See discussions, stats, and author profiles for this publication at: [Link]

net/publication/227378929

Are the Four Factors Indicators of One Factor? An Application of Structural

Equation Modeling Methodology to NBA Data in Prediction of Winning
Percentage

Article in Journal of Quantitative Analysis in Sports · January 2012

DOI: 10.1515/1559-0410.1355 · Source: RePEc

CITATIONS READS

27 3,656

1 author:

Tarek Al Baghal
University of Essex
47 PUBLICATIONS 508 CITATIONS

SEE PROFILE

All content following this page was uploaded by Tarek Al Baghal on 25 November 2014.

The user has requested enhancement of the downloaded file.

 Journal of Quantitative Analysis in
Sports
Manuscript 1355

Are the “Four Factors” Indicators of One

Factor? An Application of Structural Equation
Modeling Methodology to NBA Data in
Prediction of Winning Percentage

Tarek Baghal, University of Nebraska–Lincoln

©2012 American Statistical Association. All rights reserved.

 Are the “Four Factors” Indicators of One
Factor? An Application of Structural Equation
Modeling Methodology to NBA Data in
Prediction of Winning Percentage
Tarek Baghal

Abstract
Significant work has gone into the development of team and individual statistics in the NBA;
for example, the team measures of the “Four Factors.” Less work has been conducted using
multivariate analyses of these metrics, including identifying possible new statistical techniques to
analyze these data. In particular, this research examines the feasibility of using structural equation
modeling (SEM) for multivariate analyses of NBA Four Factors data. SEM consists of both
confirmatory factor analysis (CFA) and path modeling. Before SEM is employed, this research
first examines the effects of offensive and defensive Four Factors in a linear regression model,
expanding previous research and providing a baseline for the SEM. In doing so, the data show the
importance of effective field goal percentage. Next, structural equation modeling is employed. The
CFA finds that offensive Four Factors are indicators of a single latent factor, labeled “offensive
quality.” The “defensive quality” latent factor is estimable when replacing opposing teams’ free
throw rate with steals per possession. The SEM is extended to regress winning percentage on
latent offensive and defensive quality as well as salary. Salary is an important and often
overlooked part of multivariate models examining team statistics, but it is easily incorporated in
SEM. The explained variance for the regression in the SEM is similar to that of the linear
regression model and indicates the importance of both offensive and defensive quality, with
offensive quality having a larger effect. Team salaries are related to offensive quality, but not
defensive quality or winning. As such, a second structural equation model is proposed where the
effect of salary on winning is mediated by its relationship with offensive and defensive quality.
Since salary is related to offensive quality but not defensive quality, and offensive quality is more
important to winning percentage, this suggests that money spent is done so for offensive
performance and affects winning through the performance paid for. These results suggest potential
team strategies, as well as the applicability of SEM to sports analytics, not only to NBA data, but
to other sports data as well.

KEYWORDS: NBA, Four Factors, salary, structural equation modeling

Author Notes: The author would like to thank Nikki Siddiqi, Rene Bautista, Jim Bovaird, Justin
Mills, Ibad Siddiqi, Mathew Stange, Aaron D. Yates, and an anonymous reviewer for helpful
comments in developing this research.
 Baghal: Are the “Four Factors” Indicators of One Factor?

