Performance analysis in esports: modelling performance at the 2018 League of Legends World Championship

Abstract

Performance analysis is a well-established discipline in sports science, supported by decades of research. Comparatively, performance analysis in electronic sports (esports) is limited. Therefore, there is an opportunity to accelerate performance outcomes in esports by applying methods grounded in sports science. This study adopted a coach-centred approach to model performance at the 2018 League of Legends World Championship. Three expert coaches rated the proposed relationship between 43 variables and match outcomes in professional League of Legends competition using a Likert scale (1–10). The Likert scale was anchored with ‘no relationship’ at 1 and ‘very strong relationship’ at 10. The coaches’ median ratings were calculated for each variable. Variables with a median score ≥6 were retained for analyses. A total of 14 variables were collected from the 2018 League of Legends World Championship (n = 119) matches via video annotations and match histories. Generalized Linear Mixed Effects Models with binomial logit link function were implemented with respect to the Blue Side winning or losing the match, and individual teams were specified as random effects. Variables were screened for multicollinearity before using a step-up approach. The best model of performance included Tower Percentage (p = 0.006) and Number of Inhibitors (p = 0.029). This model achieved classification accuracy of 95.8%. While Tower Percentage and Number of Inhibitors contributed to winning or losing, further research is required to determine effective strategies to improve these variables, to understand the relevance of these variables across the complete time-series of the match, and to determine whether performance indicators remain stable across game updates.

Keywords

Electronic sports sport analytics video games

Introduction

Electronic sports (esports) is a modern competitive environment where players compete against each other via human-computer interactions,¹ thus, differentiating esports from the broader field of video gaming. There are numerous genres of games within esports, and a classification of Action Video Games has emerged that includes Multiplayer Online Battle Arenas, First Person Shooters, Third Person Shooters and Real-time Strategy.² One of the most widely played Action Video Games is League of Legends – a Multiplayer Online Battle Arena game released in 2009, with up to 100 million active users recorded per month.³ In League of Legends, two teams of five players work against each other to destroy the opposing team’s Nexus (i.e. the main structure located at the opposite side of the arena). The arena consists of three lanes that cross from the bottom left corner (Blue Side) to the top right corner (Red Side), with jungle areas separating each lane and a river passing from the top left to bottom right (Figure 1). At the start of the game, players select from a pool of 145 Champions (the Champion pool size as of September 2019) that they will control as their in-game avatar throughout each match, each of which has unique abilities and skills that the player can activate.

Figure 1.

League of Legends arena.

During a match, players face various enemies and objectives including the five opposing team players, waves of computer-controlled Minions, Jungle Monsters, enemy buildings called Towers, and other monsters that only spawn at specific times such as Elemental Drakes, the Elder Dragon, Rift Herald, and Baron Nashor. Attacking each of these enemies and objectives awards experience (used to “level-up” the players’ Champions), and gold (in-game currency used to purchase items), which make Champions more powerful via awarding new abilities or increasing the five primary attributes of attack damage, ability power, armour, magic resistance, and health. Further, bonus rewards are available for acquiring key objectives such as First Blood (extra gold for the player(s) who obtain the first kill in a match), First Tower (extra gold for the player(s) who destroy the First Tower), the Elder Dragon (power increases relative to the number Elemental Drakes that the team has slain), the Rift Herald (the player granted an additional powerful ability that they can use within a 4-minute time frame) and Baron Nashor (provides all surviving teammates with temporary bonus damage and an aura to amplify the power of their team’s nearby Minions). As Champions become stronger, they can overcome objectives more quickly that can ultimately aid in winning the match.

League of Legends is currently played competitively across numerous regions (Korea, North America, Europe West, Europe Nordic and East, Oceania, Russia, Turkey, Brazil, Latin America North, Latin America South, Vietnam, South-East Asia, and Japan). The number of ranked players varies between 200,000 players in emerging regions like Oceania and Russia, to more than 3.5 million ranked players in established nations like Korea (data accessed via website https://[region].op.gg/statistics/tier, 15/09/2019). Many players are employed full-time contracts and earn a salary to train and compete in League of Legends professional. Notably, the 2018 League of Legends World Championship boasted a prize pool of USD 6.45 million (retrieved 24/09/2019 from https://www.esportsearnings.com/tournaments).

