Efficiency Measurement of the English Football Premier League with a Random Frontier Model
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Efficiency Measurement of the English Football Premier League
with a Random Frontier Model
Carlos Pestana Barros
ISEG – School of Economics and Management
Technical University of Lisbon
Rua Miguel Lupi, 20
1249-078 Lisbon, Portugal
Tf.: 351-1-213922801 / Fax: 351-1-213967971
cbarros@iseg.utl.pt
Pedro Garcia-del-Barrio 1
ESIrg-Economics, Sport and Intangibles research group and
Universitat Internacional de Catalunya
c/ Immaculada 22
08017 Barcelona, Spain
Tf.: 34 932541800 / Fax: 34 934187673
pgarcia@cir.uic.es
Abstract
Using the random stochastic frontier model, this paper examines the
technical efficiency of the English football Premier League from 1998/99 to
2003/04. The model disentangles homogenous and heterogeneous variables
in the cost function, which leads us to advise the implementation of
common policies as well as policies by clusters.
Key words: Football, efficiency, random frontier models, policy implications
JEL-Code: L83, C69
1
Correspondent author. The author gratefully acknowledges financial support from the Ministerio de
Ciencia y Tecnologia (SEJ2004-04649, Spain) and (SEJ 2007-67295/ECON).Introduction
This paper combines sport and financial data to analyse, with a random frontier model (Greene
2005, 2006), the technical efficiency of the English football Premier League. This model allows
for heterogeneity in the data and is considered the most promising state-of-the-art modelling
available to analyze cost functions (Greene, 2003, 2004, 2005). The advantage of this method
over alternative models is twofold. First, it allows for the error term to combine different
statistical distributions. Second, it uses random parameters; i.e., parameters that describe factors
not linked to observed features on the cost function. The estimation of the random frontier
model disentangles heterogeneous and homogeneous explanatory variables to determine which
of them must be treated in a homogeneous way and which managed by clusters.
The scope of this research is to account for the fact that English clubs can be identified
as heterogeneous, given that various clusters exist in the league. We estimate a random frontier
model for clubs that played uninterruptedly in the Premier League between seasons 1998/99 and
2003/04. This ensures a balanced panel, which is a pre-requisite to obtain similar average scores
in the period at club level. As the analysis combines sport and financial variables, it permits
verifying if pitch success entails financial success (Cf. Szymanski and Kuypers, 2000, p. 22).
In the following section, we analyse institutional settings. Section 3 examines the
literature on sport efficiency, while Section 4 explains the theoretical framework. Then, Section
5 presents the data and the main results. Finally, Section 6 discusses the efficiency ranking-list
and concludes with a number of managerial implications.
2. Contextual Setting
The financial underpinning of the Premier League has created four sub-groups of clubs, in terms
of aspirations and likelihood of sport success. First, there is an elite group of three clubs that
dominate the league (Manchester United, Arsenal and more recently Chelsea, which has joined
this privileged status through the vast wealth of Abramovich). Second, there are four or five
aspiring teams that struggle to qualify for the remaining places in European competitions. Then,
we find nine or ten middle-table teams, whose main goal is avoiding relegation. Finally, there is
a group of teams (in which the newly-promoted clubs are usually present) that are engaged in a
fight to retain category. It is not unusual for teams to be relegated after one year, and even to
sink without trace, owing to the financial adjustments that they are obliged to make afterwards.
This emphasizes the importance of middle-rank clubs attracting players commensurate to their
aspirations, in an attempt to prevent drifting into the relegation zone.
The new financial scenario is related to the increasing success of the UEFA Champions
League. The main clubs enjoy phenomenal revenues generated through broadcasting contracts
1and global brand sponsorships. Whilst the league winners get the biggest prizes, qualification to
participate in Europe is considered a financial victory in itself. Of the European leagues, the
Premier is the richest one, attaining revenues of approximately €1.79 billion in the 2002/03
2
season. Additionally, match-day income in England still represents an important portion of
revenues (around 30%), while this figure is substantially smaller (15-18%) in Italy, Germany
and France.
