See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/286993315

Modeling the Offensive-Defensive Interaction and Resulting Outcomes in

Basketball

Article  in  PLoS ONE · December 2015

DOI: 10.1371/journal.pone.0144435

CITATIONS READS
16 373

5 authors, including:

Leonardo Lamas Felipe Luiz Santana

University of Brasília University of São Paulo
27 PUBLICATIONS   479 CITATIONS    6 PUBLICATIONS   58 CITATIONS   

SEE PROFILE SEE PROFILE

Matthew Heiner Carlos Ugrinowitsch

University of California, Santa Cruz University of São Paulo
6 PUBLICATIONS   18 CITATIONS    261 PUBLICATIONS   5,134 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Aerobic exercise program with or without motor complexity as an add-on to the pharmacological treatment of depression – a randomized controlled trial View project

Evaluation of team sports players' decisions View project

All content following this page was uploaded by Gilbert W Fellingham on 04 January 2016.

The user has requested enhancement of the downloaded file.

 RESEARCH ARTICLE

Modeling the Offensive-Defensive

Interaction and Resulting Outcomes in
Basketball
Leonardo Lamas1*, Felipe Santana2, Matthew Heiner3, Carlos Ugrinowitsch2,
Gilbert Fellingham4
1 Faculty of Physical Education, University of Brasilia, Brasilia, Distrito Federal, Brazil, 2 School of Physical
Education and Sport, University of Sao Paulo, Sao Paulo, Sao Paulo, Brazil, 3 Department of Applied
Mathematics and Statistics, University of California, Santa Cruz, Santa Cruz, California, United States of
America, 4 Department of Statistics, Brigham Young University, Provo, Utah, United States of America

* [email protected]

Abstract

OPEN ACCESS Purpose

Citation: Lamas L, Santana F, Heiner M, We analyzed the interaction between offensive (i.e. space creation dynamics -SCDs) and
Ugrinowitsch C, Fellingham G (2015) Modeling the defensive (i.e. space protection dynamics—SPDs) actions in six play outcomes (free shot,
Offensive-Defensive Interaction and Resulting
Outcomes in Basketball. PLoS ONE 10(12):
contested shot, new SCD, reset, foul, and turnover) in Spanish professional basketball
e0144435. doi:10.1371/journal.pone.0144435 games.
Editor: Jaime Sampaio, Research Center for Sports
Sciences, Health and Human Development
(CIDESD), University of Trás-os-Montes e Alto
Method
Douro, Vila Real, Portugal, PORTUGAL Data consisted of 1548 SCD-SPD-outcome triples obtained from six play-off games. We
Received: July 7, 2015 used Bayesian methods to compute marginal probabilities of six outcomes following five dif-
ferent SCDs. We also computed probabilities of the six outcomes following the 16 most fre-
Accepted: November 18, 2015
quent SCD-SPD combinations.
Published: December 14, 2015

Copyright: © 2015 Lamas et al. This is an open

access article distributed under the terms of the Results
Creative Commons Attribution License, which permits
The pick action (e.g. pick and roll, pop and pop) was the most prevalent SCD (33%). How-
unrestricted use, distribution, and reproduction in any
medium, provided the original author and source are
ever, this SCD did not produce the highest probability of a free shot (0.235). The highest
credited. probability of a free shot followed the SCD without ball (0.409). The pick was performed not
Data Availability Statement: All relevant data are
only to attempt scoring but also to initiate offenses, as it produced the highest probability
within the paper and its Supporting Information files. leading to a new SCD (0.403). Additionally, the SPD performed influenced the outcome of
Funding: This work was supported by Fundação de
the SCD. This reinforces the notion that the opposition (offensive-defensive interaction)
Amparo à Pesquisa do Estado de São Paulo, grant should be considered. To the best of our knowledge, in team sports, this is the first study to
number 2012/19164-0 (LL) and Conselho Nacional successfully model the tactical features involved in offense-defense interactions. Our analy-
de Pesquisa, grant number 304205/2011-7 (CU). The
ses revealed that the high frequency of occurrence of some SCDs may be justified not only
funders had no role in study design, data collection
and analysis, decision to publish, or preparation of by an associated high probability of free shots but also by the possibility of progressively
the manuscript. create more space in the defense (i.e. a new SCD as outcome). In the second case, it

PLOS ONE | DOI:10.1371/journal.pone.0144435 December 14, 2015 1 / 14

 Modeling of Offensive-Defensive Interactions in Basketball

Competing Interests: The authors have declared evidences offensive strategic features of progressive disruption of the defensive system
that no competing interests exist. through the concatenation of subsequent offensive actions.

