Identifying talent in youth football ultimately comes down to making a binary developmental prediction: Does a specific youth player have what it takes to become a professional as an adult or not? To make this prediction, which aims to support coaches’ opinions in the selection process (Sieghartsleitner et al., 2019b; Lath et al., 2020), researchers use a “shopping list of key criteria” (Williams and Reilly, 2000) to build talent identification models based on general linear models (GLM) (for an exceptional example of non-linear approaches, see Siener et al., 2021; Zuber et al., 2016; Pfeiffer and Hohmann, 2012). In essence, they break down talent into multidisciplinary components (Vaeyens et al., 2006), such as physiological (Dodd and Newans, 2018), skill-related (Murr et al., 2018), psychological (Ivarsson et al., 2020), and sociological (Reeves et al., 2018b) criteria, all of which explain some part of career outcome variability (professional vs. non-professional) (Sarmento et al., 2018; Williams et al., 2020). Once related tests are completed, all test scores are usually either converted into z-scores (Turner, 2014; Figueiredo et al., 2011; Souza-Lima et al., 2020), weighted in a scoring system (Fuchslocher et al., 2011; Höner et al., 2015), or aggregated within linear predictive models (Sieghartsleitner et al., 2019b; Höner et al., 2021). Such procedures build a total score (for example, total number of points or forecasted probability of becoming a professional in a binary logistic regression model) and position players on a talent continuum, such as ranking lists. Regardless of the calculation procedure, the ground rule of such a summative approach remains unchanged: “the better the input (observed test scores), the better the output (total score)” (Overton, 2014; Maszczyk et al., 2014).

Theoretical (Cumming et al., 2012; Baxter-Jones et al., 2005) and empirical reasons (Meylan et al., 2010; Malina et al., 2015) suggest that the input-output operations of additive-linear scientific models encounter major distortion problems during puberty. From a biological standpoint, individuals belonging to the same age group can differ strongly because some are up to 5 years older/younger than others (Malina et al., 2004). It is only consequential that inter-individual maturity differences affect indicators of athletic performance and potential, favoring either late or early maturing athletes to different degrees according to maturation stage, age, task or sport considered (Baxter-Jones et al., 2005; Mitchell et al., 2017; Hill et al., 2020; Javet et al., 2022). In sports like football, athletes who are closer to fully mature status are repeatedly shown to perform better in motor tests than others who are less developed (Malina et al., 2005; Malina and Cumming, 2004; Albaladejo-Saura et al., 2021; Sieghartsleitner et al., 2019a; with exception of the motor awkwardness hypothesis; Quatman-Yates et al., 2012). As a consequence, failing to consider issues of inter-individual maturity differences means players are erroneously positioned on talent continuums: The more mature, the better the input (motor performances), the better the output (total score, game evaluation, estimated potential, selection and success chances”; Cripps et al., 2016; Hill et al., 2021; Peña-González et al., 2021; Johnson et al., 2017; with exception of the underdog hypothesis; Smith and Weir, 2020; Cumming et al., 2018b). Therefore, while collecting motor performance data was an important first step, the next logical one is to move forward with a “game plan” to address these issues (Cumming, 2018; Baxter-Jones et al., 2005).

On the side of the coaches’ eye, some solutions such as bio-banding—grouping players by maturity status (Cumming et al., 2018a; Cumming et al., 2017; Malina et al., 2019; Reeves et al., 2018a)—or maturation-ordered shirt numbers (Lüdin et al., 2022; Mann, 2020) are already in place in different sport organizations around the world to support better rating of in-game performance. On the side of scientific models, linear (Larochelambert et al., 2022) and curvilinear (Abbott et al., 2020) correction mechanisms have been proposed to deal with inter-individual developmental differences (maturity or relative age, for the distinction see Towlson et al., 2021b). They have been implemented in alpine skiing (Larochelambert et al., 2022), athletic sprinting (Romann and Cobley, 2015; Brustio and Boccia, 2021), long jumping (Brustio et al., 2022), or swimming (Hogan et al., 2022). Correction mechanisms estimate how good adolescent athletes would be if they had competed or performed some motor tasks without developmental advances or delays. As such, they provide better informational inputs (corrected test scores, in the sense of maturity-unbiased). From a pedagogical viewpoint, corrected scores aspire to equalize the selection chances by promising (biologically) equivalent, and thus fair, inter-individual performance comparisons (Abbott et al., 2021b; Cobley et al., 2020). However, the main goal of talent identification is not (only) fairness, it is (at its core) accurate predictions of a player’s future chance of success in adulthood (Sieghartsleitner et al., 2019b; Bergkamp et al., 2019).

To our knowledge, the accuracy differences of long-term predictions from corrected and uncorrected scores are presently unexamined. Our study’s purpose provided insight into this topic. To this end, we investigated whether (1) maturity status influences our motor talent criteria (if not, no correction would be necessary). Then, for the sake of parsimony and implementation simplicity, we corrected the maturity-biased talent criteria with a linear correction mechanism and examined if (2) the correction mechanism worked (if not, the correction would be useless). Finally, we examined if (3) corrected test scores (univariate prediction) or (4) models including several corrected motor scores (multivariate prediction) make better predictions of adult performance levels of youth football players when compared with raw scores/model.