3. Review of the Literature
There are two main approaches to measure efficiency: the econometric or parametric approach
and the non-parametric or DEA approach. Unlike the econometric stochastic frontier approach,
DEA allows the use of multiple inputs and outputs, but does not impose any functional form on
the data; nor does it make distributional assumptions for the error term. Both methods assume
that the production function of the fully efficient decision unit is known. In reality this is not the
case and the efficient iso-quant has to be estimated from the sample. Under these conditions, the
frontier is relative to the sample considered in the analysis. The stochastic frontier approach has
been applied to various contexts like production (Kumbhakar and Wang, 2005; or Lothgren,
1997). However, as Table 1 shows, the studies published so far applying this methodology to
sports are scarce, which enhances the interest of this paper.
Table 1
Note that nine papers in total have used DEA, three papers adopted a deterministic econometric
frontier and two papers use the stochastic econometric frontier. In our view, this is not
sufficient for such an important issue in the context of sports management, thereby deserving
further research.
4. Theoretical Framework
The approach that we adopt here is the stochastic cost econometric frontier. In its origins, the
random frontier model was proposed by Farrell (1957), and came to prominence with
contributions from Aigner, Lovell and Schmidt (1977), Battese and Corra (1977) and Meeusen
and Van den Broeck (1977). The frontier is estimated econometrically and measures the
difference between the inefficient units and the frontier by the residuals, which are assumed to
have two components: noise and inefficiency. The general frontier cost function proposed is of
the form:
v +u
Cit = C ( X ) ⋅ e it it ; ∀ i = 1,2, … N ; ∀ t = 1,2, …T (1)
it
2
Deloitte & Touche Annual Review of Football Finance 2004. As a reference, the Italian Calcio Serie A
obtained €1,162. More recent information is available but these records are sufficiently illustrative.
2Where Cit represents a scalar cost of the decision-unit i under analysis in the t-th period; Xit is a
vector of variables including input prices and output descriptors present in the cost function.
The error term vit is assumed to be i.i.d. and represents the effect of random shocks (noise). It is
independent of uit, which represents technical inefficiencies and is assumed to be positive and to
follow a N(0, σu2 ) distribution. The positive disturbance uit is reflected in a half-normal
independent distribution truncated at zero, signifying that the cost of each club must lie on or
above its cost frontier, implying that deviations from the frontier are caused by factors
controlled by the club.
The total variance is defined as σ2 = σv2 + σu2. The contribution of the different elements
to the total variation is given by: σv2 = σ2 / (1+ λ2) and σu2 = σ2 λ2 / (1+ λ2); where λ = σu / σv ,
which provides an indication of the relative contribution of u and v to ε = u + v. Because
estimation of equation (1) yields merely the residual ε, rather than u, the latter must be
calculated indirectly (Greene, 2003). For panel data analysis, Battese and Coelli (1988)
employed the expectation of uit conditioned on the realized value of εit = uit + vit, as an estimator
of uit. In other words, E[uit / εit] is the mean productive inefficiency for club i at time t. But the
inefficiency can also be due to clubs heterogeneity, which implies the use of a random effects
model:
cit = ( β 0 + wi ) + ' x it + vit + uit (2)
where the variables are in logs and wi is a time-invariant specific random term that captures
individual heterogeneity. A second issue concerns the stochastic specification of the inefficiency
term u, for which the half normal distribution is assumed. For the likelihood function we follow
the approach proposed by Greene (2005), where the conditional density of cit given wi is:
2 ε it λε it (3)
f (cit | wi ) = φ Φ , ε it = cit − ( β 0 + wi ) − ' x it
σ σ σ
Where φ is the standard normal distribution and Φ is the cumulative distribution function.