1 Introduction

The desired outcome for team performance in National Basketball Association

(NBA) games, like in other sports, is winning. Many variables may affect game
outcomes, and statistical analyses allow for an examination of which variables are
of greatest importance. In turn, such analyses could be used in creating team
strategies, as well as direction for efficient team spending on player salaries.
Research on the NBA in this regard is in the relatively nascent stages, with recent
work appearing in order to create a common framework for analyses of NBA data
(Kubatko et al. 2007). This work has focused on the development of metrics to
assess team and player performance on a variety of offensive and defensive
dimensions.
Of particular interest in team performance analyses are the metrics the
“Four Factors”: effective field goal percentage (EFG), free throw rate (FTR),
turnovers per possession (TPP), and offensive rebounding percentage (ORP).
These can be calculated for both offensive and defensive aspects of a team. These
metrics are of importance and selected for analysis for two major reasons. First,
these measures capture the components that are thought to make up the overall
offensive or defensive performance (Kubatko et al. 2007). Second, research has
shown that the Four Factors are strongly related to winning (Teramato and Cross
2010).
Analyses of these metrics have increased the understanding of the
importance of these factors overall, and the relative importance of each factors.
Research shows that the mean of the Four Factors vary over time, with a general
overall decline in offensive responding percentage and apparently turnovers as
well (Kubatko et al. 2007). Other work has attempted to place weights on each of
the Four Factors, in order to show the relative importance in regards to team
winning (Küpfer 2005). This work suggests that effective field goal percentage is
most important, followed by turnovers per possession, with offensive rebounding
percentage and free throw rate being relatively less important (Küpfer 2005).
Multivariate analyses could further the findings using the Four Factors to
improve understanding of how it affects game outcomes. However, multivariate
analyses of these and their relation to winning has been somewhat limited. Past
research has examined team statistics for these metrics and attempted to relate
them to winning, or have developed projection systems based on individual player
projections (e.g. SCHOENE) (Doolittle and Pelton 2010).
Recently, multivariate models (i.e. ordinary least squares) were used in
estimating the relative impact of team offensive and defensive Four Factors in
winning percentage in the regular season and playoffs (Teramoto and Cross
2010). Examination of the standardized coefficients found that offensive and
defensive EFG were the most important factors in determining winning in both

1
 Submission to Journal of Quantitative Analysis in Sports

the regular season and playoffs, with defensive EFG apparently more important in
winning playoff series. Offensive and defensive FTR, while significant predictors
of winning percentage, were relatively the least important of the Four Factors
(Teramato and Cross 2010). The findings generally confirmed past arguments
about the relative importance of the factors (e.g. Küpfer 2005)
Although such research is a step in the right direction, there are some
limitations. The first limitation is the lack of cost data (i.e. team salary)
incorporated into the multivariate analyses, which fails to account for the effect of
spending on winning. Including cost allows examination of potential efficiencies
(or inefficiencies) in spending as it relates to winning. For example, if defensive
statistics are more strongly related to winning than offensive statistics, but salary
is unrelated to defensive statistics and strongly related to offensive statistics, this
denotes a possible spending inefficiency.
A second issue is that each of the metrics of the Four Factors is likely
related to the same construct. Some have argued that the Four Factors reflect
components of the offensive and defensive ratings, a metric based on points per
possession (Kubatko et al. 2007). However, this suggests that the Four Factors are
causal to the offensive rating; it does not answer what causes the Four Factors. It
is possible that the Four Factors are indeed related to and caused by the same
underlying construct of interest, team offensive and defensive quality. The Four
Factors may be driven by the underlying quality of a team, and any single
measure is limited in terms of being an imperfect (or incomplete) indicator of
offensive and defensive quality. Thus, the effect of effective field goal percentage
on winning tells a limited story, as EFG is only one component of an offense
leading to points per possession.
Therefore, it is important to estimate the underlying construct of interest.
There are several ways that this may be done; with numeric measures such as the
Four Factors, the most promising may be structural equation modeling (SEM)
(Kline 2005). SEM estimates the “true” level of a latent construct (e.g. offensive
or defensive “quality”) based on several indicators using confirmatory factor
analysis (i.e. the measurement portion of the model). In addition, SEM allows for
estimation of the relationships between variables of interest, as one would in a
regression model, by incorporating path analyses between the variables and
estimated factors (i.e. the structural portion of the model) (Kline 2005). Using this
method, it is possible to estimate the underlying quality measures of interest and
relate them to salary and winning.
There were several objectives to the current research. The first was to
conduct regression analyses on winning and the Four Factors using a larger
number of years than Teramato and Cross (2010), and compare results. Second
was to use SEM to determine if the Four Factors could be modeled as indicators
of single latent variables for offensive and defensive quality. If this could be done,

[Link] 2
 Baghal: Are the “Four Factors” Indicators of One Factor?

the third objective was to relate the identified latent quality variables to winning
percentage. Additionally, team salaries were incorporated into the SEM models in
order to estimate the relationship of cost with offensive and defensive quality as
well as winning percentage. Further, this research had the goal of introducing
SEM as an effective statistical tool in sports analytics.

