Applying different levels of practice variability for motor learning: More is not better

Participants

Sixty-six healthy participants (33 females, 33 males) took part in this study (age = 23.35 ± 3.9 years; height = 1.7 ± 0.1 m; mass = 66.3 ± 11.9 Kg). Exclusion criteria included current musculoskeletal injuries or coordination deficits that impaired participants from throwing a ball against a wall in front of them and catching the rebound. All participants were right-handed and participated in the study voluntarily.

Written informed consent was obtained from each participant prior to testing. Data were treated anonymously, and all participants were informed of the risks and benefits of the trial. The experimental procedures used in this study were in accordance with the Declaration of Helsinki and were approved by the University of Miguel Hernández of Elche Office for Research Ethics (Ethical Application Ref: DPS.FMH.01.16, DCD.FMH.01.21).

Experimental procedure and data collection

Participants performed a repetitive underhand throwing test with a tennis ball using their right hand. The task goal was to hit a target located on a wall that was at 1.65 m from the participant and at a height of 2.15 m high from the ground. During the throws, the participants stood with their feet placed at shoulder width and the foot orientation was such that the vector formed by the heels was parallel to the mediolateral axis of the platform (see Fig. 1). Participants rested their left hand on their hip during the entire test. After a throw, the participants had to catch the tennis ball as it bounced off the wall to perform the upcoming throw, using their right hand all the time. In case participants did not catch the ball, they could catch a new ball located on their right side at hip height (see Fig. 1). All the participants kept a rhythm of 40 throws per minute during the throws following a beep given by a metronome.

The protocol carried out by all participants consisted of a pre-test, a training period of 2 weeks, a post-test, and three retests (96 h, 2 weeks and 1 month after training period). Each test (pre-test, post-test, and retest) contained four sets of 20 repetitions each. The sets were divided into two blocks of two sets, with a 30 s rest between sets and 1 min rest between blocks. The training period consisted of eight sessions of eight sets of 20 trials each. The sets were divided into blocks of four sets. The resting time between sets was 30 s and the resting time between blocks was 1 min.

After the pre-test, participants were distributed according to their initial performance to homogenize the group’s performance level into four experimental groups and a control group, which did not practice in any condition (N = 10). The four experimental groups were designed according to the manipulation of the load of variable practice, referred to the amount of variable practice induced in the task conditions according to the heterogeneity of the throwing movements (Raviv, Lupyan & Green, 2022), as follows: a) constant group (CS), which trained doing constant practice in the same conditions as the evaluation tests explained above (N = 14); b) low variability group (LV), which trained the throws with variable angles and orientation, aimed to targets dispersed in a radius from 0 to of 20 cm (6.91°) maximum from the main target (i.e., the target used during the evaluation tests) (N = 13); c) medium variability group (MV), which throwed with more variable angles and orientations trying to hit targets dispersed in a radius from 0 to 40 cm (13.63°) from the main target (N = 11); and, d) high variability group (HV), which throwed with variable angles and orientations aimed to targets dispersed in a range of variable radius up to 60 cm (19.98°) maximum from the main target (N = 15). For all the variable groups, twenty targets were scattered across the circle within the corresponding maximum radius. The targets were spread out randomly and numbered to guide the participants during the practice series. In order to reduce the attentional demands of the task during variable practice, and due to the large variety of possible trajectories practiced in variable groups, the targets were distributed in two sets numbered from 1 to 10 and in two colors, red and blue, going in numerical sequence. The participants were instructed to start with targets colored in red and then those colored in blue. No participant reported any difficulty in identifying the targets or the order (see Fig. 2 to check the target features).

Each ball impact on the wall was video recorded with a “Sony HDR-SR8E” digital camera (Sony, Tokyo, Japan), at a frequency of 50 Hz, to establish the impact zone of each throw.

