This systematic review adhered to the PRISMA guidelines (Page et al., 2021) and conducted comprehensive searches in various databases, including Google Scholar, Web of Science, PubMed, Elsevier, Scopus, CNKI, and Wanfang, covering the literature from January 2000 to June 2025. The search utilized a structured strategy with Boolean operators and keywords: TS = ((“sport education” OR “cooperative learning” OR “teaching games for understanding” OR “tactical games” OR “game-based approach” OR “tactical games approach” OR “tactical approach” OR “tactical games model” OR “teaching for personal and social responsibility” OR “hybrid pedagogical model” OR “pedagogical model”) AND (“sport” OR “physical education” OR “training”) AND (“decision making” OR “skill execution” OR “motivation” OR “enjoyment”) AND (“intervention” OR “experimental” OR “quasi-experimental” OR “randomized controlled trial”)). The full electronic search strategy for Web of Science is provided in Supplementary material S1 to ensure reproducibility. Spain has a large body of literature related to pedagogical models, so only English, Spanish or Chinese articles were searched. The review protocol was not pre-registered in PROSPERO or other registries. This is acknowledged as a methodological limitation.

The study applied the PICOS framework (Population, Intervention, Comparison, Outcomes, and Study design) with specific criteria: (1) Population: Students and athletes without pre-existing injuries or medical conditions. (2) Intervention: The experimental group employed established teaching methods, including physical education, cooperative learning, teaching games for understanding, and teaching personal and social responsibility, or a combination of these methods. (3) Comparison: The control group employed traditional teaching approaches. (4) Outcomes: Evaluation of tactical ability (decision making), technical skills (skill execution), and emotional aspects (motivation and enjoyment). (5) Study Design: The included studies primarily employed quasi-experimental designs, given the structured nature of educational interventions, such as fixed class assignments. Exclusion criteria encompassed: (1) Self-controlled study designs; (2) Reviews, theses, or dissertations; (3) Studies lacking complete or sufficient data for outcome extraction.

All identified records were de-duplicated using EndNote X9. Two independent reviewers conducted a double-blind screening of titles, abstracts, and keywords according to the eligibility criteria, followed by full-text assessment. Inter-rater reliability was evaluated using Cohen’s Kappa (κ = 0.85), indicating strong agreement. Data extraction focused on: (1) General Information: First author, publication year, and study characteristics. (2) Intervention Characteristics: Study design, sample size, participant demographics (including age and gender), teaching content, and duration of the intervention. (3) Outcome Measures: Tactical ability (“decision-making”), technical ability (“skill execution”), and emotional outcomes (“motivation” and “enjoyment”).

Because primary studies reported intervention length using heterogeneous units (e.g., lessons, weeks, or months), instructional duration was standardized into total instructional minutes using a predefined conversion protocol to enable quantitative comparison of intervention “dosage.”

The following rules were applied: (1) When weekly frequency and session length were reported: Total minutes = sessions/week × minutes/session × number of weeks. (2) When duration was reported in months: Months were converted assuming 4 weeks per month, followed by the same calculation. (3) When only the number of weeks was reported without session details: A standardized exposure of 90 min/week (two 45-min PE lessons) was assumed. (4) When duration was described only qualitatively (e.g., “one semester”): The intervention was conservatively classified as ≥ 720 min, representing the minimum exposure expected in a typical semester-based PE curriculum.

This harmonization procedure allowed consistent subgroup classification while applying identical assumptions across studies to maintain comparability.

The methodological quality of the included studies was assessed using the Methodological Index for Non-Randomized Studies (MINORS), which evaluates 12 methodological aspects. Each domain was scored as 0 (not reported), 1 (reported but inadequate), or 2 (reported and adequate), with a maximum score of 24. Two reviewers independently scored each study; discrepancies were resolved by discussion. Inter-rater reliability for MINORS scoring was also high (κ = 0.89).

The validity of indirect comparisons in the network meta-analysis was based on the assumption of transitivity. Clinical and methodological comparability across studies was evaluated in terms of participant characteristics (age group and educational setting), intervention implementation (school-based PE programs), outcome constructs, and comparator conditions (primarily traditional instruction).

