The associations between physical-test performance and match performance in women’s rugby sevens players

Subjects

Thirty women’s Rugby Sevens players (age: 22 ± 5 y, height: 1.68 ± 0.05 m, mass: 69 ± 7 kg) representing five different New Zealand Provincial Union teams participated in the study. Each participant provided written informed consent and ethical approval was granted from the University of Waikato research ethics committee.

Study design

The association between physical-test performance and match performance in women’s Rugby Sevens players was examined using a descriptive correlation design. Participants performed a battery of physical tests within the two weeks before a two-day tournament. Match-running and match-action data were collected as measures of athlete performance using GPS units and video analysis. All tests employed are commonly used in Rugby [6–8].

Acceleration and maximal speed

Acceleration and maximal speed abilities were assessed over a 40-m sprint on an outdoor artificial turf. Single beam timing lights (Brower Timing System, Utah, USA) were positioned at 0, 10, 30, and 40 m. The first gate was set at a 0.5 m height, while the remaining were set at 0.75 m. Prior to performing the sprints, participants completed a 10-minute standardised warm up comprising of jogging, dynamic stretches, running drills, and three stride-outs at increasing intensity. Participants started each sprint in a standing split position 0.5 m behind the first gate and were instructed to “run as fast as possible” past the last gate. Each participant performed two maximal effort sprints, separated by three minutes of passive rest. Sprint time was measured to the nearest 0.01 second. The fastest 10-m, 30-m, 40-m, and 30–40 m times were transformed into average running speeds and used for analysis. The speed over 10 m was used as an indicator of acceleration ability, while 30-m, 40-m, and 30–40 m speeds were used as measures of maximal speed abilities. The re-test reliability (intraclass correlation coefficient, ICC) for 10-m, 30-m, and 40-m sprint times in women’s Rugby Sevens players using similar equipment was 0.90, 0.95, and 0.96 [8].

Fitness

The 1.2 km shuttle run test, also known as the Bronco test [7], was used as a measure of fitness. The test was performed outdoors, on artificial turf, in running shoes. The protocol consists of a continuous 20, 40, 60 m straight shuttle run, completed five times at maximal intensity (i.e., 20 m and back, 40 m and back, 60 m and back) [9]. Total running time was recorded with a stopwatch. Average running speed was calculated from total time and used for analysis. The ICC for Bronco time was 0.99 in men’s and women’s team-sport players combined [10].

Countermovement jump

Bodyweight countermovement jumps were assessed using dualaxis force plates (PASPORT force plate, PASCO, California, USA) and analysed using custom-made software (Weightroom, HPSNZ, Auckland, New Zealand) sampling at 100 Hz. Participants started from an upright standing position with their hands-on hips and were instructed to “bend their knees to a self-selected depth and to jump as high as possible”. Two warm up trials were given, followed by two sets of three jumps separated by two minutes rest. The best jump height (calculated from flight time) recorded by each athlete was included in the analysis. The re-test reliability of the best jump height out of three countermovement jumps was 0.97 (ICC) and 3.6% (typical error) in women’s soccer players [11].

Maximal strength

Back squat and bench press exercises were used to evaluate lower-body and upper-body maximal strength. For the back squat, athletes started from a standing position and were instructed to “lower until the thighs are parallel to the floor and then come up in the starting position”. A miss was recorded if participants failed to meet the proper depth or successfully come up in a straight position. For the bench press, athletes started with the arms fully straight, and were instructed to “lower the bar to the chest and press all the way up” while keeping the glutes in contact with the bench. If an athlete bounced the bar on the chest or failed to press the bar all the way up to a fully extended arms position, a fail was given. Each participant completed two warm up sets at sub-maximal intensities. Thereafter, participants were given three attempts to reach their 2–3 repetition maximum (RM) in each lift, with three minutes rest between attempts. One repetition maximum (1RM) from the lifts was calculated using the formula of Mayhew et al. [12]. ICCs for 1RM testing were ≥0.97 in women’s team-sport players [13].

Match performance

Match-running and match-action data were collected during the New Zealand National Rugby Sevens tournament, a two-day tournament between the New Zealand Provincial Unions where each team competes in 5 to 6 matches. Match data were considered for players that completed at least a full 7-minute half of a match; therefore, resulting in 1 to 6 files for each player and a total of 119 files included in the analysis (6 players = 1 match, 2 players = 2 matches, 2 players = 3 matches, 5 players = 4 matches, 7 players = 5 matches, 8 players = 6 matches).