Data analysis and reduction

The ball impacts were digitalized using Kinovea (version 0.8.15). An ad-hoc Matlab (version 7.11; Mathworks, Natick, MA, USA) routine was used for the calculation of real-space Cartesian coordinates of the ball rebounds. The video analysis to identify the impact zone was made manually by two different trained researchers, following the procedure used in previous studies (Moreno et al., 2023, 2020). Each participant’s accuracy was measured through the mean radial error (MRE), as a global accuracy index, computed as the average of the vector distance magnitude (cm) of the ball impact place from the target position of all the throws. The average of the absolute error in the vertical (AE_V) and horizontal (AE_H) axis were also calculated and included in the Supplemental Material. In order to compute the absolute amount of learning (AAL) the differences between the post-test and each of the three retention tests compared to the pre-test were calculated. The computations were made as follows: a) differences between pre-test (MRE_PRE) and post-test (MRE_POST); b) differences between pre-test (MRE_PRE) and retest 1 (MRE_RET1); c) differences between pre-test (MRE_PRE) and retest 2 (MRE_RET2); d) differences between pre-test (MRE_PRE) and retest 3 (MRE_RET3). However, Pearson’s correlations revealed that the different ways to compute the absolute amount of learning by the differences between the post-test and the three retention tests compared to the pre-test were positively correlated with each other (Table 1). Thus, to simplify the results, it was decided to use the amount of learning computed by the differences in MRE between the pre-test and the retest 3, representing learning after 1 month. In addition, the correlational analysis revealed a relationship between the initial performance and the absolute amount of learning (r = 0.308*, p < 0.05). This is, participants who displayed lower performance (higher MRE) achieved higher absolute amount of learning (higher AAL). Then, to avoid the potential bias on the amount of learning caused by the initial performance, a linear regression method was computed following the procedure used in previous studies (Barbado et al., 2017; Caballero, Moreno & Barbado, 2021). For that, a linear regression was estimated between initial performance (MRE_PRE) and the absolute amount of learning (AAL) (see Fig. 3A). Finally, participants’ residual scores were used to compute the residual amount of learning (RAL), which was used as the amount of learning index (Fig. 3B).

In this study, the standard deviation of the ball impact position was not used as the participant’s initial intrinsic variability index because it was highly correlated (r > 0.9) with error parameters. Instead, the Lag-1 autocorrelation coefficient of the time series data from the vertical and horizontal positions of the ball impacts in the pre-test was calculated to assess the initial intrinsic variability of the participants. Lag-1 autocorrelation coefficient allows for addressing the relationship between consecutive attempts in tasks, such as hitting a target, and assessing how each attempt is influenced by the one before it. The autocorrelation function measures the average strength of each data point with data point k time steps away (Hunt, 2016). This approach is supported by the study of Skurvydas, Brazaitis & Kamandulis (2010), who applied autocorrelation as a measure of the linear dependencies of successive trials, reflecting the variability of motor outcomes. Therefore lag-1 autocorrelation is valued for providing insights into the corrections a participant makes over successive trials, offering a complement to conventional accuracy and variability measurements. The average of the autocorrelation coefficient of the positions of the ball impacts on each axis was determined as a global variable of each participant’s initial intrinsic variability.

Statistical analysis

Prior to any statistical analysis, three subjects were removed from the original sample due to measurement issues. Thus, the participants analyzed were 63, distributed in the different groups as it was specified above in the Experimental procedure and Data Collection. The normality of the variables was evaluated using the Kolmogorov-Smirnov test with the Lilliefors correction. Intention-to-treat analyses were carried out for all parameters. Only one participant from the control group did not complete all the measurements. Specifically, that participant could not perform retest 1, so missed data were imputed following a multiple imputation method based on a linear regression estimation. A mixed ANOVA was conducted for accuracy, with practice as a within-subject factor (five levels: pre-test, post-test, retest 1, retest 2, retest 3) and group as a between-subject factor (five levels: control, constant, low variability, medium variability, high variability). A Bonferroni adjustment for multiple comparisons was performed to ascertain differences between groups for the different test evaluations. Hedge’s g was calculated as the effect size index. Hedge’s g algorithm adjusts Cohen’s d score, which is based on sample averages and gives a biased estimate of the population effect size (Hedges & Olkin, 1985), especially for small samples (N < 20). Values of effect sizes ≥0.8 were considered large, from 0.8 to 0.5 were considered moderate, and <0.5 were considered small effect sizes (Cohen, 1988). Pearson’s correlations were carried out to extract underlying relationships between the initial intrinsic variability, the level of variable practice load, and the amount of learning.

Do different loads of variable practice induce different amount of learning?

