STOCHASTIC FRONTIER

ANALYSIS

1
 MOTIVATION

• Usual textbook presentations treat producers as successful

optimizers. They maximize production, minimize cost, and
maximize profits.

• Conventional econometric techniques build on this paradigm

to estimate production/cost/profit function parameters using
regression techniques where deviations of observed choices
from optimal ones are modeled as statistical noise.

• However though every producer may attempt to optimize, not

all of them may succeed in their efforts. For example, given
the same inputs, and the same technology, some will produce
more output than others, i.e., some producers will be more
efficient than others.

• Econometric estimation techniques should allow for the fact

that deviations of observed choices from optimal ones are
due to two factors:
failure to optimize i.e., inefficiency
due to random shocks

• Stochastic Frontier Analysis or SFA is one such technique to

model producer behavior.

2
 USEFULNESS OF SFA

• SFA produces efficiency estimates or efficiency scores of

individual producers. Thus one can identify those who need
intervention and corrective measures.

• Since efficiency scores vary across producers, they can be

related to producer characteristics like size, ownership,
location, etc. Thus one can identify source of inefficiency.

• SFA provides a powerful tool for examining effects of

intervention. For example, has efficiency of the banks
changed after deregulation? Has this change varied across
ownership groups?

3
STRUCTURE OF THIS PRESENTATION

• Part 1: Theory: Illustrate the basics of SFA mainly with

analysis of cost efficiency.

Concept of efficiency

Estimation

Identification of sources of inefficiency

• Part 2: Empirics:

How to use FRONTIER program to estimate different

types of efficiency models

An application of SFA to Indian Banking (if time

permits)

• References
1. Kumbhakar, S.C. and Lovell, C.A.K (2000), Stochastic
Frontier Analysis, Cambridge University Press, U.K.

2. Coelli, T.J.; Rao, D.S. Prasada, and Battese, G.E. (1998), An

Introduction to Efficiency and Productivity Analysis, Kluwer
Academic Publishers, Boston/Dordrecht/London.

4
 TECHNICAL EFFICIENCY

• Production Function: YM = ƒ(x; β) shows the maximum

output YM producible from a given vector of inputs (x).
Here β are the production function parameters.

ƒ(L,K; β) = 2L0.5K0.5

ƒ(9,16; β) = 2.90.5.160.5 = 24

• Actual output, Y could be less than maximum output. In

fact, any output equal to or less than YM can be produced.

Y ≤ ƒ(x; β) = YM

• Figure 1

• TE = Y/YM 0 ≤ TE ≤ 1

• Y = YM. TE = ƒ(x; β) . TE

• Characterization: Y = ƒ(x; β)exp(-u) u≥0 (1)

5
 STOCHASTIC FRONTIER

• In (1) the frontier is deterministic. All deviations from

maximum output are ascribed to inefficiency.

• However sometimes maximum output itself might be lower

(higher) due to exogenous shocks. The production frontier itself
may be shifting.

• Figure 2

• Y = ƒ(x; β).exp(v).exp(-u) v ≤ 0 and u ≥ 0

ƒ(x; β) deterministic kernel

exp(v) effect on output of exogenous shocks
exp(-u) inefficiency
ƒ(x; β).exp(v) stochastic frontier

• TE = Y/ ƒ(x; β).exp(v)
= ƒ(x; β).exp(v).exp(-u)/ ƒ(x; β).exp(v)
= exp(-u)

• Y = ƒ(x; β).exp(v – u) (v – u) composite error term

6
 COST EFFICIENCY OR ECONOMIC
EFFICIENCY

• Ability to produce observed output at minimum cost, given

input prices.

• A producer may be technically efficient, but yet cost inefficient

because he fails to choose correct input combination.

Allocative inefficiency

• Figure 3

• Of course, a producer may be both technically inefficient as

well as allocatively inefficient.

Cost (or economic) inefficiency = technical inefficiency +

allocative inefficiency

• Figure 4

• Theoretically this notion is well defined but empirically it is

involving to segregate these two sources of inefficiency.

• Similarly, we can define profit efficiency, but its decomposition

into technical, and allocative inefficiency is even more
challenging. Will, therefore, concentrate on cost-efficiency in
the remainder of the lecture.

7
 COST EFFICIENCY
•
Ei = c(yi, wi; β) exp {vi + ui} (1)

– c(yi, wi; β) is the deterministic kernel

– vi random noise, takes positive and neg-
ative values
– ui captures inefficiency, takes only posi-
tive values
– Note the positive sign before ui

• Under this formulation cost efficiency can be

calculated as
c(yi, wi; β) exp {vi}
CEi = = exp {−ui}
Ei
(2)
•
0 ≤ CEi ≤ 1

1
 Estimation
• First rewrite as
ln Ei = ln c(yi, wi; β) + ui + vi (3)

• Estimating equation (3) requires:

– specification of a functional form for the
deterministic kernel c(yi, wi; β),

– an assumption about the distribution of

the random variable vi, and

– an assumption about the distribution of

the random variable ui.
• Given a particular specification for the ran-
dom variables ui and vi, the Maximum Like-
lihood (ML) technique is used to estimate
the unknown parameters.

