The present study

This study pursued two objectives with a focus on speed abilities and technical skills while also considering anthropometric and relative age characteristics. First, the predictive validity of motor tests, anthropometric variables, and relative age were investigated over a long-term period from shortly following the entrance into the German TID program (U12) until after the transition into adult soccer (> U19). Second, this study analyzed the long-term predictive power of the two theoretical (latent) performance factors (i.e., speed and technique) with simultaneous consideration of maturation-related characteristics such as relative age and anthropometric variables.

Sample and design

This prospective cohort study investigated the data from N = 14,178 male competence centers athletes (11.32 ± 0.28 years old) who belong to the top 4% of the boys playing soccer at this age [28]. All individuals participated in motor diagnostics—which were implemented nationwide in the German TID program in 2004 (see [29])–in the U12 age group in the fall of 2004, 2005 or 2006. The U12 motor tests’ results served as predictors for players’ later success at the adult level in the 2014/15 soccer season leading to a prognostic period of eight, nine or ten years, respectively. The players’ height, weight, and relative age were considered as covariates (descriptive sample characteristics are presented in Fig 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Descriptive and inferential statistics for the diagnostics in U12 separated by players’ adult performance level.

*p < .05, **p < .01, ***p < .001. The motor tests (and relative age) are negatively coded, i.e. lower values indicate better performances (or an older age) for these predictors.

https://doi.org/10.1371/journal.pone.0182211.g001

Before entering the TID program, players’ parents provided written informed consent for the recording and scientific use of the data collected in the motor diagnostics. DFB staff members (coaches with at least UEFA B-License) conducted the motor tests and assessed the predictor variables of the present study. The DFB provided the authors with players’ data of three birth cohorts (1993, 1994, and 1995). The university’s ethics department and the scientific board of the DFB approved the implementation of this study.

Measures

Statistical analysis

Data were analyzed using SPSS version 22 and Mplus version 7.1. With regard to the prognostic relevance of the motor diagnostics, the significance of mean differences between players achieving different APLs in the 2014/2015 season was determined by a one-way MANOVA including the five single motor tests as dependent variables (objective 1). Additionally, a corresponding MANCOVA was conducted incorporating height, weight and relative age as maturation-related covariates.

In order to provide robust results regarding the predictors’ prognostic relevance, the dataset of three birth cohorts (1993, 1994, 1995) was accumulated, so that the sample of the highest selection levels (APL1) achieved a sufficient number. However, the cohorts may confound the analysis of motor performances’ differences between selection levels [33]. To investigate this assumption, a two-way MANOVA was conducted testing whether there was an interaction effect of APL x cohort on the five motor tests. Because of the non-significant interaction (F(24; 48,328) = 0.96, p = .51) and non-significant differences for players from different birth cohorts with regard to height (F(2; 13,616) = 0.31, p = .74), weight (F(2; 13,909) = 2.43, p = .09), as well as relative age (F(2; 14,175) = 1.33, p = .27), the cohort variable was not further considered as a confounder in the following analysis.

To gain insight into the prognostic relevance of the single predictors (five motor tests, three maturation-related characteristics), univariate ANOVAs were conducted analyzing mean differences between the different performance levels for each predictor. Post hoc tests (Tukey HSD) were conducted to contrast the three APLs within multiple group comparisons. Effect sizes were provided independently from their significance, as the group sample sizes for the three APLs differed strongly. Cohen’s d and η² were calculated as effect sizes and classified in accordance to [34]. In this process, Cohen’s d was computed as the mean difference divided by the pooled standard deviation:

Additionally, players’ future success was predicted in a logistic regression model by the two correlated latent performance factors SA and TS with simultaneous consideration of players’ height, weight and relative age that served as manifest covariates (objective 2). Thereby, the prediction of APL—as a criterion variable with ordered categorical outcome levels—was based on a threshold model with an underlying continuous latent response variable (for more information, see [35, 36]). For a clearer view of a player’s relative chances depending on several predictors, odds ratios (for non-standardized variables) were computed as effect sizes. In accordance to Muthén [37], R² was examined to quantify the amount of variance explained by the logistic regression model and refers to the explained variance proportion in the underlying continuous latent response variable.