Match running

Match running was measured using GPS units (VX Sport 220, Visuallex Sport International, Wellington, New Zealand) sampling at 10 Hz. The validity and reliability of devices with a similar sampling rate have been investigated previously [14, 15]. Each athlete wore the same GPS unit in every match in a fitted vest under the playing jersey. Data were downloaded and analysed post-tournament using the manufacturer’s software (VX View software, Sport International, Wellington, New Zealand). Match files were trimmed to include only the time players were on the field. The variables analysed were based on women’s Rugby Sevens research [3, 4, 16] and were described as the frequency of efforts or cumulative distance covered in different speed zones (see Table 1). Sprints were defined as running efforts that required an increase of ≥0.70 m∙s⁻¹ within a second and that reached ≥2.8 m∙s⁻¹ [17]. High-intensity accelerations and decelerations represented the the total number of accelerations and decelerations performed ≥2.0 m∙s⁻².

Match actions

The first author coded match actions using video analysis. The match actions included in the analysis and their operational definitions are presented in Table 2. These measures were chosen to represent different areas of the game in agreement with previous Rugby Sevens studies [6, 18]. Intra-rater reliability for the analysis was evaluated by re-analysing 10 random matches four weeks apart and calculating the percentage error, as described by Hughes et al. [19]. Errors observed for all match activities were within a 5% error limit, which was deemed acceptable.

Statistical Analysis

Data were analysed with the Statistical Analysis System (University Edition of SAS Studio, version 9.4, SAS Institute, Cary NC). Pearson correlations between each pair of physical-test variables were derived as a correlation matrix, and the variables were ordered to reveal clusters of similar variables (higher correlations within clusters than between clusters). The same analyses were performed for matchperformance variables.

For measures of match performance that were counts or proportional to counts, the association between each physical characteristic and the measure was analysed with Poisson regression using the generalised linear mixed model procedure (Proc Glimmix) with a log link. This procedure allows modelling of count variables and accounting for repeated match-performance measurements on the same player. Each physical characteristic (predictor) presented in Table 3 was entered in the model separately as a numeric linear fixed effect to allow estimation of the effect of a two standard-deviation (2-SD) difference in the predictor on match performance [6, 20]. The number of the match played by each athlete in the tournament, and the log of total match time (as a fraction of a 14-min match) for each player in each match, were included as numeric linear fixed effects to estimate the tournament trend of the dependent variable and to adjust each player’s score to a mean match time, respectively. Random effects in the model were nominal variables representing player identity (to adjust for between-player differences in means), match identity (to adjust for between-match differences in means), and team identity (to adjust for between-team differences in means). Over- or under-dispersion of the residual variance was estimated, which was particularly important for measures representing counts of running bouts. Distance covered >7.5 m∙s⁻¹ produced an unrealistic estimate of over-dispersion, so it was analysed with the general linear mixed-model procedure (Proc Mixed) after log transformation, with values of 0 first set to half of the smallest non-zero value. The fixed and random effects were the same as for the Poisson-regression model. The same general linear mixed model was used to analyse the only measure of match performance that was not a count or count of bouts (maximal speed), which was also log transformed. Estimates for the tournament trend were also obtained from the generalised and the general linear mixed models without a physical-test predictor.

The qualitative magnitude of the effects was assessed using standardisation, with threshold values for small, moderate, large, very large, and extremely large calculated as 0.2, 0.6, 1.2, 2.0, and 4.0 of the observed between-player SD [20], derived by combining the variances represented by player identity (true differences between players) and the residual (within-player between match variance), and adjusted for small samples [21]. Effect magnitudes for tries were also determined as the factors associated with an increase in the number of tries scored by a team to give the team 1, 3, 5, 7, and 9 extra wins every 10 matches, representing small, moderate, large, very large, and extremely large effects [20]. The factors were estimated using a simulation based on points scored by all the teams in the tournament. There was a mean of 2.7 tries scored per team per match (5 points per try), with a 52% probability of converting a try (2 points per conversion). There were no field goals or penalties. We assumed a team had 10 try-scoring opportunities in a match on the basis of our experience (The resulting magnitude thresholds were not sensitive to the number of opportunities). An opponent team was assumed to have an unchanging probability of scoring a try per opportunity equal to 27% (2.7 tries per 10 opportunities). The factor associated with a given increase in wins per 10 matches allowed for the affected team to have a try-scoring probability per trial <27% before the factor was applied (probability = 0.27/√factor), but it increased to >27% after the factor was applied (probability = 0.27*√factor). Simulations were performed to generate scores in 10,000 matches for the two teams before and after the factor was applied, then wins were scored as 1 and loss or draw as 0. The factors giving 1, 3, 5, 7, and 9 extra wins every 10 matches were found by “trial and error”. Corresponding magnitude thresholds for factor increase/decrease were 1.20/0.83, 1.75/0.57, 2.60/0.38, 4.10/0.24, and 11.0/0.09. The spreadsheet of simulations is available on request.