To date, most of the research has focused on video gaming and the relationship between participation and domain-general perceptual-cognitive and perceptual-motor skills. For example, there is some evidence that video gaming is associated with enhanced working memory capacity and information processing skills.^4,5 Furthermore, expertise in multiplayer online battle arena games is related to domain-general cognitive skills and numerical processing.⁶ However, a meta-analysis of experimental studies yielded negligible effects of video gaming on executive functions such as working memory, multitasking, nonverbal intelligence, and task switching.⁵ While there is conflicting evidence concerning the relationship between domain-general skills and video gaming, no research has investigated the in-game actions associated with successful match outcomes within professional esports competitions.

Due to the lack of available peer-reviewed scientific research, professional esports coaches and players must use information that is considered poor scientific quality (e.g. anecdotal observations, individual experiences, and unvalidated statistics). Anecdotal reports reveal that professional Oceanic esports coaches and players rely on knowledge derived from other disciplines that are considered somewhat similar such as chess and sports (personal communication, [April 2019]), while some researchers have also drawn links between esports and chess.^1,7 In traditional sports, performance analysts and coaches collect data via manually annotating video footage, which can form part of the feedback loop that informs coaching decisions.⁸^–¹¹ Annotating video footage is a manual and time-consuming process, and coaches/performance analysts must decide what actions are the most relevant to capture. In contrast, esports games often record all in-game actions quantitatively and automatically, with data accessible via public websites. The increased accessibility to online data has the potential to accelerate performance analysis outcomes and the coaching feedback loop within esports. Although automatically logging in-game data should theoretically be more accurate than manual video annotation methods (e.g. during sports), the automated data are captured, aggregated, stored and visualised by proprietary software systems, and the validity of the data should not be accepted without investigation. Additionally, any software system can be subject to errors caused by bugs, software and hardware changes, or oversight of an engineer or developer i.e. there is still a level of human error involved. Ultimately, validating esports data before implementing data-driven performance analysis is essential. Therefore, the first aim of the current study was to conduct a preliminary analysis to assess the validity and inter-rater reliability of automated Match History statistics in professional League of Legends competitions, by drawing on methodologies grounded in sports science.^12–14 Secondly, the research aimed to engage with expert coaches to determine which in-game actions they believed were associated with successful match outcomes. Finally, the research aimed to determine which of the variables suggested by expert coaches were associated with successful outcomes at the highest level of competition – the 2018 League of Legends World Championship.

Methods

Before undertaking the coach survey and performance analysis, the validity and inter-rater reliability of Match Histories (official summary statistics) and video annotations were assessed by using 30 randomly selected League of Legends matches from the 2019 North American and European professional competitions. Given that the true measure of the action variables cannot be captured via the match histories due to the potential for software bugs and inaccurate data handling processes, the video footage was considered as the criterion measure as it provided a way to observe the underlying data directly, with no subjective interpretations or ratings. The Match Histories were copied from the publicly accessible repository (https://matchhistory.na.leagueoflegends.com), while each of the authors independently viewed videos on the public Video On Demand repository (https://watch.lolesports.com/vods). The authors encoded each action variable into a spreadsheet (see Table 1 for a list of variables available in both Match Histories and video footage). Three of the authors had experience playing League of Legends (255, 1401 and 1596 hours; data extracted from user accounts via the website: https://wol.gg/ data extracted 03/08/2019), as well as many hours viewing professional competitions and two years’ experience consulting with professional League of Legends teams and coaches. The most experienced author’s data were compared with the Match Histories as an assessment of validity, and between the three authors as a measure of reliability. Krippendorff’s Alpha was used with an acceptable agreement of α ≥ 0.8 according to prior recommendations (literature,¹⁵,¹⁶ p. 241) via the IRR package within R.¹⁷

Table 1.