Introduction
Basketball strategies are designed based on the coaching staff’s expected probability of success
of the specified set of plays. Match analysis techniques provide quantitative evidence of the ade-
quacy of the strategy, which is often based on box-score information. The information derived
from box-scores (e.g. effective field goal percentage; turnover percentage; offensive and defen-
sive rebound percentage; free throw efficiency) has been systematically improved over the
years as researchers have sought to quantify the relationship between these indices and the
probabilities of winning matches [1–5]. Nonetheless, these box-score derived indices are based
on the frequency of outcomes that occur at the end stages of ball possessions, limiting coaches’
ability to fully understand the causes of the quantified outcomes [6, 7]. Thus, the actual playing
pattern of a team may not be fully understood, increasing the error rate in the coach’s decision
process [8].
Alternatively, we have seen a trend toward the development of analytical tools that are sensi-
tive to the team’s dynamics (e.g. the patterns of ball circulation and players actions) and the
outcome obtained, with the aim of predicting a team’s performance [9–11]. Besides these ana-
lytical improvements, the relationship between a priori planning and execution of players
actions has been modeled [12]. According to this model, a priori planning, henceforth called
team strategy, was defined as a discrete dynamic system, which is described by a set of states.
Each state of this dynamic system is described by categorical variables containing tactical speci-
fications that are sufficient to characterize different game circumstances (e.g. actions directly
related to the ball).
The annotated states are restricted to a time interval in which the observed tactics (i.e., indi-
vidual or collective team players actions) have a high influence on team success. In basketball,
this interval occurs when the offensive team tries to create space in the opponent’s defensive
system to obtain a shooting opportunity. Usually, basketball teams have a limited number of
offensive sets, characterized by specific sequences of states, that they use multiple times in a
match [13]. In these sequences of states, there exists a large but finite set of possible actions
(e.g. offensively: pick and roll; defensively: trap on ball player) that can be performed to achieve
a limited set of outcomes (e.g. a free shot, a turnover). Thus, the relationship between
sequences of states a team performs and the final outcome, on each play, is affected by the
opponent’s actions. Consequently, one should aim to assess the content of the offense-defense
interaction to interpret the possible causes of the achieved outcome.
Categorization of offensive and defensive tactical classes of actions in basketball was devel-
oped, respectively, by Lamas et al. (2011) and Santana et al. (2015). These authors defined sets
of equivalence classes of actions that can be planned and, consequently, performed to create
and protect space in basketball matches, denominated as space creation dynamics (SCDs) and
space protection dynamics (SPDs), respectively. The compromise between increasing tactical
resolution through labeling players’ actions and limiting sample size as a consequence of man-
ual annotation dependency was positively solved as classes demonstrated to satisfatorily dis-
criminate between analyzed teams [13–15].
An extension of previous work on SCD and SPD classes is the application of statistical mod-
els that take into consideration the sequential nature of basketball match events to determine

PLOS ONE | DOI:10.1371/journal.pone.0144435 December 14, 2015 2 / 14

 Modeling of Offensive-Defensive Interactions in Basketball

the probability distribution of actions leading to positive or negative outcomes, given the
actions performed by the adversary. Thus, considering the SCD-SPD-outcome triple can be an
interesting approach to represent the basic structure of offensive-defensive interactions in the
match and to understand match dynamics. Therefore, the present work aims to analyze the
performance in professional basketball matches through modeling the interaction between the
SCDs, SPDs and outcomes obtained.

Materials and Methods

Sample
The data were gathered using video from six consecutive matches from the same team (Barce-
lona F.C.), in the same tournament (Liga ACB—Spanish championship, 2010–2011), against
semi-final and final adversaries (i.e. Caja Laboral and Bilbao Biskaia). SCD-SPD couples were
noted every time they occurred, along with the outcome. Ball possessions that terminated
before any interaction occurred were not considered. The Ethical Committee of the School of
Physical Education and Sport of the University of Sao Paulo approved all experimental proce-
dures (protocol: 2009-10). S1 Dataset.

SCD and SPD classes

Lamas et al. (2011) presented a set of offensive classes, which describes the possible actions
used to create space in the adversary’s defensive system. These offensive classes or SCDs were
used as a reference to propose corresponding defensive classes or SPDs [14]. In these studies,
there were defined equivalent classes that encompass all possible variations of the action that
define a given SCD or SPD class. For this purpose, authors followed a very rigorous validation
process. Herein, the sampling structure considered an offensive action (i.e. an SCD) as a trigger
event for a sequence of interest (i.e. offense-defense-outcome). Thus, the execution of a SPD
was conditioned by the occurrence of a previous SCD. Fig 1 provides a pictorial view of the
SCD-SPD couples.
In some cases the SPD will be described by two words. For example, the defense of a pick
could be described as a second (the defensive action on the ball handler) show (the defensive
action on the screener). Another defensive action that we characterize but do not show in Fig 1
is defense out of position.