2
 SPECIFICATION
• Deterministic Kernel
– Cobb-Douglas (in log form)
– Translog (a flexible functional form)
• Random Variables vi and ui
– vi ∼ iidN (0, σv2)
– ui ∼ iidN +(0, σu2 )
– vi and ui are distributed independently
of each other, and of the regressors.
• Given these assumptions, the log-likelihood
function for the sample of size I
i λ 1 X
ln L = K − I ln σ + ln Φ( ) − 2 2i .
X

i σ 2σ i
(4)
where i = ui + vi, σ 2 = (σu2 + σv2), λ = σσuv
and Φ(.) is the standard normal cumulative
distribution function. We substitute ln Ei −
ln c(yi, wi; β) in place of i in the likelihood
function.
3
DERIVATION OF LIKELIHOOD FUNCTION
• The density function of vi is
1 vi2
f (vi) = √ exp(− 2 )
2πσv 2σv
• The density function of ui is
2 vi2
f (ui) = √ exp(− 2 )
2πσv 2σv
• Given the independence assumption, the joint
density function of ui and vi is the product
of their individual density function, and so,

2 u2i vi2
f (ui, vi) = exp(− 2 − 2 )
2πσuσv 2σu 2σv
• Since i = vi + ui, the joint density function
for ui and i is:
2 u2i (i − ui)2
f (ui, i) = exp(− 2 − 2
)
2πσuσv 2σu 2σv

4
• The marginal density function of i is then
obtained by integrating ui out of f (ui, i)
which yields

Z
∞
f (i) = 0 f (ui , i )dui
2 −iλ −2i
= √ [1 − Φ( )] exp( 2 )
2πσ σ 2σ
• The likelihood function of the sample is then,
by independence, the product of the density
functions of the individual observations.

i=I
L(sample) = f (i)
Y

i=1
• And then taking log of the likelihood func-
tion yields the log-likelihood equation.
• Battese and Corra (1977) re-parameterization
σu2
–γ= σu2 +σv2
→ bounded between 0 and 1
σu
–λ= σv → any non-negative value

5
• If we use the γ parameterization, the log-
likelihood function is given by:

1 X 2
ln L = K−I ln σ+ ln[1−Φ(zi)]− 2 i .
X

i 2σ i
(5)
where zi = σi 1−γ
γ
s

• The log likelihood function allows for testing

the appropriateness of the SFA
– γ → 0 when either σu2 → 0 or σv2 →
∞ → OLS cost function (i.e., average
response function) with no inefficiency.
– The test should be done using the One-
sided Generalized Likelihood-Ratio Test
to ensure correct size.

– γ → 1 when either σu2 → ∞ or σv2 → 0

→ deterministic frontier with no noise.

6
Calculating Producer Specific Efficiency
• ˆi = ln Ei − ln c(yi, wi; β̂) is a composite
ˆ vi
estimate of ui +
• But it contains information about ûi. If ˆi
is high then chances are ûi is high since ex-
pectation of vi is zero.
• The conditional distribution of ui given i
could be exploited to get estimates of pro-
ducer specific inefficiency. This was first
demonstrate by Jondrow, Lovell, Materov,
and Schmidt (1982) and since then this de-
composition is known as the JLMS tech-
nique. Either the mean or the mode of this
conditional distribution can be used.
•
φ( σiλ ) i λ
E(ui | i) = σ∗[ −i λ + ( )] (6)
1 − Φ( σ ) σ

where σ∗2 = σu2 σv2/σ 2 and we have used the

parameterization λ.
7
• Given a point estimate of E(ui | i) Pro-
ducer specific efficiency can be calculated as:
•
ˆ i = exp {−E(uˆi | i)}.
CE (7)
or using the point estimator
•
ˆ i = E(exp {−ui} | i)
CE
1 − Φ(σ∗ − µ∗i/σ∗)
= [ ]
1 − Φ(−µ∗i/σ∗)
1 2
· exp {−µ∗i + σ∗ }
2
where µ∗i = iσu2 /σ 2.
• The two formulas won’t give same results as
exponential in a non-linear function.

8
Alternative Distributional Assumptions
• Till now we have assumed that ui ∼ iidN +(0, σu2 )
• A more general formulation is that of Trun-
cated Normal Distribution ui ∼ iidN +(µ, σu2 )
• Alternative functional forms like exponen-
tial and gamma distribution could also be
applied.
• Does the distributional assumption matter?
Yes it does matter in the calculation of the
efficiency numbers.
• However, the ranking of the prodcuers are
much less sensitive to distributional assump-
tions.
• In Panel Data Models, one can estimate ef-
ficiency WITHOUT making ANY assump-
tion about the distribution of ui.

