Professional Documents
Culture Documents
Some history.
Much of the work on stochastic frontiers began in
70s. Major contributions from Aigner, Schmidt,
Lovell, Battese and Coelli and Kumbhakar.
Ordinary least squares (OLS) regression
production functions:
fit a function through the centre of the data
assumes all firms are efficient
Production Function
It is a relationship between output and a set of
input quantities.
We use this when we have a single output
In case of multiple outputs:
people often use revenue (adjusted for price differences) as
an output measure
It is possible to use multi-output distance functions to study
production technology.
q f ( x1 , x2 ,..., xN ) v
4
1 N N
ln q 0 n ln xn nm ln xn ln xm u
2 n 1 m1
n 1
Translog function is very commonly used it is a generalisation of the
Cobb-Douglas function
It is a flexible functional form providing a second order approximation
Cobb-Douglas and Translog functions are linear in parameters and can
be estimated using least squares methods.
It is possible to impose restrictions on the parameters (homogeneity
5
conditions)
Disadvantages:
simplistic - assumes all firms have same
production elasticities and that subsitution
elasticities equal 1
Disadvantages:
more difficult to interpret
requires estimation of many parameters:
K+3+K(K+1)/2
can suffer from curvature violations
Functional forms
Cobb-Douglas:
11
exp( ui )
exp( xi )
exp( xi )
Min ui ln yi xi
i
subject to ui 0.
12
ui
i
ln yi xi
subject to ui 2 0.
13
14
Production functions/frontiers
Deterministic
SFA
OLS
15
Production functions/frontiers
OLS:
qi = 0 + 1xi + vi
Deterministic :
qi = 0 + 1xi - ui
SFA:
qi = 0 + 1xi + vi - ui
where
vi = noise error term - symmetric
(eg. normal distribution)
ui = inefficiency error term - non-negative
(eg. half-normal distribution)
16
q f ( x1 , x2 ,..., xN ) v u
ln qi 0 1 ln xi vi ui
qi exp(0 1 ln xi vi ui )
17
Stochastic frontiers
deterministic frontier
qi = exp(0 + 1 ln xi)
yi
noise effect
noise effect
inefficiency effect
qB ? exp(0 + 1ln xB + vB uB)
inefficiency
effect
xA
xB
18
inefficiency
deterministic
component
In general, we write the stochastic frontier model with several inputs and
a general functional form (which is linear in parameters) as
ln qi xi vi ui
This means that if a firm is inefficient, then it produces less than what is
expected from the inputs used by the firm at the given technology.
We can define technical efficiency as the ratio of observed or realised
output to the stochastic frontier output
qi
exp(xi vi ui )
TEi
exp(ui )
exp(xi vi )
exp(xi vi )
19
Stochastic Specifications
The SF model is specified as:
ln qi xi vi ui
ui
iidN (0, u2 ).
Exponential
Gamma distribution
20
var = 1
var= 4
var = 9
-0.5
0.0
0.5
1.0
Distribution of u: ui
1.5
2.0
2.5
3.0
3.5
iidN (0, u2 ).
21
f(x)
mu = -1
mu = 0
mu = 1
0.5
mu = 2
0
0
u N ( i , u2 )
i 0 k Z ki
k
22
Estimation of SF Models
Parameters to be estimated in a standard SF model
2
2
are: , v and u
Likelihood methods are used in estimating the
unknown parameters. Coelli (1995)s Montecarlo
study shows that in large samples MLE is better than
COLS.
Usually variance parameters are reparametrized in
the following forms.
2 v2 u2
and
2 u2 v2 0.
2
2
2 and u / .
