Inclusion and exclusion criteria for this systematic review and meta-analysis can be found in Table 1.

Duplicates were identified using the reference manager software (EndNote^TM X9, Clarivate Analytics, Philadelphia, PA, USA). Two authors (FMC and HS) independently performed screening of the title, abstract and reference list of each study to locate potentially relevant studies. Additionally, they reviewed the full version of the papers in detail to identify articles that met the selection criteria and those that were excluded. A discussion was made in the cases of discrepancies regarding the selection process with the participation of a third author (AFS).

Electronic databases (PubMed, Scopus, SPORTDiscus, and Web of Science) were searched for relevant publications, from inception up to 28th January 2021. Keywords and synonyms were entered in various combinations in all fields: (youth OR young OR “child^*” OR “adolescent”) AND (“team sport” OR football OR soccer OR futsal OR handball OR volleyball OR basketball OR hockey OR rugby OR cricket OR “water polo” OR lacrosse OR softball OR korfball OR baseball) AND (“small-sided games” OR “sided-games” OR “drill-based games” OR “SSG” OR “conditioned games” OR “small-sided and conditioned games”) AND (“technical” OR “tactic*” OR “skill” OR “ability” OR “behavior*” OR “decision making”). An external expert was contacted to verify the final list of references included in this systematic review and to indicate if there was any study that was not detected through our search.

A data extraction sheet, adapted from the Cochrane Consumers and Communication Review Group's data extraction template (Collaboration, 2016), was used to assess inclusion requirements and subsequently tested on ten randomly selected studies (i.e., pilot testing). This process was conducted by two independent reviewers (FMC and HS). Any disagreement regarding study eligibility was resolved in a discussion between both reviewers and a third author (AFS). Full text articles excluded, with reasons, were recorded. The records were registered in a form created in Microsoft Excel (Microsoft Corporation, Readmon, WA, USA).

Aiming to establish consistency in data analyzing and reporting, only measures that were analyzed three or more times for different articles were included. For technical execution were considered the pre-post intervention outcomes that analyzed the skill level of the player in the specific sport, or the accuracy of skill. For tactical behavior, were considered the pre-post intervention outcomes that assessed individual ability to organize the behavior based on the tactical principles and collective dynamics of the game. The method used for the assessment of technical and tactical outcomes was also extracted. Adverse effects were also extracted as secondary outcome, in case of any reported. Additionally, the following information was extracted from the included studies: (i) number of participants (n), age (years), competitive level (if available) and sex; (ii) the SSGs format and pitch size (if available); (iii) period of intervention (number of weeks) and number of sessions per week (n/w); and (iv) regimen of intervention (work duration, work intensity, modality, relief duration, relief intensity, repetitions and series, between-set recovery).

The version 2 of the Cochrane risk-of-bias tool for randomized trials (RoB2) (Sterne et al., 2019) was used to assess the risk of bias in the included randomized-controlled trials. Five dimensions are inspected in this assessment tool: (i) bias arising from the randomization process; (ii) bias due to deviations from intended interventions; (iii) bias due to missing outcome data; (iv) bias in measurement of the outcome; and (v) bias in selection of the reported result. Using RoB2 a qualitative synthesis was performed. Two of the authors (JA and HS) independently assessed the risk of bias. Any disagreement in the rating was resolved through discussion and by a third author (FMC).

The Cochrane risk of bias in non-randomized studies of interventions (ROBINS-I) was used to assess the risk of bias in included non-randomized intervention studies (Sterne et al., 2016). Three domains are analyzed in this assessment tool: (i) pre-intervention (bias due to confounding; bias in selection of participants into the study); (ii) at intervention (bias in classification of interventions); and (iii) post-intervention (bias due to deviations from intended interventions; bias due to missing data; bias in measurement of outcomes; bias in selection of the reported results). Two of the authors (JA and HS) independently assessed the risk of bias. Any disagreement in the rating was resolved through discussion and by a third author (FMC).

