Using Age- and Size-Corrected Measures of Technical Skill to Better Assess the Performances of Youth Soccer Players

Abstract

Youth soccer academies are dominated by the older players in each annual age cohort because they are judged to be better at the time of selection. Failing to identify talented players because they are simply younger in their cohort is a problem of both discrimination and poor practice. One potential method for addressing such biases is to develop and use age- and size-corrected assessments of individual players using traits closely associated with match success. In this study, we quantified the relationship between age and size with individual passing and control performance in six different tests for 170 players between 10 and 20 years old from a Tier 1 academy in Brazil. Passing tests were significantly repeatable and performance varied among tests (df = 5; F = 432.2; p < 0.001). Overall passing performance (PC_P1)—based on all tests—was significantly positively associated with age (R² = 0.42, t = 10.67; p < 0.001), height (R² = 0.19, t = 6.13; p < 0.001) and mass (R² = 0.23, t = 6.90, p < 0.001). In addition, tests of passing and control could discriminate among groups of differing playing levels (test 1: F_(2,116) = 55.2, p < 0.001; test 3: F_(2,116) = 12.0, p < 0.001). Normative algorithms from this study can be used to compare athletes during selection trials and against an elite group, after taking age and size into account, and using such algorithms could vastly reduce the insipid age-biases that plague youth football.

1. Introduction

Talent identification programs aim to select individuals most likely to become successful professional adult players. The opinion of expert scouts who observe and evaluate potential recruits during matches and training sessions is still the dominant method of identifying talented youth players [1,2,3,4]. Because scouts must often assess players from just single matches or sessions, even the most experienced professional scouts can be vulnerable to biases and errors in judgement with such limited information [5,6]. Further complicating the task of scouts is that youth players compete within annual age cohorts where there is substantial variation in age and physical maturity among the players. For example, in an age cohort of under 13 s, those players born in the month of January will be approximately 10% older than those born in December. Older players are more likely to be judged as better simply because they are mature, and physically stronger—and, consequently, youth football academies across the world are dominated by individuals born in the first half of the year [7]. This selection bias—referred to as the relative age-effect—is one of the most insipid forms of discrimination in team sports.

Selection biases in youth football are not just a problem of discrimination and poor practice but they are likely to result in many talented young players missing out on being selected to representative teams or elite academies. Any method that increases the probability that promising players are selected, even when they are younger and less physically developed than their peers, could provide an advantage for clubs adopting this approach. Bio-banding is one attempt to mediate the effects of age-biases in team sports where players are placed into “bands” or groups based on stages of physical development rather than age for specific competitions and training [8,9,10,11,12,13]. Despite the positive attitude towards these measures, “bio-banding” is rarely, if ever, used when players are being selected for academies [8,9,10,11,12,13]. An alternative approach to addressing biases in youth selection is to develop and use age- and size-corrected assessments of individual players [14]. By comparing players using metrics of performance that are corrected for age (in days) and size, one could then provide fairer, more robust and accurate assessments of relative performances. But which metrics of relative performance should be utilised when accounting for age and size during talent identification? Assessments of individual performance in soccer-specific skill tests could be a profitable pathway for developing such a quantitative assessment tool. Individual performance in tests of closed skill are more reliable predictors of success in match-realistic games than athletic traits like speed or strength [15]. For example, individuals with higher performances in tests of closed skill are more likely to be better in 1v1 games [16], small-sided possession games [17] and 11v11 match success [15]. By quantifying how age and size affect individual performances in tests of skill associated with success in matches, or match-realistic scenarios, one could provide a tool for comparing individuals across the variation in ages and sizes observed within any age cohort. Rather than replacing the expertise of scouts in talent identification, these data could then be used to increase the information available when selecting players and reducing the probability of costly errors in selection.

Quantifying how age- and size- affect a player’s ability to control and pass the ball is one such potential tool for improving talent identification. The ability to quickly control the ball and then pass it accurately to a teammate is one of the most frequent and important technical skills performed during a football game [18]. Every professional player must be able to perform this skilled action with speed, accuracy, and consistency. Most football matches consist of more than 600 individual passes [19] and an individual’s ability to perform such an action is critical to the success of a team. Passing is also likely to be as equally important to all out-field player positions given it is the principal way for the team to move the ball towards the opponents’ goal [20]. Based on data from 6000 games and 10 million events (dribbling, passing, shooting, among others) across six European leagues, Pappalardo and Cintia [21] found the number of passes, shots and goalkeeping actions were the strongest predictors of team success, and teams that were able to produce more passes than their opponents, as well as less tackles and less fouls, were more likely to succeed. Being able to derive a series of algorithms to describe how age and size affect passing performance for an elite group of youth players will allow one to compare the performances of individuals after taking age and size into account.