Predictive value of motor tests and maturation-related characteristics (objective 1)

The U12 results for all predictors (i.e., the motor performance tests, the anthropometric characteristics and relative age) correspond to the later achieved APL (Fig 2). Players who reached APL1 performed better in all motor test than the players who made it to the semi-professional or non-professional level. The same holds true for players who became semi-professional in comparison to non-professional players. Additionally, U12 players who reached APL1 were the tallest, heaviest and slightly oldest. This also extends to the APL2 players as compared to future APL3 players.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Descriptive statistics for the diagnostics in U12 separated by players’ adult performance level (Including the effect sizes for the multiple group comparisons, see also Fig 1).

*p < .05, **p < .01, ***p < .001

https://doi.org/10.1371/journal.pone.0182211.g002

The motor diagnostics’ prognostic relevance was verified by significant mean differences concerning the three APLs (F(10; 24,174) = 31.48, p < .001, η² = .01) and this APL effect remained significant in the MANCOVA under consideration of the maturation-related covariates (F(10; 23,902) = 33.40, p < .001, η² = .01). ANOVAs demonstrated the significance for the five single motor tests and for the anthropometric variables height and weight (each p < .001; Fig 1). Only relative age failed to reach significance (F (2; 14,175) = 1.33, p = .26). However, although the motor and anthropometric variables showed a prognostic relevance for future success in adulthood, effect sizes were rather small (each η² < .02).

The multiple group comparisons revealed significant mean differences in all considered variables (except for relative age) between APL3 and the two higher levels (Figs 1 and 2). Both, future professional and semi-professional players performed better in the U12 diagnostics than the non-professionals. The effect sizes in the motor test results ranged from small (e.g., a non-significant d = 0.17 between APL1 and APL2 in agility or d = 0.18 between APL2 and APL3 in shooting) to medium effect sizes (e.g., d = 0.71 and d = 0.70 between APL1 and APL3 in dribbling and sprint, respectively). Moreover, significant mean differences between APL1 and APL2 could be found for shooting (d = 0.32) and sprint (d = 0.45), showing the most discriminative power between the two highest levels. Although the dribbling test was able to differentiate both, between APL3 and APL1 (d = 0.71) as well as between APL3 and APL2 (d = 0.47), this test narrowly failed to reach significance in separating between APL1 and APL2 (d = 0.29, p = .07).

With regard to the anthropometric variables height and weight, multiple group comparisons revealed significant differences between all the examined APLs. The effect sizes ranged from d = 0.19 (APL2 vs. APL3 regarding weight) to d = 0.56 (APL1 vs. APL3 regarding weight). In accordance to the non-significant overall comparison, the relative age of the U12 competence center players was a less meaningful predictor for future APLs (d ≤ 0.10).

Predictive value of latent performance factors (objective 2)

The complex logistic regression model considered the maturation-related characteristics as covariates and predicted APL by the two factors SA and TS. The analysis of the model yielded R² = .25, indicating that 25% of the variance in APL could be explained by the model.

Fig 3 presents the factor loadings and regression coefficients of the logistic regression model. First, the manifest variables’ single loadings on the latent variables confirmed the preassigned factorial structure of the two latent factor variables SA and TS. All single tests loaded distinctively on their respective factor (.26 ≤ λ ≤ .88; each p < .001). Furthermore, the factor loadings for the dribbling test confirmed that SA (λ = .34) as well as TS (λ = .49) are the underlying performance factors.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Players’ adult performance level predicted by two latent factors (SA and TS) with simultaneous consideration of maturation-related characteristics as covariates.