We followed previously established methods (Clemente et al., 2021). Briefly, analysis and interpretation of results were only conducted in the case of at least three studies provided baseline and follow-up data for the same measure. Pre-training and post-training means and standard deviations (SD) for dependent variables were used to calculate effect sizes (ES; Hedge's g) for each outcome measure in the SSGs and control groups. Data were standardized using post-intervention SD values. The random-effects model was used to account for differences between studies that might impact the SSG-based effect (Deeks et al., 2008; Kontopantelis et al., 2013). The ES values are presented with 95% confidence intervals (CI). Calculated ES were interpreted using the following scale: <0.2, trivial; 0.2–0.6, small; >0.6–1.2, moderate; >1.2–2.0, large; >2.0–4.0, very large; >4.0, extremely large (Hopkins et al., 2009). Heterogeneity was assessed using the I² statistic, with values of <25%, 25–75%, and >75% considered to represent low, moderate, and high levels of heterogeneity, respectively (Higgins and Thompson, 2002). The risk of bias was explored using the extended Egger's test (Egger et al., 1997). When bias was present, the trim and fill method was applied (Duval and Tweedie, 2000), in which case L0 was assumed as the default estimator for missing studies (Shi and Lin, 2019). All analyses were carried out using the Comprehensive Meta-Analysis software (version 2; Biostat, Englewood, NJ, USA). Statistical significance was set at p ≤ 0.05.

Moderated analyses were planned to use a random-effects model and independently calculated single factor analysis. When possible, the median split technique was planned (Moran et al., 2019b). Moderator analysis was considered for the total session during interventions.

The searching of databases identified an initial 803 titles. Duplicates (160 references) were subsequently removed either automatically or manually. The remaining 643 articles were screened for their relevance based on titles and abstracts, resulting in the removal of a further 593 studies. The full texts of the remaining 50 articles were examined diligently. After reading full texts, a further 43 studies were excluded owing to a number of reasons: did not develop an intervention program (n=35); lack of control group (n = 3); did not included technical/tactical outcome (n = 3); insufficient statistical data (n = 1); did not reported pre-post data (n = 1). After methodological assessment, one article was excluded by critical risk of bias. None of the included articles were written in other language than English. Six articles were eligible for the systematic review and meta-analysis (Figure 1). The included articles provided mean and standard deviation pre-post-training data for at least the main outcome.

The randomized-controlled trials included in this study were analyzed with the RoB2 tool and the results can be found in Table 2.

The assessment of the non-randomized studies intervention studies can be found in Table 3. Since one article (Práxedes et al., 2016) presented overall critical evaluation (based on the assessment criteria of the instrument), and in accordance with Cochrane manual (Higgins et al., 2019), was not included in qualitative and quantitative synthesis.

Although only three of the six included articles in the qualitative and quantitative synthesis were randomized (Radziminski et al., 2013; Chaouachi et al., 2014; Daga et al., 2020), none of the six included articles reported significant differences between control ad experimental groups at baseline. The characteristics of the six studies included in the meta-analysis can be found in Table 4. Additionally, the details of the SSGs-based programs can be found in Table 5 and the other training interventions in Table 6. The three included randomized-controlled studies involved three individual experimental groups and 43 participants, and 55 participants in the four control groups. The three included non-randomized studies involved three individual experimental groups and 39 participants, and 35 participants in the three control groups.

A summary of the included studies and results of technical execution reported before and after SSGs-based compared to control intervention are provided in Table 7.

Six studies provided data for technical execution, involving six experimental and seven control groups (pooled n = 172). Results showed a small effect of SSGs on technical execution (ES = 0.59; 95% CI = 0.29 to 0.89; p < 0.001; I² = 0.0%; Egger's test p = 0.590; Figure 2) when compared to controls.

The single training factor analysis revealed a similar (p = 0.433) moderate (ES = 0.68, four study groups) and small (ES = 0.44, three study groups) improvement in technical execution after >17 and <17 SSG training sessions, respectively.

The within-group meta-analysis showed a moderate effect on technical execution after both SSGs (six groups, pooled n = 82; ES = 0.106; 95% CI = 0.63 to 1.48; p < 0.001; I² = 75.9%; Egger's test p = 0.090; Figure 3A), and control conditions (seven groups, pooled n = 90; ES = 0.65; 95% CI = 0.26 to 1.04; p = 0.001; I² = 80.3%; Egger's test p = 0.003, unchanged after the trim and fill method was applied; Figure 3B).