In this study, we quantified the relationship between age and size with individual passing and control performance for 170 players between 10 and 20 years old from an elite academy in Brazil. For each player, we measured their passing and control performance on six different passing tests. Our aim was to describe a series of normative relationships between age and size with passing performance so that others can compare their performance with this elite group of athletes. Finally, we evaluated the ability of the passing tests to discriminate among players of different levels of competition—the test’s construct validity—by comparing the performance of our elite group with athletes from two lower ranked soccer academies in two of the passing tests.

2. Methods

2.1. Subjects

We recorded the performance of 170 junior players from an elite football academy in Brazil that competes in their state and national competitions. We also measured the performance of players from two smaller academies that compete in their state competitions only. All players and guardians gave consent to be involved in this study, which was in accordance with ethical protocols for The University of Queensland, Australia and the University of São Paulo (Ribeirão Preto), Brazil. Each player’s age (14.65 ± 2.5 years; range = 10.1–20.5 years), mass (59.3 ± 15.11 kg, range = 31.7–93.0 kg) and height (1.68 ± 0.14 m, range = 1.35–1.96 m) were recorded on the first day of assessment.

2.2. Study Design

We measured the passing performance of seven age cohorts (U12, U13, U14, U15, U16, U17, U20). Each age cohort of players attended one two-hour session. Players proceeded through their normal 15 min warm-up routine with coaches prior to conducting the tests. Depending on the number of players in a cohort, players were split into groups of three or four, with groups rotating through each station in turn. Although each group proceeded through tests in the same sequence, because each group started at a randomly assigned starting station, the order of testing differed among each group. The tests were arranged in the following sequence: Test 1, Test 3, Test 4, Test 2, Test 6 and then Test 5. Before being tested, players were allowed a brief period to familiarize themselves with each task. Players completed two trials per station, resting between each trial.

2.3. Passing and Control Tests

The player’s ability to execute specific passing skills was assessed with six passing tests. The validity and repeatability of these passing tests have been reported previously [17]. Each test required players to alternately pass a size 5 ball between two boards equally spaced 4 or 8 m from the player’s position. Players were instructed to hit the rebound boards with the ball as many times as possible during the 45 s trial, while using the specified technique for each test [17]. Four additional balls were placed approximately 1 m behind players so If a player missed the target, they could quickly continue the trial with a replacement ball. If a player missed a board 5 times, the trial immediately ended. The number of successful hits and the number of times the skill was not correctly executed (errors) were recorded. Regardless of the magnitude of the mistake, a maximum of one error could be attributed per pass attempt. For example, if a player took one or more extra (or fewer) touches than specified then this was scored as only one error. To account for errors, each player’s number of successful hits was adjusted with the formula Passingmetric = number_of_hits − (0.5 × errors). Players executed each passing test two times and for each test the average of their trials was their measure of performance for that test.

2.4. Passing Tests

Test 1—There was a 90-degree angle between rebound boards and the distance from each board to the player’s starting central position was 4 m (Figure 1). Players started with a right foot pass to the right-hand board, then controlled the ball with one touch using the inside of the left foot followed by a pass with the same foot to the left-hand board. After this, player’s controlled the ball with one touch using the inside of the right foot then passed with same the foot to the right-hand board. The test continued until the 45 s elapses.

Test 2—Identical to test 1, except the distance from each board to the player’s starting position was 8 m.

Test 3—Identical to test 1, except the angle between the two rebound boards was 135-degrees.

Test 4—Identical to test 1, except the angle between the two rebound boards was 135-degrees, and the distance from each board to the player’s starting position was 8 m.

Test 5—45-degree angle between rebound boards and distance from each board to the player’s starting position was 4 m. Players passed with the right foot to the right-hand board and then used a one-touch pass with the left foot to the left-hand board. Players continued until the trial ended. Players could commence the trial with a pass to either board.

Test 6—Identical to test 5, except the distance between each board and the player’s starting position was 8 m.

2.5. Statistical Analyses

We used principal component analysis (PCA) to characterize patterns of variation among player’s height, mass and age (birth date was converted to decimal age) creating a multivariate age and size index (ASI) [14]. All variables were first standardized to a mean of zero and a standard deviation of one. The first component (PC_ASI) accounted for 89.4% of the variance in the data (Supplementary Table S1) and all vectors loaded in the same direction with larger positive values indicating greater age and size.