As part of the Institute of Sport Science at the University of Bern’s Talent selection and development in Swiss Football project, we studied motor performances of 15-year-old male elite football players born in 1999 (N = 121; M_age = 15.13, SD = 0.34), who played at the highest Swiss national level in their age category. On the one hand, choosing this age group was motivated by biological considerations (it should still be considerably affected by maturity differences; Malina et al., 2015) and on the other hand, by (selection-relevant) cultural realities (from the age group U15 onwards, players enter a new phase in the national football-specific youth development program where performance and exposure to international competition gains in importance; Schweizer Fussballverband, 2014).

The study received approval from the Ethics Committee of the Faculty of Human Sciences of the University of Bern. All players and their parents or guardians provided their written informed consent to participate.

Based on a football-specific literature review (Williams et al., 2020) and subjective talent criteria mentioned by coaches (Jokuschies and Conzelmann, 2016; Bergkamp et al., 2022), we attempted to account for the diversity of motor abilities necessary to excel in football by including a series of eight motor performance tests (see Table 1). Their practical relevance is demonstrated by the fact that they are used by both the Swiss and German Football Association as part of their talent identification program (Höner et al., 2015; Schweizer Fussballverband, 2016). Moreover, these tests have repeatedly been found to validly predict future performance levels (Höner et al., 2021; Leyhr et al., 2018; Höner et al., 2017; Höner and Votteler, 2016; Sieghartsleitner et al., 2019a).

We operationalized player maturity status at age 15 with 6 validated pragmatic somatic prediction methods (Mirwald, Moore-1, Moore-2, Fransen-1, Fransen-2) and attained percentage of predicted adult height (%PAH) (see Table 2 for equations and references). Like other research groups (Leyhr et al., 2020a), we call these prediction methods “pragmatic” because they can be used easily and prospectively. Six years later (at the age of 21), all players were re-contacted and asked to complete a questionnaire which aimed to collect their real (in the sense of observed as opposed to predicted) adult height. In 37% of the cases, they did not respond, so we searched transfermarkt. Based on the principle of crowd wisdom (Peeters, 2018), online platform transfermarkt is regularly used as a source of information on talent (Leyhr et al., 2020b; Skorski et al., 2016; Doyle and Bottomley, 2018) and sport management research (Herm et al., 2014; Prockl and Frick, 2018; Müller et al., 2017). By knowing players’ real adult height, we calculated a seventh maturity indicator retrospectively—the attained percentage of the real adult height (%RAH). Although we are aware that %RAH represents only a sub-dimension of biological maturity (the somatic dimension; Baxter-Jones, 2017), we consider it a more reliable and objective method compared with other indicators since “the one closer to adult height is more mature than the other who is further from adult height” (Malina et al., 2015).

We collected adult performance level in the 2020–2021 season in the follow-up questionnaire. By knowing adult performance level, we objectively classified players into two groups: professionals (n = 47) who played in the 1st to 3rd league within Switzerland or held professional contracts abroad and non-professionals (n = 74).

Due to absence, injury, missed, incorrect, or aborted test series, there were missing scores in the study variables (range: 0.8%–6.6%). Little’s test (1988) showed these data were not randomly missing (χ² = 172.91, df = 137, p = 0.020). Since there was no clear indication the MAR hypothesis could not be retained, we used EM-imputation based on all study variables to complete the dataset. We analyzed the imputed dataset.

Pearson’s correlations served as the measure of association between raw or uncorrected test scores and each maturity indicator. Accordingly, we identified motor performance tests confounded by maturation and requiring correction.

To correct maturity-biased raw scores, we used a four-step correction mechanism. Table 3 illustrates these steps using the example of the sprint scores, which were corrected using %RAH as a maturity indicator. Table 3 also shows how the same sprint score (6.20 sec) of three fictitious players with different maturity statuses (85%, 95%, and 100%, respectively) differed after correction (5.41, 6.17, and 6.63 sec, respectively).

Step 1 (linear regression): We conducted a simple linear regression that predicted raw test score (RS) of specific motor tasks based on maturity status (MS, Table 2).

Step 2 (create new variables): We stored parameters (b₀ and b₁) defining the corresponding regression line. The coefficient b₀ is the constant in the equation. The coefficient b₁ indicates the change in the expected raw score associated with an increase of one unit in the maturity status (MS). By means of these two parameters (b₀ and b₁), we computed two new variables:

Step 3 (calculate correction factor): We divided RS of players by biologically expected RS (ERS, see step 2) for an individual correction factor (CF).

Step 4 (generate corrected test score): We multiplied CF by ESA for corrected scores (CS). Since this calculation was based on the ESA, it led to a biological uniformization of the group and provided test scores an exact replica each player with normal development was expected to achieve. By “normal,” we mean average as opposed to early or late.