All the experimental groups improved their throwing performance after the intervention compared to the control group. Even though all practice conditions improved throwing accuracy, it was expected that the amount of learning would not be similar between groups. Therefore, which were the best conditions concerning variable practice to improve the learning process? One of the main ideas that guided this study is that applying variable practice could ease the individual’s exploratory behaviors or enhance the individual’s sensitivity to their own motor error, promoting adaptative behaviors and a faster learning rate (Barbado et al., 2017; Manor et al., 2010; Wu et al., 2014; Zhou et al., 2013). Inducing variability would provide learners with more opportunities to search and explore their environment to create a range of effective and adaptable movement solutions, leading to a higher probability of generalizing outside the movements practiced (Tenenbaum & Griffiths, 2001). However, according to the results of the current study, it cannot be confirmed that variable practice was better than constant practice since the variability groups did not show significantly better results than the constant group. It has been emphasized that in dynamic environments, conditions are never identical. The time or the space needed to perform a task continuously changes, whether the skill is performed in a high or low predictable environment. Therefore, based on the principle of ‘repetition without repetition’ (Bernstein, 1967), every movement is different, even in closed skills. Adapting to the changing conditions provided by the neuromuscular system in the interaction with the environment is a key factor for success. In this sense, previous researchers have highlighted the need to adjust the amount of variability through the manipulation of individual, environmental, and task constraints. Renshaw et al. (2016) suggest the need to find the optimum level of variability. They proposed that a low amount of variable practice would not promote additional motor pattern forming or an adequate system reorganization, but, on the other hand, too much variability would make the environment unmanageable for the learner.

Based on the rationale mentioned above, it was hypothesized it would be possible to observe an inverted U-Shape between the different loads of variable practice and the amount of learning in adaptation (post-test) and consolidation tests (retests). Therefore, it was checked if low-to-medium levels of variability could be optimal to improve the amount of learning as suggested by previous studies (Caballero, Luis & Sabido, 2012; Harbourne & Stergiou, 2009; McEwen & Lasley, 2002; Moreno & Ordoño, 2015; Ranganathan & Newell, 2010). According to the findings of this study, no significant differences in the amount of learning were observed between practice groups, so initially, the inverted U-Shape cannot be confirmed. Nevertheless, from the authors’ point of view, the fact that the low variable practice group showed larger benefits in terms of effect size (0.93 ≤ d ≤ 1.00) compared to the other group (Constant group: 0.55 ≤ d ≤ 0.69; Medium variability group: 0.58 ≤ d ≤ 0.72; High variability group: 0.71 ≤ d ≤ 0.80) makes it not entirely possible to reject the hypothesis. Figure 4C shows that the lowest variability conditions (i.e., constant practice group) and the highest variability conditions (i.e., high variability and even medium variability groups) did not cause such a large learning effect size in this throwing task. The lower effect size observed in the constant practice group could be caused by a reduced practice stimulus that did not maximize the learning process. Therefore, some increased level of motor variability seems to be helpful in improving the learning process. Supported by the idea that the CNS regulates intrinsic motor variability to enhance the ability to adapt (Barbado et al., 2012; Davids et al., 2003; Lamoth & van Heuvelen, 2012; Mandelblat-Cerf, Paz & Vaadia, 2009; Moreno & Ordoño, 2010; Wu et al., 2014; Zhou et al., 2013), the variable practice could stimulate the intrinsic variability to keep exploring the possible solutions. However, the results from the medium and high variability groups do not support the application of high levels of variability to foster the learning process. Based on these results, high variable practice conditions might hinder the search for a stable motor solution to fulfill the demands of the throwing task proposed in this study. In spite of these noticeable findings, it must be reinforced that the sample size does not allow us to confirm the inverted U-shape hypothesis. Therefore, more research is needed to elucidate whether there is an optimal variability load for promoting a higher amount of learning depending on the task and individual features.

Do the individual’s intrinsic features modulate the effect of variable practice?

Quite possibly, the differences found regarding the effect of the different levels of variability can be related to the fact that not all individuals respond in the same way to a similar training stimulus. Therefore, the critical point is that the individual’s characteristics modulate the role of motor variability during the learning process (Caballero et al., 2017; Newell & Vaillancourt, 2001; Vaillancourt & Newell, 2002, 2003). Renshaw et al. (2016) proposed that the amount of variable practice needed to be matched to the performer’s skill level. Beginners could benefit from low loads of variable practice, which guide their exploration towards reduced functional solutions. Conversely, the most expert performers would need more significant enhanced variability during practice to promote more dexterous behaviors (García-Herrero et al., 2016). Another important learner feature is the intrinsic motor variability itself. Previous studies have shown that a higher motor variability (i.e., higher motor exploration) at the baseline is related to a higher amount of learning (Barbado et al., 2017; Wu et al., 2014). Based on this idea, it could be expected that people with higher intrinsic motor variability do not need additional variable practice to promote a higher amount of learning. On the other hand, people with low intrinsic motor variability would benefit from variable practice.