Estimates of repeatability for each test were calculated using intra-class correlation coefficients (two-way mixed effects model) with the “psych” package of R. For summary statistics, a Welch’s ANOVA followed by Games–Howell post hoc test [22] detected differences in performance among the passing tests, and paired correlations were estimated with the “Hmisc” package of R [23]. To confirm that each passing test could discriminate among different age groups, we ran separate ANOVA’s followed by Tukey Honest Significant Difference (95% confidence intervals) [24] to detect differences in performance among teams. Prior to correlation and ANOVA analyses, data from tests one, two and three were corrected with a square transformation to achieve normality. All tests were then included in a PCA to characterize patterns of variation among the correlated measures of passing performance. Prior to PCA, data from each passing test were standardized to a mean of zero and a standard deviation of one. The first component of the PCA (PC_P1) explained 56.1% of the variation observed in the data (Supplementary Table S2); because all vectors of PC_P1 loaded in the same direction, and larger positive values were indicative of higher passing performance over all tests, PC_P1 represents overall passing and control performance. The second component of the PCA (PC_P2) explained 11.2% of the variation.

To determine the independent effects of age, height, mass and ASI on performance, we ran separate linear models [24] for each of these variables, estimating their effect on overall passing and control performance (PC_P1). Passing performance (PC_P1) was first corrected with a square transformation to achieve normality.

To evaluate the passing tests’ construct validity, we compared the performance of athletes from the elite Brazilian academy with athletes from two lower ranked academies from Brazil, referred to subsequently as sub-elite A and sub-elite B. Both sub-elite clubs completed test 1 and test 3 as described earlier, and the age of players were recorded on the day of testing—sub-elite A (n = 51, 12.96 ± 1.30, range = 10.61–15.10 years) and sub-elite B academy (n = 51, 16.35 ± 1.97, range = 12.69–19.79 years). To compare performance among the academies, we needed to account for the varying distribution of ages among them. To achieve this, we first ran separate linear models estimating the effect of age on test 1 performance (square transformed) and test 3 performance (square transformed), using only data from the elite academy. We then used parameter estimates from these models to calculate residuals, separately for each test, for players from all three academies. These residuals indicate the difference between actual player performance and the performance that is expected for their age. Then, an ANOVA followed by Tukey Honest Significant Difference (95% confidence intervals) identified differences among the academies on residual performance, separately for each test.

Normality of residuals used in analyses was assessed using the Shapiro–Wilk test and visual inspections of Q-Q plots using R. We used power analyses to determine if we had sufficient statistical power to detect significant effects for both continuous traits (regression) and among groups (ANOVA) [24]. We assumed a medium effect size in each case, an alpha level of 0.05 and power of 0.80. Results indicated that we had sufficient statistical power across all analyses conducted. Data are presented as means ± standard deviations.

3. Results

Overall passing performance (PC_P1) based on all tests was significantly positively associated with age (R² = 0.42, t = 10.67; p < 0.001), height (R² = 0.19, t = 6.13; p < 0.001) and mass (R² = 0.23, t = 6.90, p < 0.001) (Figure 2). A player’s age and size index (ASI) was also significantly positively associated with their overall passing performance (R² = 0.31, t = 8.29, p < 0.001) (

Table 1. Independent effects of age, height, mass and ASI on overall passing performance (PC_P1). Each pairing of intercept and β represents a separate linear model. Passing performance (PC_P1) was square transformed [y = (x + 10)²] prior to analyses. * p < 0.001.

	Intercept	β	R2
Age (years)	−37.867 *	9.515 *	0.42
Height (cm)	−82.415 *	1.103 *	0.19
Mass (kg)	35.093 *	1.152 *	0.23
ASI	102.83 *	12.038 *	0.31

).

All individual passing tests were significantly repeatable (Supplementary Table S3) and performance significantly varied among the passing tests (df = 5; F = 432.2; p < 0.001) (

Table 2. Performance on each passing test for the elite academy. Differences among tests identified with Welch’s ANOVA and the Games–Howell post hoc test.

	n	Mean ± SD (Passes/Test)
Test 1	161	22.84 ± 4.60
Test 2	164	12.45 ± 2.38
Test 3	159	19.58 ± 4.08
Test 4	164	11.07 ± 2.70
Test 5	165	27.36 ± 5.38
Test 6	164	12.39 ± 3.18

Post hoc significant differences between Test 1 and Test 2, 3, 4, 5, 6; Test 2 and Test 3, 4, 5; Test 3 and Test 4, 5, 6; Test 4 and Test 5, 6; Test 5 and 6. All significant at p < 0.001.