We repeated this procedure for each motor performance test and maturity indicator. In this way, there are corrected test scores for each player in each motor performance test using Mirwald, Moore-1, Moore-2, Fransen-1, Fransen-2, %PAH, or %RAH as maturity indicators.

To ensure the correction mechanism served its purpose, we calculated Pearson’s correlations between each maturity indicator and their corresponding corrected test scores. If r = 0, then R² = 0, thus implying that the proportion of the variance that a maturity indicator and a motor performance test previously shared was successfully partialed out.

To determine whether each corrected talent criterion was more closely linked with adult performance level (professional vs. non-professional) than each raw talent criterion, we computed point-biserial correlations—a special case of the ordinary Pearson correlation. If the correlation coefficients had noticeable descriptive differences, we tested them for significance using the R package cocor (Diedenhofen and Musch, 2015).

Answering the fourth question required three main analytical steps: (1) regression modelling, (2) estimating receiver operating characteristic curves, and (3) applying the DeLong-Test.

Regression modelling: We estimated eight binary logistic regression models that included eight motor talent criteria (Table 1; z-values). Model one served as the reference model; raw test scores were used as input (raw motor performance model). Depending on results from research question 1, the seven other models included motor performance scores corrected based on each corresponding maturity indicator (e.g., model with test scores corrected according to Mirwald). We compared predictions from these multivariate regression models with a baseline model—a model without predictors (Field, 2018). We tested the significance of the coefficients (improvement over the baseline model) with the omnibus test of the model coefficients (Zeileis and Hothorn, 2002). Its null hypothesis (H₀) is that both models (that is, the baseline model and a regression model with eight predictors) predict the adult performance level of the players equally well. We tested appropriate calibration with the Hosmer-Lemeshow test (Hosmer et al., 2013). The model fit was quantified using Nagelkerke-R² (R²_N). Each regression model generated the probability of being classified as professional or non-professional for all players based on a combination of (raw or corrected) motor performance. If a player’s predicted probability exceeded 0.5, the player was classified as professional.

Receiver operating characteristic analysis (ROC): The predicted probabilities we obtained in each regression analysis were used to create non-parametric ROC curves. In other words, each ROC curve represented one of the regression models. The area under each curve (AUC) can be used as an index of the classification’s (discriminative ability) overall quality since it reflects the relationship between sensitivity (correctly identified professionals) and specificity (correctly identified non-professionals) (Hanley, 1989). The more players are correctly classified by a regression model, the closer the AUC is to 1. The following guidelines have been suggested (Hosmer et al., 2013) to facilitate the interpretation of the AUC: 0.9–1.0 (outstanding discrimination), 0.8–0.9 (excellent discrimination), 0.7–0.8 (acceptable discrimination), and 0.5–0.7 (poor discrimination).

DeLong-Test: We compared the AUC of each of the seven corrected regression models along with the raw regression model by employing the DeLong-Test. The H₀ states that the raw and each corrected model do not differ in their ability to predict a dichotomous criterion (professional vs. non-professional) over 6 years (H₀: ΔAUC = 0).

We set alpha levels for all significance tests to 0.05.

However, in view of the seven model comparisons, we adjusted the significance level for research question 4 using Bonferroni’s method (α = 0.05/7 = 0.007). We analyzed data using IBM SPSS version 27.

We provide means and standard deviations of all variables in the supplementary material (Table S1). According to all maturity indicators except %PAH (r = 0.21; p = 0.023), at the age of 15, the maturity status of the 47 professionals and 74 non-professionals did not differ significantly (Table 4).

Table 5 confirms the need for a correction mechanism by displaying significant maturational influences (0.16 ≤ | r | ≤ 0.72; ps < 0.05), which were more pronounced in functional capacities than in motor skills and followed the same principle: the more mature, the better the player performed. All negative correlations referred to motor tasks with reverse coding (the shorter the time needed, the faster and the better) and therefore do not contradict this principle.

As expected, all significant correlations in Table 5 were reduced to (almost) zero and became insignificant (0.00 ≤ r ≤ 0.03) after applying the correction mechanism.

Table 6 illustrates significant correlations between raw motor talent criteria of sprinting, agility, dribbling, juggling, and shooting and adult performance level (0.16 ≤ | r | ≤ 0.29, ps < 0.05). Thus, some motor variables appear useful for making predictions. In most cases, the predictive validity of these variables barely changed after we applied the correction mechanism. We only observed slight descriptive improvements or deteriorations in the correlations with adult performance levels after application of the correction mechanism (e.g., sprint raw: r = – 0.24 vs. sprint corrected with %RAH: r = – 0.29 vs. sprint corrected with %PAH: r = – 0.13). However, none of these differences were significant.

