these estimates also relied on indirect comparisons.

3.6.2. RPP

Apart from the comparisons of HMB, VD, and CR with PLA, only the comparison of HMB with the WS node had a 95% CrI excluding 0, with an MD of 1.09 W·kg⁻¹ (95% CrI: 0.21–1.97). No direct comparative study was available for these two interventions.

3.6.3. VJ

Beyond the comparison with PLA, the MDs for FE versus BI, CAF, PS, and VD were 7.70, 7.98, 6.73, and 7.52 cm, respectively, with corresponding 95% CrIs of 1.39–14.08, 0.89–15.18, 0.43–13.06, and 1.23–13.83 cm. The MD for AOX+VE versus BI was 3.74 cm (95% CrI: 0.10–7.46). None of these comparisons had direct studies, and the study information they shared was not treated as repeated independent confirmation.

3.6.4. CMJ

The MDs for BET versus BALA, BI, and sodium bicarbonate (SB) were 3.43 cm (95% CrI: 0.79–6.09), 4.15 cm (95% CrI: 0.78–7.58), and 3.74 cm (95% CrI: 0.14–7.32), respectively; none of these comparisons was supported by direct studies. The MD for CR versus BI was 2.69 cm (95% CrI: 0.04–5.30), with 1 direct study available. All other comparisons with 95% CrIs excluding 0 have been reported in the results versus PLA.

3.7. Subgroup analyses and network meta-regression

Subgroup analyses and meta-regression covered five dimensions: sex, sport, training level, intervention duration, and concomitant training. Of the 20 outcome–categorical covariate combinations across the four outcomes, interaction terms were estimable for 19; the concomitant training interaction in VJ was not identifiable. A continuous-duration model was additionally fitted for each of the four outcomes. Missing covariate data resulted in different numbers of studies being included in the individual models. Missingness affected only the relevant analysis and did not lead to the removal of studies from all primary networks (Supplementary Material 9).

3.7.1. Intervention duration

Within the supplement class, the change in MD per doubling of duration was 3.77 W for PP (95% CrI: −12.00 to 19.54), −0.13 W·kg⁻¹ for RPP (95% CrI: −0.30 to 0.05), 0.26 cm for VJ (95% CrI: −0.52 to 1.05), and 0.34 cm for CMJ (95% CrI: −0.02 to 0.69). All four intervals included 0. With >56 days as the reference, the intervals for all estimable interaction coefficients for the ≤7-day, 8–28-day, and 29–56-day stages also included 0. For CMJ, the coefficient for ≤7 days versus >56 days was −2.58 cm (95% CrI: −5.32 to 0.28).

3.7.2. Sex

The interaction coefficients for female versus male participants were −24.61 W for PP (95% CrI: −82.98 to 32.81), −0.50 W·kg⁻¹ for RPP (95% CrI: −1.25 to 0.21), 0.96 cm for VJ (95% CrI: −2.10 to 4.20), and 0.06 cm for CMJ (95% CrI: −1.51 to 1.53). The intervals for the coefficients comparing mixed-sex populations with male participants also all included 0. Differences between intervals for individual treatments in the subgroup plots did not translate into clear evidence of sex interactions.

3.7.3. Training level

In RPP, the interaction coefficient for McKay Tier 5 versus Tier 1 was 1.31 W·kg⁻¹ (95% CrI: 0.20–2.43); the 95% CrIs for all other estimable tier comparisons included 0. This high-tier category in RPP was informed by only 1 study, and the coefficient represented an exploratory association at the study level.

3.7.4. Sport

In RPP, the interaction coefficient for combat athletes versus power athletes was 1.15 W·kg⁻¹ (95% CrI: 0.04–2.17); the 95% CrIs for all other estimable sport comparisons included 0. Some sport categories in CMJ were completely confounded with treatment allocation, and no separate interaction estimates were reported. Study-level meta-regression was not adjusted for multiple testing, and the findings for training tier and sport category were retained as exploratory.

3.7.5. Concomitant training

The interaction coefficients for the presence versus absence of concomitant training were 33.33 W for PP (95% CrI: −58.74 to 125.67), 0.37 W·kg⁻¹ for RPP (95% CrI: −1.06 to 1.76), and 1.18 cm for CMJ (95% CrI: −1.19 to 3.57). In VJ, the category without concomitant training involved only 1 CAF study, and training status was completely confounded with the intervention node. Accordingly, only conditional effects supported by observations were presented, and the non-identifiable interaction coefficient was not reported.