9
 Analyzing Efficiency Behaviour
• Two questions:
– What is the behavior of efficiencies over
time? Are they increasing, decreasing or
constant?
– What explains the variations in ineffi-
ciencies among producers and across time?
• Time behavior
– Following Kumbhakar (1990), Battese and
Coelli (1992) proposed a simple model
that can be used to estimate the time
behavior of inefficiencies.

uit = {exp[−η(t − T )]}ui, (8)

where the ui ∼ N +(µ, σ 2), and η is a
parameter to be estimated.
– inefficiencies in periods prior to T depend
on the parameter η.
– uiT = ui reference or benchmark
10
 – inefficiency prior to period T is the prod-
uct of the terminal year’s inefficiency and
exp {η(T − t)}.
If η is positive, inefficiencies fall over time
If η is negative, inefficiencies increase over
time.
– Efficiency behavior is monotonic
– Ordering of firms in terms of inefficiencies
time-invariant
– Good for understanding aggregative be-
havior
– Battese and Coelli “Model 1” in FRON-
TIER program
• Interesting Hypothses to Test
– Test η = 0 → Implies Time invariant
efficiency
– Test η = µ = 0 → Time invariant effi-
ciency with Half-Normal Distribution

11
 Explaining efficiency
• Certain factors influence the environment in
which production takes place
– degree of competitiveness
– input and output quality
– network characteristics
– ownership form
– changes in regulation,
– management characteristics
• Two ways to handle them
– Include them as variables in the produc-
tion process as control variables. Using
this interpretation, these variables influ-
ence the structure of the technology by
which conventional inputs are converted
into outputs, but not efficiency.

12
 E = c(w, y; γ) exp {v + u} (9)
The parameter γ now includes cost pa-
rameters as well as environmental param-
eters.
– Associate variation in estimated efficiency
with variation in the exogenous variables.
• Early papers adopted two stage approach
– Stage 1: In the first stage a stochastic
frontier equation was estimated (exclud-
ing the exogenous variables), typically by
MLE under the usual distributional and
independence assumptions, and the re-
gression residuals were decomposed using
the JLMS technique.
– Stage 2: The estimated inefficiencies were
regressed on exogenous variables to ex-
plain/locate the source of inefficiency.

13
• Econometrically inconsistent because the iid
assumption necessary to use the JLMS tech-
nique is contracticted in the second stage.
• Kumbhakar, Ghosh, and McGuckin (1991),
and Reifcheneider and Setvenson (1991) ap-
proach. All the parameters of the stochastic
frontier function as well as those of the inef-
ficiency function was estimated together in
a single MLE procedure.
• The model

ln Eit = lnC(wit, yit; β) + vit + uit (10)

uit = δ 0zit + it (11)
uit captures the effect of economic ineffi-
ciency, which has a systematic component
δ 0zit associated with the exogenous variables
and a random component it.

14
• The non-negativity requirement that uit =
(δ 0zit −it) ≥ 0 is modeled as it ∼ N (0, σ2)
with the distribution of it being bounded
below by the variable truncation point −δ 0zit.
• This is implemented as ”Model 2” in FRON-
TIER program.
• Interesting hypotheses to Test
– Test: γ = δ0 = δ1 = · · · = δm = 0 →
Implies no ineffciency
– Test: δ1 = · · · = δm = 0 → Implies the
Truncated Normal Distribution

15
FRONTIER 4.1 -- A PROGRAM FOR ESTIMATING
STOCHASTIC FRONTIER PRODUCTION AND COST
FUNCTION

1
 THE FRONTIER PROGRAM

• This was developed by Tim Coelli at the Centre for Efficiency

and Productivity Analysis (CEPA). CEPA was initially house
at the University of New England, but has recently moved at
the University of Queensland.

• Website: http://www.uq.edu.au/economics/cepa/software.htm

• Read the paragraph on FRONTIER 4.1. Click the link given

immediately below

• And download the file “front41-xp.zip”

• User’s Guide: Coelli, T.J. (1996), “A Guide to FRONTIER

Version 4.1: A Computer Program for Stochastic Frontier
Production and Cost Function Estimation”, CEPA Working
Paper 96/7, Department of Econometrics, University of New
England, Armidale NSW Australia.

• We are going to use the examples given in this guide to

understand how to use the FRONTIER program.

2
REQUIREMENTS OF THE PROGRAM

• The FRONTIER program requires you to prepare TWO files

for estimating stochastic frontier production and cost
functions.

• The Data File which should preferably have the a .dta

extension like bank.dta

• The Instruction File which should preferably have a .ins

extension bank.ins

3
 The Data File

• Data should be listed in an text file. Format of this file is very

important.