None of the six included studies presented tactical behavior-related outcomes. In screening only one study presented tactical behavior outcome (Práxedes et al., 2016), although was excluded based on a critical score obtained after using the ROBINS tool.

SSGs were effective for improving players' technical execution in the selected studies. This result is in line with the assumption that SSGs might nurture youth sports players' technical development, as they provide players with a game-based environment in which they are constantly required to apply motor actions to solve emerging problems. Interestingly, all the selected studies (Radziminski et al., 2013; Chaouachi et al., 2014; Jastrzebski et al., 2014; Trajkovic et al., 2017; Arslan et al., 2020; Daga et al., 2020) adopted formats ranging from 1 vs. 1 to 3 vs. 3. These smaller SSG formats were reported to increase players' ball engagement and provide more variable stimuli (Clemente et al., 2019; Caso and van der Kamp, 2020), which might explain the performance enhancements. The current review, therefore, reinforces the positive role that SSGs play in technical development and encourages coaches to adopt this tool in their training programs.

Although positive adaptations in technical execution were reported when comparing control vs. intervention groups, the magnitude of these differences was small. We argue that the type of measurement regarding technical execution might explain this reduced statistical effect. Specifically, most of the adopted tests are based on measuring agility with the ball or change-of-direction abilities (Radziminski et al., 2013; Chaouachi et al., 2014; Jastrzebski et al., 2014; Arslan et al., 2020; Daga et al., 2020), while others adopted a battery of closed-environment tests (Radziminski et al., 2013; Jastrzebski et al., 2014). In both cases, the measurements were taken out of the game context. Contrarily, in SSGs, technical executions are always integrated with tactical decisions, so larger differences between control and intervention groups could be expected if the measurements were taken in a more representative game-based scenario. Interestingly, a previous study found no significant correlations between in-game tactical-technical dribbling performance and players' dribbling performance in a technical test without the game context (Praça et al., 2015), reinforcing the previous rationale.

Also, the performance in the agility test is influenced by a cognitive aspect (perceptual and decision-making processes) and by players' abilities to quickly change direction (Young et al., 2015; Pojskic et al., 2018). Therefore, technical execution is not directly measured. Consequently, players might experience an improvement in the testing performance, which is not fully explained by an improvement in ball dribbling technical execution. Therefore, the magnitude of the difference between control and experimental groups is expected to be reduced. For both mentioned reasons, future studies should look to develop and adopt valid and reliable in-game-based instruments to measure improvements in technical execution.

Moderate pre-to-post improvements were observed within control and intervention groups. While improving technical execution is an expected outcome for the intervention group, the same did not apply to the control groups. Besides the abovementioned issues regarding the tests and measurements, we argue that this unexpected outcome is linked to the way the training programs were designed. Specifically, most of the selected studies included game-based activities (in which technical executions were required) in the training program of the control group as players took part in sport-specific team training (Radziminski et al., 2013; Chaouachi et al., 2014; Jastrzebski et al., 2014; Trajkovic et al., 2017; Arslan et al., 2020). In the selected studies, however, there was no information regarding the content of such sport-specific team training.

Considering how training programs are usually developed, we might expect players to engage in technical execution training, which explains the improvement in the control group. However, as the magnitude of the difference was higher in the intervention groups than in the control ones (ES 1.06 and 0.65, respectively), the SSGs boosted the technical execution enhancements attained through regular training regimens, which reinforces their positive role. However, as high heterogeneity between the studies was observed in the control and intervention groups, generalizations are not recommended and future studies are required.