3.8. Assessment of heterogeneity and inconsistency

The posterior medians of the between-study SD $\tau$ for PP, RPP, VJ, and CMJ were 6.523 W (95% CrI: 0.297–24.155), 0.181 W·kg⁻¹ (95% CrI: 0.008–0.560), 0.596 cm (95% CrI: 0.028–1.929), and 0.518 cm (95% CrI: 0.025–1.372), respectively. The corresponding descriptive $I^2$ medians were 3.6%, 14.6%, 7.2%, and 11.1%, with upper 95% CrI bounds reaching 34.1%, 62.1%, 45.0%, and 46.7%, respectively. Central estimates of heterogeneity were small, but the posterior ranges remained compatible with larger between-study differences (Supplementary Material 10).

The random-effects design-by-treatment interaction test accounting for within-study covariance did not detect global inconsistency: PP, Q = 8.51, df = 14, p = 0.861; RPP, Q = 7.52, df = 5, p = 0.185; VJ, Q = 2.19, df = 6, p = 0.901; and CMJ, Q = 11.50, df = 10, p = 0.320. In the paired numerical-integration analyses, DIC increased relative to the consistency models by 17.74, 0.47, 8.61, and 6.99, respectively, in the design-specific models.

Among comparisons retaining an independent indirect pathway after removal of the corresponding direct studies, 29 Bayesian node-splitting analyses were completed: 13 for PP, 1 for RPP, 6 for VJ, and 9 for CMJ. All 95% CrIs for the differences between direct and indirect effects included 0. The only comparison eligible for node splitting in RPP was CR versus BALA, for which the direct-minus-indirect effect was −0.71 W·kg⁻¹ (95% CrI: −3.97 to 2.55). Assessment of local inconsistency in RPP was therefore limited to this identifiable comparison.

Loop-specific analyses identified a small number of unadjusted local signals. The BI–DNO–PLA loop in PP had p = 0.049; the CR–CR+BALA–PLA loop in RPP had p = 0.010; and two related loops involving CAF, GRE, CAF+GRE, and PLA in CMJ both had p = 0.006. After Holm adjustment within each outcome, the smallest p values were 0.063 for RPP and 0.097 for CMJ, neither below 0.05. Global and local tests differed in their coverage, and some direct comparisons lacked separable indirect information.

3.9. Small-study effects and sensitivity analyses

3.9.1. Comparison-adjusted funnel plots and small-study effects

Formal tests of small-study effects included 33 contrasts from 27 studies for PP, 24 contrasts from 23 studies for RPP, 18 contrasts from 18 studies for VJ, and 35 contrasts from 34 studies for CMJ. In Egger-type GLS regression, p values for the standard error term were 0.313, 0.480, and 0.485 for PP, RPP, and VJ, respectively. For CMJ, the coefficient of the standard error term was −0.72 (95% confidence interval: −1.21 to −0.23), with p = 0.006 and a Holm-adjusted p value of 0.023 across the four outcomes, indicating that less precise studies reported smaller relative effects Figure 4.

Figure 4.

SUCRA heatmap of dietary supplements across four outcome networks.

In CMJ, p values for this test after sequential omission of individual studies ranged from 0.001 to 0.069. After excluding Pereira et al. (2026, study ID 38), the coefficient was −0.54 (95% confidence interval: −1.12 to 0.04), with p = 0.069, indicating that the assessment of small-study effects was influenced by a single study. Funnel-plot findings were not directly equated with the presence or absence of publication bias. Results under different test restrictions and the corresponding plots are provided in Supplementary Material 11 (Figure 5).

Figure 5.

Comparison-adjusted funnel plots for PP, RPP, VJ, and CMJ.

3.9.2. Pre–post and between-condition crossover correlations

For each outcome, 25 combinations of r and q were examined, with r = q = 0.5 as the baseline; correlation-dependent sampling uncertainty was re-estimated in each scenario. All VJ data came from parallel-group trials, so changing q alone did not alter the data. Results obtained with different q values at the same r value were therefore not treated as additional independent evidence.

In PP, the posterior mean for HMB ranged from 62.42 to 68.87 W across combinations, but interval widths varied considerably. At r = q = 0.1, the estimate was 62.42 W (95% CrI: −11.05 to 135.88), and the 95% CrI changed from excluding 0 in the primary model to including 0. The mean for BALA ranged from 38.63 to 49.17 W, with intervals also including 0 under some combinations. In RPP, the mean for HMB ranged from 0.83 to 0.86 W·kg⁻¹, and the lower 95% CrI bound remained above 0 under every combination; the intervals for CR and VD crossed 0 under some combinations.

In VJ, the mean for FE ranged from 6.54 to 7.07 cm, and the lower 95% CrI bound varied from −1.16 to 3.42 cm according to the correlation assumptions. In CMJ, BET retained a positive 95% CrI excluding 0 under every combination, whereas CAF, CAF+GRE, CR, MIPS, and PEA had intervals including 0 under some combinations. Different assumptions primarily altered precision, although mean estimates also changed for nodes such as CR and CAF. After recalculating rankings within the common set of nodes in each scenario, the maximum absolute changes in SUCRA were 0.185, 0.106, 0.119, and 0.126 for PP, RPP, VJ, and CMJ, respectively.