Outcomes
All possible outcomes of an SCD-SPD couple were grouped in six classes, a) free shot—the
shooting player is not closely guarded, b) contested shot—the shooting player is guarded, c)
new SCD—the previous SCD is concatenated with a subsequent SCD, d) reset—offense is neu-
tralized and needs to re-start with a new play, e) foul—the offensive player is fouled, and f)
turnover—the offense loses ball possession.

Reliability procedures
In the experiment, a single researcher collected all sample data. His reliability in assigning an
SCD, an SPD and an outcome in a given offense-defense interaction was assessed. To evaluate
the intra-observer consistency a set of 200 offensive-defensive couples of actions were ran-
domly selected from ball possessions of high-level basketball games from different leagues (e.g.
NBA, EuroLeague). The game analyst annotated each of the triple elements (i.e. offensive
action, defensive action and outcome) and classified them according to the SCD, SPD and out-
come classes, on three different occasions, one week apart each other. Reliability ratios were

PLOS ONE | DOI:10.1371/journal.pone.0144435 December 14, 2015 3 / 14

 Modeling of Offensive-Defensive Interactions in Basketball

Fig 1. Patterns of SCD-SPD couples in 1 on 1, 2 on 2, and 3 on 3 interactions. Continuous arrows indicate player movement without the ball, zigzag
arrows indicate player movement with the ball, dotted arrows indicate a pass, and a T indicates a screen. Diagrams 1A and 1B show a 1 on 1 SCD on the
perimeter where the SPD position is neutral in 1A and oriented in 1B. Diagrams 2A and 2B show a 1 on 1 SCD in the post where the SPD position in 2A is
neutral and in 2B is oriented. Diagrams 3A and 3B show 2 on 2 interactions where the SPD emphasis is on the player without the ball. In 3A the SPD position
is away and in 3B the SPD position is close. Diagrams 4A and 4B show the pick SCD with SPD positions focusing on the ball handler. In 4A—“x” the SPD is
second (fight through), in 4A—“y” the SPD is third (middle), and in 4A—“z” the SPD is fourth (behind). In diagram 4B the SPD position is deny. In diagrams
5A-D the SPD defender actions focus on the player setting the screen. In 5A the SPD position is show, in 5B the SPD position is open, in 5C the SPD position
is sustain, in 5D the SPD position is away. Diagrams 6A and 6B present combined defense between defenders of ball handler and screener. In 6A and 6B,
the SPD positions are, respectively, switch and trap. In diagrams 7–9 the SCD is a screen. In 7A-B the SPD positions focus on the player for whom the
screen is being set. In 7A the SPD position is deny. In 7B—“x” the SPD position is second (fight through), in 7B—“y” the SPD position is third (middle), in 7B
—“z” the SPD position is fourth (behind). In 8A-D the SPD positions focus on the player setting the screen. In 8A the SPD position is show, in 8B the SPD
position is open, in 8C the SPD position is away, in 8D the SPD position is sustain. Diagram 9A displays combined defense between defenders of the player
on whom the screen is being set and the screener. The SPD position is the switch. Adapted from Santana et al. (2015), for illustrative purposes only.
doi:10.1371/journal.pone.0144435.g001

evaluated according to the levels of agreement for the Kappa value [16]: <0 less than the
chance agreement, 0.01—0.20 slight agreement, 0.21—0.40 fair agreement, 0.41—0.60 moder-
ate agreement, 0.61—0.80 substantial agreement and 0.81—0.99 almost perfect agreement.

Data Analysis
Analysis was conducted using 1,548 sequences, each consisting of an SCD-SPD-outcome triple.
All sequences were pooled from the six matches to obtain adequate counts for statistical infer-
ence and were analyzed from two perspectives using Bayesian methods. The first perspective
considered the marginal association of SCDs to outcomes, while the second perspective related
SCD-SPD combinations to outcomes.

PLOS ONE | DOI:10.1371/journal.pone.0144435 December 14, 2015 4 / 14

 Modeling of Offensive-Defensive Interactions in Basketball

In the first analysis (perspective), counts of SCDs leading to each of the six outcomes were
tabulated. For example, the free shot outcome had 77 sequences that were initiated with the 1
on 1 perimeter SCD, 116 sequences initiated with the on ball screen SCD, and so on for each of
the five distinct SCDs. These counts of different SCDs leading to a free shot were then modeled
as a multinomial random vector in which n1, n2, . . ., n5 represent these counts, and p1, p2, . . .,
p5 represent the probabilities of each SCD preceding a free shot. Specifically, p1 is the probabil-
ity that a 1 on 1 perimeter SCD leads to a free shot with corresponding n1 = 77, p2 is the proba-
bility that a 1 on 1 post leads to a free shot with corresponding n2 = 22, etc. The multinomial
model is then given by Eq 1:
ðS5i¼1 ni Þ! n1 n2 n3 n4 n5
f ðnjpÞ ¼ p1 p2 p3 p4 p5 ð1Þ
P5i¼1 ni !