Conditioned on wi , the T observations for club i are independent and their joint density is:
T
2 ε it λε it
f (ci1 ,..., ciT | wi ) = ∏ φ Φ (4)
t =1 σ σ σ
The unconditional joint density is obtained by integrating the heterogeneity out of the density,
T
2 ε it λε it T
2 ε it λε it (5)
Li = f ( ci1 ,..., ciT ) = ∏σ φ σ
Φ
σ
g ( wi )dwi = E w ii ∏σ φ σ
Φ
σ
wi t =1 t =1
The log likelihood is then maximized with respect to β0, β, σ, λ and any other parameter
appearing in the distribution of wi. Even if the integral in expression (5) will be intractable, the
right hand side of (5) leads us to propose computing the log likelihood by simulation. Averaging
3the expectation over a sufficient number of draws from the distribution of wi will produce a
sufficiently accurate estimate of the integral shown in (5) to allow estimation of the parameters
(see Gourieroux and Monfort, 1996 and Train, 2003). The simulated log likelihood is then given
by the expression:
N
1 R T
2 ε it | wir λε it | wir
log Ls ( β 0 , , λ , σ , θ ) =
i =1
log
R r =1
∏σ φ
t =1 σ
Φ
σ
(6)
where θ includes the parameters of the distribution of wi and wir is the r-th draw for observation
i. Based on our panel data, Table 4 presents the maximum likelihood estimators of model (1) as
found in recent studies carried out by other authors (Greene, 2004 and 2005).
5. Data and Results
To estimate the cost frontier, we use a balanced panel. The sample comprises the twelve clubs
that were uninterruptedly competing at the Premier League from 1998/99 to 2003/04. Frontier
models require identifying inputs (resources) and outputs (transformation of resources). This is
accomplished using the usual criteria: availability of data, findings of previous studies, and
opinions of professionals. The variables have been transformed as described in Table 2, where
monetary variables are expressed in £'000, deflated by GDP deflator and denoted at prices of
2000.
Table 2
We estimate the stochastic generalized Cobb-Douglas cost function, with three input prices and
three outputs (sales, points and attendance) and the Translog Frontier model. We adopt the log-
log specification to allow for non-linearity of the frontier. In order to capture the specificity of
the two types of capital (funds used and the premises) that clubs require for developing their
activity, we disentangle the analysis into capital-premises and capital-investment. Then, we
impose linear homogeneity in input prices, restricting the parameters in the estimated function:
LgCost it = β 0 + β1Trend + β 2 LgPLit + β 3 LgPK 1it + β 4 LgPK 2 it + β 5 LgSalesit + β 6 LgPoinit + β 7 LgAtt it + ( vit + u it ) (7)
where PL, PK1 and PK2 are respectively the prices of labor, capital-Premises, and capital-
investment. This is the cost frontier model, known in Coelli, Rao and Battese (1998) as the
Error Components Model, as it accounts for causes of efficiency controlled by management.
Next, Table 3 presents the results obtained for the stochastic frontier, using GAUSS and
assuming a half-normal distribution specification for the costs function frontier.
Table 3
Regularity conditions require for the cost function to be linearly homogeneous, non-decreasing
and concave in input prices (Cornes, 1992). Attending to the number of observations and
4exogenous variables, we use the Cobb-Douglas model with a half-normal distribution, a choice
that is supported by the data. Then, the error components model (Coelli et al., 1998) is adopted.
Having estimated two rival models, homogeneous and heterogeneous Cobb-Douglas frontier
models, we follow the Likelihood test to select the most adequate functional form. The test
compares models with different numbers of parameters by means of the Chi-square distribution,
indicating that the heterogeneous frontier is preferred to the standard model. We also computed
the Chi-square statistic for the general model specification. It also advocates the heterogeneous
frontier, thereby supporting the relevance of adding the variables.
Finally, in order to differentiate between the frontier model and the cost function, we
consider the sigma square and the lambda variables of the cost frontier model. They are
statistically significant, meaning that the traditional cost function is unable to capture adequately
all the dimensions of the data. Furthermore, the random cost function fits the data well, since
both the R2 and the overall F-statistic (of the initial OLS used to obtain the starting values for
the maximum-likelihood estimation) are higher than the standard cost function. The value of the
parameter lambda is positive and statistically significant in the stochastic inefficiency effects
and the coefficients of the variables have the expected signs. Cost increases alongside the trend,
which indicates that there were no technological improvements during the period to drive costs
down. Moreover, costs significantly increase with the price of labour, the price of capital-
premises and attendance. It also rises with the price of capital-investment and sales; a
relationship that is statistically significant only for the random frontier model. The significant
random parameters vary along the sample. The identification of the mean values of random
parameters implies taking into account heterogeneity when implementing policies for cost
control.