• The data must be listed by observation.

• There MUST be 3+k[+p] columns in the following order:

1. Firm number (an integer in the range 1 to N)

2. Period number (an integer in the range 1 to T)
3. Yit dependent variable
4. to 3 + k x1it … xkit k number of
regressors
3+k+1 to 3+k+p z1it … zpit p number of
optional regressors

4
• The program assumes a linear functional form. Thus for
estimating a Cobb-Douglas production function, all the input
and output quantities should be logged.

• The observations can be listed in any order but the columns

must be in the stated order.

• There must be at least one observation on each of the N firms

and there must be at least one observation in time period 1
and in time period T.

• If you are using a single cross-section of data, then column 2

(the time period column) should contain the value “1”
throughout. N

5
 EXECUTING THE PROGRAM

• Type “front41” to begin execution. The program will then ask if

the instructions will come from a terminal (interactive mode) or
from a file (batch mode).

• If the interactive (terminal) option is selected, questions will be

asked in the same order as they appear in the instruction file.

6
 THE INSTRUCTION FILE

• We now discuss the structure of the instruction file.

• As Coelli write in his Guide, “The best way to describe how to use
the program is to provide some examples.” So we discuss the
examples given in the users Guide. The data, program, output
used in this presentation are all taken directly from the Guide.

• The estimation of five SF models are discussed.

1. A Cobb-Douglas production frontier using cross-sectional data
and a assuming a half-normal distribution.
2. A Translog production frontier using cross-sectional data and
assuming a truncated normal distribution.
3. A Cobb-Douglas cost frontier using cross-sectional data and
assuming a half-normal distribution.
4. The Battese and Coelli (1992) specification (Model 1).
5. The Battese and Coelli (1995) specification (Model 2).

• For simplicity there are only two production inputs in all cases. In
the cross-sectional examples there are 60 firms, while in the panel
data examples, there are 15 firms and 4 time periods.

7
 EXAMPLE 1

• (4.1) A Cobb-Douglas production frontier using cross-sectional

data and assuming a half-normal distribution.

ln(Qi) = β0 + β1ln(Ki) + β2ln(Li) + (Vi - Ui),

• where Qi, Ki and Li are output, capital and labour, respectively, and
Vi and Ui are assumed normal and half-normal distributed,
respectively.

Table 1a - Listing of Data File EG1.DAT

__________________________________________________
1. 1. 12.778 9.416 35.134
2. 1. 24.285 4.643 77.297
3. 1. 20.855 5.095 89.799
.
.
.
58. 1. 21.358 9.329 87.124
59. 1. 27.124 7.834 60.340
60. 1. 14.105 5.621 44.218

8
__________________________________________________

Table 1b - Listing of Shazam Instruction File EG1.SHA

__________________________________________________
read(eg1.dat) n t y x1 x2
genr ly=log(y)
genr lx1=log(x1)
genr lx2=log(x2)
file 33 eg1.dta
write(33) n t ly lx1 lx2
stop
__________________________________________________

Table 1c - Listing of Data File EG1.DTA

_________________________________________________
1.000000 1.000000 2.547725 2.242410 3.559169
2.000000 1.000000 3.189859 1.535361 4.347655
3.000000 1.000000 3.037594 1.628260 4.497574
.
.
.
58.00000 1.000000 3.061426 2.233128 4.467332
59.00000 1.000000 3.300419 2.058473 4.099995
60.00000 1.000000 2.646529 1.726510 3.789132
________________________________________________________

9
Table 1d - Listing of Instruction File EG1.INS
________________________________________________________________
1 1=ERROR COMPONENTS MODEL, 2=TE EFFECTS MODEL
eg1.dta DATA FILE NAME
eg1.out OUTPUT FILE NAME
1 1=PRODUCTION FUNCTION, 2=COST FUNCTION
y LOGGED DEPENDENT VARIABLE (Y/N)
60 NUMBER OF CROSS-SECTIONS
1 NUMBER OF TIME PERIODS
60 NUMBER OF OBSERVATIONS IN TOTAL
2 NUMBER OF REGRESSOR VARIABLES (Xs)
n MU (Y/N) [OR DELTA0 (Y/N) IF USING TE EFFECTS MODEL]
n ETA (Y/N) [OR NUMBER OF TE EFFECTS REGRESSORS (Zs)]
n STARTING VALUES (Y/N)
IF YES THEN BETA0
BETA1 TO
BETAK
SIGMA SQUARED
GAMMA
MU [OR DELTA0
ETA DELTA1 TO
DELTAK]

NOTE: IF YOU ARE SUPPLYING STARTING VALUES

AND YOU HAVE RESTRICTED MU [OR DELTA0] TO
BE ZERO THEN YOU SHOULD NOT SUPPLY A
STARTING VALUE FOR THIS PARAMETER.
_______________________________________________________________