2 Data and Methods

Annual team data were obtained from [Link] (2010) between

the 1995-1996 and 2008-2009 seasons. The start date was selected as it coincided
with the first year of play for the Vancouver Grizzlies and Toronto Raptors,
increasing the number of teams in the league to 29. In 2004, the Charlotte Bobcats
joined the league. Treating each year as an independent observation leads to a
total sample size of 411. Salary data came from Patricia Bender’s collection of
NBA data (Bender 2010).
Team possessions and Four Factors statistics were calculated using
formulas commonly accepted by other NBA researchers (Kubatko et al. 2007). In
particular, these are:

(1) EFG = (FGM + 0.5 × 3PM)/FGA.

(2) FTR = FTM/FGA
(3) TPP = TO t /POSS t 1

(4) ORP = OREB t /(OREB t+ DREBo)

Where FGM is field goal made, FGA is field goal attempted, 3PM is three point
field goals made, FTM is free throw made, TO is turnovers, OREB is offensive
rebounding, DREB is defensive rebounds, and POSS is possessions. Subscript t
indicates team of interest, subscript o indicates opposing team.
For use in the SEM, the Four Factors were all multiplied by 100 to
indicate percentage rather than proportion for EFG, DEFG, ORP, and DORP,
number of turnovers per 100 possessions, and number of free throws made per
100 field goal attempts. This transformation makes interpretation easier and
consistent with other measures, such as offensive and defensive ratings (Kubatko
et al. 2007). The mean of team salaries increases rapidly with each passing year,
e.g. NBA mean salary in 1995-1996 =$24.97 million, mean salary in 2008-2009 =
$71.66 million. Therefore, these data were transformed to a ratio of annual team
salary for that year and NBA mean salary (i.e. salary of team in year/NBA overall
mean in year). The result allows for a consistent metric across all years,
interpretable as the relative size of team salary compared to the entire league, with
1
POSS = FGAt + 0.44 × FTAt – OREBt + TOt

3
 Submission to Journal of Quantitative Analysis in Sports

a ratio of 1 indicating team salary for a given year was equal to the NBA overall
mean salary for that year. For example, in 2008-2009 the New York Knicks had a
team salary about 1.35 times larger than the league average of $71.66 million. All
structural equation models were estimated using Mplus (Muthen and Muthen
2009).

3 Results

The means and standard deviations of the offensive Four Factors years across all
teams and years are presented in Table 1. As the means across all teams and years
for offensive and defensive Four Factors are calculated as such to be nearly
identical, only offensive Four Factor means are presented. As can be seen, when
taking three point shots into account, teams approach fifty percent effectiveness in
field goal attempts. One way to interpret the mean FTR is that for each shot a
team takes, on average it makes 0.23 free throws (i.e. 10 attempts yields slightly
more than 2 made free throws). Slightly more than 15% of possession leads to
turnovers, while teams get about 29% of the rebounds from their own missed
shots.

Table 1: Offensive and Defensive Four Factors Means, All Teams and Years

Variable Mean
(SD)
EFG 0.484
(0.021)
FTR 0.237
(0.026)
TPP 0.156
(0.014)
ORP 0.287
(0.028)
EFG = Effective Field Goal %, FTR = Free Throw Rate, TPP =
Turnovers Per Possession, ORP= Offensive Rebound Percent

Previous research has regressed winning percentage on these offensive

and defensive Four Factors, i.e. effective field goal percentage (EFG), free throw
rate (FTR), turnovers per possession (TPP), and offensive rebounding percentage
(ORP) (Teramato and Cross 2010). This was replicated as a first step in the
current research. However, the number of years, and thus number of cases, was
increased in the current research from 10 to 14 (with the sample increasing from
295 to 411). Increasing the years and sample size serves as a check on past results

[Link] 4
 Baghal: Are the “Four Factors” Indicators of One Factor?

while increasing certainty and how applicable these are over time. Results of the
current ordinary least squares (OLS) regression of winning percentage on
offensive and defensive Four Factors are presented in Table 2 (defensive
counterparts of the offensive Four Factors are preceded by D in the abbreviation).
A White test of the residuals (White 1980) suggested that heteroskedasticity was
not an issue ( χ 442 = 49.50, p = .2631), and thus unadjusted standard errors are
reported.

