On Successful Team Formation – TU Wien

Loading...

Taking too long?

Reload document

| Open in new tab

On Successful Team Formation

Nataliia Pobiedina
Vienna University of
Technology
Austria
pobiedina@ec.tuwien.ac.at

Julia Neidhardt
Vienna University of
Technology
Austria
neidhardt@ec.tuwien.ac.at

Maria del Carmen
Calatrava Moreno
Vienna University of
Technology
Austria

maria.moreno@tuwien.ac.at

Laszlo Grad-Gyenge
Vienna University of
Technology
Austria
grad-
gyenge@tuwien.ac.at

Hannes Werthner
Vienna University of
Technology
Austria
werthner@ec.tuwien.ac.at

ABSTRACT
Teamwork plays an important role in many areas of today’s
society. Thus, the question of how to form an effective team
is of increasing interest.
In this paper we use the team-
oriented multiplayer online game Dota 2 to study cooper-
ation within teams and the success of teams. Making use
of game log data, we choose a statistical approach to iden-
tify factors that increase the chance of a team to win. The
factors that we analyze are related to the roles that players
can take within the game, the experiences of the players and
friendship ties within a team. Our results show that such
data can be used to infer social behavior patterns.

Categories and Subject Descriptors
J.4 [Social and Behavioral Science]: Sociology; H.5.m
[Information Interfaces and Presentation]: Miscella-
neous

Keywords
Online game, Online community, Team formation, Social
ties, Statistical analysis

INTRODUCTION

1.
Successful teams and how to build them is a hot topic in
a wide range of fields. In every day’s work, in professional
sports as well as in leisure activities there is a growing in-
terest in how people cooperate and which factors influence
team performance in a positive way. Plenty of articles are
published that provide tips for managers on how to compose
effective teams (e.g., [8]). As research has shown, teamwork
also plays an increasing role in the production of high impact
science [2].

As more and more human activities are moving to the Web,
the Web has become a mirror of modern society. Data is
easily available and can be used to study human relations
and behavioral patterns. Thus, the Web provides an un-
precedented opportunity to observe social interaction on the
large scale.

Our aim in this context is to study cooperation within teams
and success of teams. To do this, we use the multiplayer
online game Dota 2. In this game two teams, consisting of
five members each, play against each other, and the task
is to defeat the opposed team. To achieve this goal, close
cooperation and intelligent interaction between the members
of the team are needed – hence, a challenge that is present in
many “real world” situations as well. Particularly, we want
to find out whether or not the success of a team is influenced
by a) the distribution of the roles that are chosen by the
players in the game, b) previous experiences of the team
members, and c) friendship ties within a team. Furthermore,
related to the chosen roles of the players, we investigate
which of the two factors has a higher influence on the team’s
success: either the individual choices of the roles by the
players or the combination of the role choices of the entire
team.

Online games have already widely been used to study so-
cial interactions. By now research mainly focused on so-
called Massively Multiplayer Online Role-Playing Games.
Although in these types of games cooperation is possible,
each player makes his/her own progress and has individual
tasks. Dota 2, in contrast, is a game in which the players
are always assigned to a team and thus have common goals
and interests. Furthermore, most of the existing studies are
qualitative, using surveys and questionnaires to learn about
the behavior and the motivations of users. As opposed to
this, our analysis is based on log data, i.e., data that is
automatically generated to record the events that are hap-
pening during a game. Our quantitative approach provides
the opportunity to draw a global picture that can lead to
new insights and a better understanding on what is needed
to make a team successful.

cial abilities are a set of four unique spells specific to each
hero (for example, there are such spells as “Enchant totem”,
“Greevil’s greed”, “Nature’s guise” and so on). Both at-
tributes and abilities are enhanced with experience accu-
mulated over the course of the game. Through the com-
bination of initial attributes and special abilities different
heroes are suited for different strategies (in Dota 2 they say
game “roles”) and can be played in a variety of ways (e.g.
“Pusher”, “Carry”, “Nuker”, etc.). Each player chooses a
strategy not only based on the selected hero, but also on
the heroes of the other members of the team. Through the
choices of these strategies the flexibility of the team is in-
creased and facilitating the formation of more competitive
teams.