P
5
where 0  pi  1 for all i and pi ¼ 1
i¼1
Posterior inference for the pi requires a prior distribution over these parameters. A Dirichlet
prior was chosen for this model. The density for this prior is given by Eq 2:
GðS5i¼1 ai Þ a1 1 a2 1 a3 1 a4 1 a5 1
pðpjaÞ ¼ p1 p2 p3 p4 p5 ð2Þ
P5i¼1 Gðai Þ

where αi > 0 and Γ() is the gamma function. The Dirichlet prior is computationally convenient
in a multinomial model because of conjugacy, that is, application of Bayes’ theorem over (1)
[17] and (2) yields a posterior density of the form presented in Eq 3:
GðS5i¼1 ðni þ ai ÞÞ n1 þa1 1 n2 þa2 1 n3 þa3 1 n4 þa4 1 n5 þa5 1
pðpjn; aÞ ¼ p1 p2 p3 p4 p5 ð3Þ
P5i¼1 Gðni þ ai Þ

which is again Dirichlet with updated α parameters. In this model, the α parameters can be
considered as prior “counts,” where larger values of αi weight the prior density in favor of pi.
Small values of α correspond to a weak prior, resulting in the counts in the data, the ni, domi-
nating the posterior inference for the pi.
In this analysis, the values αi = 0.098 for i = 1, . . ., 5 were chosen in accordance with recom-
mendations by de Campos and Benavoli (2011), who presented a method of finding weak prior
α values for which Bayes point estimators of the pi are comparable to the maximum likelihood
estimator in mean squared error. Furthermore, the five αi values were set equal to one another
so as to not favor any one pi over the others a priori.
This model was replicated six times, with one model for each outcome (i.e. free shot, con-
tested shot, etc.). The estimated parameters in each model represented the probabilities of each
of the five SCDs preceding the specified outcome. In each model, comparative inference for the
pi parameters was carried out using Monte Carlo simulation from the full joint posterior (Eq
3). Posterior simulations used 50,000 independent draws, resulting in a standard deviation of
less than 0.003 in Monte Carlo approximations of posterior probabilities. Point and interval
estimates for the individual pi parameters were obtained from their exact marginal posteriors,
which are beta distributions [18, 19]. These marginal posterior distributions for all six models
are summarized with box plots in Fig 2. Each box plot shows the posterior median point esti-
mates and 95% credible intervals for the pi are reported next to the box plots. Note that the
numbers reported at the top of each plot near the title indicate the total number of data points
contributing to the corresponding model.
In the second analysis, counts of outcomes resulting from each SCD-SPD combination were
tabulated. For example, the SCD-SPD combination 1 on 1 post (SCD) to out of position (SPD)

PLOS ONE | DOI:10.1371/journal.pone.0144435 December 14, 2015 5 / 14

 Modeling of Offensive-Defensive Interactions in Basketball

Fig 2. Plots showing point and interval estimates of the probability of each SCD leading to a certain outcome. Plots show median, 50 and 95%
credible intervals (CI). Medians and 95% credible intervals are shown on the right axis.
doi:10.1371/journal.pone.0144435.g002

was followed by a free shot in n1 = 7 sequences, a contested shot in n2 = 2 sequences, and so on

for the six outcomes. These six counts were modeled as a multinomial vector with parameters
p1, . . ., p6, which are interpreted as the probabilities of the 1 on 1 post with an out of position
defense combination leading to each of the six outcomes. The same method was employed in

PLOS ONE | DOI:10.1371/journal.pone.0144435 December 14, 2015 6 / 14

 Modeling of Offensive-Defensive Interactions in Basketball

this analysis as in the first, with prior α = 0.061 adjusted to reflect use of six parameters. This
model was replicated for each SCD-SPD combination. Only the couples with frequencies of
occurrence  25 events were used for statistical modeling.
For the purpose of this paper, when a probability is classified as ‘significant’, it means the
posterior probability of the stated event (πpost) exceeded 0.95. We recognize that we are not
using the word ‘significant’ in its classical sense, that is we are not stating and then rejecting a
series of null hypotheses. Rather, since we are considering these problems from a Bayesian
standpoint, we are simply attempting to categorize the strength of a posterior probability. Fur-
ther, we also recognize that the generalizability of the study may be suspect, since our work
focuses on one team. Nonetheless, we feel that the use of statements about the posterior proba-
bility of the various events are useful in characterizing offensive and defensive strategies in bas-
ketball. More extensive posterior summaries for several of the models are reported in Figs 3–7.