Table 2: Regression of Winning Percent on Offensive and Defensive Four

Factors

Variable Coefficient (SE) Standardized

Coefficient
EFG 4.740* (0.160) 0.620
FTR 0.782*(0.131) 0.125
TPP -4.410* (0.280) -0.373
ORP 1.564*(0.130) 0.267
DEFG -3.666* (0.163) -0.450
DFTR -0.875* (0.128) -0.155
DTPP 3.274* (0.270) 0.280
DORP -1.666* (0.158) -0.249
Constant 0.196 (0.119) 0
R2=0.859, N=411, *p<0.01
EFG = Effective Field Goal %, FTR = Free Throw Rate, TPP = Turnovers Per Possession, ORP=
Offensive Rebound Percent, D indicates Defensive aspects of measures

The results are similar to past results using a more restricted data set
(Teramato and Cross 2010), with all coefficients significant and in the expected
direction. Although all of the offensive and defensive Four Factors are important
in predicting winning percentage, the importance of effective field goal
percentage is apparent. Unlike the past work in Teramoto and Cross, who found
that offensive and defensive EFG are nearly the same in the effect on regular
season winning, this analysis shows that offensive field goal percentage carries
more importance in predicting winning percentage. A contrast between absolute
values of offensive and defensive effective field goal percentage shows that the
coefficient for offensive EFG is significantly larger, F(1,402) = 25.47, p<0.01.
The second most important measure was TPP, followed by DTPP, ORP, and
DORP, with defensive and offensive free throw rate, respectively, found to be
relatively the least important of the Four Factors.
As previously noted, each of the offensive and defensive Four Factors are
conceptually linked to the overall offensive and defensive performance of the
team. While the above analysis shows the relative importance of each of the

5
 Submission to Journal of Quantitative Analysis in Sports

individual metrics, it is possible that these metrics are also only observed
outcomes of overall offensive and defensive quality. Considering a single
measure such as EFG may be limited because it is only one aspect of offensive or
defensive quality. This quality is latent, directly unobservable, only manifested
through measures such as the Four Factors. Therefore, each of the offensive
(defensive) Four Factors is driven by the level of offensive (defensive) quality.
To test whether these Four Factors are indeed indicators of latent
characteristics, as well as examine the effects of these latent traits on winning,
confirmatory factor analysis (CFA) was conducted. The offensive set of Four
Factors loaded onto a single factor, indicating these all measure one construct,
labeled “offensive quality”. However, the defensive Four Factors failed to load
onto one factor; further analysis indicated this was due to the defensive FTR
measure. This seems plausible, as free throws made by the other team are not
completely in the control of the defense, only the number of free throws
attempted. Other measures of defensive quality were identified in order to keep
equal number of indicators for both the latent constructs. Steals per possession
(SPP) (M=0.083, SD =0.010) was selected as it is a defensive statistic that is
within the control of the defense.
The standard OLS regression model reported in Table 2 was largely
unchanged when SPP replaced DFTR (R2=0.844), with the standardized beta for
SPP = 0.065 (p<.10), somewhat smaller in effect than defensive free throw rate.
In addition, when SPP replaced defensive FTR in the SEM, a single latent
variable was identified, labeled “defensive quality”. For both offensive and
defensive latent factors, the loading for ORP and DORP was constrained to 1, as
one observed variable (or latent variance) must be constrained to have an
estimable model (Kline 2005). Selecting observable variables as those constrained
is consistent with the Unit Loading Identification (ULI) frequently used for model
estimation (Kline 2005).
Modification indices output by the program indicated that allowing
residuals of offensive FTR to covary with the residuals of TPP, as well as
allowing the residual of defensive ORP to covary with defensive EFG residual
would improve model fit. Since it is conceivable logical that both of these
covariations could occur, the model included these covariations. That is, it seems
reasonable that the amount of free throws made in relation to number of shots
taken is related to turnovers (which may also affect field goals attempted, part of
FTR). Similarly, how many offensive rebounds the other team gets may affect
that team’s field goal percentage (e.g. shots closer to/farther from the basket).
Salary (annual salary/league mean salary ratio) entered the model first as a
correlated variable to the latent offensive and defensive factors. Next, winning
percentage was regressed on offensive and defensive “quality” latent factors as
well as salary. Using common convention for the diagramming of SEM, Figure 1