23
25
FRONTIER output
the final mle estimates are :
coefficient
beta
beta
beta
beta
beta
beta
0
1
2
3
4
5
standard-error
t-ratio
0.27436347E+00
0.15110945E-01
0.53138167E+00
0.23089543E+00
0.20327381E+00
-0.47586195E+00
0.39600416E-01 0.69282978E+01
0.67544802E-02 0.22371736E+01
0.79213877E-01 0.67081892E+01
0.74764329E-01 0.30883101E+01
0.44785423E-01 0.45388387E+01
0.20221150E+00 -0.23532883E+01
beta 6
0.60884085E+00
beta 7
0.61740289E-01
beta 8
-0.56447322E+00
beta 9
-0.13705357E+00
beta10
-0.72189747E-02
sigma-squared 0.22170997E+00
0.16599693E+00 0.36677839E+01
0.13839069E+00 0.44613038E+00
0.26523510E+00 -0.21281996E+01
0.14081595E+00 -0.97328160E+00
0.92425705E-01 -0.78105703E-01
0.24943636E-01 0.88884383E+01
gamma
0.88355629E+00 0.36275231E-01
mu is restricted to be zero
eta is restricted to be zero
log likelihood function = -0.74409920E+02
0.24357013E+02
26
SF Models - continued
Predicting Firm Level Efficiencies:
Once the SF model is estimated using MLE method, we
compute the following:
ui* (ln qi xi ) u2 / 2
and
*2 v2 u2 / 2 .
TE i E exp(ui ) qi *
*2
ui*
exp ui* .
*
2
SF Models - continued
Industry efficiency:
Industry efficiency can be computed as the average of
technical efficiencies of the firms in the sample
Industry efficiency can be seen as the expected value of a
randomly selected firm from the industry. Then we have
u2
TE E exp(ui ) 2 u exp .
2
28
FRONTIER output
technical efficiency estimates :
firm
eff.-est.
1
0.77532384
2
0.72892751
3
0.77332991
341
0.76900626
342
0.92610064
343
0.81931012
344
0.89042718
mean efficiency = 0.72941885
Tests of hypotheses
e.g., Is there significant technical inefficiency?
H0: =0 versus H1: >0
Test options:
t-test
t-ratio = (parameter estimate) / (standard error)
LR test safer
30
Distributional assumptions
the truncated normal distribution
N(,2) truncated at zero
More general patterns
Can test hypothesis that =0 using t-test
or LR test
The restriction =0 produces the halfnormal distribution: |N(0,2)|
33
Scale efficiency
For a Translog Production Function (Ray,
1998)
An output-orientated scale efficiency measure
is:
SEi = exp[(1-i)2/2]
where i is the scale elasticity of the i-th firm
and
K K
jk
j 1 k 1
uit 1 exp(bt ct ) ui
3.
uit exp (t T ui
37
0
1
10
11
12 13
K90 ( =
14
15
= -.02)
BC92 ( = .1)
over
time. These trends are also independent of any other data on the
firms. There is scope for further work in this area.
38
N ( zit , u2 )
39
Current research
There is scope to conduct a lot of research in this area.
Some areas where work is being done and still be done
are:
Modelling movements in inefficiency over time
incorporating exogenous factors driving inefficiency
effects
Possibility of covariance between the random disturbance
and the distribution of the inefficiency term
Endogeniety due to possible effects of inefficiency and
technical change on the choice of input variables
Modelling risk into efficiency estimation and
interpretation
40
Current research
We have seen how technical efficiency can be computed,
but it is difficult to compute standard errors.
Peter Schmidt and his colleagues have been working on a
number of related topics here.
Bootstrap estimators and confidence intervals for
efficiency levels in SF models with pantel data
Testing whether technical inefficiency depends on firm
characteristics
On the distribution of inefficiency effects under
different assumptions
Data:
912 Aged Care Facilities
30% response rate response bias
Data validation
Actual observations used: 787
Methods:
DEA and SFA Methods
Peeled DEA
43
Preferred model:
Output variables:
High care weighted bed days
Low care weighted bed days
Input variables:
Floor area in square meters
Labour costs
Other costs
44
General findings
Average technical efficiency was calculated to be 0.83. 17%
cost savings could be achieved if all ACFs operated on the
frontier.
Results consistent between SFA and DEA models
Found variation across ACFs in different states (lowest 0.79
in Victoria and 0.87 in NSW/ACT).
Privately run ACFs has higher mean TE of 0.89.
Average scale efficiency of 0.93 was found.
An average size of 30 to 60 beds was found to be near optimal
scale.
A second stage Tobit model was run to see the factors driving
inefficiency.
Potential cost savings was calculated to be $316 m for the
number of firms in the sample greater savings for the
industry!!
Economies of Scope were examined using new methodology
developed but no significant economies were found.
45