Agreement between the most experienced player and Match History; and between three experienced players.

	Experienced player vs match history			Agreement between three experienced players
	α	ICC	RMSE	α
Match Duration	0.995	1.000	5.027	0.999
First Blood Time	0.998	0.999	4.590	0.998
First Tower Time	0.990	0.996	12.412	0.993
Match Outcome	1.000	1.000	0.000	1.000
First Blood Team	0.868	0.866	0.258	0.911
First Tower Team	0.934	0.935	0.183	0.955
Rift Herald Team	1.000	1.000	0.000	1.000
First Baron Team	1.000	1.000	0.000	0.962
First Inhibitor Team	1.000	1.000	0.000	0.861
First Elder Dragon Team	1.000	1.000	0.000	0.899
Level	0.998	0.999	0.408	0.997
Gold	1.000	1.000	0.177	0.995
Creep Score	0.999	0.992	26.311	0.998
Kills	1.000	1.000	0.000	0.992
Deaths	1.000	1.000	0.000	0.992
Assists	1.000	1.000	0.000	0.987
Wards Placed	0.863	0.906	17.595	0.997
Wards Destroyed	0.883	0.913	9.670	0.992
Towers	0.997	0.999	0.129	0.937
Inhibitors	0.995	0.994	0.129	0.975
Elemental Drakes	1.000	1.000	0.000	1.000
Barons	0.934	0.933	0.224	0.866
Elder Dragons	1.000	1.000	0.000	0.903

Note: α = Krippendorff’s Alpha; Creep Score = a combination of Minion and Monster kills; RMSE = Root Mean Square Error; ICC = Intra-class Correlation Coefficien.

The results showed that the most experienced author was in acceptable agreement with the Match Histories (α = 0.863–1.000), while the three experienced author were in acceptable agreement with each other (α = 0.861–1.000) (Table 1). However, the Wards Placed and Wards Destroyed had a near-perfect agreement between the three experienced authors but a relatively low agreement between the most experienced author and the Match Histories. Therefore, there might be a systematic error between Match Histories and video footage and all vision-related variables were removed from the subsequent performance analysis. Additional research would be required to ascertain the reasons for these errors, however, it is likely that there is some disparity between when the game is technically won (Nexus destroyed) and when the final vision variables are captured by the software. Aside from the vision-related variables, the Match Histories and the experienced author’s video annotations could be used interchangeably to facilitate performance analysis in League of Legends. Using video annotations might be useful in situations where the Match History statistics are not available, and a match of interest must be coded manually via video footage – a scenario that was encountered twice during the current study. Complete methodology and analysis is available in preprint.¹⁸

Coach analysis of performance variables

A survey was created by three of this study’s authors who had experience in League of Legends as per the preliminary analysis. Variables were included based on the collective agreement of these players via open discussion, and if the variables could be collected from Match Histories and/or video annotation as per the preliminary analysis. The final survey identified 43 potential performance variables and asked coaches to “select how strong they believed a list of 43 performance variables related to match outcome in professional League of Legends”. Three expert coaches completed the survey. Two were head coaches and one was an assistant coach at the time of data collection, each with 3–4 years’ experience coaching, and 7–9 years playing League of Legends. Two coaches had coached at the highest level in their region (Professional) and one had coached at the second highest level (Academy). Coaches used a 1–10 Likert scale when rating the relationship between the performance variable and match outcome (i.e. 10 options resulting in no middle/neutral option and enforcing a non-neutral response). The Likert scale contained two anchors: 1 = “no relationship” and 10 = “very strong relationship”. The authors also asked the coaches to list any additional performance indicators that they felt would influence match outcome. Table 2 shows the complete list of performance variables included in the survey.

Table 2.

Variables include in coach survey.