[Link] 6
 Baghal: Are the “Four Factors” Indicators of One Factor?

illustrates the final model described here, where boxes represent observed
variables, circles represent latent variables, the number one represents the ULI
constraint, and e indicates the residuals for a given endogenous variable.

Figure 1. SEM of Winning Percent on Offensive and Defensive Quality

Latent Variables

e1 ORP 1

e2 FTR Off.
Quality
e3 TPP

e4 EFG

WIN% e9
e5 DORP 1

e6 SPP
Def.
Quality
e7 DTPP SALARY

e8 DEFG

EFG = Effective Field Goal %, FTR = Free Throw Rate, TPP = Turnovers Per Possession, ORP= Offensive Rebound
Percent, D indicates Defensive aspects of measures, SPP= Steals Per Possession

This model was estimable, although many of the model quality indicators
did not meet standard levels (Hu and Bentler 1999): model χ 292 =720.410 p<.001,
CFI=0.578, RMSEA = 0.241, 90%CI RMSEA = 0.226, 0.256, SMSR = 0.118.
Parameter and variance estimates are presented in Table 2. Panel A presents the
measurement portion of the model (the estimation of the latent factors) with Panel
B displaying the results of the structural portion of the model (the path/regression
model).
Turning to the measurement portion of the model, it is important to note
that based on the loadings, the scaling of the offensive and defensive quality latent
indicators is similar, and thus interpretable in a similar metric. The parameters’
paths from the latent variables to the indicators are interpreted as in a regression,
with an increasing of one unit in the latent variable leading to an expected change
in the indicator variable.
Generally, all parameters in the measurement model are in the expected
directions, e.g. offensive quality is directly related to EFG and defensive quality
inversely related to DEFG. The exceptions may be that higher ORP and DORP

7
 Submission to Journal of Quantitative Analysis in Sports

(see standardized coefficients) are related to lower offensive and defensive

quality, respectively. This may be for a number of factors, possibly related to the
shot previous or following the rebound. For example, if most offensive rebounds
come from long shots with rebounds being further the basket, follow-up shots
may not be as optimal as expected. It also may be simply related to the current
model specification, where ORP and DORP loadings are constrained to unity.
However, when this constraint is removed, and the unity constraint is placed on
the latent variance (UVI, see Kline 2005), the same pattern emerges.

Table 3: SEM Regressing Winning Percentage on Offensive and Defensive

Quality and Salary

Panel A: Measurement Model

Loading Standardized Error Factor
(S.E.) Parameter Variance Variance/
Disturbance
Offensive Quality 0.753
ORP 1.00 -0.316 6.798
(0.00)
FTR 0.406* 0.137 6.454
(0.169)
TPP -0.788* -0.504 1.375
(0.145)
EFG 1.730* 0.715 2.157
(0.292)
FTR with TPP 1.274

Defensive Quality 0.979

DORP 1.00 0.412 4.783
(0.00)
SPP 0.940* 0.943 0.108
(0.116)
DTPP 1.146* 0.824 0.608
(0.138)
DEFG -0.286* -0.143 3.812
(0.115)
DEFG with DORP 1.152

* indicates p<.05. Bold indicates latent variables.