Results
Reliabilty results for the game analyst responsible for data collection indicated a high intra-
rater consistency between days (agreement level > 0.92). Initially, we present the posterior
probabilities of the outcomes for each SCD. Next, the posterior probabilities of outcomes for
SCD-SPD combinations are given.
The relative frequencies of the SCDs observed in the matches analyzed were the following: i)
1 on 1 perimeter—19% (diagrams 1A and 1B in Fig 1); ii) 1 on 1 post—8% (diagrams 2A and
2B in Fig 1); iii) without ball—22% (diagrams 3A and 3B in Fig 1); iv) pick—33% (diagrams
4A-6B in Fig 1); v) screen—18% (diagrams 7A-9A in Fig 1). The SCDs that produced the two
highest median posterior probabilities of occurrence of the predefined outcomes were the fol-
lowing (see Fig 2): i) free shot—without the ball (0.409) or a pick (0.235); ii) contested shot—1
on 1 perimeter (0.367) and 1 on 1 post (0.285); iii) new SCD—a pick (0.403) and a screen
(0.268); iv) reset—a pick (0.457) and a 1 on 1 perimeter (0.186); v) foul—a pick (0.329) and
without ball (0.277); vi) turnover—a pick (0.366) and without ball (0.257).
Figures below present the probabilities of outcomes following SCD-SPD couples. The credi-
ble intervals are given in the box plots of each figure. Fig 3 presents the probabilities of the out-
comes when the interactions start with an action of the ball player on the perimeter (i.e. 1 on 1
on the perimeter). The contested shot (0.316) was the most likely outcome when a non-ori-
ented defense (Panel A) was used. When an oriented defense was used (Panel B), the contested
shot (0.262) and new SCD (0.262) had higher probabilities of occurrence than the other out-
comes. When the defense was out of position (Panel C), a free shot (0.400) had the highest
probability. These data offer compelling evidence that defense out of position produced a sig-
nificantly higher probability of a free shot than oriented defense (πpost > 0.99). Similarly, non-
oriented defense produced a significantly higher probability of a contested shot than defense
out of position (πpost > 0.99).
Fig 4 presents the probabilities of the outcomes when interactions started in the post (1 on 1
in the post). A contested shot (0.421) was the most likely outcome in the case of non-oriented
defense (Panel A). Similarly, the contested shot (0.530) was the most likely outcome for ori-
ented defense (Panel B). Both non-oriented (0.166) and oriented (0.101) defenses produced
significantly lower probabilities of a free shot than a contested shot (πpost > 0.99, in both
cases).
Fig 5 presents the probabilities of the outcomes when interactions started with space crea-
tion without ball. A free shot (0.638) had the highest probability of occurrence in an away
defense (Panel A). The probability of a free shot was significantly lower (0.206) in a close
defense (Panel B). The close defense had a significantly higher probability (0.298) of a reset

PLOS ONE | DOI:10.1371/journal.pone.0144435 December 14, 2015 7 / 14

 Modeling of Offensive-Defensive Interactions in Basketball

Fig 3. Outcome probabilities for an SCD of 1 on 1 in the perimeter. The SPDs are: Panel A) not oriented; Panel B) oriented; Panel C) defense out of
position. X-axis probability; Y-axis—possible outcomes; second Y-axis—Posterior median estimate of outcome probability (95% credible intervals).
doi:10.1371/journal.pone.0144435.g003

than either an away defense (0.140) or an out of position defense (0.034), (πpost = 0.98 and πpost
> 0.99, respectively).
In Fig 6 we present the probabilities of the outcomes when interactions started with a pick.
Free shot (0.344), reset (0.316) and new SCD (0.247) had the highest probabilities among the
possible outcomes with a second-away defense (Panel A). In a second-show defense (Panel B),
new SCD (0.451) had the highest probability of occurrence, followed by reset (0.293). Regard-
ing the switch defense, reset (0.351), new SCD (0.288), and contested shot (0.204) had the high-
est probabilities of occurrence. A new SCD (0.439) had a higher probability of occurrence than
any other outcome except free shot (0.298) in the third-open SPD. Second-away (0.344) had a
significantly higher probability of leading to a free shot than either second-show (0.151) or
switch (0.099) (πpost > 0.99, in both cases). Alternatively, second-show (0.451) produced a
higher probability of occurrence of a new SCD than either second-away (0.247) or switch
(0.288) (πpost = 0.97 and πpost > 0.99, respectively). Overall, these SPDs presented low probabil-
ities of leading to a contested shot. Finally, there were observed moderate and similar probabili-
ties of leading to a reset among the defenses.
In Fig 7 we present the probabilities of the outcomes when interactions started with a screen.
Switches produced moderate and similar probabilities (0.304 in each case) of a free shot, a new