Category	Performance Indicator	Median Coach Score	Retained After Coach Survey
	Champion Level	7	Yes	No
	Creep Ratio (own creeps vs opposition creeps)	6	Yes	Yes
Champion Power	Creep Score	6	Yes	No
Champion Power	Creeps per Minute	7	Yes	Yes
	Gold per Minute	6	Yes	No
	Gold Ratio (own gold vs opposition gold)	6	Yes	Yes
	Level Ratio (own level vs opposition level)	6	Yes	No
	Total Gold	6	Yes	Yes
	Assist Ratio (own assists vs opposition assists)	1	No	No
	Death by First Blood	1	No	No
	Death Ratio (own deaths vs opposition deaths)	4	No	No
	First Blood	3	No	No
	First Blood Assist	2	No	No
Kills,Deaths &Assists	KDA (Kills + Assists)/Deaths	1	No	No
	KDA Ratio (Own KDA vs opposition KDA)	1	No	No
	Kill Ratio (own kills vs opposition kills)	5	No	No
	Number of Assists	2	No	No
	Number of Assists per Minute	1	No	No
	Number of Deaths	4	No	No
	Number of Deaths per Minute	1	No	No
	Number of Kills	2	No	No
	Number of Kills per Minute	1	No	No
	Baron Ratio (own Barons vs opposition Barons)	9	Yes	No
	Dragon Ratio (own Dragons vs opposition Dragons)	6	Yes	No
	First Baron	9	Yes	Yes
	First Dragon	5	No	No
	First Elder Dragon	8	Yes	No
	First Inhibitor	8	Yes	No
	First Tower	7	Yes	Yes
	Number of Cloud Drakes	6	Yes	Yes
Objectives	Number of Elder Dragons	6	Yes	Yes
	Number of Elemental Drakes	7	Yes	Yes
	Number of Infernal Drakes	7	Yes	Yes
	Number of Inhibitors	8	Yes	Yes
	Number of Mountain Drakes	8	Yes	Yes
	Number of Ocean Drakes	6	Yes	Yes
	Number of Towers	10	Yes	No
	Rift Herald	8	Yes	Yes
	Tower Ratio (own towers vs opposition towers)	9	Yes	No
	Vision Ratio (own vision vs opposition vision)	7	Yes	No*
Vision	Vision Score	7	Yes	No*
	Wards Destroyed	8	Yes	No*
	Wards Placed	6	Yes	No*

Note: *denotes variable that coaches identified for inclusion but were excluded due to data quality concerns.

2018 League of Legends World Championship

The 2018 World Championship was comprised of 24 teams from 14 regions. The game version was v8.19. Due to historical performance variations across regions, top-performing regions were permitted to each enter their best three teams, while lower-performing regions each entered their single best team (the winning team at yearly regional finals). A total of 119 matches were played throughout the World Championship, with the early stages of the tournament following a round-robin format, while knockout stages such as the finals followed a best-of-five format. Due to these factors, regions and teams did not play an equal number of games (average = 10 ± 6 games per team; range = 4 [Brazil, Oceania, South-East-Asia] to 47 [China and Europe]). Given that draws are not possible in League of Legends, no matches needed to be excluded from analysis when examining win/loss outcomes.

Performance indicators

Video footage for the 2018 League of Legends World Championship was accessed and viewed via the public League of Legends Video On Demand repository, and the Match Histories from the public League of Legends Match History repository as per the preliminary analysis. For each match, the region and name of the winning and losing team were recorded, as well as the team sides (Blue vs. Red) and match duration. Performance indicators determined via the coach survey were recorded within three categories: 1) frequency data (e.g. number of kills, deaths and assists for each player); 2) time-dependent data (e.g. the exact time that the first tower was destroyed); and 3) categorical data (e.g. which team destroyed the first tower). Table 3 displays the complete list of variables. Institutional ethics approval was received prior to undertaking this study.

Table 3.

Descriptive team-based statistics of the 2018 League of Legends World Championship.