6. Conclusion
Common policies can be defined for English clubs based on the average values of the
homogeneous variables; whereas individual policies by clusters may be prescribed to account
for heterogeneous variables. The model does not specify how many clusters exist in the sample,
an issue which has to be established by non-empirical means, but it identifies their
heterogeneous nature. Given that the scale parameters of the heterogeneous variables are
statistically significant, we recognize their heterogeneous nature, which entails managerial
insights and policy implications.
5References
Aigner, D.J.; C.A.K. Lovell, and P. Schmidt, 1977, Formulation and estimation of stochastic frontier
production function models, Journal of Econometrics 6, 21-37.
Audas, R.; S. Dobson and J. Goddard 1999, Organizational Performance and Managerial Turnover,
Managerial and Decision Economics 20, 305-318.
Barros, C.P., 2003, Incentive Regulation and Efficiency in Sports Organisational Training Activities, Sport
Management Review 6 (1), 33-52.
Barros, C.P. and A. Santos, 2003, Productivity in Sports Organisational Training Activities: A DEA Study,
European Journal of Sport Management Quarterly 1, 46-65.
Barros, C.P. and A. Santos, 2005, Les relations entre performance sportive et la performance financiere:
application au cas du football Portugais, in: Bolle, G. and Desbordes M., eds., Marketing et Football:
une perspective internationale, Presses Universitaires du Sport, (Voiron, France) 347-374.
Barros, C.P. and S. Leach, 2006a, Performance evaluation of the English Premier League with data
Envelopment analysis, Applied Economics 38 (12), 1449-1458.
Barros, C.P. and S. Leach, 2006b, Analyzing the Performance of the English Football League with an
Econometric Frontier Model, Journal of Sport Economics (forthcoming).
Barros, C.P. and S. Leach, 2006c, Technical Efficiency in the English Football Association Premier
League, Applied Economic Letters (forthcoming).
Battese, G.E. and G.S. Corra, 1977, Estimation of a production frontier model with application to the
pastoral zone of eastern Australia, Australian Journal of Agriculture Economics 21, 169-179.
Battese, G.E. and T.J. Coelli, 1988, Prediction of firm-level technical efficiencies with a generalised
frontier production function and panel data, Journal of Econometrics 38, 387-399.
Coelli, T.J.; P. Rao and G.E. Battese, 1998, An Introduction to Efficiency and Productivity Analysis,
Kluwer Academic Press.
Cornes R., 1992, Duality and modern economics. (Cambridge University Press, Cambridge, UK).
Dawson, P.; S. Dobson and B. Gerrard, 2000, Stochastic Frontier and the Temporal Structure of Managerial
Efficiency in English Soccer, Journal of Sports Economics 1, 341-362.
Farsi, M; M. Filippini and M. Kuenzie, 2005, Unobserved heterogeneity in stochastic cost frontier models:
an application to Swiss nursing homes, Applied Economics 37 (8), 2127-2141.
Farrell, M.J., 1957, The Measurement of Productive Efficiency, Journal of the Royal Statistical Society,
Series A, 120 (3), 253-290.
Fizel, J.L. and M.P. D’Itri, 1996, Estimating Managerial Efficiency: The case of college basketball coaches,
Journal of Sport Management 10, 435-445.
Fizel, J.L. and M.P. D’Itri, 1997, Managerial efficiency, managerial succession and organizational
performance, Managerial and Decision Economics 18, 295-308.
Gourieroux, C. and A. Monfort, 1996, Simulation Based Methods: Econometric Methods. (Oxford
University Press, Oxford, UK).
Greene, W., 2003, Econometric Analysis, 5th ed. (Prentice Hall, Upper Saddle River, NJ).
Greene, W., 2004, Distinguishing between heterogeneity and efficiency: stochastic frontier analysis of the
World Health Organisation’s panel on national health care systems, Health Economics 13, 959-980.
6Greene, W., 2005, Fixed and random effects in stochastic frontier models, Journal of Productivity Analysis
23, 7-32.
Haas, D.J., 2003A, Technical Efficiency in the Major League Soccer. Journal of Sport Economics 4 (3),
203- 215
Haas, D.J., 2003B, Productive Efficiency of English Football Teams – A Data Envelopment Approach.
Managerial and Decision Economics 24, 403-410.