*p < .05; **p < .01; ***p < .001. The regression coefficient for relative age was estimated as β = -.003.

https://doi.org/10.1371/journal.pone.0182211.g003

The logistic regression model revealed a significant influence on APL for the latent factor variable SA (β = .07, p < .05) and TS (β = .46, p < .001). In this case, a notable correlation (r = .33) between SA and TS leads to a rather small contribution of SA to APL. Nevertheless, the technical components represented by TS possess an even more incremental predictive power compared to SA containing predominantly speed-related abilities. With regard to the two latent factor variables SA and TS, the ORs e^β of the logistic regression model (OR_SA = 4.58, OR_TS = 6.72) indicated that a one standard deviation (SD_SA = 0.09, SD_TS = 0.50) increase in the respective latent factor variable improved the chances of achieving the professional level by a factor of 1.15 (= 4.58^0.09) for SA and by a factor of 2.59 (= 6.72^0.50) for TS, respectively.

Furthermore, a significant impact of players’ weight was verified by the model, indicating that heavier U12 players had higher chances of achieving a better APL (β = -.22, p < .001). In addition to this, height (p = .32) and relative age (p = .84) revealed no further predictive power.

Predictive value of motor tests and maturation-related characteristics (objective 1)

For interpretation of the prognostic relevance, one must consider that design features strongly affect the estimates of predictors’ prognostic relevance. For example, the duration of the investigated prognostic period and its relative developmental phase play an important role, because discriminating factors among youth players vary between the different development stages of adolescence [38]. In addition, the sample size (determining the statistical test power) and the performance levels at which the predictors and criteria were assessed (leading to different “ranges of talent”, [39]) are further potential moderators of the effect size.

Because of the strong influence of design features, the current results are best comparable with findings from Höner and Votteler [7]. That study also analyzed the prognostic relevance of motor diagnostics within the German soccer TID in early adolescence, but for the cohorts 1993 to 1997 and over a mid-term period from the onset of puberty to middle-to-late adolescence (including the transition from basic to elite talent promotion). Future success was operationalized with four performance levels in which the highest (youth national team) and the lowest level (not further selected for the TID in U16-U19) were most similar to APL1 and APL3 in the present study. The present results confirm the former mid-term prognostic results for an even longer period that additionally includes the entrance into adult soccer as a further crucial career transition. Both studies detected similar effect sizes. Multiple group comparisons between the highest and lowest performance levels revealed the highest effect sizes for sprint (d = 0.67 and d = 0.70 for the mid- and long-term study, respectively) and dribbling (d = 0.64, d = 0.71). Furthermore, the studies found medium effect sizes for agility (d = 0.45, d = 0.41) and ball control (d = 0.59, d = 0.54). However, with regard to the shooting test, the effect size in the present study (d = 0.53) was considerably higher than in the mid-term study (d = 0.30). This is quite remarkable given the longer period, as one would assume that, all else unchanged, the association between the predictor and the criterion variable decreases over a longer prognostic period.

For reasons of comparability in the following paragraphs, the results from other studies that refer as much as possible to the age group U12 are reported. Moreover, the present study’s differing effects between APL1 and APL3 are compared to similar findings from the other studies (e.g., youth national team vs. non-selected). Effect sizes were translated in terms of Cohen’s d on the basis of the descriptive statistics provided by the original studies (in cases where a study did not provide Cohen’s d as an effect size).

Regarding the prognostic relevance of speed related predictors, this study revealed slightly larger effect sizes for sprint and nearly identical effect sizes for agility compared to other prospective studies dealing with players at this age. Whereas Le Gall et al. [17] found no prognostic relevance for the achieved adult performance level of French U14 academy players, Gonaus and Müller [6] revealed significant small to medium effect sizes for sprint (d = 0.36) and agility (d = 0.49) in their 3-year prospective study with Austrian U15 youth academy players. In their 2-year study with elite Belgian U12 players, Deprez et al. [21] found small to medium effect sizes for their tests in 5 m sprint, 30 m sprint and agility (d = 0.3, d = 0.6 and d = 0.5, respectively), where only the effect for the 30 m sprint reached significance. Further studies [14, 40–42] confirmed the prognostic validity of speed and agility tests. However, various design parameters of these studies differed from the presented study (e.g., other performance criteria, lower performance level), so that detailed comparisons of the effect sizes become superfluous.