3.9.3. Risk of bias, study design, and handling of standard deviations

Excluding only results at high risk of bias did not change interval classifications among retained nodes in any of the four networks. When only low-risk results were retained, 23, 15, 14, and 18 studies remained for PP, RPP, VJ, and CMJ, respectively, and the numbers of nodes decreased to 28, 12, 21, and 18. Nodes lost after restriction were recorded separately and were not assigned zero effects.

Among low-risk studies, the estimate for HMB in PP remained 64.66 W (95% CrI: 13.34–115.67), whereas that for BALA was 18.11 W (95% CrI: −36.68 to 70.80). In RPP, the estimate for HMB remained 0.89 W·kg⁻¹ (95% CrI: 0.30–1.52), whereas that for VD was 0.57 W·kg⁻¹ (95% CrI: −0.49 to 1.60). In CMJ, the estimates for CAF, CAF+GRE, CR, and MIPS were 2.17 (95% CrI: −0.49 to 4.31), 2.69 (95% CrI: −0.18 to 5.17), 0.67 (95% CrI: −3.48 to 4.81), and 0.91 cm (95% CrI: −2.81 to 4.57), respectively; all intervals changed from excluding 0 to including 0. No additional changes in interval classification occurred among retained nodes in VJ, but FE had been removed under this restriction, so its result could not be considered stable in low-risk studies on that basis. BET and PEA were also absent from the CMJ network restricted to low-risk studies, so their effects could not be estimated under this restriction. A stricter restriction requiring both the assessor judgement and the tool recommendation to indicate low risk yielded the same principal pattern of changes.

After restricting the analysis to randomised parallel-group trials, the MD for BALA in PP was 36.34 W (95% CrI: −0.68 to 73.54), and that for CR in RPP was 0.35 W·kg⁻¹ (95% CrI: −0.01 to 0.68); both intervals included 0. The PP evidence for HMB came from crossover trials, so its effect could not be estimated in this scenario. After restriction according to the source of the change-score SD, the intervals for BALA in PP, CR in RPP, and CAF in CMJ also included 0, whereas HMB in both PP and RPP retained positive intervals excluding 0. This restriction reduced both studies and nodes, and the estimates reflected the combined changes in SD handling and evidence composition.

In an additional re-estimation using the SDs originally recorded in the extraction table in place of those used in the primary approach, the MD for CR versus PLA in CMJ was 0.97 cm (95% CrI: −1.26 to 3.20), with the 95% CrI including 0. No changes in interval classification occurred among retained comparisons in PP, RPP, or VJ. This scenario retained the study set and was compared separately from analyses excluding studies according to SD source.

3.9.4. Leave-one-study-out analyses and sample size restrictions

Sequential omission by original study ID yielded 155 outcome-level re-estimations: 47 for PP, 33 for RPP, 31 for VJ, and 44 for CMJ. All strata and related records from the same study were removed simultaneously. In VJ, no retained comparisons changed in terms of whether their 95% CrIs included 0, although nodes supported by a single study could disappear when that study was removed.

In PP, after excluding Durkalec-Michalski et al. (2017), the estimate for HMB was 50.35 W (95% CrI: −16.10 to 116.76); the interval for BALA included 0 after excluding studies with IDs 30, 65, or 94. In RPP, the interval for CR included 0 after excluding IDs 5, 7, 18, or 96, and that for VD included 0 after excluding IDs 42 or 44. In CMJ, after excluding Nobari et al. (2021), the estimate for BET was 3.21 cm (95% CrI: −0.03 to 6.45); after excluding Pereira et al. (2026), that for CR was 0.61 cm (95% CrI: −1.76 to 2.98); and after excluding Puente-Fernández et al. (2025), that for MIPS was 1.96 cm (95% CrI: −1.10 to 5.00). The interval for CAF also crossed 0 in several leave-one-study-out scenarios, with some changes occurring near the boundary where the lower interval bound was close to 0.

After restricting studies to those with sample sizes at or above the median for each outcome, the estimate for BALA in PP was 15.22 W (95% CrI: −38.48 to 67.06), that for CR in RPP was 0.17 W·kg⁻¹ (95% CrI: −0.39 to 0.59), and that for MIPS in CMJ was 1.25 cm (95% CrI: −2.49 to 5.00); all intervals included 0. These results indicated that some average effects depended on contributions from a small number of studies or studies with smaller sample sizes.