10
Table 1e - Listing of Output File EG1.OUT
________________________________________________________
Output from the program FRONTIER (Version 4.1)

instruction file = eg1.ins

data file = eg1.dta

Error Components Frontier (see B&C 1992)

The model is a production function
The dependent variable is logged

(… initial estimation outputs based on OLS)

the final mle estimates are :

coefficient standard-error t-ratio

beta 0 0.56161963E+00 0.20261668E+00 0.27718331E+01

beta 1 0.28110205E+00 0.47643365E-01 0.59001301E+01
beta 2 0.53647981E+00 0.45251553E-01 0.11855501E+02
sigma-squared
0.21700046E+00 0.63909106E-01 0.33954545E+01

gamma 0.79720730E+00 0.13642399E+00 0.58436004E+01

mu is restricted to be zero

eta is restricted to be zero

log likelihood function = -0.17027229E+02

LR test of the one-sided error = 0.28392402E+01

with number of restrictions = 1

[note that this statistic has a mixed chi-square distribution]

11
(… more output )

technical efficiency estimates :

firm eff.-est.

1 0.65068880E+00
2 0.82889151E+00
3 0.72642592E+00
.
.
.
58 0.66471456E+00
59 0.85670448E+00
60 0.70842786E+00

mean efficiency = 0.74056772E+00

_____________________________________________________

12
 EXAMPLE 2

• (4.2) A Translog production frontier using cross-sectional data

and assuming a truncated normal distribution.

ln(Qi) = β0 + β1ln(Ki) + β2ln(Li) + β3ln(Ki)2 + β4ln(Li)2 +

β5ln(Ki)ln(Li) + (Vi - Ui),

• where Qi, Ki, Li and Vi are as defined earlier, and Ui has truncated
normal distribution.

Table 2a - Listing of Data File EG2.DAT

_________________________________________________
1. 1. 12.778 9.416 35.134
2. 1. 24.285 4.643 77.297
3. 1. 20.855 5.095 89.799
.
.
.
58. 1. 21.358 9.329 87.124
59. 1. 27.124 7.834 60.340
60. 1. 14.105 5.621 44.218
________________________________________________

13
Table 2b - Listing of Shazam Instruction File EG2.SHA
__________________________________________________
read(eg2.dat) n t y x1 x2
genr ly=log(y)
genr lx1=log(x1)
genr lx2=log(x2)
genr lx1s=log(x1)*log(x1)
genr lx2s=log(x2)*log(x2)
genr lx12=log(x1)*log(x2)
file 33 eg2.dta
write(33) n t ly lx1 lx2 lx1s lx2s lx12
stop
__________________________________________________

14
Table 2d - Listing of Instruction File EG2.INS
________________________________________________________
1 1=ERROR COMPONENTS MODEL, 2=TE EFFECTS MODEL
eg2.dta DATA FILE NAME
eg2.out OUTPUT FILE NAME
1 1=PRODUCTION FUNCTION, 2=COST FUNCTION
y LOGGED DEPENDENT VARIABLE (Y/N)
60 NUMBER OF CROSS-SECTIONS
1 NUMBER OF TIME PERIODS
60 NUMBER OF OBSERVATIONS IN TOTAL
5 NUMBER OF REGRESSOR VARIABLES (Xs)
y MU (Y/N) [OR DELTA0 (Y/N) IF USING TE EFFECTS MODEL]
n ETA (Y/N) [OR NUMBER OF TE EFFECTS REGRESSORS (Zs)]
n STARTING VALUES (Y/N)
IF YES THEN BETA0
BETA1 TO
BETAK
SIGMA SQUARED
GAMMA
MU [OR DELTA0
ETA DELTA1 TO
DELTAK]

NOTE: IF YOU ARE SUPPLYING STARTING VALUES

AND YOU HAVE RESTRICTED MU [OR DELTA0] TO
BE ZERO THEN YOU SHOULD NOT SUPPLY A
STARTING VALUE FOR THIS PARAMETER.
_____________________________________________________________

15
 EXAMPLE 3

• (4.3) A Cobb-Douglas cost frontier using cross-sectional data and

assuming a half-normal distribution.
ln(Ci/Wi) = β0 + β1ln(Qi) + β2ln(Ri/Wi) + (Vi + Ui)

• where Ci, Qi, Ri and Wi are cost, output, capital price and labour
price, respectively, and Vi and Ui are assumed normal and half-
normal distributed, respectively.