Physical-test scores are shown as means and standard deviations (SDs). Means of the dependent variables are shown as the back-transformed least-squares means with SDs in back-transformed ± percent units (when <30%) or × /÷ factor units (when >30%) derived from a model without a physical-test predictor. Effects are presented in percent units with uncertainty expressed as ± 90% compatibility (or confidence) limits (CL), when either the effect or the ± 90% CL were <30%; otherwise, factor effects with × /÷90% CL are reported. Decisions about magnitudes accounting for sampling uncertainty were based on one-sided interval hypothesis tests, according to which a hypothesis of a given magnitude (substantial, non-substantial) is rejected if the 90% compatibility interval falls outside that magnitude [22, 23]. P-values for the tests were therefore the areas of the sampling t-distribution of the effect falling in the hypothesised magnitude, with the distribution centred on the observed effect. Hypotheses of inferiority (substantial negative) and superiority (substantial positive) were rejected if their respective p-values (p₋ and p₊) were <0.05; rejection of both hypotheses represents a decisively trivial effect in equivalence testing. The hypothesis of non-inferiority (non-substantial-negative) or non-superiority (non-substantial-positive) was rejected if its p-value (p_N– = 1 – p₋ or p_N+ = 1 – p₊) was <0.05, representing a decisively substantial effect in minimal-effects testing. A complementary Bayesian interpretation of sampling uncertainty was also provided, when at least one substantial hypotheses was rejected: the p-value for the other hypothesis is the posterior probability of a substantial true magnitude of the effect in a Bayesian analysis with a non-informative prior [22–24], and it was interpreted qualitatively using the following scale: >0.25, possibly; >0.75, likely; >0.95, very likely; and >0.995, most likely [20]. The probability of a trivial true magnitude (1 – p₋–p₊) was also interpreted with the same scale. Possible or likely magnitudes are categorised as some evidence for those magnitudes; very likely and most likely are categorised as good evidence. Probabilities were not interpreted for unclear effects: those with inadequate precision at the 90% level, defined by failure to reject both substantial hypotheses (p₋ >0.05 and p₊ >0.05). Effects on magnitudes and probabilities of a weakly informative normally distributed prior centred on the nil effect and excluding extremely large effects at the 90% level were also investigated [24, 25].

Effects with adequate precision at the 99% level (p₋<0.005 or p₊<0.005) are highlighted in bold in tables, since these represent stronger evidence against substantial hypotheses than the 90% level and therefore incur lower inflation of error with multiple hypothesis tests. For practitioners considering implementation of a treatment based on an effect in this study (e.g., training to improve try scoring by increasing jump height), the effect needs only a modest chance of benefit (at least possibly increased try scoring, p₊>0.25) but a low risk of harm (most unlikely impaired try scoring, p₋<0.005). Substantial effects highlighted in bold therefore represent potentially implementable effects. However, it is only for effects on tries scored assessed via match winning that the outcomes have direct relevance to benefit and harm (winning and losing matches); these effects were also deemed potentially implementable when the chance of benefit outweighed an otherwise unacceptable risk of harm (the odds ratio of benefit to harm >66.3) [26]. For these effects, the potential for benefit and harm was also investigated for realistic changes in physical-test measures (less than 2 SD).

Physical tests and match performance

Mean values and between-subjects SD of physical-test scores are presented in Table 3, while means of the dependent variables with the within-player, between-player, and observed SD are reported in Table 1. The within-player SD represents the match-to-match within-player variation, the between-player SD is the true difference between players, and the observed SD is the combination of the within- and between-player SDs representing the observed between-player SD in a typical match.