3.9.5. Comparator definitions, concomitant training, and study populations

After separately excluding BI arms or entire studies containing BI, the 95% CrI for BALA in PP included 0, and the interval for CAF in CMJ also included 0. When BI and PLA were merged into a single control node, the intervals for CR, CR+BALA+CIT, DNO, and PRO in PP changed to excluding 0, whereas those for CR and VD in RPP changed to including 0. Merging controls changed the composition of the reference underlying the estimates; these values were reported separately from the primary results retaining PLA and BI as two nodes.

When only studies involving concomitant training were included, the interval for CR in RPP included 0. In VJ, rematching controls according to the exercise conditions in study ID 61 did not change interval classifications among retained comparisons, but the maximum absolute change in SUCRA among common nodes was 0.267. Restricting the population to McKay Tiers ≥2 or ≥3 resulted in intervals including 0 for BALA in PP and CR and VD in RPP; the same change occurred for MIPS in CMJ, with the CAF interval additionally including 0 at Tier ≥3. These population restrictions changed the direct evidence and network connectivity, and their results were not used as substitutes for formal between-group interaction tests.

3.9.6. Intervention duration and combination nodes

Excluding single-day regimens did not change interval classifications among commonly retained comparisons in any of the four networks. Restriction to >28 days resulted in intervals including 0 for BALA in PP and CR and VD in RPP, whereas HMB in PP and RPP remained positive with intervals excluding 0. When only single-day regimens were retained, PP and VJ each had only 1 study remaining, RPP had no estimable network, and CMJ retained only 6 studies and 5 nodes, with intervals for both CAF and MIPS including 0. This restriction could not provide a comparison of acute and long-term effects covering all supplements.

Excluding the MIPS node, excluding combination-ingredient nodes, or splitting MIPS by study did not change interval classifications among commonly retained comparisons. Splitting key nodes supported by data into single-day and multiple-day regimens also did not change interval classifications among comparable nodes. However, deleted or redefined combination nodes no longer represented the same comparisons as in the original primary model; these changes are documented in the records of node loss and redefinition.

3.9.7. VJ test procedures

Further restriction of existing VJ records by test procedure substantially reduced network size. Retaining verified SJ records yielded 4 studies, 6 nodes, and 113 participants; retaining only records explicitly labelled CMJ in the original reports yielded 7 studies, 11 nodes, and 221 participants; and retaining only records named Sargent yielded 3 studies, 6 nodes, and 82 participants. After excluding records with unclear or unmatched procedures, 15 studies, 16 nodes, and 453 participants remained. Each restriction used the test reported in the original study, and overall network results were not interpreted as common effects across all jump procedures.

In the network restricted to Sargent records, the estimate for FE was 6.55 cm (95% CrI: −1.48 to 14.12). The mean was close to the primary result, but the 95% CrI included 0. FE was not retained in the estimable network in scenarios restricted to SJ, restricted to explicitly identified CMJ, or excluding records with unclear procedures. Conversely, after excluding SJ, explicitly identified CMJ records, or repeated jumps, the interval for FE excluded 0 in each network in which it remained.

For Vargas-Molina et al. (2022), substituting another jump test reported in the same study yielded interval classifications consistent with the baseline when SJ data were used. When Abalakov jump data were used, the MD for CR versus PLA was 1.82 cm (95% CrI: 0.30–3.22), and the 95% CrI changed to excluding 0. This change arose from substituting the selected test within the same study and was presented separately from the effect of removing the entire study.

3.9.8. Doses and administration methods within the same supplement

Among records with verifiable ingredients, numerical doses, units, and regimens, 18 outcome–supplement models with dose-specific nodes were constructed, with additional analyses including sequential omission of individual dose regimens. There were 101 pairwise comparisons between dose nodes, of which 12 had direct studies and 89 relied exclusively on indirect evidence. Except for one CR regimen comparison in VJ, the 95% CrIs for all between-dose comparisons included 0.

In VJ, the MD for CR at 0.25 g·kg⁻¹·d⁻¹ versus 30 g·d⁻¹ was −7.77 cm (95% CrI: −15.03 to −0.49), with no direct comparative study available. In PP, the BALA regimen of 0.1 g·kg⁻¹·d⁻¹ versus PLA yielded an MD of 100.32 W (95% CrI: 0.92–199.68), but intervals for comparisons with other BALA doses did not exclude no difference. These findings did not provide a consistent body of evidence supporting a monotonic dose relationship or identifying an optimal dose.

When only records with verifiable doses of the target supplements were used, the pooled MD for CR in RPP was −0.04 W·kg⁻¹ (95% CrI: −0.53 to 0.44). In CMJ, the estimates for CAF and CR were 0.75 cm (95% CrI: −0.70 to 2.16) and 0.30 cm (95% CrI: −2.33 to 2.94), respectively. For all three comparisons, the 95% CrIs changed from excluding 0 in the primary analysis to including 0. In VJ, FE retained a positive interval when dose restrictions and node splitting were applied separately to CR, PRO, and VD, but this did not constitute a test of the dose relationship for FE itself.