All eight regression models (one with raw scores, and seven with corrected scores when needed; Table 7) were statistically significant (omnibus test of model coefficients: ps < 0.05) and appropriately calibrated (Hosmer-Lemeshow test: ps > 0.05). The models explained between 19% and 28% (Nagelkerke’s R²) of the variance among adult performance level and correctly classified 64.5% to 68.6% of the cases.

However, DeLong tests revealed that no model based on the corrected scores was significantly better than those based on raw ones (see Table 8). Consequently, all regression models (raw or corrected) demonstrated similar predictive validity (– 0.02 ≤ ΔAUC ≤ 0.03, ps > 0.05).

Our results replicate existing data and information about maturity-associated variations: Maturity status correlates with motor performance scores and affects functional capacities, such as sprinting and jumping, more strongly than motor skills, such as juggling (Meylan et al., 2010; Albaladejo-Saura et al., 2021). For scientific talent identification models operating within a summative approach, our results imply that input-output operations are confounded—the more mature the player, the better input (test scores) and output (total score: predicted probability of being classified as professional in a binary logistic regression). In line with previous studies (e.g., Romann and Cobley, 2015; Abbott et al., 2021b; Larochelambert et al., 2022), our correction mechanism served its purpose: Any correlation between motor performance test scores and the confounder (maturity status) was removed. Yet contrary to our expectations, univariate and multivariate predictions of adult performance level did not substantially improve with corrected scores.

An answer to this question emanates from two lines of thought: (1) characteristics of the variables included in the correction and prediction process (initial predictive validity of the motor variable, the initial maturity bias and the validity of the maturity indicators) and (2) the selection system in which predictions have been made.

In the case of univariate prediction, our results indicate that the correction mechanism failed generating better predictive variables. We suggest three possible reasons for this failure (Figure 1): 1) the variables have no predictive validity per se (the correction procedure cannot improve something, i.e. the predictive value, that doesn’t exist); 2) the corrected and raw test scores are very similar (similar scores do not entail different predictive results; y-axis); 3) the corrected scores are not meaningful (they still contain at least the same number of errors as the raw scores; x-axis).

The first reason mentioned seems conceivable (the motor variables considered have no predictive validity), however, we can hardly argue in this direction—on the one hand, because most of our results (Table 6), and on the other hand, because the empirical discourse related to our motor variables does not support it (Leyhr et al., 2018; Höner et al., 2021; Höner et al., 2017; Sieghartsleitner et al., 2019a). Consequently, we elaborate more deeply on the two other reasons that came to mind by means of Figure 1.

The first reason mentioned seems conceivable (the motor variables considered have no predictive validity), however, we can hardly argue in this direction—on the one hand, because most of our results (Table 6), and on the other hand, because the empirical discourse related to our motor variables does not support it (Leyhr et al., 2018; Höner et al., 2021; Höner et al., 2017; Sieghartsleitner et al., 2019a). Consequently, we elaborate more deeply on the two other reasons that came to mind by means of Figure 1.

The y-axis of Figure 1 implies that the similarity of raw and corrected scores impact their differences in predictive validity. The lower the correlations between motor performance tests and maturity indicators (Table 5), the smaller the confounding influence and (need for) adjustment and by extension, the greater the similarity between raw and corrected scores and their predictive validity. In other words, no maturity-bias does not require correction, no correction means unchanged scores, and unchanged scores imply unchanged prediction. Accordingly, since the sprint test is the only one displaying strong correlations with maturity status (see Table 5), it is probably the best candidate for raw and corrected scores that differ sufficiently to show a noticeable change in predictive validity.

The x-axis of Figure 1 suggests that a positive change in predictive validity depends on meaningful correction mechanism foundations (validity of the chosen maturity indicator). An invalid maturity indicator is unlikely to generate a valid (meaningful) corrected score to make a valid prediction. Even though some literature shows that pragmatic maturity indicators are sufficient to economically inform coaches about youth player maturity status (Romann et al., 2017; Leyhr et al., 2020a; Malina et al., 2012), serious validity concerns as a function of age and stage of maturation have also been raised (Malina and Kozieł, 2014; Teunissen et al., 2020; Parr et al., 2020). The lower the validity of the chosen maturity indicator (left side, Figure 1), the greater the loss of the correction and predictive process with respect to meaningful informative value. Since our sample characteristics (15 years-old, highly selective, more mature than same-aged non-athletes) match the criteria when pragmatic maturity indicators (especially the offset methods) become less accurate (Malina and Kozieł, 2014; Malina et al., 2015), the loss in meaningful informative/predictive value may be substantial. Notably, the predictive validity of all the scores corrected with a pragmatic maturity indicator tends to be slightly lower than that of the raw scores (see Table 6). Hence, if one is seeking a talent criterion that is expected to exhibit a noticeable increase in its predictive validity after applying the correction mechanism, it will probably be found among the ones that have been corrected with more objective and reliable methods (here, %RAH; or in general, skeletal age or real age at PHV; Malina, 2017).