To get more details about each attribute, ability as well as
game role, we recommend to address the official wiki page of
Dota 2 [5]. In Table 1 we present three heroes from Dota 2:
“Treant Protector”, “Phantom Lancer” and “Lina”. We also
provide the values of their main three attributes (“Strength”,
“Intelligence” and “Agility”) as well as class to which they
belong and game roles in which these heroes can be played.

Dota 2 is a team-oriented game in which strategy and team
coordination is decisive to achieve a victory. Communication
between team members is a vital part of the game, acting
as a binding force that makes a team function. Players can
communicate through typing, voice chat, pinging the map
and writing on the minimap.

Valve has built a social network around Dota 2 utilizing
Valve’s Steam software [20] in order to provide social and
community functionality for the game. Steam accounts save
personal files and settings on the online accounts. The play-
ers can set up private games with friends or join public
games. In private games, teams might, however, be formed
not only by humans but also by Artificial Intelligent (AI)
bots. In this case other players in the community are locked
out and the game is played with computer-controlled heroes,
who can also interpret simple commands of human players.
Dota 2 has not been publicly released yet. Even if its beta
version limits its test early access, it is currently one of the
cornerstone games at several electronic sports tournaments,
and considered one of the best and highest e-sport games [9,
17, 10].

On the basis of the data from the game we want to inves-
tigate team formations and which factors influence the suc-
cess of the team in the game. We formulate the following
hypotheses:

1. team selection of heroes influences the game outcome;

2. overall gaming experience of players influences the game

outcome;

3. playing with friends increases the chance to win;

4. individual selection of heroes as well as the combina-
tion of selected heroes influence the game outcome.

3. RELATED WORK
Virtual worlds are playing an important role in the study
of diverse fields such as sociology [11], psychology [23, 6],

Figure 1: The Map of Dota 2.

The rest of the paper is organized as follows: In section 2
we describe the game in more detail. In section 3 the re-
lated work is presented. In section 4 and section 5 descrip-
tive statistics of the data are provided and the preprocessing
steps for further analysis are discussed. In section 6 the anal-
ysis is presented and the results are shown. Our conclusions
and plans for future work are presented in section 7.

2. THE GAME AND ITS COMMUNITY
Dota 2 [19] is a so-called multiplayer online battle arena
(MOBA) video game developed by Valve [21]. Each player
controls a character called “hero”, who participates in a team
combat with the objective to demolish the opposing team’s
fortified stronghold. We are aware that this is a very cruel
terminology, but we stick to it since it stems from the game
creators.

Players are pitted against each other as two distinct fac-
tions of five players each, the Radiant and the Dire. Their
strongholds, called base towers, are located at opposing ends
of a geographically balanced squared map (see Figure 1).
These are connected by three main lanes, which are guarded
by defensive towers and weaker computer-controlled units,
called creeps. Killed heroes revive in the corresponding
area of their base after a waiting time proportional to their
level and the game time. Through the destruction of en-
emy forces, heroes may gain both experience and gold. The
former accumulates to gain higher levels that enhance the
hero’s attributes and abilities. The latter is the currency
of the game, which is distributed to the team members ac-
cording to their accomplishments. Gold also accumulates
periodically to each hero. It is mainly used to acquire items
that substantially complement or alter abilities, as well as
to buy an instant revival of the hero.

Each player selects one hero out of 96 available in Dota 2.
These heroes are unique characters that differ in their ini-
tial attributes and special abilities. On the one hand, ini-
tial attributes categorize heroes primarily according to their
strength, agility and intelligence. On the other hand, spe-

Table 1: Examples of heroes with some of their characteristics.

Strength

Agility

Intelligence

Class

Strength

Agility

Intelligence

Game role

25

15

17

Durable
Initiator
Lane support
Disabler

18

23

21

Carry
Escape
Pusher

18

16

27

Nuker
Disabler
Support

economy [13, 16], etc. These studies raise the question of
how the mechanisms of human behavior are being translated
and developed in an artificial environment.