When the pooled data from Vargas-Molina et al. (2026) were replaced separately by each caffeine administration condition, the MD for CAF versus PLA in CMJ was 1.31 cm (95% CrI: 0.10–2.42) under the capsule condition, 1.14 cm (95% CrI: −0.16 to 2.29) under the chewing-gum condition, and 1.42 cm (95% CrI: 0.22–2.53) under the mouth-rinse condition. The 95% CrI included 0 in the chewing-gum substitution scenario but excluded 0 in the other two scenarios. These were results from separate re-estimations of the entire network, not direct effect differences between the three administration methods.

3.9.9. Prior specifications and ranking uncertainty

Setting the scale of the half-normal heterogeneity prior to 0.25S or 1S, using a half-t prior, or setting the treatment-effect prior SD to S or 4S did not change interval classifications among retained comparisons in RPP, VJ, or CMJ. In PP, WS had intervals with lower bounds slightly above 0 under the narrower heterogeneity prior and the wider treatment-effect prior, with MDs of 70.41 W (95% CrI: 0.04–140.74) and 71.17 W (95% CrI: 0.11–142.21), respectively; the interval for this node included 0 in the primary analysis. These changes occurred mainly near the statistical boundary.

All types of sensitivity analysis presented effect magnitudes, credible intervals, node loss, and rankings among common nodes. The estimate for HMB in RPP remained positive with an interval excluding 0 under the examined correlation assumptions, low-risk-of-bias restrictions, SD sources, and approaches to handling controls. HMB in PP, FE in VJ, and several CMJ comparisons were more sensitive to correlations, individual studies, measurement procedures, or node handling.

3.10. CINeMA certainty of evidence

After assessment of all six domains, none of the 1,636 pairwise comparisons had high-certainty evidence; 17 had moderate, 55 had low, and 1,564 had very low certainty. Of the 703 PP comparisons, 16 had moderate, 42 had low, and 645 had very low certainty. The corresponding counts among the 231 RPP comparisons were 1, 5, and 225. All 351 VJ comparisons had very low certainty. Of the 351 CMJ comparisons, 8 had low and 343 had very low certainty (Supplementary Material 12).

Among the 110 comparisons with PLA, 3 had moderate, 9 had low, and 98 had very low certainty. In PP, HMB and BI versus PLA had moderate certainty, and BALA had low certainty. In RPP, HMB had moderate, CR had low, and VD had very low certainty. In VJ, FE versus PLA had very low certainty. In CMJ, BET and PEA had low certainty, whereas CAF, CAF+GRE, CR, and MIPS had very low certainty. Higher SUCRA values did not correspond to higher certainty: LC and KA in PP, LEU and ARA in RPP, and FE in VJ all had very low certainty.

The reporting bias domain was rated as some concerns for all current comparisons. Networks in which tests of small-study effects did not meet the testing threshold were not automatically judged to have no reporting bias. Other concerns were determined comparison by comparison from contributions of within-study bias, indirectness, and the positions of intervals relative to the reference range. The distribution-based reference thresholds used to assess imprecision and heterogeneity were 29.94 W, 0.31 W·kg⁻¹, 1.28 cm, and 1.02 cm for PP, RPP, VJ, and CMJ, respectively; these were not treated as validated minimal clinically important differences. The evidence profiles and comparison-specific rationales for downgrading used the same final ratings.

4. Discussion

4.1. Principal findings

This review included 119 studies and evaluated the effects of dietary supplements on PP, RPP, VJ, and CMJ separately. Compared with placebo, HMB and β-alanine showed favourable effects in the PP network; HMB, vitamin D, and creatine showed favourable effects in the RPP network; and iron supplementation showed a favourable effect in the VJ network. In the CMJ network, the 95% credible intervals for betaine, palmitoylethanolamide, caffeine combined with Rhodiola rosea, creatine, multi-ingredient supplements, and caffeine all excluded zero. These findings complement previous evidence syntheses focusing on individual sports or supplements [2–6] and should be interpreted in the context of the original trials, supplementation regimens, and overall findings of the existing literature.

4.2. Interpretation of findings by outcome

4.2.1. Peak power

The MD in PP for HMB versus placebo was 64.70 W (95% CrI: 19.15–110.50), suggesting an association between HMB supplementation and higher Wingate peak power in the included studies. Portal et al. observed improvements in anaerobic performance following HMB supplementation in adolescent volleyball players, and Durkalec-Michalski et al. also reported favourable changes in power in highly trained combat sport athletes receiving supplementation for 12 weeks [47,63]. Most of these trials evaluated supplementation in the context of ongoing training, and their findings can be understood in relation to training adaptation and the maintenance of muscle function. As a metabolite of leucine, HMB is involved in muscle protein metabolism and post-exercise recovery, but these physiological actions cannot be directly equated with increases in peak power [147]. The meta-analysis by Sanchez-Martinez et al. found no clear overall benefits of HMB for strength or body composition in trained or competitive athletes [148]. These bodies of research assessed different outcomes: previous null findings for strength or body composition do not directly invalidate the power estimate in this review, and the present power findings are likewise insufficient to establish broad benefits of HMB for muscle growth or exercise performance.