Table 3a - Listing of Data File EG3.DAT

__________________________________________________
1. 1. 783.469 35.893 11.925 28.591
2. 1. 439.742 24.322 12.857 23.098
3. 1. 445.813 34.838 14.368 16.564
.
.
.
58. 1. 216.558 26.888 7.853 10.882
59. 1. 408.234 20.848 9.411 23.281
60. 1. 1114.369 32.514 14.919 29.672
_________________________________________________

16
Table 3b - Listing of Shazam Instruction File EG3.SHA
_____________________________________________________
read(eg3.dat) n t c q r w
genr lcw=log(c/w)
genr lq=log(q)
genr lrw=log(r/w)
file 33 eg3.dta
write(33) n t lcw lq lrw
stop
________________________________________________________

Table 3c - Listing of Data File EG3.DTA

_______________________________________________________
1.000000 1.000000 3.310640 3.580542 -0.8744549
2.000000 1.000000 2.946442 3.191381 -0.5858576
3.000000 1.000000 3.292668 3.550709 -0.1422282
.
.
.
58.00000 1.000000 2.990748 3.291680 -0.3262144
59.00000 1.000000 2.864203 3.037258 -0.9057584
60.00000 1.000000 3.625840 3.481671 -0.6875683
_______________________________________________________

17
Table 3d - Listing of Instruction File EG3.INS
_______________________________________________________________
1 1=ERROR COMPONENTS MODEL, 2=TE EFFECTS MODEL
eg3.dta DATA FILE NAME
eg3.out OUTPUT FILE NAME
2 1=PRODUCTION FUNCTION, 2=COST FUNCTION
y LOGGED DEPENDENT VARIABLE (Y/N)
60 NUMBER OF CROSS-SECTIONS
1 NUMBER OF TIME PERIODS
60 NUMBER OF OBSERVATIONS IN TOTAL
2 NUMBER OF REGRESSOR VARIABLES (Xs)
n MU (Y/N) [OR DELTA0 (Y/N) IF USING TE EFFECTS MODEL]
n ETA (Y/N) [OR NUMBER OF TE EFFECTS REGRESSORS (Zs)]
n STARTING VALUES (Y/N)
IF YES THEN BETA0
BETA1 TO
BETAK
SIGMA SQUARED
GAMMA
MU [OR DELTA0
ETA DELTA1 TO
DELTAK]

NOTE: IF YOU ARE SUPPLYING STARTING VALUES

AND YOU HAVE RESTRICTED MU [OR DELTA0] TO
BE ZERO THEN YOU SHOULD NOT SUPPLY A
STARTING VALUE FOR THIS PARAMETER.
___________________________________________________________

18
 EXAMPLE 4

• (4.4) The Battese and Coelli (1992) specification (Model 1).

ln(Qi) = β0 + β1ln(Ki) + β2ln(Li) + (Vi - Ui)

• Data on 15 firms observed over 4 time periods.

19
Table 4a - Listing of Data File EG4.DAT
________________________________________________________
_____________
1. 1. 15.131 9.416 35.134
2. 1. 26.309 4.643 77.297
3. 1. 6.886 5.095 89.799
4. 1. 11.168 4.935 35.698
5. 1. 16.605 8.717 27.878
6. 1. 10.897 1.066 92.174
7. 1. 8.239 0.258 97.907
8. 1. 19.203 6.334 82.084
9. 1. 16.032 2.350 38.876
10. 1. 12.434 1.076 81.761
11. 1. 2.676 3.432 9.476
12. 1. 29.232 4.033 55.096
13. 1. 16.580 7.975 73.130
14. 1. 12.903 7.604 24.350
15. 1. 10.618 0.344 65.380

data for the 2nd and 3rd firm

1. 4. 11.583 4.551 36.704

2. 4. 31.612 7.223 89.312
3. 4. 12.088 9.561 29.055
4. 4. 13.736 4.871 50.018
5. 4. 19.274 9.312 40.996
6. 4. 15.471 2.895 63.051
7. 4. 23.190 8.085 60.992
8. 4. 30.192 8.656 94.159
9. 4. 23.627 3.427 39.312
10. 4. 14.128 1.918 78.628
11. 4. 11.433 6.177 64.377
12. 4. 4.074 7.188 1.073
13. 4. 23.314 9.329 87.124
14. 4. 22.737 7.834 60.340
15. 4. 22.639 5.621 44.218
_____________________________________________________________________

20
Table 4b - Listing of Shazam Instruction File EG4.SHA
_______________________________________________
read(eg4.dat) n t y x1 x2
genr ly=log(y)
genr lx1=log(x1)
genr lx2=log(x2)
file 33 eg4.dta
write(33) n t ly lx1 lx2
stop
__________________________________________________

Table 4 c - Listing of Data File EG4.DTA

__________________________________________________
1.000000 1.000000 2.716746 2.242410 3.559169
2.000000 1.000000 3.269911 1.535361 4.347655
3.000000 1.000000 1.929490 1.628260 4.497574
.
.
.
13.00000 4.000000 3.149054 2.233128 4.467332
14.00000 4.000000 3.123994 2.058473 4.099995
15.00000 4.000000 3.119674 1.726510 3.789132
____________________________________________________