Finally, a similar small-to-moderate improvement in technical execution was observed after >17 and <17 SSG training sessions. Although longer interventions are expected to generate greater improvements, two issues might explain the current results. While one study (Trajkovic et al., 2017) included 36 sessions, all the other >17 training session interventions varied from 18 to 24 sessions. Therefore, the magnitude of the difference in the length of the training program was likely to be not enough to elicit different adaptations. Moreover, none of the selected studies introduced non-arbitrary pedagogical strategies to control and progress the training contents in the intervention group. Adjusting the task complexity to players' current levels is expected to facilitate the learning (Machado et al., 2020). Therefore, the absence of a control condition for the training process might have reduced the expected impact of the longest intervention programs on players' technical execution.

While the abovementioned rationale seems to explain the current results, the two longest intervention programs (Trajkovic et al., 2017; Daga et al., 2020) were coincidently the two most statistically weighted studies in the meta-analysis (see Figure 2), as they showed the most relevant intervention impact on the responses (see hedge's G and 95% CI) and the lowest variance across the studies.

In team sports, learning is not a linear process, and differences are age and maturity-dependent (Fransen et al., 2017). For this reason, although the current analysis did not show differences between >17 and <17 SSG training session programs, short training programs might not be enough to favor the development of technical skills in late-developed players, who require longer training periods to achieve such improvements.

No selected study examined the effects of SSGs on tactical behavior. This result indicates the need for further research on this topic for an evidence-based prescription. We acknowledge that the absence of evidence must not be understood as evidence of absence, as previous game-based approaches that are strongly based on small-sided games have been effective in nurturing team sports players' tactical skills (Kinnerk et al., 2018). There are two main possible causes for this absence of studies: the difficulty in choosing instruments to evaluate tactical behaviors and the challenge of designing interventions with high external validity in team sports.

Previous instruments used to evaluate tactical behaviors are largely based on observational designs (Oslin et al., 1998; Costa et al., 2011; González-Víllora et al., 2015). Although this low-cost strategy allows performance indicators to be obtaining in team sports, there are issues regarding their reliability, as they account for discrete events. Specifically, low-frequency events are less likely to present a good test-retest reliability, which increases the variability of the responses and reduces the power of the analysis. For example, weak intersession intraclass correlation coefficients were reported for the frequency of tactical principles in a previous study in soccer (Bredt et al., 2016).

Moving toward contemporary technologies, such as positional analysis, might allow the researchers to access more reliable measures, which will encourage the development of intervention studies. Previous studies adopted this perspective using linear and nonlinear modeling to detect players' and teams' behaviors in game-based contexts (Ramos et al., 2017; Ribeiro et al., 2020) and showed interesting pathways for future research.

This review provides interesting practical applications for coaches and researchers. Specifically, because significant and beneficial effects of using SSG training interventions for improving technical execution were observed, coaches are encouraged to systematically include SSGs in training programs to nurture players' technical skills. During SSGs, players experience high variability in their actions, which may also provide players with a more adaptative technical pattern in line with the demands of the matches in team sports.

On the other hand, limitations must be accounted for. From the selected studies, five addressed improvements in technical execution in soccer, while only one included volleyball athletes. Therefore, although our conclusions seem to be more generalizable to soccer, studies in other sports are required. Also, the studies did not include players of different age groups and did not include elite youth players. As the technical development of high-level athletes might be different from novice youth-academy players, future studies in elite groups are recommended. Also, in-game technical execution assessments are still required, as there is no evidence that the SSGs are more effective than control training for improving in-game performance. Finally, constantly developing valid and reliable instruments to assess players' behaviors related to game-based tasks will benefit researchers and practitioners by increasing the quality of the information used to adjust training contents and complexity.

The only selected study that considered tactical behaviors was excluded due to methodological issues. The lack of control for confounding variables might have biased the results, which reduced the methodological quality and demands caution when interpreting such results. As there were no selected studies regarding players' tactical behavior development after SSGs-based training programs, we encourage researchers to search for suitable experimental designs to address this issue.

Finally, integrated analysis covering the impacts of SSG-based on programs in multiple dimensions (e.g., physiological/physical, technical, and tactical) must be considered, namely integrating into the analysis the dose-response effects by monitoring the load imposed (Gomez-Carmona et al., 2018; Gómez-Carmona et al., 2019) and the accumulated effects of technical/tactical responses to specific drills (Fernández-Espínola et al., 2020).