Gameplay data and player characteristics are drawing the
attention of recent studies in the field of social computing
and web science [14]. In particular the study of Massively
Multiplayer Online Role-Playing Games (MMORPGs) are
gathering most of the attention. This is due to their nature
that allows players’ cooperation and competition on a large
scale, as well as interaction assuming the role of a character
whose actions can be controlled, in the case of MMORPGs.
The social interactions that take place in them are well ex-
plored demonstrating the crucial role that they play. Cole et
al. [3] examine them through the analysis of online question-
naires that interrogate about social interactions that occur
both within and outside MMORPGs. Their results show
that these are extremely social games that favor the possi-
bilities of players making life-long friends and partners. Re-
cent studies analyze log data of this kind of games with the
aim to build models of human features and behavior, such
as activities, interactions and cooperations [18]. Ka(cid:32)lu˙za et
al.[11] carried out their research of an MMORPG. They use
World of Warcraft as a case study that they analyze from a
sociological viewpoint. In their study they identify players’
communication as as a driver for community engagement.
Likewise they relate gamers’ intercultural communication to
its influence on players’ behavior and group organization in
such artificial communication environment. Their descrip-
tive research concludes that origin, culture and language are
important factors of player attractiveness that have an effect
on the creation of national guilds, communication problems
and generalization of players’ behavior based on the coun-
try of origin. Group formation of gamers is also examined in
one of the latest studies of Keegan et al. [12] who collected
data about characters and accounts from the Sony Online
Entertainment’s MMORPG EverQuest II.

Cooperation and competition in online games is examined
from a different aspect by Yuan et al.[24] by conducting a

quantitative study and analyzing game logs. Their results
show that the selection based on in-game score level of part-
ners to cooperate with is important for the players, while
choosing the opponents is slightly biased.

The categorization MOBA (Multiplayer Online Battle Arena
games), also known as Dota-like games, often refers to games
with two teams of players competing against each other and
controlling a single character in the battlefield. Although
this genre emphasizes a more cooperative team-play, the lit-
erature about it is very scarce. A very recent paper analyzes
the relationship between real life leadership styles (author-
itarian, democratic or laissez-faire) and game roles of two
MOBA games, Dota 2 and Heroes of Newerth [15]. The
method used was a close-ended questionnaire to examine
daily life and gameplay behaviors.

Dota 2 as a game differs from MMORPGs since this is first
of all a team game and only afterwards a game with ele-
ments of traditional MMORPGs, and the team perspective
is the main focus of our analysis. We are not aware of any
recent work using game log data to analyze the behavior and
interaction of MOBA players, and with this study we aim
to cover the gap.

4. THE GAME AND ITS DATA
The data set used in this study has been retrieved in XML
from Steam and Dota 2 utilizing their Web APIs [22, 4] and
later migrated to PostgreSQL. The Dota 2 data was made
public by the community of Dota 2 players. It contains the
match history and the details of the matches played in the
year 2011.

Using the Steam API we incorporate additional informa-
tion of those players that appear in the match history of
the Dota 2 data. This information is extracted from the
players’ profiles in the Steam platform. Such profiles con-
tain user information such as name, country, sign up date,
last log off date, etc. The list of friends is also extracted,
as well as their type of relationship and starting friendship

date. The visibility of this data is, however, dependent on
the confidentiality of the user profile, which can public or
private.

The entire database contains information on 885,228 matches.
Since we focus on team aspects, we need details about both
the matches and the players involved, such as start time and
duration of the match, its outcome (i.e., which team wins),
the number of human players (there are also teams that in-
clude AI bots – see Section 2), the difficulty or “skill” of the
match, account ids of the players, the heroes they choose,
the performance of those heroes in the match (i.e., how of-
ten they are defeated, how many others they damage, how
much gold they acquire and so on). For the majority of the
matches in the database not all this information is provided,
so we filter them out. Also we keep only the matches that
contain two teams with five players in each team and all
players have public profiles.