21
Table 4d - Listing of Instruction File EG4.INS
________________________________________________________________
_____
1 1=ERROR COMPONENTS MODEL, 2=TE EFFECTS MODEL
eg4.dta DATA FILE NAME
eg4.out OUTPUT FILE NAME
1 1=PRODUCTION FUNCTION, 2=COST FUNCTION
y LOGGED DEPENDENT VARIABLE (Y/N)
15 NUMBER OF CROSS-SECTIONS
4 NUMBER OF TIME PERIODS
60 NUMBER OF OBSERVATIONS IN TOTAL
2 NUMBER OF REGRESSOR VARIABLES (Xs)
y MU (Y/N) [OR DELTA0 (Y/N) IF USING TE EFFECTS MODEL]
y ETA (Y/N) [OR NUMBER OF TE EFFECTS REGRESSORS (Zs)]
n STARTING VALUES (Y/N)
IF YES THEN BETA0
BETA1 TO
BETAK
SIGMA SQUARED
GAMMA
MU [OR DELTA0
ETA DELTA1 TO
DELTAK]

NOTE: IF YOU ARE SUPPLYING STARTING VALUES

AND YOU HAVE RESTRICTED MU [OR DELTA0] TO
BE ZERO THEN YOU SHOULD NOT SUPPLY A
STARTING VALUE FOR THIS PARAMETER.
_______________________________________________________________

22
 EXAMPLE 5
• (4.5) The Battese and Coelli (1995) specification (Model 2).

ln(Qi) = β0 + β1ln(Ki) + β2ln(Li) + (Vi - Ui)

Table 5a - Listing of Data File EG5.DAT

______________________________________________________
1. 1. 15.131 9.416 35.134 1.000
2. 1. 26.309 4.643 77.297 1.000
3. 1. 6.886 5.095 89.799 1.000
.
.
.
13. 4. 23.314 9.329 87.124 4.000
14. 4. 22.737 7.834 60.340 4.000
15. 4. 22.639 5.621 44.218 4.000
________________________________________________________

Table 5b - Listing of Shazam Instruction File EG5.SHA

________________________________________________________
read(eg5.dat) n t y x1 x2 z1
genr ly=log(y)
genr lx1=log(x1)
genr lx2=log(x2)
file 33 eg5.dta
write(33) n t ly lx1 lx2 z1
stop
________________________________________________________

23
Table 5c - Listing of Data File EG5.DTA
_____________________________________________________________________
1.000000 1.000000 2.716746 2.242410 3.559169 1.000000
2.000000 1.000000 3.269911 1.535361 4.347655 1.000000
3.000000 1.000000 1.929490 1.628260 4.497574 1.000000
.
.
.
13.00000 4.000000 3.149054 2.233128 4.467332 4.000000
14.00000 4.000000 3.123994 2.058473 4.099995 4.000000
15.00000 4.000000 3.119674 1.726510 3.789132 4.000000
_____________________________________________________________________

24
Table 5d - Listing of Instruction File EG5.INS
________________________________________________________________
2 1=ERROR COMPONENTS MODEL, 2=TE EFFECTS MODEL
eg5.dta DATA FILE NAME
eg5.out OUTPUT FILE NAME
1 1=PRODUCTION FUNCTION, 2=COST FUNCTION
y LOGGED DEPENDENT VARIABLE (Y/N)
15 NUMBER OF CROSS-SECTIONS
4 NUMBER OF TIME PERIODS
60 NUMBER OF OBSERVATIONS IN TOTAL
2 NUMBER OF REGRESSOR VARIABLES (Xs)
y MU (Y/N) [OR DELTA0 (Y/N) IF USING TE EFFECTS MODEL]
1 ETA (Y/N) [OR NUMBER OF TE EFFECTS REGRESSORS (Zs)]
n STARTING VALUES (Y/N)
IF YES THEN BETA0
BETA1 TO
BETAK
SIGMA SQUARED
GAMMA
MU [OR DELTA0
ETA DELTA1 TO
DELTAK]

NOTE: IF YOU ARE SUPPLYING STARTING VALUES

AND YOU HAVE RESTRICTED MU [OR DELTA0] TO
BE ZERO THEN YOU SHOULD NOT SUPPLY A
STARTING VALUE FOR THIS PARAMETER.
________________________________________________________________

25
 APPENDIX - PROGRAMMER'S GUIDE

A.1 The FRONT41.000 File

The start-up file FRONT41.000 is listed in Table A1. Ten values may be altered in
FRONT41.000. A brief description of each value is provided below.