Hadley, L.; M. Poitras; J. Ruggiero and S. Knowles, 2000, Performance evaluation of National Football
League teams, Managerial and Decision Economics 21, 63-70.
Hoefler, R.A. and J.E. Payne, 1997, Measuring efficiency in the National Basketball Association,
Economics Letters 55, 293-299.
Hoefler, R.A. and J.E. Payne, 2006, Efficiency in the National Basketball Association: A Stochastic
Frontier Approach with Panel Data, Managerial and Decision Economics 27 (4), 279-85.
Kumbhakar, Subal C. and Hung-Jen Wang, 2005, Estimation of Growth Convergence Using a Stochastic
Production Frontier Approach, Economics Letters 88 (3), 300-305.
Kumbhakar, S.; E. Tsionas; B. Park and L. Simar, L., 2006, Nonparametric stochastic frontiers: A local
maximum likelihood approach, Journal of Econometrics (forthcoming).
Lothgren, M., 1997, Generalized Stochastic Frontier Production Models, Economics Letters 57 (3), 255-59.
Meeusen, W. and J. Van den Broeck, 1977, Efficiency estimation from a Cobb-Douglas production
function with composed error, International Economic Review 18, 435-444.
Porter, P. and G.W. Scully, 1982, Measuring managerial efficiency: The case of baseball. Southern
Economic Journal 48, 642-650.
Scully, G.W., 1994, Managerial efficiency and survivability in professional team sports. Managerial and
Decision Economics 15, 403-411.
Szymanski, S. and T. Kuypers, 2000, Winners and Losers: The Business Strategy of Football. (Viking
Books, London, 1999; softcover Penguin Books, 2000).
Train, K., 2003, Discrete Choice Methods with Simulation (Cambridge University Press, Cambridge, UK).
Zak, T.A.; C.J. Huang and J.J. Siegfried, 1979, Production efficiency: The case of professional basketball,
Journal of Business 52, 379-392.
7Table 1. Survey of the Literature on Frontier models in Sports.
Papers Method Units Inputs Outputs Prices
Stochastic production NBA association Ratios of: field goal %, Actual number of wins
frontier model clubs, 2001-2002 free throw %, offensive
Hoefler and Payne
and defensive rebounds,
(2006)
assists, steals, turnover,
blocked shots difference
Technical efficiency Soccer clubs in the Operational cost Points, attendance, Price of labour,
effects model English Premier turnover. price of capital-
Barros and Leach
League Contextual factors: players, price
(2006c)
population, income, of capital-
European premises
Stochastic frontier model Soccer clubs in the Operational cost points, attendance Price of:
Barros and Leach
English Premier labour, capital,
(2006b)
League and stadiums,
DEA-CCR and BCC Soccer clubs in the Players, wages, net Points, attendance and
Barros and Leach
model English Premier assets and stadium turnover
(2006a)
League facilities
DEA-CCR Model and Soccer clubs in the Supplies & services Match, membership,
DEA-BCC model Portuguese First expenditure, wage TV and sponsorship
Barros and Santos Division expenditure, receipts, gains on
(2005) amortization players sold, financial
expenditure, other costs. receipts, points won,
tickets sold
DEA-CCR and DEA- 12 US soccer clubs Players wages, coaches Points awarded,
Haas (2003A) BCC model observed in year wages, stadium number of spectators
2000 utilization rate and total revenue
DEA-CCR and DEA- 20 Premier League Total wages, coach Points, spectators and
Haas(2003B) BCC model clubs observed in salary, home town revenue
year 2000/2001 population
DEA-Malmquist index 18 training activities Number of Trainers, Number of
of sports federations, trainers reward, number participants, number
Barros and Santos
1999-2001 of administrators, of courses, number of
(2003)
administrators reward approvals
and physical capital
DEA-Allocative model 19 training activities Number of Trainers, Number of Price of:
of sports federations, number of participants, number trainers,
Barros (2003)
1998-2001 administrators, physical of courses, number of administrators,
capital approvals and capital
DEA-CCR model in first 147 College Player talent, opponent Winning percentages
Fizel and D’Itri
stage and regression basketball teams, strength,
(1997)
analysis in second stage 1984-1991
Fizel and D’Itri DEA-CCR model Baseball managers Player talent, opponent Winning percentages
(1996) strength,
A linear program-ming Major League Team hitting and team Team percent wins
Porter and Scully
technique (possibly DEA- baseball teams, pitching
(1982)
CCR) 1961-1980
Stochastic Cobb-Douglas Sample of English Player age, career Winning percentages
frontier model football managers, league experience,
1992-1998 career goals, num. of
Dawson, Dobson
previous teams, league
and Gerrard (2000)
appearances in previous
season, goals scored,
player divisional status
Hadley,Poitras, Deterministic frontier US NFL teams, 24 independent Team wins
Ruggiero and model 1969/70-1992/93 variables describing
Knowles (2000) attack and defence.