PLOS ONE | DOI:10.1371/journal.pone.0144435 December 14, 2015 8 / 14

 Modeling of Offensive-Defensive Interactions in Basketball

Fig 4. Outcome probabilities for an SCD of 1 on 1 in the post. The SPDs are: Panel A) not oriented defense; Panel B) oriented defense. X-axis
probability; Y-axis—possible outcomes; second Y-axis—Posterior median estimate of outcome probability (95% credible intervals).
doi:10.1371/journal.pone.0144435.g004

Fig 5. Outcome probabilities for a 2 on 2 SCD where the SPD emphasis is on the player without the ball. Panel A) away defense; Panel B) close
defense; Panel C) defense out of position. X-axis probability; Y-axis—possible outcomes; second Y-axis—Posterior median estimate of outcome probability
(95% credible intervals).
doi:10.1371/journal.pone.0144435.g005

PLOS ONE | DOI:10.1371/journal.pone.0144435 December 14, 2015 9 / 14

 Modeling of Offensive-Defensive Interactions in Basketball

Fig 6. Outcome probabilities for a 2 on 2 pick. Panel A) second-away; Panel B) second-show; Panel C) switch; Panel D) third-open. X-axis probability; Y-
axis—possible outcomes; second Y-axis—Posterior median estimate of outcome probability (95% credible intervals).
doi:10.1371/journal.pone.0144435.g006

SCD, and a reset. The most likely outcome following a third-open defense was a free shot
(0.402). Third-open produced a significantly higher probability of a free shot than second-
away (0.221) or second-sustain (0.235) (πpost = 0.95 and πpost = 0.96). A contested shot had low
and similar probabilities among the different types of defense. New SCD had a significantly
higher probability for second-away (0.481) and second-sustain (0.527) than for third-open
(0.252) (πpost = 0.97 and πpost > 0.99). For second-away and second-sustain the probability of a
new SCD was significantly higher than any other outcome (πpost > 0.99).

Discussion
This study aimed to model the relationship between SCDs, SPDs and outcomes obtained from
the analyses of professional basketball matches. Overall, the novel finding of the present study
encompasses the ability to determine SCDs favoring either positive attack outcomes (e.g. free
shot) or negative attack outcomes (e.g. turnover) when taking into consideration all possible
SPDs.
The presented work is complementary to several studies that have developed variables to
assess the collective dynamic features of the opposition between teams [20–22]. Differently
from previous studies, the semantic of the actions was the focus of the analyses and match situ-
ations were distinguished based on the variables that describe each game state in a sequence of

PLOS ONE | DOI:10.1371/journal.pone.0144435 December 14, 2015 10 / 14

 Modeling of Offensive-Defensive Interactions in Basketball

Fig 7. Outcome probabilities for a 3 on 3 screen. Panel A) second-away; Panel B) second-sustain; Panel C) switch; Panel D) third-open. X-axis
probability; Y-axis—possible outcomes; second Y-axis—Posterior median estimate of outcome probability (95% credible intervals).
doi:10.1371/journal.pone.0144435.g007

events. This type of analysis provides criteria for selecting information that are relevant to
explain teams’ tactics and its relation to team strategy.
The analysis of SCD-SPD couples showed the outcome obtained with every SCD was influ-
enced by the type of SPD performed by the adversary. This reinforces the notion that the oppo-
sition should be analyzed taking into consideration the offensive-defensive interaction. Thus,
the present study was an improvement on previous work in which only the SCDs of a single
team were considered in the analysis [13] or in which a predictive modeling was missing [14].
The pick was the most prevalent SCD (33%). However, this SCD did not produce the high-
est probability of a free shot (0.235). The highest probability of a free shot followed a without
ball SCD (0.409). The pick seemed to be used to not only lead to a scoring opportunity but also
to initiate offenses, as it produced the highest probability leading to a new SCD (0.403). The
without ball also presented the second greatest probability of leading to a foul (0.277) or a turn-
over (0.257). The mentioned probabilities of succesful outcomes (i.e. free shot or foul) and
unsuccesful outcome (i.e. turnover) of the without ball characterize this SCD as possibly pro-
viding a high reward even though a non-neglectable risk also exists. Although frequent (19%),
1 on 1 perimeter presented the highest probability of a contested shot (0.367), which corrobo-
rates with the empirical notion that the difficulty for the ball player to free himself without any
help increases in modern basketball. The second hightest probability of a contested shot was