The MD in PP for β-alanine was 40.96 W (95% CrI: 5.10–76.78). This finding may be related to its role in increasing muscle carnosine stores and enhancing intracellular buffering during high-intensity exercise, although sustained exercise capacity should be distinguished from instantaneous peak output. The meta-analysis by Hobson et al. found more evident benefits of β-alanine during exercise lasting 1–4 min, whereas exercise lasting less than 60 s showed no clear improvement [149]. Saunders et al. subsequently included 40 studies, finding a small overall effect and showing that exercise duration and test characteristics influenced the estimates. Benefits were more evident in exercise capacity tests, whereas performance in fixed-task tests did not improve to the same extent [150]. These reviews support a role for β-alanine in some forms of high-intensity exercise, but the buffering mechanism alone cannot explain the Wingate peak power finding in this review. The prediction interval for this comparison crossed zero, and the estimate was attenuated when restricted to studies at low risk of bias. The current finding is therefore more appropriately described as a possible benefit under specific training and supplementation conditions.

4.2.2. Relative peak power

The MD for HMB in the RPP network was 0.84 W·kg⁻¹ (95% CrI: 0.42–1.28). The 95% prediction interval was also above zero, and the effect remained favourable across multiple sensitivity analyses. This finding was supported by 4 direct comparison studies and was consistent with findings from some trials in volleyball and combat sport athletes [47,63]. Unlike PP, RPP is influenced by both changes in power and body mass; its improvement should therefore not be attributed solely to increased muscle power production. Given the inconsistent findings for HMB in the broader literature on strength and body composition and its current classification in AIS Group C, the moderate-certainty rating in this review should be understood as an assessment of the present HMB–RPP comparison. It cannot yet be translated into supplementation recommendations applicable to all trained populations [13,147,148].

The MD in RPP for creatine was 0.36 W·kg⁻¹ (95% CrI: 0.03–0.66). Creatine can increase skeletal muscle creatine and phosphocreatine stores, supporting ATP resynthesis during high-intensity contractions. When sustained supplementation is combined with resistance training, it may also promote adaptation by supporting training load [151]. This physiological basis is consistent with the direction of the power improvement observed in this review. The systematic review and meta-analysis by Kazeminasab et al. evaluated both strength and power and found that creatine supplementation could improve some strength measures and vertical jump performance, suggesting that its potential benefits are not confined to a single test [152]. However, RPP is normalised to body mass, and changes in body mass during supplementation also contribute to the final outcome. The present estimate supports a small improvement under the included regimens and should not be interpreted as indicating identical effects across all doses and supplementation durations.

The MD in RPP for vitamin D was 0.66 W·kg⁻¹ (95% CrI: 0.09–1.21). Unlike creatine, vitamin D supplementation primarily concerns the restoration of nutritional status. Jung et al. studied vitamin D-insufficient taekwondo athletes during winter training, whereas Kujach et al. examined changes in exercise capacity after supplementation in young men with low baseline vitamin D levels [69,71]. These study contexts suggest that some benefits may have arisen from correcting pre-existing insufficiency. The broader evidence provides an essential comparison: Sist et al. included 11 randomised trials and found no clear overall improvement in maximal strength or jump power in athletes; the updated meta-analysis by Han et al. likewise found no statistically significant overall strength effect, although the quadriceps strength subgroup showed a favourable result [153,154]. Restoration of nutritional markers and improvement in exercise performance therefore do not necessarily occur together. The present RPP finding provides an additional observation of performance changes in the context of vitamin D insufficiency; it does not support the claim that individuals with normal vitamin D status would obtain the same benefit from additional supplementation.

4.2.3. Vertical jump

The MD for iron supplementation in the VJ network was 6.81 cm (95% CrI: 0.84–12.83). The original study supporting this comparison involved young female athletes with iron deficiency without anaemia, making their nutritional context central to interpretation [135]. Iron is involved in oxygen transport and multiple energy metabolism processes; correcting depleted iron stores may help maintain training capacity and indirectly influence performance. The meta-analysis by Burden et al. found that iron treatment could improve iron status and aerobic capacity in iron-deficient, non-anaemic endurance athletes [155]. The systematic review by Rubeor et al., however, showed that performance did not improve in all studies of iron deficiency without anaemia, with favourable results occurring more often in studies involving lower baseline iron stores [156]. Together, these findings support interpreting iron supplementation in the context of nutritional insufficiency rather than as a general means of enhancing power.