EFG = Effective Field Goal %, FTR = Free Throw Rate, TPP = Turnovers Per Possession, ORP= Offensive Rebound
Percent, D indicates Defensive aspects of measures, SPP= Steals Per Possession

[Link] 8
 Baghal: Are the “Four Factors” Indicators of One Factor?

Panel B. Structural Model

Coefficient Standardized R2
(S.E.) Parameter
Winning Percentage 0.820
ON
Offensive Quality 17.694* 0.956
(3.127)
Defensive Quality 6.782* 0.418
(1.188)
Salary 0.473 0.067
(0.326)
Off. Quality with 0.457*
Salary (0.141)
Def. Quality with -0.094
Salary (0.117)
* indicates p<.05. Bold indicates latent variables.

Still, the other indicators follow relationships with the underlying latent
variable as expected. For example, a one unit increase in offensive quality leads to
an expected increase in 1.73% in effective field goal percent (EFG), and a one
unit increase in defensive quality leads to an expected increase leads to an
expected increase of 1.15 other team turnovers per one hundred possessions.
Increases in offensive quality also leads to increases in own team free throw rate,
with decreases in own team turnover per possession. Defensive quality increases
also leads to increases in own team’s steals per possession and decreases in other
teams effective field goal percentage.
In regards to the structural portion of the model, there are clear effects of
offensive and defensive quality latent factors on winning percentage. Of
importance is that the amount of variance explained in winning percent
(R2=0.820) is approximately the same as the OLS regression discussed above
(R2=0.859 for the original model, R2=0.844 for the model replacing DFTR with
SPP). This suggests that, at the least, the explanatory power of the model is not
decreased significantly by using SEM. Further, both offensive and defensive
quality are significant effects, leading to large gains in winning percentage.
However, offensive quality leads to greater increases in winning percentage than
offensive quality, as seen in both the unstandardized and standardized
coefficients. This is consistent with the OLS regressions above and those in
Teramoto and Cross, which show the possible relative importance of offensive
traits in the regular season.

9
 Submission to Journal of Quantitative Analysis in Sports

However, unlike those regressions, the SEM show how each of the Four
Factors come together to have an overall effect as opposed to examining the
constituent effects. The current analysis suggests that a one unit increase in
offensive quality will lead to a 17.69% increase in winning percent, and a one unit
increase in defensive quality leads to an expected 6.78% increase in winning
percent. This also suggests that increases in a unit in the offensive and defensive
quality are not easily accomplished. Although offense quality may be relatively
more important than defensive quality, at least in the regular season, it is clear that
defense plays a large role winning games.

Figure 2. SEM of Winning Percent on Offensive and Defensive Quality

Latent Variables, with Salary a Prior Causal Effect

e1 ORP 1

e2 FTR
Off.
Quality
e3 TPP

e4 EFG

SALARY WIN% e9
e5 DORP 1

e6 SPP
Def.
e7 DTPP Quality

e8 DEFG

EFG = Effective Field Goal %, FTR = Free Throw Rate, TPP = Turnovers Per Possession, ORP= Offensive Rebound
Percent, D indicates Defensive aspects of measures, SPP= Steals Per Possession

An interesting pattern of relationships emerge between salary and other

variables, as indicated in Table 3. First, salary is not significantly correlated with
defensive quality (r=-0.094, p=0.43), but is correlated with offensive quality
(r=0.457 p<0.01). This suggests that money is spent on offensive quality, but not
necessarily on defensive quality. Additionally, the effect of salary on winning
percentage is also found to be non-significant at standard levels (although the
coefficient is in the expected direction, with more salary leading to increases in
expected win percent). It is possible this is due to the relationship with offensive
quality, which also has a larger impact on winning percentage. That is, salary has
an indirect effect; salary is directly related to offensive quality (and possibly

[Link] 10
 Baghal: Are the “Four Factors” Indicators of One Factor?

defensive quality), which in turn helps predict winning (i.e. the effect of salary on
winning is mediated through its effects on team quality measures). A benefit of
SEM is that such a relationship can be modeled. This model is presented in Figure
2, and the estimated coefficients in Table 4. Since the measurement portion of the
model remains unchanged, only the structural portion of the model is reported in
Table 4.