In this study, the motor task performed was a relatively common throwing and catching ball task, but participants displayed different initial performance levels as well as different initial intrinsic variability. The participants were distributed into the experimental groups so there were no differences between groups at the beginning of the treatment. Nonetheless, it also implies that, in every group, there were participants with a different initial skill level and different initial intrinsic variability, so this should have been considered. The correlational analyses showed that those participants with higher errors in the pre-test showed a higher amount of learning. Thus, the amount of learning is highly determined by the initial performance because less skillful individuals have more room for improvement than the more skilled ones. With the purpose of avoiding the potential bias on the amount of learning caused by the initial performance, a residual amount of learning was computed using a linear regression method (Barbado et al., 2017; Caballero, Moreno & Barbado, 2021) (see the Statistical Analysis section for more details). The participants’ initial intrinsic variability by the global autocorrelation of the time series data of the ball position was also measured, providing information about the internal relationships between the successive attempts to hit the target. Bivariate correlational analyses showed that more autocorrelated variability was not related to the amount of learning when all the participants, independently of the training group, were included in the analysis. However, once the sample was separated according to the training groups, those variables were positively correlated for the group in which participants trained with a low load of variable practice. The autocorrelation of the error time series data has been used to identify the way in which the performer varies trying to adjust his/her movement to an external target, and it has been related to the intention to adjust the movement to a target (Urbán et al., 2019). In this sense, the behavior of individuals who showed higher autocorrelated fluctuations (i.e., lower number of changes in the error direction) performed fewer movement adjustments (Amoud et al., 2007; Wang & Yang, 2012; Zhou et al., 2013), which in turn, would be an index of poor explorative behavior or lower motor error sensitivity (Barbado et al., 2017). Based on this feature, these participants would be more benefited by variable practice, to promote intrinsic exploration behavior. A variable environment would help them find an optimal motor solution to adjust their movement to the target and foster their learning. These results are in coherence with those found in an advanced stage of learning, in which exploitation rather than exploration behavior is utilized (Herzfeld & Shadmehr, 2014; Wu et al., 2014). It must be pointed out that this correlation was not observed when higher variable practice was applied (MV and HV groups). In these cases, individuals with lower intrinsic exploration behavior did not benefit by an “excessive” variable practice. These results also seem to support the idea that an optimal load of variable practice must be found to maximize the amount of learning according to each participant’s features. However, this rationale should be confirmed by future studies as the low sample size for each group could bias the correlation results.

Limitations

The conclusions of this study have to be considered with caution due to some limitations. The main limitation was the small sample size, making it difficult to find differences between groups despite the large effect sizes observed. Additionally, the groups were balanced according to the initial performance of the participants, but groups were heterogeneous in the distribution of the autocorrelated variability of the participants. Considering this and following previous suggestions (Woolley & Doupe, 2008; Wu et al., 2014), a larger sample size would have allowed the groups to be divided according to their intrinsic variability and performance levels to confirm how the intrinsic characteristic of the learner could mediate the effect of the different load of variable practice.

In order to diminish the attentional demands that can be attributed to selecting the targets to throw at in the variable practice conditions, the targets were carefully numbered and colored to facilitate their identification. Nevertheless, it can be considered an additional limitation of the study.

Another limitation is related to the fact that we did not assess the performance during the learning practice period, as well as the intrinsic variability within each experimental group throughout the acquisition phase. The evolution of performance during practice can be informative about the participants’ learning processes within each group. However, it was not assessed as the performance during practice trials could not be compared because each group practiced throwing toward different dispersed targets. The intrinsic variability of the participants could be used to analyze the extent of variability load they were exposed to, and it could be considered in future studies.

It also has to be taken into consideration that in this experiment, the load of variable practice has been chosen arbitrarily, according to the limits and possibilities of the task. The highest level was determined by the maximal angle in throwing that allowed the participants to catch the ball repeatedly without moving their feet from the marks on the floor. Therefore, the classification of low, medium, and high variability levels has to be taken with caution because these were not adjusted to the level of the learner. Moreno & Ordoño (2015) proposed that the tasks should be designed with a load level higher than the demands to which the learner is adapted to facilitate a new level of adaptation. To the best of the authors’ knowledge, no previous literature has addressed objective procedures to measure the load of variable practice in experimental learning procedures. Future studies should be directed to manipulate the variable practice level adapted to the learner’s intrinsic variability.

Finally, a possible relation between the effect of different loads of variable practice and the initial intrinsic variability of the learner measured by the autocorrelation of the time series data from the error of their trials has been found. The autocorrelation function measures linear dependencies between data points but not nonlinear dependencies. It cannot distinguish the structure of the variability to allocate the complex, chaotic, or random characteristics (Hunt, 2016). Due to the time series data length analyzed, other nonlinear tools that assess long range autocorrelation, like detrended fluctuation analysis, were not used. Nevertheless, linear autocorrelation can be used to analyze the fluctuations and dependencies of data points in a time series and it is applicable in short time series data.