The jump finding in this review concerns a different motor task from those examined in the endurance literature; improvements in aerobic capacity cannot be taken directly as mechanistic evidence for increases in vertical jump height. This comparison had only 1 direct study, and its credible interval crossed zero when the analysis was restricted by jump procedure. This relatively large MD is therefore more appropriately viewed as a preliminary observation in a specific population. Its practical implication is to highlight assessment of iron status, rather than to support routine iron supplementation in iron-replete individuals seeking to improve jump height.

4.2.4. Countermovement jump

The MD in CMJ for creatine was 2.26 cm (95% CrI: 0.46–3.94). The meta-analysis by Kazeminasab et al. found an improvement in vertical jump height of approximately 1.48 cm, similar in direction to the present finding, although its mix of jump tests was not identical to the network constructed here after explicitly distinguishing CMJ [152]. Pereira et al. evaluated creatine gummies in female beach volleyball players, providing direct study data on performance changes in a training context [65]. In conjunction with the established physiological and performance evidence for creatine, the current result can be interpreted as an average performance difference arising from the combined effects of supplementation and training adaptation [151]. Its magnitude remains influenced by the particular studies included. The prediction interval in this review crossed zero, indicating that a new study under similar conditions may not observe the same improvement.

The MD in CMJ for caffeine was 1.25 cm (95% CrI: 0.02–2.37). The umbrella review by Grgic et al., which synthesised 21 meta-analyses, found that caffeine could exert small favourable effects on several types of exercise performance, including muscular strength, power, and jumping [157]. The ISSN position stand also supports improvements in exercise performance with appropriate caffeine intake protocols, with its principal actions involving adenosine receptor antagonism and central nervous system stimulation [158]. The small positive effect observed in this review is consistent with the overall direction of these reviews, but the lower credible interval bound was close to zero, indicating continued imprecision in the estimated magnitude. The practical value of caffeine should be assessed in relation to intake timing, dose, and individual tolerability, rather than determined solely by network rankings.

The MD in CMJ for caffeine combined with Rhodiola rosea was 2.30 cm (95% CrI: 0.31–4.17), based on 1 direct study. Dou et al. used 4 weeks of Rhodiola rosea supplementation with caffeine ingested before testing in male collegiate soccer players; the study included both single-ingredient and combined-supplementation groups [106]. Caffeine-related improvements were mainly evident in jumping performance, whereas Rhodiola rosea was more closely associated with perceived exertion and maintenance of performance during repeated heading tasks; the combined regimen showed a broader range of performance changes. This design combined acute stimulation with sustained supplementation, providing a specific context for interpreting the combined regimen. Considered alongside the existing review evidence for caffeine, the current finding supports further investigation of the formulation’s practical value [157,158]. However, the larger point estimate for the combined regimen versus placebo does not itself establish superiority over caffeine alone or synergy between the ingredients.

The MD in CMJ for betaine was 3.72 cm (95% CrI: 1.24–6.18). Nobari et al. used a 14-week regimen of 2 g per day in adolescent professional soccer players and observed favourable CMJ changes during regular training and competition [77]. Cholewa et al. also investigated the effects of betaine on body composition and exercise performance in the context of resistance training [110]. The osmoprotective and methyl-donor functions of betaine provide a physiological basis for effects on cellular function, although the more direct support currently comes from training intervention studies. The meta-analysis by Zawieja et al., which included 17 studies, found a favourable effect of long-term betaine supplementation on strength. The jump result reached statistical significance after exclusion of one low-quality study, whereas measures such as cycling sprint power showed no clear benefit [159]. The direction of the jump improvement in this review was similar to that in the meta-analysis, but estimates were attenuated after exclusion of individual studies. A stable effect magnitude across training contexts therefore cannot yet be established.

The MD in CMJ for palmitoylethanolamide was 2.87 cm (95% CrI: 0.41–5.29). Huschtscha et al. used 300 mg per day during 8 weeks of resistance training and observed an improvement in CMJ performance, but no between-group difference in lean body mass gain [113]. This finding suggests that the jump improvement should not be directly attributed to muscle hypertrophy. By contrast, Schouten et al. found no improvement in strength, jump recovery, or muscle soreness with palmitoylethanolamide in an exercise-induced muscle damage model [160]. These trials addressed sustained training and post-exercise recovery, respectively, indicating that the current literature has not established a consistent explanation of performance benefits. The favourable estimate in this review primarily reflects observations under the former type of training regimen and requires confirmation in independent trials.