Against the backdrop of these reflections, our expectation to systematically create notably better predictive variables by simply applying a correction mechanism may have been unrealistic. Not all test variables were predisposed to a noticeable improvement in their predictive validity after applying the correction mechanism. In fact, combining the three determining factors (1. Has the variable some predictive validity? 2. How similar are corrected and uncorrected scores? 3. How meaningful is the foundation of the correction procedure?), it appears that—among all test variables—the sprint test when corrected according to our most objective and reliable method (%RAH) was the only candidate where an improvement in predictive validity after the correction mechanism could be envisaged. Unsurprisingly, it is also the only corrected variable with a (descriptively) better predictive validity than its raw version (see Table 6). Although this predictive benefit remains trivial and non-significant, our data and reasoning suggest the closer to the top right corner of Figure 1, the more promising the application of a correction mechanism might be.

The selection system in use when we conducted our study might also be responsible for a lack of predictive improvement in the multivariate context (research question 4; Table 8). It is clear that the selection-related decision-making process depends primarily on evaluations of match performance and subjective insights from coaches and scouts (Roberts et al., 2019; Christensen, 2009). Nevertheless, it seems reasonable to assume that coaches’ selection decisions are based on available information—the players’ characteristics (Lath et al., 2021; Bergkamp et al., 2022; Lüdin et al., 2022). Thus, information systematically collected (for selection recommendation purposes) or visually noticeable (in training and competition contexts) probably have more influence (in the sense of decision-making power) on coach selection decisions than uncollected or unnoticed information (Lüdin et al., 2022). This distinction is particularly important because raw motor performances are observed by coaches on a day-to-day basis and systematically tracked as part of the talent identification program in Switzerland (Schweizer Fussballverband, 2016; Fuchslocher et al., 2011), while the corrected scores were nonexistent during the study duration. In other words, it was simply impossible for corrected scores to influence the probability of survival of early and late maturing players in the elite-sport system. Contrastingly, through the (maturity-biased) tracks raw performances left behind (the more mature, the better the raw scores, the better the impression; Hancock et al., 2013; Hill et al., 2021; Cripps et al., 2016; Furley and Memmert, 2016), they could actively shape coaches’ opinions, selection chances, final decisions, and by extension, chances to access continuous professional support (Meylan et al., 2010; Malina et al., 2015). However, whether and the extent coaches were actually influenced by maturity-biased raw motor performances is not the focal point here. The core issue is raw scores, unlike corrected scores, had decision-making power. Unfortunately, developmental research has no way of experimentally manipulating the temporal component to compare whom coaches would select if they had access to sources of information not confounded by maturity status. Yet, it seems difficult to refute the hypothesis that maturity-biased and -unbiased performance considerations may lead to different conclusions about players and thus, to different selection decisions (Lüdin et al., 2022). In turn, it suggests that raw scores had a perceptible (and non-compensable) advantage over corrected scores. Is this advantage somehow reflected statistically? We suggest it is—with more decision-making power, raw scores carried more (predictive) information than corrected scores.

The potential of a multivariate regression model to discriminate between future professionals and non-professionals (AUC) is expected to improve, the better the model fits the data (Steyerberg, 2019). Therefore, our attention should be focused on the extent, if at all, the correction mechanism leads to better-fitting models (Nagelkerke-R²). Typically, adding relevant information, such as additional predictors, to models leads to improvements in model fit (Field, 2018). The more information a statistical model can work with, the better it is expected to perform (Field, 2018). For corrected scores, it is rather unfortunate because the correction mechanism does not add, but rather extracts everything related to biological maturation in each motor performance test (b + e; see Figure 2) (i.e., the correlation between corrected score and maturity status is equal to zero; research question 2). Figure 2 highlights three important points:

1. Before applying the correction mechanism, the motor scores contained information related to maturity status—the maturity and motor information overlap (in the sense of confounded; raw scores = a + b + e + f).

2. After applying the correction mechanism, the maturity and motor information are separated—the motor performance tests are free from this confounding influence (corrected scores = a + f).

3. The dependent variable (career outcome: professional vs. non-professional) remains unchanged during the entire process (a + b + c + d).

If we accept these three points, we have to assume that the extent of change in the model fit (R²) after the correction process primarily relies on the correction-related absence of components b and e (however, no one is really interested in component e, because e has nothing to do with the career outcome). Accordingly, the bigger component b is (the more biological maturation embedded in raw scores has predictive validity), the more the motor scores lose their predictive power after applying the correction mechanism when compared with raw scores. Since existing literature generally acknowledges that early maturers are more likely to be selected in football academies than late maturers (Johnson et al., 2017), the size of component b is likely to be substantial. Thus, as long as the selection process is (even slightly) biased by the maturity status, the model fit of a regression model using corrected scores will theoretically always be worse than the one using the raw scores only. Does our data support this view? It seems so.