	Win	Loss
Match Duration (min)	32.27 ± 5.96
Rift Herald (%)*	64.7	31.9
First Tower (%)	68.1	31.9
First Baron (%)*	80.7	13.4
Level	77.7 ± 7.0	72.2 ± 8.1
Level Percentage	51.9 ± 1.3	48.1 ± 1.3
Creep Score	1114 ± 211	1052 ± 222
Creep Score Per Minute	34.6 ± 2.7	32.6 ± 2.4
Creep Score Percentage	51.5 ± 2.3	48.5 ± 2.3
Towers Taken	9.0 ± 1.9	2.6 ± 2.5
Tower Percentage	79.4 ± 18.2	20.6 ± 18.2
Inhibitors Taken	1.4 ± 0.8	0.1 ± 0.3
Dragons Taken	2.0 ± 1.1	1.1 ± 1.1
Dragon Percentage	65.5 ± 32.7	34.5 ± 32.7
Elder Dragons Taken	0.06 ± 0.27	0.04 ± 0.20
Barons Taken	1.0 ± 0.5	0.2 ± 0.5
Gold	61992 ± 10434	51306 ± 12347
Gold Per Minute	1934 ± 142	1580 ± 196
Gold Percentage	55.1 ± 4.0	44.9 ± 4.0

Note: all variables are team measures i.e. calculated as a sum of the five individual team players. * = variable was not obtained in all games i.e. values for winning and losing sides do not add to 100%.

Statistical analysis

Coach survey

Data from the coaching survey were exported from Google Forms as a spreadsheet. For each variable, the Median value of the three coaches’ responses was calculated using Microsoft Excel. Variables with a Median value ≥6 were retained in the final analysis of 2018 League of Legends world championship. A median value of 6 was chosen as it represents a variable that is believed to be relatively strong in its association with match outcomes, while a value of 5 is believed to be relatively low in its association with match outcomes as there was no middle/neutral option available.

2018 League of Legends World Championship

Variables retained after the survey were manually entered into a spreadsheet for each match by one of the experienced players/authors via manually copying data from the Match Histories or video annotation from the Video on Demand. Variables that contextualized the frequencies (e.g. own team gold vs opposition gold) were expressed as percentages of total match values (e.g. own team gold/[own team gold + opposition gold]) for modelling purposes as per previous recommendations for variable normalization in performance analysis.¹⁰ Analyses were conducted using the R statistical framework.¹⁹ Variables were initially examined for multicollinearity using a correlation matrix. If a correlation of ≥0.8 was observed between two variables, only the variable with the greatest correlation with respect to match outcome was retained for analysis.

Given that there was an uneven distribution of matches played by each team, generalized linear mixed effects models were deemed appropriate to examine the relationships between performance indicators and match outcomes, as some correlations within groups were expected. The “glmer” function within the “lme4” package was used²⁰ and the Blue Side Team Name was specified as a random effect, while all other independent variables were specified as fixed effects. All matches were analysed with respect to the Blue side, and a binomial link function was adopted to analyse the data with respect to the Blue side winning or losing the match. While all four measures of Vision were retained after the coach survey, the previous data quality concerns regarding variables relating to vision (e.g. Wards Placed and Wards Destroyed; see preliminary analysis above), resulted in a decision to remove those four variables. Therefore, 24 variables remained for further analysis. The 24 variables were assessed for Multicollinearity, and a further 10 variables were removed, leaving a final 14 variables. Given that there were 14 variables remaining and only 119 observations, a step-up approach was adopted when modelling the generalized linear mixed-effects models. In this process, a null model was firstly specified. Subsequently, each variable was added one-by-one as a fixed effect to identify the variable that explained the most variance by comparing the new model against the null model via the anova function in R to view the alpha value and Akaike Information Criterion (AIC). Models were also assessed by comparing the predicted vs. observed values to calculate the classification accuracy of the model. The residuals for each model were also examined via QQ plots to assess normality. Following the identification of the best model with one fixed effect, a second fixed effect was added one-by-one as per the first model and assessed via the same method until no further model improvement was observed.