PLOS ONE | DOI:10.1371/journal.pone.0144435 December 14, 2015 11 / 14

 Modeling of Offensive-Defensive Interactions in Basketball

found for 1 on 1 post (0.285) and this SCD presented the lowest frequency of ocurrence (8%).
Besides the fact that the post area is naturally well protected for its proximity to the basket,
favoring the shot be contested, low recurrence of 1 on 1 post aligns with the defensive effort for
keeping the ball away from the post. Finally, there was not a great difference in outcome when
the SCD was a screen.
In the 1 on 1 perimeter SCD, an out of position defense yielded the greatest probability of a
free shot (0.400). Out of position defense is often the consequence of a previous SCD or a
defensive mistake. Hence, from an offensive standpoint, this result supports the notion that the
concatenation of disruptive actions is useful in progressively creating sufficient space. Other
contextual elements, such as the tactical expertise of the offensive player being guarded, may
indicate which SPD should be performed.
In the 1 on 1 in the post SCD, both oriented (0.503) and non-oriented (0.421) defenses pro-
duced high probabilities of contested shots. Due to the proximity to the basket, the defense is
always close in post actions, but often these contested shots have a high probability of scoring
points. Additionally, in a non-oriented defense, the need to guard the post player by concen-
trating defensive players close to the basket often leads to a new SCD (0.217).
The without the ball SCD was more efficiently guarded by a close defense (free shot probabil-
ity of 0.206) than both away defense (0.638) and out of position defense (0.647). An away defense
may be a consequence of the displacement of a defensive player to another area of the court to
help a teammate. Also, an out of position defense may represent a stronger commitment of one
defender in helping another one. The outcomes of these SPDs against a without ball SCD high-
light a strategic concern regarding help defense, as the commitment in helping other defensive
players seems to reduce the ability of the defenders to guard their assigned attackers.
In relation to the pick SCD, the second-show SPD (0.151) was significantly more effective
than either the second-away (0.344) or the third-open SPD (0.298) in decreasing the probabil-
ity of free shots (πpost > 0.99 and πpost = 0.98, respectively). Among these three SPDs, second-
show puts the greatest pressure on the ball player. All three of these SPDs, second-show, sec-
ond-away and third-open, lead to a reset with similar probabilities (0.293, 0.316 and 0.204).
Second-show also yielded a high probability of leading to a new SCD (0.451), indicating a pos-
sible trade-off between avoiding a shot and enabling the offense to concatenate a new SCD
because of the high commitment of the defenders involved. The defender performing the show
is momentarily allocated in the defense of the ball player and his attacker may move to start a
new SCD.
In regards to the screen, the third-open SPD led to the highest probability of free shots
(0.402). This is expected because this SPD provides enough space for the screen receiver to per-
form an immediate shot. For the other SPDs, second-away and second-sustain, there was a
high probability of a new SCD occurring, 0.481 and 0.527 respectively. The high probability of
a new SCD seems to be due to the defensive demands in guarding an offensive action involving
three players, which requires the commitment of the defenders in helping each other.
In the present study, the sample data were gathered from three high-level basketball teams
only. Thus, one should consider this fact when attempting to generalize the practical findings
in the present study to other teams or leagues. Nonetheless, previous work has shown that sim-
ilar situations arise at different competition levels indicating that our findings may be applica-
ble to different samples [23]. Additionally, considering the current technological
improvements that have enabled acquiring larger amounts of data from a game or even a com-
plete season [24], these findings may be helpful for labelling events on a large scale supported
by pattern recognition methods.
In summary, to the best of our knowledge, in team sports, this was the first study to model
the tactical features of the offense-defense interaction. Our analyses revealed that SCDs with

PLOS ONE | DOI:10.1371/journal.pone.0144435 December 14, 2015 12 / 14

 Modeling of Offensive-Defensive Interactions in Basketball

high frequency of occurrence did not necessarily lead to high probabilities of free shots. How-
ever, SCD’s often lead to a new SCD as an outcome, evidencing offensive strategic features of
progressive disruption of the defensive system through the concatenation of subsequent offen-
sive actions.

Supporting Information
S1 Dataset. Triples of SCD-SPD-outcome from the sample.
(XLSX)

Author Contributions
Conceived and designed the experiments: LL GF CU. Performed the experiments: FS LL. Ana-
lyzed the data: MH GF. Contributed reagents/materials/analysis tools: CU MH. Wrote the
paper: CU LL GF FS.