Apart from one outlier (motor performance model corrected with %RAH; R²_N = .28), which might just be an example of capitalizing on chance, our model fits always decreased when using corrected (.19 ≤ R²_N ≤ .24) instead of raw test scores (R²_N = .24) to predict player career outcomes. In concrete terms, the stronger the correlation between maturity indicator and adult performance level (component b in Figure 2; Table 4), the greater the decrease in model fit (Table 7). In line with this, the largest reduction we observed in model fit (R²_N = .19) after applying the correction mechanism was in the multivariate regression model using test scores corrected with maturity indicators having the highest correlation with adult performance level (%PAH, Table 4). Consequently, it suggests a methodological artifact: Raw scores had a (methodological and non-compensable) predictive advantage over corrected scores.

To sum up, expecting the application of a correction mechanism to automatically improve predictions from multivariate regression models was unrealistic. It overlooked existing selection system influence during the study duration, which produced a methodological artifact inflating raw score predictive power. The methodological artifact can seemingly be avoided only if the career outcome (professional vs. non-professional; a + b + c + d in Figure 2) emerges from an intervention manipulating the selection system during puberty (control vs. intervention group).

In the control group, players are selected annually according to their raw scores (without applying the correction mechanism). Players in the intervention group are selected annually according to their corrected scores (with applying the correction mechanism). Since the selection process is not affected by maturity-related selection errors in the intervention group, the right players—the players with the highest (motor) potential—are chosen. Accordingly, on average, players selected from the intervention group are expected to reach a higher adult performance level than players selected from the control group. However, to implement such an intervention, clubs and federations would need to relinquish control over talent selection decisions and agree to use a unidimensional talent identification model based on motor performance, which is inconceivable and definitively outdated (Williams et al., 2020). Therefore, the approach adopted in this study may be the only way to empirically introduce and theoretically discuss potential benefits associated with corrected scores.

Taken together, our results indicate that correction mechanism produce principally-sound corrected scores as valid as raw scores from a predictive point of view (Table 6 and Table 8). What practical implications can be drawn from this predictive equivalence, or rather, is a correction recommended or not?

Because coaches are encouraged to use scientific models in the talent identification process (Sieghartsleitner et al., 2019b; Williams et al., 2020) and corrected scores are available, it becomes necessary to reflect upon what kind of scientific test scores (raw or corrected) should be relied on. From a methodological perspective, it can be argued that compared with raw scores, corrected scores have limited utility since more work from implementing a correction mechanism is needed for the same overall predictive output. This violates the principle of parsimony in building models (Field, 2018), which “tells us to remove what is unnecessary” (Sober, 1981, p. 145) and speaks against their application.

Although one can see the glass as half empty (same output for more work), one can also see it as half full (same output, but one fewer problem). In the words of Karl Popper (2000):

Accordingly, from a critical rationalist point of view, corrected scores should be favored because they predict at least everything successfully predicted by their predecessors (raw scores). In addition, they promise to avoid the acknowledged problem of maturity-biased inter-individual performance comparisons in scientific models and make new testable predictions (for example, see Figure 1). By correcting, and thus changing each player’s scores, the correction mechanism changes the position of each player on any kind of talent continuum based on one of several summative approaches (z-scores addition, Turner, 2014; Souza-Lima et al., 2020; weighted scoring system, Fuchslocher et al., 2011; Höner et al., 2015; regression, Sieghartsleitner et al., 2019a). Those who previously (and erroneously) did not emerge as candidates for selection could receive this status after applying a correction mechanism (and vice versa). At the same time, it may increase the acceptance of scientific forecasting in practice. Indeed, the more our models avoid mispositioning players on the talent continuum due to early or late maturation, the more the models will be taken seriously.

Interestingly, on some level the correction mechanism is not drastically different from the bio-banding strategy, which has garnered a reputation among applied sports scientists in recent years (Cumming et al., 2017; Cumming et al., 2018a; Reeves et al., 2018a; Towlson et al., 2021a). Both approaches “correct” something to solve a maturity-related problem for one of the complementary parts of the talent identification process, that is coaches and scientific models (Lath et al., 2020; Sieghartsleitner et al., 2019b). Bio-banding “corrects” the playing environment and allows coaches to observe how good each player would be if they played in an environment without maturity-related (dis-) advantages (Cumming et al., 2017; Malina et al., 2019; Rogol et al., 2018). It is supposed to enable coaches to better detect talented players. The correction mechanism seeks to achieve a similar goal. It corrects the test scores mathematically and allows scientific talent identification models to estimate how good each player would be in each test, if he was as biologically developed as the others. In turn, it is supposed to help scientific models better detect talented players in the future.