). Differences in performance among the tests that were identified with Welch’s ANOVA were examined using the Games–Howell post hoc test (Table 2). Players completed the most passes on average in test 5 (27.4 ± 5.38 passes) where they turned through 45 degrees when passing to a board 4 m from the player. In the identical test over 8 m (test 6), the number of passes completed was significantly lower (12.4 ± 3.2 passes). Players completed 22.8 ± 4.6 passes when turning through 90 degrees and the rebound board was 4 m from the player (test 1), which was significantly greater than the same test over 8 m (12.5 ± 2.4 passes). In addition, players completed 19.6 ± 4.1 passes when turning through 135 degrees and the rebound board was 4 m from the player (test 3), which was significantly greater than the same test over 8 m (11.1 ± 2.7 passes).

Significant positive correlations were found among all paired tests, ranging from 0.39 to 0.59 (Supplementary Table S4). Significant differences among the cohorts were found on all tests, indicating that each test could discriminate among different age groups (

Table 3. Performance on passing tests (passes/test) for each age group from the soccer academy. Differences among age groups identified with ANOVA and Tukey multiple comparisons. Post hoc comparisons between ages groups for each test were conducted, and significant differences (p < 0.05) to U12 are denoted by ^a, U13 denoted by ^b, U14 denoted by ^c, U15 denoted by ^d, U16 denoted by ^e, U17 denoted by ^f.

Team	Test 1	Test 2	Test 3	Test 4	Test 5	Test 6
U12	15.8 ± 4.6 (n = 22)	10.9 ± 2.3 (n = 24)	15.7 ± 4.9 (n = 21)	9.26 ± 2.0 (n = 24)	22.6 ± 5.0 (n = 24)	15.7 ± 4.9 (n = 21)
U13	21.7 ± 2.6 a (n = 22)	11.0 ± 2.0 (n = 22)	20.6 ± 3.0 a (n = 22)	9.05 ± 2.4 (n = 22)	23.4 ± 5.4 (n = 21)	20.6 ± 3.0 (n = 22)
U14	22.4 ± 3.1 a (n = 22)	11.6 ± 2.3 (n = 22)	19.3 ± 3.2 (n = 21)	10.7 ± 2.2 (n = 22)	27.7 ± 3.8 ab (n = 24)	19.3 ± 3.2 (n = 21)
U15	24.3 ± 2.2 a (n = 24)	12.7 ± 2.3 (n = 24)	18.7 ± 3.5 (n = 24)	11. 5 ± 2.3 ab (n = 24)	28.4 ± 2.3 ab (n = 24)	18.7 ± 3.5 (n = 24)
U16	25.0 ± 3.8 ab (n = 19)	12.9 ± 2.1 (n = 19)	20.2 ± 5.5 a (n = 19)	11.3 ± 3.2 b (n = 19)	30.9 ± 5.1 ab (n = 20)	20.2 ± 5.5 (n = 19)
U17	23.4 ± 3.7 a (n = 20)	13.2 ± 2.3 ab (n = 20)	20.5 ± 2.8 a (n = 20)	12.2 ± 2.5 ab (n = 20)	30.6 ± 4.3 ab (n = 20)	20.5 ± 2.8 abcde (n = 20)
U20	26.1 ± 3.4 abcf (N = 32)	14.2 ± 1.5 abc (N = 33)	21.3 ± 3.0 a (n = 32)	12. 9 ± 2.0 ab (n = 33)	28.3 ± 4.5 ab (n = 32)	21.3 ± 3.0 abcde (n = 32)

). On all tests, the older cohorts tended to perform better than the younger cohorts (Table 3).

The analyses of residuals of passing performance indicate that test 1 (F = 55.2, p < 0.001) and test 3 (F = 12.0, p < 0.001) can discriminate between elite and sub-elite groups (

Table 4. Effect of age on passing test 1 and 3. Each pairing of intercept and β represents a separate linear model. Only data from the elite academy were included in these models and data from both tests were square transformed.

	Intercept	β	R2
Test 1	−174.17 *	48.39 **	0.35
Test 3	106.14	19.79 **	0.10

* p < 0.05, ** p < 0.001.

). On both tests, the elite group’s performance was significantly better than both the sub-elite groups (Figure 3). The elite group had significantly higher passing scores than the two sub-elite groups on both tests 1 and 2 (p < 0.001). The average scores for the elite group were 22.84 ± 4.60 and 19.58 ± 4.08 for tests 1 and 3, respectively. The average scores for sub-elite group A and B for test 1 were 19.73 ± 2.71 and 19.80 ± 2.32, respectively. The passing scores for sub-elite group A and B for test 3 were 16.58 ± 3.09 and 18.93 ± 2.55, respectively.