References
1. Kubatko J, Oliver D, Pelton K, Rosenbaum D. A Starting Point for Analyzing Basketball Statistics. Jour-
nal of Quantitative Analysis in Sports. 2007; 3(3). doi: 10.2202/1559-0410.1070
2. Oliver D. Basketball on paper. Dulles: Potomac Books; 2004.
3. Baghal T. Are the “Four Factors” Indicators of One Factor? An Application of Structural Equation Model-
ing Methodology to NBA Data in Prediction of Winning Percentage. Journal of Quantitative Analysis in
Sports. 2012; 1(8).
4. Cade W. Basketball Analytics: Optimizing the Official Basketball Box-Score (Play-by-Play). In: SESUG;
2013. p. 1–7.
5. Ruiz M, Lopez-Hernández F, Martinez J, Castellano A. The relationship between concentration of scor-
ing and offensive efficiency in the NBA. Journal of Quantitative Analysis in Sports. 2014; 1(10):27–36.
6. Page GL, Fellingham GW, Reese CS. Using box-scores to determine a position’s contribution to win-
ning basketball games. Journal of Quantitative Analysis in Sports. 2014; 4(3):Article 1.
7. Correia V, Araújo D, Vilar L, Davids K. From recording discrete actions to studying continuous goal-
directed behaviours in team sports. Journal of Sports Sciences. 2013; 31:546–553. doi: 10.1080/
02640414.2012.738926 PMID: 23140581
8. Franks I, Miller G. Eyewitness testimony in sport. Journal of Sport Behavior. 1986; 9:38–45.
9. Štrumbelj E, Vračar P. Simulating a basketball match with a homogeneous Markov model and forecast-
ing the outcome. International Journal of Forecasting. 2012 April–June; 28(2):532–542. doi: 10.1016/j.
ijforecast.2011.01.004
10. Oh M, Keshri S, Iyengar G. Graphical Model for Baskeball Match Simulation. In: MIT Sloan Sports Ana-
lytics Conference; 2015. p. 2–8.
11. Lucey P, Bialkowiski A, Carr P, Yue Y, Matthews I. How to Get an Open Shot: Analyzing Team Move-
ment in Basketball using Tracking Data. In: MIT Sloan Sports Analytics Conference; 2014.
12. Lamas L, Barrera J, Otranto G, Ugrinowitsch C. Invasion team sports: strategy and match modeling.
International Journal of Performance Analysis in Sport. 2014; 14(1):307–329.
13. Lamas L, D DRJ, Santana F, Rostaiser E, Negretti L, Ugrinowitsch C. Space creation dynamics in bas-
ketball offence: validation and evaluation of elite teams. International Journal of Performance Analysis
in Sport. 2011; 11(1):71–84.
14. Santana F, Rostaiser E, Sherzer E, Ugrinowitsch C, Barrera J, Lamas L. Space protection dynamics in
basketball: validation and application to the evaluation of offense-defense patterns. Motriz. 2015; 21
(1):34–44.
15. Crum P. Space creation dynamics in basketball: a comparison between British and Spanish leagues.
University of Chester; 2013.
16. Landis JR, Koch GG. The measurement of observer agreement for categorical data. Biometrics. 1977;
1(33):159–174. doi: 10.2307/2529310
17. de Campos C, Benavoli A. Inference with multinomial data: Why to weaken the prior strength. In: Inter-
national Joint Conference on Artificial Intelligence; 2011. p. 2107–2112.

PLOS ONE | DOI:10.1371/journal.pone.0144435 December 14, 2015 13 / 14

 Modeling of Offensive-Defensive Interactions in Basketball

18. R Core Team. R: A language and environment for statistical computing. Vienna, Austria: R Foundation
for Statistical Computing; 2015. Available from: http://www.R-project.org/.
19. Martin AD, Quinn KM, Park JH. MCMCpack: Markov Chain Monte Carlo in R. Journal of Statistical Soft-
ware. 2011; 42(9):1–21. doi: 10.18637/jss.v042.i09
20. Davids K, Araujo D, Shuttleworth R. Applications of dynamical systems theory to football. In: Reilly T,
Cabri J, Araújo D, editors. Science and Football V: The proceedings of fifth world congress on science
and football. Routledge; 2005. p. 537–550.
21. Bourbousson J, Seve C, McGarry T. Space–time coordination dynamics in basketball: Part 2. The inter-
action between the two teams. Journal of Sports Sciences. 2010; 3(28):349–358. doi: 10.1080/
02640410903503640
22. Vilar L, Araújo D, Travassos B, Davids K. Coordination tendencies are shaped by attacker and
defender interactions with the goal and the ball in futsal. Human Movement Science. 2014; 33:14–24.
doi: 10.1016/j.humov.2013.08.012 PMID: 24576704
23. Erčulj F, Štrumbelj E. Basketball Shot Types and Shot Success in Different Levels of Competitive Bas-
ketball. Plos One. 2015; 6(10):1–14.
24. Cervone D, D’Amour A, Bornn L, Goldsberry K. Predicting points and valuing decisions in real time with
NBA optical tracking data. In: MIT Sloan Sports Analytics Conference; 2014. p. 1–9.

PLOS ONE | DOI:10.1371/journal.pone.0144435 December 14, 2015 14 / 14

 Copyright of PLoS ONE is the property of Public Library of Science and its content may not
be copied or emailed to multiple sites or posted to a listserv without the copyright holder's
express written permission. However, users may print, download, or email articles for
individual use.

View publication stats