This study aimed to determine the influence of task constraints on goal kicking performance in Australian football. The study also compared the applicability of three analysis techniques: logistic regression, CBA and conditional inference trees. The different analysis techniques had similar accuracy levels. Given the similar performance of the different analysis techniques, consideration of their levels of applicability and ability to guide different types of decision-making should guide their respective use in the applied setting.
The constraints of location, pressure and shot type all influenced shot outcome. The likelihood of scoring altered as location changed, this was demonstrated across each analysis techniques applied (Table 2, Figs. 2, 4, 5, 6, 7). A potential explanation for this influence is the change of technique based on kick distance [17]. These findings align with research in other invasion sports, where location has also been identified as a predictor of kicking success in Rugby Union [7]. Further, two types of pressure, corralling and none, had the least influence on goal kicking accuracy in the logistic regression model (Table 2). Both the logistic regression and conditional inference trees showed closing and physical pressure to have a negative impact on shot outcome (Fig. 7). Location and defensive pressure have been shown to influence shot outcome in basketball [21, 23]. Shot type did not create a branch within the general play conditional inference tree (Fig. 7). However, differences between set shot and general plays shot success are demonstrated (Figs. 4, 6, 7).
Areas of equal opportunity in AF goal kicking success have been calculated in AF [37]. These should not only be calculated by location, but also include additional constraints such as physical pressure and shot type. The CBA and conditional inference trees (Figs. 5, 6, 7) demonstrated the similarity between combinations of constraint types and likelihood of goal kicking accuracy. Areas of equal opportunity are shown in Fig. 5C, D where a shot under physical pressure from X(0,10) and Y(20,30) has a confidence level of 0.38 in contrast with a shot under no pressure from X(40,50) and Y(20,30) which has a confidence level of 0.55. This is also demonstrated by the CBA as a general play shot taken from the same location has a very different likelihood of a goal based on the type of pressure (Fig. 6). This demonstrates how it may be beneficial to move the ball wider and further from goal, to avoid taking a shot under physical pressure. This information could be applied to inform decision-making around shot selection and education around the concept of areas of equal opportunity so that players understand that simply being closer to goal does not increase the likelihood of scoring, but the context of pressure and the shot type available will impact the outcome, thus this style of analysis provides a potential educational tool for athletes and coaches.
The analysis techniques utilised in this study displayed similar levels of accuracy and present the influence of constraints differently. The logistic regression model was more accurate at predicting a goal, whereas in contrast, the CBA and conditional inference models were better at predicting no goal (Table 1). For the logistic regression models, the influence of different constraints is evident, such as the impact of experiencing physical pressure compared with no pressure (Table 2); however, the interaction between constraints is difficult to observe. For instance, this model displays constraints independently and any interactions cannot be explored nonlinearly like in other methods. In contrast, the CBA and conditional inference trees models permit nonlinear interaction of constraints and their combined influence on goal kicking as a part of their inherent design as demonstrated (Figs. 4, 5, 6, 7). Whilst, each analysis technique suggests similar the patterns of constraint influence how these results and interaction is visualised varies.
The applicability of an analysis technique is due to more than its accuracy, but also how easily it can be interpreted and implemented. Benefits of the CBA and conditional inference tree techniques are both their nonlinear nature and visual output. These have the potential to demonstrate the interaction of multiple constraints. For example, a snap shot at goal is more likely based on a location close to the boundary line, whereas a shot from 50 m from goal whilst under chasing pressure would be a drop punt kick on the run. This may inform the design of goal kicking drills that better replicate competition, as they can consider both the frequency and prevalence of constraints. Dadzie and Rowe [38] suggest that visualisations may enhance the understanding of data, leading to the ability to enable instinctive and effective knowledge discovery. This is partly due to the decreased cognitive work required to interpret visualisations, as visuals take advantage of innate human perception [13, 38, 39]. In this, whilst the accuracy of each technique in determining the outcome of a shot on goal was similar, the CBA and conditional inference tree techniques may have an advantage over the logistic regression in regards to applicability, clear visuals may aid in-game and post-game assessment of shot selection and execution, although this was not formally investigated in this study.
Utilising multiple analysis techniques allows for the demonstration of variation and importance in model outputs and visuals. This is critical as findings may not be always be interpreted accurately and used effectively to inform decision-making [40, 41]. Further, when providing results to coaches, their willingness to accept and apply findings is critical [42]. Thus, a less accurate model, such as CBA which had the lowest F1 value, may be utilised over a slightly more accurate technique, due to the reduced complexity and higher interpretability of the model output. If results are too complex to interpret, then the likelihood of the findings being implemented is minimal. Thus, multiple analysis techniques can provide benefit in offering varied options to display results which can be aligned to the individual users. For instance, for some coaches, an understanding of how each individual constraint influences a shot on goal may help narrow down focus areas within training drills or guide language cues. Yet, other visualisations such as the heat map style of the CBA may enable a different perspective of potential kick success. They may help a coach understand how a team or individual is performing under certain circumstances in time-restricted environments. For example, in competition, being able to quickly understand how a shot success is being influenced by multiple constraints may help with decision-making in regards to personnel changes, tactics or messaging to players. Furthermore, if coaches are able to see differences in outcomes for kicks under the same constraints in training and competition settings, it may help to better inform drill design [43, 44]. The conditional inference tree visualisation provides a clearer grouping of similar opportunities and may aid coaches in educating athletes around their decision-making and what shot opportunities have an increased likelihood of success. For instance, a player may be passing the ball off to someone they think is in a better location to shoot from; however, they may be equally likely to be successful, therefore, they should take the shot themselves and not increase the chance of a turnover by making an additional pass. Ultimately, a coach may have a preference on how data are presented and being able to visualise multiple techniques allows for the customised presentation of results to suit coaches needs at a given time. Using the examples above, a coach may prefer a heatmap during a game or in post-game reporting to quickly demonstrate how the team performed under given match, whereas they may prefer to use a tree-like visual to further understand areas of equal opportunity to help develop an attacking game plan and team structures during planning sessions. It has also been suggested that appropriate staff should be embedded within professional clubs to aid in the statistical interpretation and applicability in industry settings, however, producing analysis in practitioner friendly formats is also of use [45].
Future research could include additional constraints within the models. Examples of additional constraints may include, exploring the game context such as time remaining and score margin as well as individual traits such as playing position and preferred foot [7, 46]. Additional data and the identification of key constraints which influence goal kicking could lead to more accurate models. This may help improve model accuracy to levels to make appropriate inferences from these data. Further data would enable the field to be divided into smaller regions to create more specific findings, as well as the potential to develop individual or team specific models. This would have a major impact in improving the accuracy and applicability of each model. Improved data capture may reduce subjectivity which currently exists in the measurement of currently collected constraints (for example see, Behendi et al. [47], Nibali et al. [48] and Victor et al. [49]). For instance, a constraint such as pressure could be measured on a continuous scale or as via a density metric [50].