All analyses were conducted using R Studio (version 4.4.3) and Stata 15.1 (StataCorp, TX, USA). Bayesian Network Meta-Analysis (NMA) was performed in R using the gemtc (version 0.8–2) and rjags (version 4–13) packages, with data management via dplyr. Stata 15.1 was used for subgroup analyses.

Continuous outcomes were summarized as Standardized Mean Differences (SMDs) with 95% Credible Intervals (CrIs). A Bayesian random-effects model was applied to account for between-study variability, using non-informative priors to minimize prior influence on posterior estimates.

Markov Chain Monte Carlo (MCMC) sampling was conducted with four chains for 50,000 iterations, with the first 20,000 iterations discarded as burn-in. Convergence was assessed using the Potential Scale Reduction Factor (PSRF), with values approaching 1.0 indicating convergence. Trace plots and posterior density plots were visually inspected to confirm adequate chain mixing.

Local inconsistency was evaluated using node-splitting analyses comparing direct and indirect evidence, with p > 0.05 indicating no significant inconsistency. The ranking of pedagogical models was estimated using the Surface Under the Cumulative Ranking curve (SUCRA), ranging from 0% (least effective) to 100% (most effective).

Publication bias was assessed through comparison-adjusted funnel plots generated using the netmeta package and quantified using Egger’s regression test.

Between-study heterogeneity was examined using both conventional I² statistics and the posterior median of the between-study standard deviation (τ) with corresponding 95% CrIs. To explore potential sources of variability, subgroup analyses stratified by instructional duration and sport type were conducted using random-effects models in Stata 15.1.

A comprehensive search across seven databases identified 1,250 articles. After removing duplicates, 922 articles remained. Title and abstract screening excluded 368 articles, and 489 were excluded after full-text review. Ultimately, 65 articles met the inclusion criteria and were included in the analysis. The study selection process is illustrated using a PRISMA 2020 flow diagram (Figure 1).

Disagreements during the screening process were resolved through discussions between two independent researchers. A total of 65 quasi-experimental studies (pre- and post-test) published in English were included, encompassing 8,027 participants across 16 countries. The interventions assessed various pedagogical models, including SE, TGfU, CL, TPSR, and HPMs, in comparison to traditional teaching methods. Geographically, the majority of studies were conducted in Europe (63.0%, primarily in Spain), followed by Asia (18.5%) and North America (10.8%). Participants ranged from K-12 to higher education, with middle school students constituting the largest subgroup at 47.4%. The duration of interventions varied, with the most common range being 720–1,440 min (37.0%). Detailed characteristics of the interventions and outcome measures are summarized in Table 1 and Supplementary Table S1.

The quality of the included studies was assessed using the Methodological Index for Non-Randomized Studies (MINORS) tool, with the results summarized in Supplementary Table S2. Of the 65 studies, 35 (53.8%) were rated as high quality (scores: 19–24), representing a total sample of 4,203 participants (52.4%). Twenty-six studies (40.0%) were classified as medium quality (scores: 16–18), with a combined sample of 3,617 participants (45.0%). The remaining four studies (6.2%) were considered low quality (scores: 13–15), involving 207 participants (2.6%). Studies receiving lower scores were typically characterized by small sample sizes, non-equivalent baselines, lack of blinding procedures, and incomplete reporting of dropout rates (Andrianto, 2023; Wallhead and Ntoumanis, 2004).

A total of 65 papers were included in this study. As shown in Figure 2, the thickness of the connecting lines in the network plot indicates the number of corresponding studies, i.e., the thicker the line, the more comparative studies between the two groups of interventions. In all the literature, the frequency of the different outcome indicators was as follows: decision-making was used in 18 studies with 23 data sets; skill execution was used in 26 studies with 44 data sets; and motivation was used in 33 studies with 37 data sets; Enjoyment was used in 22 studies with 26 data sets.

Figure 2 illustrates that, among the aforementioned outcome indicators, most studies have focused on direct comparisons between instructional models and control groups. In contrast, the number of two-by-two comparisons between different instructional models was relatively small. Based on the thickness of the connecting lines in the network plot, it can be further observed that, for the decision-making and skill execution indicators, more studies were comparing TGfU with the control group, whereas for the motivation and enjoyment indicators, the number of studies comparing SE with the control group was the number of studies on motivation and enjoyment between SE and the control group was more significant.