Compared to the research of speed related predictors, there are considerably fewer prospective studies covering the prognostic relevance of technical skills [25]. In this regard the observed medium effect sizes for dribbling, ball control, and shooting fit into the broad range of effect sizes found in previous research. In their 2-year prospective studies, Figueiredo et al. [14] and Deprez et al. [21] analyzed football-specific technical skills’ prognostic relevance in the age group U12. Performances in a wall pass and shooting accuracy test were non-significant predictors regarding the prospective playing status of Portuguese regional club players. By contrast, a ball control test (d = 1.08) and a dribbling test (d = 1.20) showed prognostic relevance in the same study [14]. To a smaller extent, dribbling was also a significant predictor (d = 0.3) for Belgian elite players’ future playing status [21]. Furthermore, albeit for older Dutch youth academy players in late adolescence (U16-U19), Hujigen et al. [20] detected medium effect sizes indicating lower performances in different dribbling features for players who were deselected from the youth academies one year later. Whereas the prognostic effects reached significance for the peak shuttle dribble (d = 0.66) as well as for the repeated shuttle dribble test (d = 0.60), they failed significance for the slalom dribble test (d = 0.37).

The present results emphasize the importance of the anthropometric variables height (d = 0.50) and weight (d = 0.55) in early adolescence for more or less successful adult players. These medium effect sizes are quite similar to the prognostic value displayed by Gil et al. [40] for weight (d = 0.41), but also lower than the prognostic relevancies for height reported by Figueiredo et al. [14] (e.g., d = 1.03, in U12/U13) and Gil et al. [40] (e.g., d = 0.81, in U15) in their studies with regional club players (prognostic periods ≤ 2 years). However, they are higher than the effects of the prospective studies conducted by Deprez et al. [21], Gil et al. [41], Gravina et al. [42] and Le Gall et al. [17], who all found non-significant impact of these variables in early adolescence for future success.

Still, it remains unanswered whether differences in the anthropometric variables are caused by maturation-related biases in the talent selection (e.g., biologically more mature players are more often selected) or whether these variables have to be considered as performance factors in adulthood. Indeed, the 89 APL1 players were, as adults, on average 182.20 cm tall (based on the data provided by [31]). Thus, the “successful” group of this study was slightly taller than the corresponding German population of 18 to 25 year-old males who are on average 181 cm tall [43]. As the difference between the APL1 and APL3 players in the U12 age (3.37 cm) was noticeably larger, this might indicate an advanced biological maturation for future APL1 players in the U12 age group. However, for clarification of this important issue, more sophisticated and valid measures for biological maturity (e.g., skeletal age or ages at peak height velocity; [44]), differentiated analyses regarding the influence of anthropometric variables and playing position, (e.g., [45]) and a repeated measurement lasting over different developmental phases would be necessary.

Several causes might have led to the non-significant influence of relative age. The main reason is probably that the U12 players who participated in this study were already selected for the German TID program. These players ranked among the top 4% of their age cohort and the influence of relative age on performance factors is reduced in already selected samples [18]. Thus, a relative age effect already existed within this study sample (i.e., 60.4% of players were born in the first half of the year). The rate of the APL1 players (61.8%) is quite comparable, leading to non-significant effect sizes in the multiple group comparisons for relative age. However, the fact that the extent of the relative age effect at the beginning of a TID program is similar to its magnitude in adult professional soccer does not imply that the relative age effect is irrelevant for the TID process. For example, there is a considerably higher relative age effect in middle-to-late adolescence. More than 70% of the players that are selected for German youth academies and youth national teams are born in the first half of the year (for similar data in France, Spain or Switzerland, see [46–48]). Upon entrance into the adult professional level the relative age effect is reduced (but still present), because at this point previously selected relatively younger players have even higher chances of becoming a professional [49].

Predictive value of latent performance factors (objective 2)

Overall, the ANOVA analyses with manifest variables indicate a small magnitude of explained variance regarding the three adult performance levels (η² < .02). However, despite the small explained variance, the quite relevant effect sizes derived by the multiple group comparisons are a first hint that the tests obtained prognostic relevance. The results of the SEM confirm these general findings of the “traditional” approaches, but also go beyond these findings with regard to the 1) overall explained variance, 2) role of the covariates, 3) consideration of the underlying performance factors and 4) attributes of the measurement model.