Table 4: SEM Regressing Winning Percentage on Offensive and Defensive

Quality with Salary as Prior Cause

Structural Model

Coefficient Standardized R2
(S.E.) Parameter
Offensive Quality 0.077
ON
Salary 0.126* 0.278
(0.024)
Defensive Quality 0.001
ON
Salary -0.012 -0.030
(0.021)
Winning Percentage 0.790
ON
Offensive Quality 13.952* 0.856
(0.759)
Defensive Quality 4.498* 0.248
(0.811)
* indicates p<.05. Bold indicates latent variables.

As with the previous model, the relationship between salary and defensive
quality is not significant. However, there is a clear relationship between salary
and offensive quality, with increases in salary leading to increases in the offensive
quality. Evidently, when teams spend on players, they do so in regards to
offensive characteristics rather than defensive. This may be an effective spending
strategy, as can also be seen from Table 4. Although the effects of offensive and
defensive quality are diminished when salary is used as causally prior compared
to the original model, the effects are still strong and similar in nature. Increasing
offensive quality leads to greater gains than does defensive quality. Since salary is
more clearly linked to offensive quality, spending on offense indirectly leads to a

11
 Submission to Journal of Quantitative Analysis in Sports

greater winning percentage. However, although the model does indicate a possible
scenario for predicting winning percentage, it is important to note that the
explained variance of winning percentage (R2) did decrease slightly. This decline
was minor, however, and the overall R2 was still in the range of other models
(0.790).

4 Discussion

Generally, this research shows the potential use of structural equation models on
NBA data, and more generally, in sports analytics. SEM appears to be a useful
tool to consider, especially given the need for more multivariate analyses and in
situations when many indicators of an underlying dimension are present.
Although past research was expanded and replicated, it examined each offensive
and defensive measures separately, rather than as a whole. The SEM models
presented here allow for estimation of the relation between the latent
characteristics of interest (quality) and to outcomes of interest, i.e. winning.
Further, such models may be useful in analysis of other sports statistics, especially
those where team data uses different metrics for different positions on offense and
defense (e.g. football) may be especially open to the use of SEM. For example,
rushing yards per attempt, pass yards per attempt, and yards after the catch are in
many ways distinct but may all be driven by an underlying quality of the team as
a whole, including coaching strategy and player skills.
Some limitations and areas of further research should be pointed out. First,
although SEM appears to be a useful tool to be used in sports statistical analyses,
potentially across a number of sports, the models examined in this paper may be
improved. Neither model achieved overall model fit statistics generally used as
standards is structural equation modeling. However, the explained variance (i.e.
R2) for the variable of interest, winning percent, was consistently high and at
levels found in other analyses examining the same outcome. Next, the reasons
why some measures did not fit well into the latent model, such as defensive free
throw rate, or why offensive rebounding percentage and the defensive counterpart
loaded in unexpected ways should be examined further. Some reasons were
posited as to why in this paper, but further studies should examine these using
some empirical methods.
Additional research should also focus how salary affects winning and the
best way to conceptualize this effect. This can include statistical issues or those on
how to use results from statistical models in decision-making. On the statistical
side, it may be that salary determines and is determined by offensive quality (i.e. a
circular pattern). For example, it may be that more spending brings in better
players, who then based on greater incentive to perform, perform at a higher level,
potentially driving up contract demands. This information could be used by teams

[Link] 12
 Baghal: Are the “Four Factors” Indicators of One Factor?

to better develop strategies for what types of players to spend money on.
Similarly, more research is needed on the player level in addition to team level
data. Structural equation models, as well as other statistical models such as
multilevel models, can examine player level impacts on outcomes of interest such
as offensive quality or winning. However, individual level data leads to the
potential problem of nonindependence of observations. Specifically, it is difficult
to ascertain how much of the player performance is due to teammates or players
defending. The adage “he makes his teammates better” is a potential statistical
problem, one that should be explored.