The validity of the network meta-analysis depends on the assumption of consistency between direct and indirect evidence. Comparison of the Deviance Information Criterion (DIC) between the consistency model and the unrelated mean effects (UME) model showed negligible differences (< 5 points) across all outcomes, indicating no significant inconsistency within the network. Accordingly, the consistency model was adopted for all analyses.

Model convergence was evaluated using the Potential Scale Reduction Factor (PSRF), with all parameters stabilizing at 1.00, indicating satisfactory convergence of the Markov chains. Visual inspection of trace plots demonstrated stable sampling without drift, and posterior density plots showed smooth, unimodal distributions, confirming adequate chain mixing and posterior stability (Supplementary Figure S1).

To further validate the network structure, node-splitting analysis was performed to evaluate local inconsistency by comparing direct and indirect evidence within closed loops. The analysis revealed no significant discrepancies (p > 0.05) between direct and indirect estimates for any comparison, supporting the assumption of consistency underlying the network model. Detailed results of the node-splitting assessment are provided in Supplementary Table S3.

The rankings reflect the relative probability of each model being the most effective in improving each specific learning outcome. Surface Under the Cumulative Ranking (SUCRA) values were used to rank pedagogical models for each outcome (Table 2). 1) Decision making: HPM (80.1%) > TGfU (57.1%) > TPSR (54.5%) > CL (49.6%) > SE (43.0%) > CON (15.9%). 2) Skill execution: HPM (77.2%) > CL (75.5%) > SE (51.2%) > TGfU (45.0%) > CON (1.1%). 3) Motivation: TGfU (82.1%) >SE (70.7%) >HPM (61.3%) > CL (55.0%) > TPSR (30.2) > CON (0.7%). 4) Enjoyment: SE (74.0%) > HPM (66.9%) > CL (55.2%) > TPSR (49.0%) > TGfU (35.0%) > CON (1.2%).

To complement SUCRA rankings, pairwise relative effects between pedagogical models were examined using Bayesian league tables (Supplementary Tables S4–S7). Across all outcomes, most comparisons were associated with wide and overlapping 95% credible intervals (CrIs), indicating considerable uncertainty in the magnitude of differences between interventions.

For decision-making, Teaching Games for Understanding (TGfU) showed a positive effect compared with traditional instruction (CON) (SMD = 0.83, 95% CrI 0.36–1.31), whereas comparisons among active pedagogical models demonstrated overlapping CrIs. For skill execution, several models produced favourable estimates relative to CON, but nearly all inter-model contrasts included the null value. For motivation and enjoyment, CrIs for all pairwise comparisons spanned zero, suggesting similar estimated effects across pedagogical approaches.

For the 65 included studies, the risk of publication bias was assessed in this study, and the results are shown in Supplementary Figure 1. Based on the funnel plot and Egger’s test results (derived using RStudio 4.4.3 and rjags), the distribution of effect sizes appeared to be concentrated in the upper region of the funnel plot. This pattern suggests that most studies had relatively large sample sizes. Only a few studies were located at the bottom or near the edges of the funnel, indicating a small-study effect. Egger’s test further supported this observation: the p-values for decision-making (p = 0.1582), motivation (p = 0.2101), and enjoyment (p = 0.0245) were all above the 0.05 threshold, indicating no significant evidence of publication bias for these outcomes.

However, for skill execution, Egger’s test revealed a statistically significant result (p = 0.0077), suggesting potential publication bias in this domain. Consequently, the findings related to skill execution should be interpreted with caution, and future research should prioritize high-quality, well-reported studies to verify and strengthen the current evidence base.

Between-study heterogeneity was quantified using the posterior median of the between-study standard deviation (τ) with 95% credible intervals (CrIs) (Table 3).

For decision-making, skill execution, and motivation, τ values were similar (median τ = 4.97), with wide CrIs (Decision-making: 3.28–7.97; Skill execution: 3.27–7.93; Motivation: 3.28–7.86). Enjoyment showed comparatively lower heterogeneity (median τ = 2.17, 95% CrI 1.63–3.07).