The MD in CMJ for multi-ingredient supplements was 1.79 cm (95% CrI: 0.33–3.24). Combined formulations may provide energy, protein, or other physiologically active ingredients needed for training, but their effects depend on their actual composition. The review of multi-ingredient pre-workout supplements by Harty et al. noted that some studies observed improvements in muscular endurance and subjective states, whereas strength and power findings were inconsistent [161]. The formulations included in this review also encompassed post-workout supplements. For example, Puente-Fernández et al. evaluated a combined product containing carbohydrate, leucine-enriched whey protein, creatine, HMB, and vitamin D₃ during 6 weeks of resistance training [118]. Such studies concern a different intended use from pre-workout formulations primarily targeting acute stimulation. The pooled result in this review can therefore describe the average performance under the included combined regimens, but cannot be attributed to any single ingredient or extrapolated to other products with different formulations.

4.3. Overall interpretation

These findings indicate that discussion of supplement effects requires consideration of both existing evidence and actual conditions of use. Caffeine acts primarily in relation to pre-exercise intake, whereas creatine and β-alanine generally require supplementation over a period of time to increase the relevant intramuscular stores. Vitamin D and iron primarily involve the identification and correction of nutritional insufficiency [1,149–151,153–158]. An identical ingredient name therefore does not imply complete equivalence between regimens. Analyses of dose, duration, and concomitant training can help explain the composition of the evidence, but the available data are insufficient to identify an optimal regimen for each supplement. A reasonable application is to consider the comparative estimates from this review alongside established evidence-based protocols, participants’ nutritional status, and specific training goals.

The HMB finding in RPP was a relatively stable favourable result in this review, and some findings for creatine and caffeine were consistent in direction with their broader performance evidence. Findings for iron, vitamin D, and less-studied single or combined formulations require interpretation in relation to the relevant populations and regimens. Certainty was very low for most comparisons; SUCRA rankings therefore reflect only relative ranking probabilities within each network and cannot replace assessment of effect magnitude and evidence reliability. The four outcomes were derived from different study sets and network structures. Accordingly, this review describes supplement effects within each network and does not infer interactions between supplements and motor tasks from differences between network results.

4.4. Limitations

This review was retrospectively registered and included only peer-reviewed English-language publications. Some reports could not contribute to the relevant quantitative synthesis because outcome data or measures of dispersion were missing or insufficient for calculation. If data availability was associated with the direction or magnitude of the findings, the synthesised evidence may be subject to data-availability bias. The included studies differed in sport, training level, sex, and concomitant training conditions, and some interventions were supported by a single study, limiting the precision of comparisons and the interpretation of subgroup analyses. Certainty was low or very low for most comparisons, and some favourable findings were attenuated in sensitivity analyses, indicating limited stability.

The same supplement node included regimens with different doses and supplementation durations, and the available data were insufficient to fully distinguish their effects. Some studies incompletely reported baseline nutritional status, habitual caffeine intake, and training load, further limiting interpretation of effect differences. The pooled estimates therefore primarily reflect the average effects of the included regimens and cannot be used directly to identify the optimal supplementation approach.

In addition, test procedures and measurement equipment varied across VJ and CMJ studies, some change-score SDs depended on correlation assumptions, and within-participant covariance information was incomplete for some crossover trials. These factors increased uncertainty in the effect estimates. The four outcome networks comprised different study sets, and differences in effects or rankings between networks cannot be used to establish interactions between supplements and motor tasks.

5. Conclusions

This review included 119 studies and evaluated the relative effects of dietary supplements separately within the PP, RPP, VJ, and CMJ networks. Compared with placebo, favourable effects were observed for HMB and β-alanine in PP; HMB, creatine, and vitamin D in RPP; iron in VJ; and betaine, palmitoylethanolamide, creatine, caffeine, and some combined or multi-ingredient regimens in CMJ. Among these findings, the HMB effect in RPP was relatively stable across multiple sensitivity analyses and had moderate CINeMA certainty, although the number of direct studies remained limited. Certainty was low or very low for most other comparisons, and the prediction intervals for some favourable effects crossed zero or the effects were attenuated under different analytical conditions.

Exploratory meta-regression identified associations in the RPP comparisons by training level and sport, but these were insufficient to establish consistent patterns of effect modification. Sensitivity analyses of jump test type, dose, correlations, and risk of bias further showed that some findings depended on the particular studies included and the analytical assumptions. The four outcome networks differed in their studies, participants, and interventions; these findings therefore support only comparisons within each network and cannot establish that outcome type modifies supplement effects. Supplement use should take account of effect magnitude, certainty of evidence, nutritional status, and specific conditions of use, and individualised regimens should not be based solely on SUCRA rankings. The key favourable findings require validation in independent direct comparison studies.

Supplementary Material

R_code.docx

Supplementary materials

Supplementary materials

Supplementary materials

Disclosure statement

No potential conflict of interest was reported by the author(s).

Funding

This study received no funding.

Supplementary material

Supplemental data for this article can be accessed at https://doi.org/10.1080/15502783.2026.2740221.