Correlations within and between groups of variables

Correlation matrices for the physical-test and match measures are shown in the Supplementary Tables 1–4, where clusters of variables have been identified as those with correlations ≥0.50 within clusters and <0.50 between clusters. The clusters of variables in each correlation matrix are delineated in the Supplementary Table 4, because correlated variables were expected to have similar effects. There were two overlapping clusters of physical-test variables for maximal-speed running measures, one with 10-m speed and one with Bronco; there was also a well-defined cluster for strength measures (Supplementary Table 1). The correlation matrix for measures of match running revealed four clusters, representing running at high intensities, running at lower intensities (with distance >5.5 m∙s⁻¹ in both clusters), total running (with sprinting, accelerations, and decelerations contributing to this concept), and high-intensity changes in speed (Supplementary Table 2). Fewer and less well-defined clusters of variables were found for match actions, with a cluster for tries and line breaks, and a cluster (with one correlation of 0.47) for work rate, carries, and tackle breaks (Supplementary Table 3). No clusters contained match-running and match-action variables, but correlations of match running with match actions (Supplementary Table 4) revealed similar magnitudes within the running and action clusters; in particular, tries and line breaks had negative correlations with measures of total running and positive correlations with high-intensity running.

Effects of physical-test scores on match performance

The effects of a 2-SD difference in physical-test scores on match running and match actions are presented in Table 4. Compatibility intervals and Bayesian probabilities are shown for a minimally informative prior, since appreciable shrinkage occurred with the weakly informative prior for only one effect, jump height on tries scored. In this instance, the factor effect reduced to 3.1, 90% compatibility limits × /÷2.6, but the magnitude remained large, very likely substantial, and potentially implementable with the odds-ratio assessment of benefit and harm defined by thresholds for match winning.

Match running

Large positive effects were observed for jump height on match maximal speed and distance covered at high intensity, and for acceleration on match maximal speed, where the effects had sufficient precision for the true magnitudes to be very likely or most likely substantial. Acceleration, speed, Bronco, and strength scores displayed consistent moderate positive effects on distance covered at high intensity during matches, with some effects showing sufficient precision for the true magnitudes to be very likely substantial. Moderate positive effects where precision was sufficient for the true magnitude to be very likely substantial were also evident for 30-m speed on match maximal speed, for jump height on distance >7.5 m∙s⁻¹, and for Bronco on match total distance, decelerations, and high-intensity decelerations.

Measures where the effects had adequate precision but were only likely substantial included small and moderate positive effects for speed, Bronco, and back squat on match maximal speed, and for speed and Bronco on match distance covered >7.5 m∙s⁻¹ and on some measures of total running. In contrast, small to moderate negative effects were observed for strength and jump height on some variables contributing to total match running and high intensity changes in speed. Precision was consistently inadequate for speed on match distance 3.5–5.0 m∙s⁻¹, accelerations, high-intensity accelerations, high-intensity decelerations, and for strength and jump height on match decelerations and high-intensity decelerations.

Match actions

Moderate and small positive effects were observed for jump height and back squat on tries scored when assessed using standardisation, with adequate precision at the 90% and 99% levels respectively; both were likely substantial. Greater positive effects characterised the same predictors for tries scored when assessed via match winning, and the effects became very likely substantial; both effects had adequate precision only at the 90% level, hence the risk of harm was too high (p₋>0.005) for a conservative assessment of implementability, but they were potentially implementable when considering the odds ratio of benefit to harm (3700 and 4200, respectively). For these predictors, changes as small as 0.25 SD predicted tries scored that were at least possibly beneficial and with negligible risk of harm, but changes of 0.2 SD were unlikely to be beneficial.

A number of predictors had small or moderate positive effects but were only likely or possibly substantial: back squat and jump height on line breaks, work rate, carries, tackle breaks; 10-m speed on line breaks; maximal speed on work rate; Bronco on work rate and carries; and strength on tackles. On the other hand, back squat and jump height displayed small negative effects on turnovers won with adequate precision, but the effect was only possibly or likely substantial. Precision was inadequate for several predictors on various match-action measures, including the observed small positive effects of measures of running speed and bench press on match winning. Changes in measures of running speed and in bench press smaller than 2 SD were either unclear or at least likely trivial.

Tournament trend

When considering the effects obtained without a physical-test predictor, the measures of match total running and high-intensity changes in speed displayed small to moderate negative trends across the tournament, with adequate precision but only possibly or likely substantial magnitudes. The trend for the match measures of high-speed running ranged from small likely reductions through to small possible increases, but three of the six measures had inadequate precision. There was a similar pattern for match actions, with seven of the 11 measures lacking adequate precision. The tournament trends sometimes changed substantially with different physical-test predictors in the model, but overall the trends were similar to those without predictors.