Results

Coach survey

Of the 43 variables analysed, 28 had a median score ≥6, and therefore were retained for subsequent analysis. These variables are represented in Table 1 and are related to Champion Power, Objectives, and Vision. No variables relating to Kills, Deaths and Assists were given a Median score ≥6 by the three coaches. Only one coach responded to the question asking them to identify any other potential performance indicators. This coach noted that “gold lead in relation to the time in the game” would be important and would be rated 9/10 on the Likert Scale. However, given that time-series analysis was not within the scope of this study and that the other two coaches did not include a similar comment, this variable was not included.

2018 League of Legends World Championship

During the assessment of multicollinearity, Tower Percentage was the strongest variable, resulting in the removal of seven variables that were multicollinear with Tower Percentage (Level Percentage, Towers Taken, Inhibitor Percentage, First Inhibitor, Barons, Gold Percentage, and Gold Per Minute). Descriptive statistics for all variables are available in Table 3 as mean ± SD. Additionally, Tower Percentage was the strongest variable in the generalized linear mixed effects model, becoming the variable retained at the first step of the step-up approach. Subsequently, the number of Inhibitors Taken was retained in the model at the second step of the step-up approach and the inclusion of additional variables thereafter did not improve the model any further. The final model that included Tower Percentage and number of Inhibitors Taken as fixed effects was significantly better than the null model (p < 0.001), reducing the AIC from 148.8 to 37.4 and achieving a prediction accuracy of 95.8% when comparing the predicted values to the observed values for match outcome. Specifications of the final model are presented in Table 4 and estimates have been exponentiated to allow for interpretations as odds ratios.

Table 4.

Best model of performance for the 2018 League of Legends World Championship.

	Estimate	95% CI lower	95% CI upper	Standard Error	z value	p	Odds Ratio
Intercept	–6.674	–14.825	–3.866	1.830	–3.646	<0.001	0.001
Tower Percentage	0.084	0.032	0.186	0.031	2.738	0.006	1.087
Inhibitors Taken	2.568	0.681	7.616	1.179	2.178	0.029	13.036

Discussion

The current study is the first to explore the validity of automated match statistics in esports, and to undertake an analysis of in-game performance indicators within a professional esports tournament (i.e. the 2018 League of Legends World Championship). The study aimed to apply methods grounded in sports science to model esports performance, which was achieved via assessing the validity of automated match statistics and inter-rater reliability of video annotation, engaging with expert coaches, and conducting performance analysis facilitated by match statistics and video. This study utilised generalized linear mixed-effects models and identified that the percentage of Towers Taken was most strongly related to match outcome, while the number of Inhibitors Taken also contributed to the best model of performance (95.8% classification accuracy). Specifically, an improvement of 1% in Tower Percentage resulted in 8.7% greater odds of winning the match, while with each additional Inhibitor Taken, there was a thirteen-fold increase in the odds of winning the match.

This study is the first to provide evidence for the importance of Towers and Inhibitors for match outcomes in League of Legends. However, neither of these findings were surprising, given that Towers are the main structures that prevent players from progressing towards the opponent’s Nexus. Towers cause significant damage to players, especially during the early phases of the game when Champions have limited defences and relatively low health. They also prevent access to key objectives such as the opponent’s Inhibitors and the Nexus. Additionally, Inhibitors provide an important effect that helps teams to end the game (stronger Minion waves are deployed towards the opponent’s Nexus to help the team destroy it). While Towers and Inhibitors are crucial for successful performance, further research is required to determine effective strategies to improve Tower Percentage and destroy Inhibitors. Specifically, to improve Tower Percentage, teams need to destroy opposing Towers while simultaneously protecting their own Towers. Additionally, given that there are 11 Towers on each side, there may be optimal strategies for targeting Towers at specific moments of the game, which future research should explore. Coaches should work with their players to develop a strategy that maximises their Tower Percentage and ability to destroy enemy Inhibitors throughout each match.