The resulting data set comprises 87,204 matches, which are
played by 138,101 individuals. Since there are ten play-
ers per match, each player participates on average in 6.3
matches. 18.7% of the players (25,812) take part in more
than 10 matches; and the highest number of matches that
is played by one person is 94. On average a match lasts 45
minutes; and 50% of the matches take between 37 and 45.7
minutes. Each of the 138,101 individuals spends on average
284.1 minutes resp. 4.7 hours playing the game.

There are four different difficulty levels of matches. In the
dataset variable “skill” indicates the difficulty of the match:
0 – low, 1 – normal, 2 – high, 3 – very high. Unfortunately,
in the filtered dataset we have only two matches of the very
high difficulty (see Table 2).

Table 2: Amount of matches of different difficulty
levels.

Skill
Amount

Low
40216

Normal
37481

High
9505

Very high
2

In the dataset under consideration, approximately half of
the players (52% resp. 71,869) provide information on their
country; 232 distinct countries are indicated. This informa-
tion will be used in our future studies. In addition, in our
dataset each player has on average 34.6 friends. The maxi-
mum number of friends a player has is 302; and around 1770
players don’t gave any friends at all.

Within the 87,204 matches only 66 distinct heroes out of 96
are chosen by the players. The overall distribution of the
heroes’ frequency is quite balanced, the most popular hero
is chosen in 3.6% of all cases; and 75.8% of the heroes (50
out of the 66) are chosen in more than 1% of the cases.

5. THE GAME AND ITS PLAYERS
As we mention in Section 2, each player has to select one
hero at the beginning of the game. This selection influences
the role (or equivalently, the strategy) which the player will
follow during the game. Since we are interested in the in-
fluence of role distribution on the success of the team, we

perceive hero selection as an indicator of role distribution
inside the team and consider it as an important factor for
the team success. However, there arises a difficulty: there
are 5 players in each team, and to study role distribution
would mean studying all possible combinations of 5 heroes
out of 66. That results in 677,040; thus, this approach is not
feasible.

Another approach would be to classify heroes and then to
study the combinations of classes. There are two possible
classifications of heroes: the first one considers the main
three attributes of heroes (namely, “Strength”, “Agility” and
“Intelligence”); and the second one is based upon the game
roles (for example, “Carry”, “Pusher”, “Supporter” and so
on; there are more than 8 game roles) in which heroes can
be played. However, both classifications have shortcomings:
the first approach does not consider all the variability of hero
attributes (there are in total 17 basic attributes), and in the
second approach each hero can be played in more than one
of the defined game roles (see, for example, Table 1).

Thus, we introduce a new approach to consider the influ-
ence of hero selection on the team success. We use the data
about initial attributes of heroes; and for the sake of clarity
and better interpretation, we apply a dimension reduction
algorithm to receive one single score for each hero. We use
logistic regression for this purpose. We take the data about
played heroes in a match and its outcome (872,040 data
points for 10 players in 87,204 matches) and train the logis-
tic regression model on 70% of observations of our dataset,
and then apply the calculated coefficients to the rest.

Logistic regression has been chosen since it can be perceived
as a linear transformation of initial attributes with regard
of their influence on the match outcome. The coefficients
of the trained model indicate that not only strength, agility
and intelligence of the heroes are decisive, but also other
attributes are very important (especially, movement speed,
ranges of day sight and night sight). Thus, for each hero out
of 66 we obtain a unique hero score.

Another factor which we assume to influence the team suc-
cess is the experience of players. As for the gaming expe-
rience of players, we consider not only information about
the amount of previously played and won matches, played
time, but also information about performance in previous
matches. After the match finishes each player receives statis-
tics about his/her performance which includes 13 different
measures such as #kills, #deaths, spent gold, final level and
so on. So, in total we get again (like for hero selection) 17
different attributes related to the gaming experience of each
player. Following the same reasoning as for hero selection,
we use logistic regression on the same dataset (but instead
of using attributes of selected heroes by players we use at-
tributes of their experience) to calculate an experience score
for each player in a specific match.