5 Conclusions

The current research produces several important findings. First, based on the SEM
models, it is found that the Four Factors are indicators of more general team traits,
labeled offensive and defensive quality. This suggests that the quality of the team
can be measured and drives other aspects of the team, such as team efficiency and
winning percentage. Interestingly, defensive free throw rate did not load on the
defensive quality latent variable. It is suggested here this may be due to the fact
that free throws made by the other team is only affected by the defense through
the number of free throws allowed. Once the foul is committed, whether the free
throw is made or not is out of the hands of the defensive team. As such, it may not
be as good of an indictor of defensive quality. Once this measure was replaced by
steals per possession, a statistic more in control of the defense, then a latent
variable for defensive quality was estimable. Additionally, offensive rebounding
percentage and its defensive counterpart both seemingly are affected by the
quality latent variables in ways opposite than expected. The reasons for this are
less clear, potentially related to the shot prior to or following the offensive
rebound.
Second, the research indicates the potential greater importance of offense
than may have been thought previously. Previous research and research using
techniques such as ordinary least squares suggest that although offense may be
more important in winning percentage, at least in the regular season, there is not
much difference between offensive and defensive statistics. The above findings
suggest that possible overall team offensive quality increases winning much more
than defensive quality does. This may explain why teams perceived to have high
powered offense and poor defense (e.g. Phoenix Suns) have had success in the
regular season.
Third, spending on team salary seems to be mainly related to offensive
quality, while directly unrelated to defensive quality and winning percentage.
However, spending on team salary seems to have a moderated effect on winning,
whereby spending leads to greater offensive quality, which in turn leads to greater

13
 Submission to Journal of Quantitative Analysis in Sports

winning percentage. Further, since offensive quality is relatively more important

to winning percentage than defensive quality, this may suggest that spending is
directed to the correct team characteristics. However, since spending is unrelated
to defensive quality, which does have a significant impact on winning, this
possibly suggests more spending should focus on defensive characteristics,
although not necessarily at the expense of offensive quality. Using the
relationships between salary, quality factors, and the observed indicators, teams
could also begin to examine potential characteristics to target in search of a higher
winning percentage. For example, offensive quality may be best increased by
improving effective field goal percentage to the team, but decreasing turnovers
does not follow far behind in regards to potential importance.

References

[Link]. 2011. [Link] accessed

2/28/2011
Bender, P. 2011. “Patricia’s Various NBA Basketball Stuff”
[Link] accessed 2/28/2011
Doolittle, B. and Pelton, K. 2010 Pro Basketball Prospectus 2010-2011, 2010.
Hu, L. and Bentler, P.M. 1999.“Cutoff criteria for fit indexes in covariance
structure analysis: Conventional criteria versus new alternatives.”
Structural Equation Modeling vol. 6, pp. 1-55
Kline, R.B. 2005. Principles and Practices of Structural Equation Modeling, 2nd
Edition, New York, NY
Kubatko, J., Oliver, D. Pelton, K. and Rosenbaum, D.T. 2007 “A Starting Point
for Analyzing Basketball Statistics” Journal of Quantitative Analysis in
Sports, vol. 3, iss. 3, Article 1.
Küpfer, E. 2005. “Team Similarity.” APBRMetrics forum,
[Link]
25dbe120f6c6f86e596d0d6dabbb8, accessed 2/28/2011
Muthén, L. K., & Muthén, B.O. (2007). Mplus user’s guide (5th Ed.). Los
Angeles, CA: Muthén & Muthén.
Teramoto, M. and Cross, C.L. 2010.“Relative Importance of Performance Factors
in Winning NBA Games in Regular Season versus Playoffs” Journal of
Quantitative Analysis in Sports, vol. 6, iss. 3, Article 2
White, H. 1980. “A heteroscedasticity-Consistent Covariance Matrix Estimator
and a Direct Test for Heteroscedasticity” Econometrica vol. 48, pp. 817-
838.

[Link] 14

View publication stats