While the final model of performance produced extremely high prediction accuracy by using only two variables, this does not mean that they are the only variables coaches and players should focus on, or that other variables are not also important. Instead, this only reflects that other variables added no further value to the prediction model once Tower Percentage and Inhibitors were included. Importantly, seven variables were removed before the modelling process due to collinearity with Tower Percentage (Level Percentage, Towers Taken, Inhibitor Percentage, First Inhibitor, Barons, Gold Percentage, and Gold Per Minute). Therefore, further consideration of these variables is warranted. In particular, the causal pathways between each of these variables should be studied. For example, does a team having more gold cause them to take more towers, or does taking more towers cause a team to obtain more gold? Knowledge of the game suggests that both of these are true, and the causal pathway is likely bidirectional. However, further studies using time-series analysis could provide confirmation and greater clarity on how to use these relationships strategically. Secondly, Table 3 shows potentially substantial differences between winning and losing teams for Rift Herald (64.7% vs 31.9%), First Tower (68.1% vs 31.9%), First Baron (80.1% vs 13.4%), Barons Taken (1.0 ± 0.5 vs 0.2 ± 0.5), and all measures of Gold. Univariate analysis of variance was not undertaken for each individual variable, but these data could contribute to the development of hypotheses for further studies.

As noted above, the difference between winning and losing teams for First Tower appears large, yet the First Tower objective did not contribute to the model of performance, likely due to the high variance taken up by Tower Percentage and Inhibitors. Further research should determine whether Towers at various locations are of greater importance than others. It should be noted that two of the Towers are located next to the Nexus and are often destroyed during the final moments of the game. Therefore, the Tower Percentage may be somewhat artificially inflated as a measure of performance in the final statistics, given that the game state in which the winning team finally overcame their opponents may have occurred during a final team fight, i.e. prior to taking these final two Towers. Time-series analysis was beyond the scope of this study but is an area that requires further investigation. These findings based on the objective data are in support of the coaches’ subjective opinions, who also indicated that number of Towers Taken is the most important variable (the only variable with a median value of 10/10), while they also rated Tower Ratio as 9/10.

Limitations

This research should be interpreted considering several limitations. First, only three coaches completed the survey, and therefore, the survey results cannot be generalised. Future research would benefit from seeking engagement from more industry experts, however, the findings of this study supported the opinions of the three expert coaches. It should be noted that one survey responder disclosed via personal communication that contracts are a particularly sensitive topic within esports and that the informed consent form was likely perceived as a form of contract, ultimately deterring responders. While this is somewhat speculative, researchers may benefit from a more sensitive approach when engaging the esports community in future. Second, due to the inclusion bias towards previously high-performing regions (e.g. China, Europe, North America and Korea) and the knockout style format of competition, there may be inherent bias within the data towards teams who played more games (i.e. the results may reflect specific gameplay styles from China and Europe who each played 47 games). While this was accounted for within the mixed-effects model, it could not be accounted for within the summary statistics. Given that the League of Legends World Championship encompasses the top teams across all competitive global regions, it is accepted that the summary statistics encompass current performances at the highest level of competitive League of Legends. Third, League of Legends receives relatively minor updates roughly every two weeks (e.g. minor changes to champion skills, cooldowns, or adding new champions to the game), while more extensive updates typically occur once per calendar year, (e.g. changes to the way the Rune system works [abilities that players select prior to each match] or adding new defences to the Towers to change how players strategize). As noted in this study, the patch version for the 2018 World Championships was v8.19 and future research should be conducted to determine performance indicators that are longitudinally related to successful performance. Finally, due to the limited sample size, it was not feasible to include many independent variables or to withhold a sample to facilitate predictions on unseen data. Future research should aim to acquire larger samples so that more variables can be included, and the predictive power of the models can be assessed. This could be facilitated via the Riot Games Application Program Interface (API); however, validation of the API data quality is required prior to implementation.

Conclusion

The current study applied traditional performance analysis methods that were learned in the sports science domain, to help understand esports performance. At the highest level of League of Legends competition, Tower Percentage and Number of Inhibitors were identified as having the strongest relationships with performance. Therefore, further research should seek to identify strategies to effectively target these objectives across the time-series of the match, and to determine whether there are differences between World Championship competition and other competitions e.g. Professional level and Academy level within individual regions.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

Andrew R Novak

Kyle JM Bennett

References

Pluss MA, Bennett KJM, Novak AR, et al. Esports: the chess of the 21st century. Front Psychol 2019; 10: 156–155. DOI: 10.3389/fpsyg.2019.00156.