To have the same base for both scores, we transformed them
to be in the interval [0, 1000]. Figure 2 shows the results of
Spearman correlation test and scatter plots between hero
score, experience score and win on the dataset which we use
to calculate the scores. Variable win is binary and indicates
whether player was in the team which won (value 1) or in the

for the team success. In Section 5 we show how to obtain
hero score using a supervised dimension reduction method.
We calculate the team hero score for each team in a specific
match as the average of hero score for the heroes selected by
team members. Remember that in Section 5 we show that
this score is the indicator of role distribution inside the team.
We normalize the team hero score across all matches and
teams. Afterwards, we form two samples: the first contains
scores for the winning team (win-team) and the second has
scores of the loosing team (loss-team). Both samples still
follow normal distribution.

To show that hero selection influences the team success, we
need to test whether the mean difference of team hero score
is equal to zero.
In case it is not zero, that would mean
that team hero score is different for win-team and loss-team,
and thus, hero selection is an important success factor. We
formulate the corresponding null-hypothesis.

H0: win-team and loss-team have the same means of team
hero score.
H1: members of win-team make better selection of heroes
which results in a higher team hero score compared to loss-
team.

We perform a paired t-test to test the null-hypothesis. How-
ever, F -test shows that we cannot claim that both samples
have equal variances with 95% confidence level. That is why
we perform a paired Welch’s t-test for these two samples.
Our test is paired since both samples are aligned according
to the matches.

We reject the null hypothesis with p-value= 1.802e − 06
which is significantly lower than the significance level and
accept the alternative hypothesis. This means that a win-
team has on average a higher team hero score than a loss-
team. This observation leads to the conclusion that team
hero score influences the team success.

Since there are matches of different difficulty levels, we per-
form a two-factor analysis to identify whether team hero
score is influential in matches of all difficulty levels. Re-
member that factor “skill” (see Table 2) indicates the diffi-
culty level: there are three difficulty levels from 0 (low) to 2
(high). We perform ANOVA test with team hero score be-
ing the dependent variable and the two independent factors
win and skill.

The results of ANOVA confirm our previous ones of signifi-
cant dependence of the team hero score on the team success
(loss or win). Figure 3 demonstrates the results of ANOVA
by using Tukey multiple comparison. The full comparison
chart includes 15 levels, but for the sake of interpretation
we include only 3 levels which are relevant to our study. We
explain the meaning of labels on y-axis by the first row. La-
bel 0 : 0 − 1 : 0 means that the difference of team hero score
for teams 0 : 0 (teams which lost in matches of low diffi-
culty level) and teams 1 : 0 (teams which won in matches of
low difficulty level) is compared. So, the first number in the
level encoding shows the value of factor win and the second
number corresponds to the value of skill factor.

According to Figure 3, there is a significant difference in the

Figure 2: Correlation between hero score, experience
score and win.

team which lost (value 0). As we see, there is a significant
high correlation coefficient (63%) between experience score
and hero score. This observation can be explained by the
fact that the higher gaming experience of the player, the
better he/she is at choosing heroes. However, the correlation
with variable win is, though significant, but very low.
It
turns out and we will show in the following section that
the team composition is an important factor for the match
outcome.

6. ANALYSIS AND RESULTS
In this section we show how different factors influence the
team success in a virtual environment, using the prepro-
cessed data as described in Sections 5 and 4. We use statis-
tical methods from R [1] to test our hypotheses about the
influence of role distribution, experience and social ties on
the team success.

To test the difference of scores for teams we perform paired t-
test or Mann-Whitney-Wilcoxon test which depends whether
the provided samples are normally distributed or not. We
apply χ2-test to test the dependence of the team success on
the number of friends in the team.

To make sure that the factors under consideration are im-
portant for matches of different difficulty levels (factor skill ),
we perform two way ANOVA.