Conventional heterogeneity statistics from pairwise analyses were consistent with these findings, with I² values generally exceeding 70% (Supplementary Table S8).

Subgroup analyses were conducted in Stata 15.1 using random-effects models stratified by instructional duration and sport type (Table 4). Interventions exceeding approximately 720 min tended to show larger effect estimates than shorter-duration interventions (<720 min). Differences between team and individual sport contexts were also observed in selected outcomes, although estimates were imprecise in analyses with limited studies.

The included studies predominantly employed quasi-experimental designs, focusing largely on K–12 and college students, with research heavily concentrated in Europe. The most significant contribution of this network meta-analysis is its definitive rejection of the quest for a universal pedagogical model. Instead, our findings establish a new paradigm of specialized effectiveness. While traditional PE often relies on a one-size-fits-all approach, this study demonstrates that pedagogical models act as specialized tools for distinct educational domains.

However, this differentiation should be interpreted as probabilistic rather than definitive, as the league table analyses revealed substantial overlap in credible intervals between models, indicating uncertainty in the magnitude of these differences.

Hybrid Pedagogical Models (HPMs) appear particularly suited to supporting tactical cognition and decision-making, Cooperative Learning (CL) shows potential for facilitating technical skill development, Sport Education (SE) demonstrates strong alignment with motivational outcomes, and Teaching Personal and Social Responsibility (TPSR) contributes meaningfully to enjoyment and emotional engagement.

These patterns should not be interpreted as strict hierarchies of effectiveness but rather as tendencies observed within a network largely informed by indirect comparisons with traditional instruction.

This findings suggest a shift in professional practice: from “model adherence,” where instruction is constrained by a single pedagogical framework, to “goal-oriented selection,” where the specific learning objective guides pedagogical choice.

To move beyond empirical correlation and understand the causal logic behind these tendencies, we examine the psycho-pedagogical architecture of the models showing comparatively favourable rankings.

A pivotal finding of this NMA is the identification of instructional duration as a key moderator of efficacy, establishing a clear dose–response relationship. Our subgroup analysis reveals a bifurcation point at 720 min, which serves as a critical threshold for pedagogical impact.

From a practical perspective, 720 min corresponds approximately to one school term of physical education in many educational systems (e.g., two 40–45 min lessons per week over 8–10 weeks). This indicates that the threshold identified in the analysis is not an exceptional intervention dosage but rather aligns with the duration of a sustained instructional unit or season-based curriculum.

Furthermore, instructional duration was associated with reductions in heterogeneity in selected comparisons in the global analysis. Notably, in the >1,440 min subgroup for decision-making, heterogeneity was completely eliminated (I² = 0.0%), confirming that the apparent inconsistency in previous research is largely an artifact of inadequate implementation length rather than model incoherence.

Theoretically, this confirms a period of “pedagogical latency.” Deep learning—whether automatizing complex motor skills or internalizing tactical structures—cannot be rushed. Short, “unit-based” trials fail to allow the core social and psychological mechanisms (e.g., the peer trust in CL or the role identity in SE) to mature sufficiently to impact performance (Psotta and Martin, 2011; Nathan, 2016). Consequently, sufficient dosage (≥720 min) must be regarded not as a variable, but as a prerequisite for success.

Importantly, this finding highlights a logistical consideration for practitioners: short instructional blocks or fragmented units may be insufficient for pedagogical models to activate their intended social, cognitive, and motivational mechanisms. Thus, the “720-min rule” should be interpreted as a realistic planning benchmark rather than an experimental ideal.

Synthesizing these insights, we propose a practical decision framework to guide evidence-informed pedagogical selection. When applying this framework, educators should consider not only model selection but also whether sufficient instructional continuity can be scheduled to meet the minimum effective duration identified in this study. Without this temporal commitment, even theoretically well-aligned models may fail to produce meaningful learning gains.