4. Discussion

We found that an individual’s passing and control performance was positively significantly associated with their age, height, mass and the combined multivariate description of these traits (ASI). Among these traits, the best predictor of passing and control performance was the age of the individual athlete, and we could estimate performance with the formula passing = −37.87 + (9.52 × age). This algorithm can now be used to control for the effects of age on passing performances when assessing and comparing talented youth players during any selection trials. These age-corrected metrics represent an important pathway for reducing the age-discrimination biases that plague player selections in youth football. Focusing on age-corrected metrics for these specific passing tests is particularly promising because previous studies show that they correlate highly with success in game-realistic possession games [17]. For example, individual passing and control performance explained approximately 60% of the variance in a player’s average number of successful passes and percentage passing success in 3v1 possession games [17]. In addition, individuals that were better at tests of passing and dribbling performance were more likely to have greater passing success in small-sided games [17]. Previously developed tests of passing and control performance, such as the Loughborough Soccer Passing Test (LSPT) [25] and those used by the Portuguese Football Federation [26] and the Football Association of Finland [27], could also be utilized in a similar way to provide age-corrected comparisons. However, at this time, no normative relationships describing the effects of age and size on passing and control performance have been developed for these alternative tests.

Passing and controlling the ball during a match is a highly dynamic task, as players must quickly change their body orientation to receive the ball, control it and position themselves to pass the ball in the desired direction. “Rondo” possession games are used in almost all professional and youth training sessions to improve the ability of players to undertake this complex activity, which is routinely required by players in elite professional football. For this reason, a closed skill where high performance is associated with an ability to rapidly control and pass the ball under pressure is an excellent starting point for developing a comprehensive quantitative framework for age- and size-corrected player assessments. The obvious utility of these simple, well-designed closed skill tests is that they can be easily reproduced so that each player can be tested under almost identical conditions. However, improvements in the performance of individuals in these closed skill tests may not directly lead to better performances in similar tasks in matches. This is a key limitation of using closed skill tests, and it remains to be tested whether there is a causal link between performance in the tests and in matches. However, we do know that the closed skill tests are repeatable measures of individual performance [20] and, as demonstrated in this study, can discriminate among groups of different standards. The performance of athletes from three clubs were compared in two of the six passing tests and showed those athletes from the elite academy consistently had better performances than the other two sub-elite academies.

The utility of such technical performance tests for player development and coaching will be markedly improved if one could also incorporate elements of unpredictability and decision-making into the test. Testing a player under increased cognitive loads can better replicate the game-realism of a testing environment and help improve a player’s capacity to replicate the motor task under the stress of a game. Dual-action motor tasks have been used previously to help players improve their capacity to perform motor tasks under cognitive loads [28]. Utilizing such designs could simultaneously quantify each individual’s capacity to perform under stress and provide an ideal environment to track their improvements through time. One potential avenue for building cognitive complexity into passing performance tasks would be to have multiple boards where the order of the boards is randomized using lights embedded into the targets. By utilizing a set-up where players need to constantly scan among multiple potential targets and then provide a subsequent pass towards a randomly activated board would provide the additional complexity and cognitive load that is closer to match relevant scenarios.

Identifying those traits associated with an individual’s success in passing and control could potentially help coaches improve their team’s success by targeting and developing those traits that directly improve performance. Here, age was a better predictor of passing performance than ASI, while others [14] found that ASI was a better predictor of kicking speed than age, height or mass alone. This suggests that ASI may better predict power-based activities such as kick speed and sprinting, while age alone better predicts technical traits like passing and control or dribbling. While physical conditioning and strength training may improve kicking and sprinting performance, it may have limited influence on passing and control ability. Instead, identifying and training the underlying traits associated with passing and control success is likely to increase performance.

Data-driven selection practices are now common at the elite professional level of football but are rarely used at youth levels. The age- and size-biases that plague the selection of players at youth level are unlikely to change without the adoption of data-driven approaches. By comparing players using metrics of performance that are corrected for age and size, one could then provide fairer, more robust and accurate assessments of relative performances. In this study, we presented a testing protocol with normative data to assess the passing and control performance of youth soccer players. With the equations we provide, one can compare a players’ passing performance against the expected performance based on their age and physical development. Rather than replacing the expertise of scouts in talent identification, these data could be used to increase the information available when selecting players and reduce the probability of costly errors in selection. Combining assessments of players that are free from the biases of age and size that dominate youth selection should become standard practice for fairness and the long-term success of elite academies. As with every innovation in player coaching and development, it is likely to be only the first clubs and associations to adopt such practices that will substantially benefit—the others will be left behind and need to play catch up.