Since we also want to verify that the selection of hero by
each player is important, we also use log-linear analysis with
likelihood ratio test (an adjustment of ANOVA).

All tests are performed with 95% confidence level. Addi-
tionally, whenever possible all the assumptions of the cor-
responding test are verified (for example, normality, equal
variances, homogeneity of variances, independece).

6.1 Hero Selection
One of the first hypotheses which we want to test is that the
role distribution (i.e., selecting a specific hero) is important

Figure 3: Means of team hero score with regard to
team success (win 1 and loss 0) and difficulty levels
(low 0, normal 1 and high 2).

Figure 4: Means of team hero score across matches
of different difficulty levels (low 0, normal 1 and high
2).

means of team hero score only for matches of the lowest skill
(skill = 0): a win-team has a higher score than a loss-team.
As for the difficulty levels 1 (normal) and 2 (high), we also
see that team hero score is higher for the win-team, but this
difference in scores is not statistically significant.

samples using the mean and standard deviation calculated
previously. The correlation analysis shows that team hero
score (which we studied previously) correlates significantly
with the individual scores having the highest correlation co-
efficient with the second ranked hero score in the team.

Going a step further, we look at the dependency of team
hero score and skill of match. We do this to obtain further
insights how well a team selects heroes based on the difficulty
level of the match. We state the following null-hypothesis
and its alternative.

Then, we perform paired Mann-Whitney-Wilcoxon-tests for
samples of hero score of the same rank from win-team and
loss-team, in total we do 5 tests. All 5 tests conclude that
there is a significant mean difference in hero scores for win-
team and loss-team, and win-team has a higher score.

H0: team hero score is the same across matches of different
difficulty levels.
H1: the higher the skill of the match, the better teams select
heroes.

The results of ANOVA test show that we do not reject the
the null-hypothesis. In Figure 4 we present again the Tukey
multiple comparison chart. The label 1 − 0 on y-axis means
that we compare the means of team hero score for matches of
difficulty level 1 and 0. The results imply that the selection
of heroes is balanced across all difficulty levels of matches.

This analysis concludes that team hero score is an important
factor for the team success irrelevant of the difficulty level
of match. It provides support for our first hypothesis.

6.2

Individual Hero Selection vs Team Hero
Selection

We want to check what is more influential for the match out-
come: very good selection of heroes by individual players or
very good combination of heroes selected by team. For this
purpose we extract hero score for each player in the team
and sort them in the increasing order. We obtain 5 samples
for win-team and loss-team separately: each sample corre-
sponds to hero score of players in the ranking order either
from win-team or loss-team. We normalize the scores in all

To consider also interaction between hero scores of differ-
ent ranks we perform log-linear analysis, that is we use logit
model to fit our data and for this model we perform ANOVA
with likelihood ratio test. By interaction we mean all possi-
ble combinations of hero scores for 5 players in the team. In
total there are 26 combinations: 10 pairwise, 10 three-wise,
5 four-wise and 1 total combinations. In Table 3 we present
the results of ANOVA with likelihood ratio test where each
term corresponds either to hero score of a specific rank (from
p1 being the weakest to p5 being the strongest) or a spe-
cific combination of hero score for players (from p1:p2 till
p1:p2:p3:p4:p5).

From Table 3 we see that the most influential hero score is
the one for the strongest hero selected in the team which
we call team leader (term “p5”). However, further analy-
sis of deviance uncovers that the weakest (term “p1”) and
the fourth strongest (term “p4”) heroes matter as well as
different combinations of heroes in the team (for example,
terms “p2:p4”, “p1:p2”, “p1:p2:p3:p4”). Thus, we may con-
clude that the team success depends significantly on the
team leader as well as successful combination of heroes in
the team. This conclusion provides support for our fourth
hypothesis that individual selection of heroes as well as their
combination is important for the team success.

Table 3: ANOVA results using likelihood ratio test.