Hazard functions English prof. soccer, Match result, league Duration (measured
Audas, Dobson and 1972/73-1996/97, position, manager age, by the number of
Goddard (1999) match-level data manager experience, league matches
player experience played)
Stochastic production 27 NBA teams, Ratios of: field goal %, Actual number of wins
frontier 1992-1993 free throw %, turnover,
Hoefler and Payne offensive rebounds,
(1997) defensive rebounds,
assists, steals, difference
in blocked shots
Deterministic and 41 Basketball Team hitting and team Win percent
Scully (1994) stochastic Cobb-Douglas coaches, 1949/50 to pitching
frontier model 1989/90
Cobb-Douglas NBA teams 10 variables of pitch Ratio of final scores
Zak, Huang and
deterministic frontier performance: ratio of
Siegfried (1979)
model steals, ratio of assists…
8Table 2. Descriptive Statistics of the Data
Standard
Variable Description Minimum Maximum Mean
Deviation
Logarithm of operational cost in Euros at constant price
LgCost 6.6685 8.9475 7.4633 0.4104
2000=100
Logarithm of price of workers, measured by dividing total
LgPL 4.61378 6.8152 5.7316 0.3782
wages between the number of workers
LgPK1- Logarithm of price of capital-premises, measured by the
0.00453 0.3959 0.0689 0.0486
premises amortizations divided by the value of the total assets
LgPK2- Logarithm of price capital-investment, measured by the cost
3.07E-06 2.1188 0.2438 0.3603
investment of long term investment divided by the long term debt
Logarithm of the sales of each club in pound at constant
LgSales 5.6367 8.3703 7.2507 0.4537
price 2000=100
Logarithm of the number of points obtained by each club in
LgPoin 1.4313 1.9542 1.7216 0.0988
the league
LgAtt Logarithm of the number of attendees
3.9469 4.9410 4.4003 0.2302
9Table 3. Stochastic Cobb-Douglas panel cost frontier (Dependent Variable: Log Cost)
Variables Random Frontier model Non Random Frontier Model
Non-random parameters Coefficient (t-ratio) Coefficient (t-ratio)
Constant (β0) 1.0380 (5.480) 1.1940 (1.442)
Trend (β1) 0.0269 (5.709) 0.0270 (2.680)
LgPL (β2) 0.6993 (19.61) 0.6809 (5.232)
LgPK1 (β3) 0.5401 (5.141) 0.5513 (2.248)
LgPK2(β4) − 0.0490 (0.409)
LgSales(β5) 0.0540 (2.018) 0.0521 (0.461)
LgPoin (β6) − 0.2350 (0.793)
LgAtt (β7) − 0.2881 (1.694)
Mean for Random Parameters
LgPK2 (β4) 0.6022 (3.957) −
LgPoin (β6) 0.1975 (2.219) −
LgAtt (β7) 0.3388 (6.256) −
Scale Parameters for Distributions of Random Parameter
LgPK2 (β4) 1.4281 (10.41) −
LgPoin (β6) 0.0202 (4.459) −
LgAtt (β7) 0.0115 (6.453) −
Statistics of the model
σ = [σ V2 + σ U2 ]
1/ 2 0.1362 (29.03) 0.1225 (1.079)
λ = σU /σV 0.2532 (2.706) 0.8094 (2.132)
Log likelihood 75.169 72.010
Chi Square 144.338 132.214
Degrees of freedom 3 3
Probability 0.000 0.000
Observations 72 72
t-statistics in parentheses (* indicates that the parameter is significant at 1% level).
10You can also read