First, the logistic regression model led to a considerably higher magnitude of explained variance (R² = 25%), so that SA and TS explain with simultaneous consideration of anthropometric characteristics and relative age a quarter of the APLs. To the authors’ knowledge no study in soccer exists using a latent variable approach for the estimation of the prognostic relevance of performance factors. Thus, the results for objective 2 cannot be compared to other studies. Nevertheless, there are two main reasons for the increase in explained variance in the latent variable model. One, the latent variable model corrects for attenuation due to measurement error of single manifest variables (e.g., [26]). The correction of attenuation is conducted by including several indicators of the latent exogenous TS and SA factors. Within the latent structural model, regression coefficients are unbiased and error free. Two, the manifest discrete outcome variable with its levels (APL1-3) is approximated by an underlying latent continuous variable y*, which is linked by a threshold model to the observable outcome y. The underlying latent continuous variable y* represents a more realistic “true ability” factor which leads to the manifest outcome via a threshold parametrization (for technical details of the Mplus latent variable framework see [50]). As a result of these two advances, the latent variable model approach better explains the relationship between the variables.

Second, regarding the role of the covariates, only weight (β = -.22, p < .001) was a significant anthropometric covariate, indicating that heavier players have a greater chance of achieving a higher APL. Furthermore, height (β = .11, p = .32) provides no further explanation of the APL and the relative age effect had no additional influence on the APL. However, due to the strong association of weight and height (r = .73), it would obviously be wrong to conclude that height in early adolescence is unimportant for future success. According to the results for objective 1, the prognostic relevance of the two anthropometric variables was quite similar. However, the insights into the impact of maturity characteristics on the TID process are limited in this study, as its focus laid on the relevance of speed and technical performance factors with simultaneous consideration of maturation-related covariates. Therefore, the anthropometric variables and relative age were entered on the same level within the logistic regression as equal covariates and no underlying latent maturity factor was integrated into the model nor was height conceptualized as a predictor for weight (that then might have served as a mediator on APL). For a more concrete clarification of the impact of maturity characteristics on the TID process more valid measures for maturity, repeated measurements and a theoretical model including core assumptions of the mechanisms would be necessary.

Third, the analyses revealed significant path coefficients for both underlying performance factors (β = .07, p < .05 for SA; β = .46, p < .001 for TS). It was observed that TS had a stronger effect than SA leading to the interpretation that technical skills seem to be stronger predictors than speed abilities. However, some parts of the difference between the effects might be due to the collinearity between TS and SA, which means that TS and SA have variance in common. Therefore, the difference in the effects should be interpreted cautiously.

Fourth, the attributes of the measurement model validated the particular role of dribbling as it was influenced by both theoretical factors (λ = .34 for SA and λ = .49 for TS). The reason for this is that speed-related as well as technical skills impact dribbling performance in this test. The TS-factor influences the dribbling and ball control performances to nearly the same relevant extent (λ ≈ .50). The considerably lower coefficient on shooting (λ = .26) is attributable to a high proportion of measurement errors, because of the shooting test’s low reliability. Regarding the SA-factor loadings on the indicators sprint (λ = .49), agility (λ = .88) and dribbling (λ = .49), this study confirms the factor solution found by Höner et al. [30] with data from other cohorts (1999, 2000) from the age group U12 in German soccer TID (with factor loadings λ = .50, λ = .92, and λ = .35 for sprint, agility and dribbling, respectively). The SA-factor mainly affects agility, whereas the sprint performance seems to be influenced by other factors as well. In accordance with this, the estimated variances of the measurement errors (ϴ^δ = .76 for 20 m sprint and ϴ^δ = .22 for agility) illustrate that in particular agility performance is explained by the SA-factor (ca. 78%). As both tests have similar (excellent) reliabilities, these differences in the factor loadings and variance of the measurement errors are not caused by different measurement errors. Confirming the findings from Höner et al. [30] the impact of the SA-factor on the dribbling performance is lower. This is caused by the additional influence of the TS-factor as well as by the lower reliability of this football-specific test compared to the sprint and agility test.