Term
p1
p2
p3
p4
p5
p1:p2
p1:p3
p2:p3
p1:p4
p2:p4
p3:p4
p1:p5
p2:p5
p3:p5
p4:p5
p1:p2:p3
p1:p2:p4
p1:p3:p4
p2:p3:p4
p1:p2:p5
p1:p3:p5
p2:p3:p5
p1:p4:p5
p2:p4:p5
p3:p4:p5
p1:p2:p3:p4
p1:p2:p3:p5
p1:p2:p4:p5
p1:p3:p4:p5
p2:p3:p4:p5
p1:p2:p3:p4:p5

Deviance
18.09
0.35
1.56
21.57
267.25
104.54
7.74
0.005
52.93
161.59
101.58
70.25
65.50
56.17
51.51
41.87
4.84
2.90
66.17
1.46
0.42
0.14
3.50
11.92
2.40
21.70
0.0007
0.64
1.83
1.18
0.15

p-value
2.096e − 05 ***
0.552
0.210
3.410e − 06 ***
4.504e − 60 ***
1.533e − 24 ***
0.005 **
0.941
3.443e − 13 ***
5.059e − 37 ***
6.841e − 24 ***
5.212e − 17 ***
5.808e − 16 ***
6.618e − 14 ***
7.115e − 13 ***
9.726e − 11 ***
0.027 *
0.0884 .
4.126e − 16 ***
0.225
0.512
0.702
0.0612 .
0.0005 ***
0.121
3.181e − 06 ***
0.978
0.421
0.175
0.276
0.694

Sign. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1

6.3 Experience of Players
In Section 5 we obtain an experience score by using a su-
pervised dimension reduction method. As in Section 6.1, we
calculate a team experience score for each team in a specific
match as an average of experience scores of team members.
We then normalize team experience score across all matches
and teams. Again we form two samples: the first contains
scores for the team which won (win-team) and the second
has scores of the team which lost (loss-team). Both samples
still follow normal distribution (null hypothesis of normality
test accepted).

In order to show the dependence of team success on expe-
rience score, we formulate the following null-hypothesis and
its alternative.

H0: win-team and loss-team have the same team experience
score.
H1: members of win-team have higher team experience score
compared to loss-team.

Like in the case of hero score, we perform a paired Welch’s
t-test for the two samples with team experience score. The
results of the test lead us to the rejection of the null hy-
pothesis with p-value=0.007, and we accept the alternative
hypothesis. This means that a win-team has on average a
higher team experience score than a loss-team, i.e. experi-
ence matters.

Figure 5: Means of team experience score with re-
gard to team success (win 1 and loss 0) and difficulty
levels (low 0, normal 1 and high 2).

Again, to obtain more insights we relate experience to the
difficulty of the match (factor “skill”). We perform ANOVA
test with team experience score being the dependent variable
and the two independent factors win (value 0 means that
team lost and value 1 means that team won) and skill.

Figure 5 visualizes the results of ANOVA with Tukey multi-
ple comparison plot. The full comparison chart includes 15
levels, but for the sake of interpretation we include only 3
levels which are relevant to our study. The meaning of labels
on y-axis is the same as in Section 6.1. As we see, win-team
has a higher team experience score than loss-team, but this
difference is not statistically significant. We need more data
and insights to investigate this fact.

To verify the dependence of team experience score on the
difficulty level of the match we again perform ANOVA test
with score as a dependent variable and skill as independent
factor. We state the null-hypothesis as follows.

H0: team experience score is the same across matches of dif-
ferent skills.
H1: the higher the skill of the match, the higher team expe-
rience score is.

Figure 6 shows Tukey multiple comparison plot for ANOVA
test, and according to the results of ANOVA test and this
figure we reject the null-hypothesis: teams in matches of
high difficulty level have considerably higher team experience
score than those of low and normal difficulties. Though, the
mean difference between difficulties “low” and “normal” can-
not be claimed significant. Obviously, the more experience
a player is, the more difficult matches he/she participates
in.

While calculating experience score we take into considera-
tion only experience of players individually. A more sophis-
ticated approach is required to calculate overall team expe-
rience, taking into account experience of pairwise com