Furthermore, from a talent development perspective, just as bio-banding changes the behaviors and experiences of each player on the field (Cumming et al., 2018a), the correction mechanism would not be impervious to side effects. Specifically, according to Vallerand and Losier’s (1999) integrative analysis of motivation, each (social) context influences the basic psychological needs (e.g., perceived competence; Deci and Ryan, 1985), which in turn increase or undermine one’s intrinsic motivation and lead to a host of consequences (e.g., dropout intention, negative or positive emotions; Balish et al., 2014; Pelletier et al., 2001; Sarrazin et al., 2002). Thus, a context which is characterized by comparisons based on corrected (instead of raw) scores might help late maturers build confidence in their ability (satisfaction of need for competence), which in turn might increase their intrinsic motivation and decrease their possible dropout intentions. In contrast, corrected scores might threaten early maturers’ experience of competence, leading to less intrinsic motivation and negative emotions such as frustration, dissatisfaction, and fear of disappointing significant others. Therefore, players should not be left alone with their corrected scores, rather coaches have to explain beforehand the meaning behind such correction (in the sense of individualized feedback). In particular, correction-related feedback should make late maturers realize that their feeling incompetent might be unwarranted and that in all likelihood they will catch up with the others. Inversely, the correction-related feedback should make early maturers realize that their feeling competent might be, to some extent, unjustified and that they need to continue to work upon a particular aspect of their game, because in all likelihood the other players will catch up. In this respect, implementing a correction mechanism responds to a recent call for strategies creating learning contexts that encourage early and late maturing boys to develop more adaptative psychological skill sets (Cumming et al., 2018b). It is worth noting that such maturity-unbiased contexts may change players’, parents’ and coaches’ expectations regarding achievement. Theoretically, these expectations are not devoid of consequences, since they may actively shape talent development paths through self-fulfilling prophecies, such as the Pygmalion, the Galatea and the Golem effect (when higher/lower other-/self-expectations lead to higher/poorer career outcomes; Hancock et al., 2013; Leonardo Filho, 2016; Babad et al., 1982; Rosenthal and Jacobson, 1968).

To sum up, similar to bio-banding strategy (Cumming et al., 2017), corrected scores provide a new piece of information and mirror reality in a restructured, more nuanced way (Larochelambert et al., 2022; Abbott et al., 2021b). They refine judgment and might affect players’ personality and expectations. As a result, corrected scores may also impact career outcomes. Given their perceived highly subjective usefulness for maturity-unbiased, inter- and intra-individual performance comparisons, and equivalent predictive validity, we recommend the use of corrected scores to complement current talent identification and development methods. In this context, one of the strengths of maturity-based correction mechanisms is their application in all sports, which responds to a recent call for scientific solutions with an impact beyond a single sport (Abt et al., 2022).

There are four noteworthy study limitations. First, we focused on isolated skills and abilities in the tests we used (see Table 1). Thus, we could not completely capture the unpredictable and complex nature of football. For instance, the test assessing dribbling skill, not unlike a slalom around stationary markers, places a high reliance on the player’s ability to accelerate. Unfortunately, it misses some aspects of the cognitive, perceptual and motor skills that are essential to dribbling the ball in match conditions (McDermott et al., 2015; Ali, 2011). The same kind of criticism applies to the other tests. However, limited ecological validity was the price to pay to maximize closeness to practice—i.e., to investigate the predictive validity of the specific tests used in German and Swiss talent identification programs (Höner et al., 2015; Schweizer Fussballverband, 2016). Furthermore, while it is true that an increase in the complexity of the test battery should improve (ecological or predictive) validity, it could at the same time impair the reliability of the tests (see reliability-validity dilemma; Slomp and Fuite, 2004; Höner and Roth, 2002). Indeed, we had serious concerns that less reliable tests would provide less trustworthy corrected scores.

Second, we considered biological maturation as a finite and linear resource, which can only have a unidirectional effect on motor performances (the more mature, the better). Yet clearly our conceptualization was reductionist and failed to consider possible problems, such as maturity-related altered motor control also known as “adolescent awkwardness,” emerging at certain points during development (Quatman-Yates et al., 2012). Indeed, some athletes experience performance drops due to accelerated periods of growth, which could not be captured or considered in our correction procedure. Therefore, it remains unclear to what extent it is legitimate to linearize the phenomenon of maturation, which is primarily regarded as non-linear (Boeyer et al., 2020) and to correct its influence on talent criteria with linear methods. Other modelling approaches may be adopted in future studies (e.g., curvilinear; Abbott et al., 2021b; Abbott et al., 2021a).

Third, we interpreted results guided by the assumption that %RAH was a highly valid indicator of the somatic biological maturation of the players. However, its validity in our study is in question. We did not directly measure somatic biological maturation; we asked about it in a follow-up survey or collected it through transfermarkt. Both data collection methods are not without problems. Unlike other studies in talent research (Leyhr et al., 2020a), we assumed players did not overestimate their current adult height. Thus, we decided not to adjust self-reported adult height according to Epstein’s equation (Epstein et al., 1995). We believe these adjustments would have had limited bearing on our study findings since they are generally small in nature. Transfermarkt relies on the principle of crowd wisdom to generate information (Peeters, 2018). Consequently, even though it provides reliable and precise information about players (Prockl and Frick, 2018), the validity of its data depends on the crowd’s wisdom. To ascertain validity, we calculated the correlation between self-reported adult height and adult height found on transfermarkt for a sub-sample of 12 players for whom data were available (r_s = 0.99). In our case, the crowd seemed wise and our assumptions about %RAH are valid. However, future studies should reflect on the necessity of Epstein’s formula in the context of talent research and further examine the margin of error associated with adult height indicated on transfermarkt.