Table A1 - The start-up file FRONT41.000

____________________________________________________________
KEY VALUES USED IN FRONTIER PROGRAM (VERSION 4.1)
NUMBER: DESCRIPTION:
5 IPRINT - PRINT INFO EVERY “N” ITERATIONS, 0=DO NOT PRINT
1 INDIC - USED IN UNIDIMENSIONAL SEARCH PROCEDURE - SEE BELOW
0.00001 TOL - CONVERGENCE TOLERANCE (PROPORTIONAL)
0.001 TOL2 - TOLERANCE USED IN UNI-DIMENSIONAL SEARCH PROCEDURE
1.0D+16 BIGNUM - USED TO SET BOUNDS ON DEN & DIST
0.00001 STEP1 - SIZE OF 1ST STEP IN SEARCH PROCEDURE
1 IGRID2 - 1=DOUBLE ACCURACY GRID SEARCH, 0=SINGLE
0.1 GRIDNO - STEPS TAKEN IN SINGLE ACCURACY GRID SEARCH ON GAMMA
100 MAXIT - MAXIMUM NUMBER OF ITERATIONS PERMITTED
1 ITE - 1=PRINT ALL TE ESTIMATES, 0=PRINT ONLY MEAN TE

26
Deregulation, Ownership, and Efficiency Change in Indian Banking:
An Application of Stochastic Frontier Analysis

Subal C. Kumbhakar
Department of Economics
State University of New York
Binghamton, NY 13902, USA
E-mail: [email protected]

and

Subrata Sarkar}
Indira Gandhi Institute of Development Research
Gen. Vaidya Marg, Goregaon (East)
Mumbai 400 065, INDIA
E-mail: [email protected]
 Table 3: Estimated Parameters of the Translog Cost Function
and the Simple Time Varying Inefficiency Function Based on
Battese and Coelli Model 1

Period: 1986- Period: 1986- Period: 1993-

2000 1992 2000
All Banks All Banks All Banks
Coeff. T-stat Coeff. T-stat Coeff. T-stat

Sigma- 0.024 8.711 0.021 8.574 0.022 5.817

squared

Gamma 0.767 29.750 0.829 35.034 0.845 40.456

Mu 0.273 9.598 0.265 8.534 0.273 8.915

Eta 0.016 3.208 0.019 1.378 0.016 1.475

Log-L 765.600 396.420 462.75

LR Test
for
One- 571.22 292.22 349.54
sided
Error
d.o.f 3 3 3

y1 = deposits, y2 = advances, y3 = investments, y4 = branches, t = time, wl = relative wage.

All variables, except time, are logged as per translog cost function. The second order terms
are obvious: y12 = y1*y2; y33 = y3*y3; etc. .

28
Table 4: Mean Efficiency of Public and Private Banks
Based on Battese and Coelli Model 1

Mean Efficiency

Year All Banks Public Private

Banks Banks

1986 0.689 0.652 0.745

1987 0.693 0.656 0.746
1988 0.698 0.661 0.747
1989 0.702 0.666 0.751
1990 0.706 0.670 0.754
1991 0.710 0.675 0.758
1992 0.715 0.679 0.761
1993 0.719 0.683 0.765
1994 0.723 0.688 0.768
1995 0.726 0.692 0.772
1996 0.730 0.696 0.775
1997 0.735 0.700 0.782
1998 0.739 0.705 0.790
1999 0.743 0.709 0.793
2000 0.747 0.713 0.796

Sample: Period 1986-2000, All banks.

29
Table 7: Estimated Parameters of the Inefficiency Function Based on
Battese and Coelli Model 2
Model A Model B Model C Model D

Intercept 0.108 0.143 0.054 0.320

Pvt -0.223 -0.155 -0.219 -0.151
t -0.006 -0.060
t*Pvt -0.005
Dereg 0.103 -0.253
Dereg*Pvt -0.035 -0.321
Dereg*t 0.051
Dereg*t*P 0.023
vt

Bold numbers denote significant coefficients at the 5% level of significance

Pvt (private) and Dereg (deregulation), are dummy variables. Dereg=1 if year > 1992.
The variable t stands for time, with t=1 for the year 1986. The interaction between
the dummy variables time and private is captured by t*Pvt., and Dereg*t*Pvt is a
three-way interaction.

30
 Figure 2: Average Efficiency of Banks by Ownership Group
1.020

1.000

0.980

0.960

0.940
eff

0.920

0.900
vvv
0.880

0.860

0.840

0.820
1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000
Year
Public Bank Privete Bank Average bank performance

31
 Deterministic frontier

YM

YA

XA X
 Stochastic Frontier

f(x) exp(vi)
vi < 0

YA

XA
X
X2

(X1A,X2A)

Actual expenditure

Minimum expenditure

Q = 100

X1
Slope = w1/w2
X2

XA
C

W Q1 = 100

Q2 = 75

B D F X1

Total inefficiency : BF
Technical inefficiency : BD
Allocative inefficiency : DF