Kowal M, Toth AJ, Exton C and Campbell MJ. Different cognitive abilities displayed by action video gamers and non-gamers. Comp in Hum Beh 2018; 88: 255–262. DOI: 10.1016/j.chb.2018.07.010.

Statista. Number of League of Legends monthly active users. (MAU) from 2011 to 2016. 2016. Available at: https://www.statista.com/statistics/317099/number-lol-registered-users-worldwide/ (9 October 2019).

Colzato L, van den Wildenberg W, Zmigrod S, et al. Action video gaming and cognitive control: playing first person shooter games is associated with improvement in working memory but not action inhibition. Psychol Res 2013; 77: 234–239. DOI: 10.1007/s00426-012-0415-2.

Powers K, Brooks P, Aldrich N, et al. Effects of videogame play on information processing: A meta-analytic investigation. Psychon Bull Rev 2013; 20: 1055–1079. DOI: 10.3758/s13423-013-0418-z.

Bonny JW and Castaneda LM. Number processing ability is connected to longitudinal changes in multiplayer online battle arena skill. Comp in Hum Beh 2017; 66: 377–387. DOI: 10.1016/j.chb.2016.10.005.

Bonny JW and Castaneda LM. Impact of the arrangement of game information on recall performance of Multiplayer Online Battle Arena players. Appl Cognit Psychol 2016; 30: 664–671. DOI: 10.1002/acp.3234.

Bennett M, Bezodis D, Shearer DA, et al. Descriptive conversion of performance indicators in rugby union. J Sci Med Sport 2019; 22: 330–334. DOI:10.1016/ j.jsams.2018.08.008.

Hughes MD and Bartlett RM. The use of performance indicators in performance analysis. J Sports Sci 2002; 20: 739–754. DOI: 10.1080/026404102320675602.

10.

Hughes M and Franks I. Notational analysis of sport. (2nd ed.). London, England: Routledge, 2004.

11.

Parmar N, James N, Hughes M, et al. Team performance indicators that predict match outcome and points difference in professional rugby league. International Journal of Performance Analysis in Sport 2017; 17: 1044–1056. DOI: 10.1080/24748668.2017.1419409.

12.

Hughes A, Barnes S, Churchill SM, et al. Performance indicators that discriminate winning and losing in elite men’s and women’s Rugby Union. International Journal of Performance Analysis in Sport 2017; 17: 534–544. DOI: 10.1080/24748668.2017.1366759.

13.

Robertson S, Gupta R and McIntosh S. A method to assess the influence of individual player performance distribution on match outcome in team sports. J Sports Sci 2016; 34: 1893–1900. DOI: 10.1080/02640414.2016. 1142106.

14.

Vaz L, Rooyen M and Sampaio J. Rugby game-related statistics that discriminate between winning and losing teams in IRB and Super Twelve close games. J Sports Sci & Med 2010; 9: 51–55.

15.

Krippendorff K. Estimating the reliability, systematic error and random error of interval data. Edu and Psycho Meas 1970; 30: 61–70. DOI: 10.1177/001316447003000105.

16.

Krippendorff K. 2004. Content analysis: an introduction to its methodology. 2nd ed. Thousand Oaks, California: SAGE.

17.

Gamer M, Lemon J, Fellows I and Puspendra S. Various coefficients of interrater reliability and agreement, 2019; v0.84.1.

18.

Novak AR, Bennett KJM, Pluss MA, et al. Performance analysis in esports: Part 1 – the validity and reliability of match statistics and notational analysis in League of Legends. SportRxiv. DOI:10.31236/ osf.io/sm3nj.

19.

R Core Team. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria, 2013. http://www.R-project.org/.

20.

Bates D, Maechler M, Bolker B, et al. Linear Mixed-Effects Models using ‘Eigen’ and S4, 2020; v1.1–23.