Scenario A: The “Integrated Performance” Priority (Tactics + Skill). For comprehensive development in complex team sports, Hybrid Pedagogical Models (HPMs) may represent a comparatively favourable option. Our data confirms HPMs as the top-ranked model for both tactical decision-making (SUCRA = 80.1%) and skill execution (SUCRA = 77.2%). By integrating the cognitive scaffolding of TGfU with the situated pressure of SE, HPMs effectively bridge the gap between “knowing” and “doing.”

Scenario B: The “Intrinsic Motivation” Priority. To combat student disengagement or passivity, Teaching Games for Understanding (TGfU) should be prioritized. Ranking highest for motivation (SUCRA = 82.1%), TGfU’s cognitively engaging questioning strategies stimulate the learner’s desire to participate and solve problems, even before technical mastery is achieved.

Scenario C: The “Belonging & Enjoyment” Priority. To build a cohesive class culture and foster emotional attachment to sport, Sport Education (SE) is optimal (Enjoyment SUCRA = 74.0%). Teachers should commit to a full “season” structure to leverage its role-based affiliation, which creates an autonomy-supportive micro-society that sustains long-term enjoyment.

Scenario D: The “Technical Refinement” Priority. For focused, high-repetition refinement of specific motor actions (e.g., isolated drills), Cooperative Learning (CL) remains a powerful tool (Skill SUCRA = 75.5%). Implementing CL with structured peer-assessment protocols activates the “distributed scaffolding” mechanism necessary for correcting technical errors.

Scenario E: The “Social–Emotional Foundation” Priority. For groups requiring emotional regulation, initiate with TPSR to establish a safe, accountable environment before layering in other models.

Crucially, the “720-Minute Rule” applies universally. Rather than representing an intensified intervention, this duration reflects the minimum sustained exposure required for pedagogical mechanisms to stabilize within authentic school timetables.

To support model-based instructional decision-making, PE teachers should adopt a goal-oriented selection framework guided by the principles of “outcome alignment” and “sufficient instructional duration”: (1) For Integrated Tactical and Skills Development: Hybrid Pedagogical Models may provide a productive structure for linking tactical understanding with applied performance, particularly in complex team sports. (2) For Enhancing Student Motivation: Teaching Games for Understanding can be used to stimulate engagement through problem-solving and contextual learning. (3) For Promoting Enjoyment and Belonging: Sport Education is well suited to longer instructional units that emphasize affiliation, identity, and shared responsibility. (4) For Focused Skill Refinement: Cooperative Learning structures can enhance feedback density and peer-supported correction during practice. (5) For Establishing Social–Emotional Foundations: TPSR can be used to cultivate responsibility and classroom climate prior to integrating other models.

Across all scenarios, adequate instructional continuity is essential. The findings suggest that pedagogical models require sustained implementation approaching a full instructional unit (≈720 min) for their mechanisms to take effect; fragmented or short-duration applications may fail to realize anticipated benefits. Rather than prescribing a universally “best” model, this framework supports adaptive alignment between pedagogical approach, intended learning outcome, and realistic school scheduling conditions.

Several limitations warrant acknowledgement. First, despite a multilingual search strategy (English, Chinese, and Spanish), the exclusion of studies in other languages may introduce language bias. Second, the predominance of quasi-experimental designs limits causal inference compared to randomized controlled trials, although it enhances ecological validity in school contexts. Third, a formal certainty appraisal system (e.g., GRADE) was not applied. Fourth, variations in implementation fidelity and teacher expertise were inconsistently reported and could not be modelled directly. Fourth, the review protocol was not prospectively registered in PROSPERO or another registry. Although the study followed PRISMA guidelines and applied predefined eligibility and analytical procedures, the absence of preregistration may increase the risk of unintentional procedural bias.

Fifth, regarding publication bias, Egger’s test indicated significant asymmetry for skill execution (p = 0.0077), suggesting that small studies with null or negative findings may be underrepresented. This limits confidence in the apparent strength of certain models for technical outcomes and reinforces the need for cautious interpretation of ranking results.

Finally, substantial heterogeneity was observed in several comparisons. While random-effects modelling and subgroup analyses addressed this statistically, the remaining variability likely reflects contextual diversity inherent in educational interventions rather than true inconsistency between pedagogical models. Notably, extended-duration interventions reduced heterogeneity, supporting the interpretation that implementation length is a critical explanatory factor.