Fourth, our sample was (a) highly selected and (b) relatively small. It resulted in a reduced dispersion of maturity status, which may have deflated, for example, the reported relationships between maturity status and motor performances. Our small sample size meant we could not split the dataset into exploratory and validation subsamples without violating important requirements of statistical analysis. For this reason, we evaluated present models using only the exploratory dataset and failed meeting current standards of machine learning practice (Till et al., 2016). Consequently, our regression models are only internally valid and may not generalize to the population (Steyerberg et al., 2010). Considering these four limitations, our results seem trustworthy, yet need replication.

On a methodological level, further research could attempt to reproduce our findings with different operationalizations of maturity status, such as with Khamis-Roche method (Khamis and Roche, 1994), sexual and/or skeletal maturity indicators (Malina et al., 2015), different sample characteristics, like those where pragmatic maturity indicators are more accurate (Malina and Kozieł, 2014), different correction mechanisms (e.g., curvilinear instead of linear; Abbott et al., 2021b or size- and/or body-mass-related instead of maturity-related; Valente dos Santos et al., 2014a; Cunha et al., 2011), or other possibly confounded talent criteria, like psychological skill (Cumming et al., 2012; Cumming et al., 2018b). It would also be interesting to extend the correction mechanism to further variables confounding the input-output operations of additive-linear scientific models and explore how the long-term predictive validity responds. Potential candidates are relative (chronological) age (Leyhr et al., 2021; Romann et al., 2018; Votteler and Höner, 2014) or training age (Johnston and Baker, 2020; Guimarães et al., 2019).

We invite sports physiologists to further disentangle the direct, indirect, and mediating effects of biological maturation on performance in motor tests (e.g., body size descriptors, such as stature and body mass, not only predict maturity status, but also motor performance, Valente dos Santos et al., 2012; Valente dos Santos et al., 2014b). Sports psychologists might be interested in determining whether knowledge of corrected scores results in changes in player evaluation (as this is the case for knowledge of maturity status on shirt number; Lüdin et al., 2022), player personality (e.g., basic psychological needs, football-related self-concept, self-efficacy, motivational characteristics), support from parents or coaches, or team role and status (Cumming et al., 2012; Schmidt et al., 2015; Cumming et al., 2006; Eisenmann et al., 2020).

The correction mechanism also has applications in training science, namely as complement to developmental fitness curves (Myburgh et al., 2020; Owen et al., 2022; Till et al., 2022; Cumming, 2018) for (a) profiling strengths and weaknesses and (b) monitoring performance development during puberty. Performance profiling is meant to guide coaches so they better tailor training to player needs (Till et al., 2018; Eisenmann et al., 2020). Figure 3 shows a fictitious (late-maturing) player who appears to have a raw score-based major sprinting weakness. After correction (in red), it becomes clear that the weakness is not problematic. His corrected sprint performance is above average. However, without information provided by the correction mechanism, strength and conditioning coaches might invest time for sprint training, which could be better utilized elsewhere. Regarding the issue of monitoring, strength and conditioning coaches struggle to separate training effects from maturation effects after training intervention, such as strength training (van Hooren and Ste Croix, 2020; Moran et al., 2017; Till et al., 2018). By making pre-post comparisons with corrected rather than uncorrected scores, coaches better capture true training effects and better understand athlete progress (Eisenmann et al., 2020; Till et al., 2018).

Finally, on a more general note and beyond the realm of youth elite sports, the correction mechanism could be applied in school sport settings. Indeed, knowing maturity status influences physical activity patterns and motor performance, which, in turn, influences personality development and creates motivational basis for lifelong sports participation (Cumming et al., 2020; Cumming et al., 2012; Conzelmann and Schmidt, 2020), is it fair or expedient when physical education teachers evaluate students according to their raw performance during puberty (the more mature the adolescent, the better the performance; Nevill et al., 2021; Vist Hagen et al., 2022)?

Thus, a series of new research questions and options for practical application arise with available corrected scores.

Our study confirmed (1) that maturity status influenced some motor talent criteria and (2) maturity-based correction mechanisms worked. Contrary to our expectations, corrected scores (3) considered one by one (univariate prediction) or (4) in combination (multivariate prediction) were not found to be better predictors of who became a professional athlete six years later than uncorrected scores. In our opinion, predictive equivalence does not imply the futility of correction mechanisms (more work for the same output), but it underlines their potential importance (same output, one fewer problem) and the need to revise our correction-related expectations according to four factors: initial predictive validity of considered motor variables, validity of chosen maturity indicators, initial maturity-bias, and analyzed cohort selection system. Even though the recommendation to use maturity-based correction mechanisms seems theoretically legitimate, its empirical basis (from a predictive point of view for talent identification) remains to be established.