Warsaw School of Economics Institute of Econometrics
Department of Applied Econometrics
Department of Applied Econometrics Working Papers
Warsaw School of Economics Al. Niepodleglosci 164 02-554 Warszawa, Poland
Working Paper No. 1-08
Bayesian analysis of growth using stochastic frontier model
Arkadiusz Wiśniowski
Warsaw School of Economics
This paper is available at the Warsaw School of Economics
Department of Applied Econometrics website at: http://www.sgh.waw.pl/instytuty/zes/wp/
frontier model
Arkadiusz Wi±niowski
Warsaw S hool of E onomi s
email: awisniowskiop.pl
Abstra t
We employ Bayesian approa h to the analysis of e onomi growth in
Poland.
The results of estimation of a sto hasti frontier modelapplied to pro-
du tion fun tion of Polish voivodships in 2000 - 2004 are presented.
Sto hasti frontier approa h allows to de ompose growth into te hno-
logi al hange,input hange ande ien y hange.In orderto ompute
the posterior hara teristi s of the growth omponents we employ the
GibbsMCMC sampler.
Keywords:Bayesiananalysis,Gibbssampler,e onomi growth,sto has-
ti frontier analysis
JEL odes: C11, C33, O49
1 Introdu tion
E onomi growth and its sour es belong to the most important issues of
e onomi s [Barro, Sala-i-Martin 1999℄, [Koop, Osiewalski, Steel 2000℄. The
sto hasti frontier model providesa formal framework to de ompose the e o-
nomi growth into three omponents: input hange, te hni al hange and
e ien y hange. The aim of this paper is to present the appli ation of the
Bayesian frameworkinthe analysisof e onomi growth.Wemodelthe growth
by means of a sto hasti produ tion fun tion, thus we use the sold industrial
produ tionasameasure of it.The analysis isperformed for16voivodships in
Poland in the period2000 - 2004.
To our analysis we apply a Bayesian framework with re ently developed
numeri al Markov Chain Monte Carlo methods. Bayesian approa h seems to
betheappropriatetoolsin eitallowsustofo us ofanyquantityofinterestby
derivingitsposteriordistribution(inparti ular the omponentsof theoutput
growth),tointegrateoutallnuisan eparameters,tohandleallrestri tionsand
regularity onditions that result from the e onomi theory, as well as to deal
with alarge numberof parameters in the model.
This resear h is based mainly on the papers by G. Koop, J. Osiewalski
and M.F.J. Steel [1999, 2000℄ and toa lesser extent on[Osiewalski 2001℄ and
[Osiewalski,Steel1998℄. Inthe next se tionwedes ribethe sto hasti frontier
model.Se tion3presentsthede ompositionofgrowth,fourthse tionprovides
the modelina Bayesian framework withthe derivation ofGibbs samplerused
toestimate the unknown parameters. In Se tion5 the resultsfollow, last two
se tionspresent on lusions, modelextensions and a summary.
2 Sto hasti frontier model
The sto hasti frontier modelwas originallyproposed by W. Meeusen and J.
vandenBroe k[1977℄andD.Aigner,C.A.K.LovellandP.S hmidt[1977℄.The
model onsists of the mi roe onomi produ tion fun tion and two error om-
ponents:one ree tingrandomnessof thefrontier itselfandonethat measures
ine ien y.
In the model we assume that all omparable agents or units (herein we
onsider all 16 voivodships in Poland) produ e a ording to a ommon te h-
nology.Thisassumptionallowsusto onstru ta ommonprodu tionfun tion
and to interpret allsystemati deviations fromit being a result of the under-
usageofinputs.Inotherwords,thevoivodship anoperateeitheronorwithin
afrontier.voivodshipsare territorialunits (provin es) with fullfreedom of in-
formation ow thus the assumption that te hnology used by ompanies, e.g.
in Podkarpa kie voivodship an be opied and used by a plant operating in
Dolno±l¡skie seems tobe reasonable 1
.
The followingpresents the produ tionfun tion under onsideration
Y
it
=f
t (K
it
;L
it )
it z
it
; (2.1)
whereY
it
is anoutput value, f
t (K
it
;L
it
)is aprodu tion fun tion with apital
and labour inputs respe tively,
it
(0 <
it
1) is a random term ree ting
produ tion e ien y (so- allede ien y indi ator) and z
it
is a random term
that aptures the general sto hasti nature of e onomi variables (whi h is a
resultofe.g.measurementerror).Subs ripti=1;:::;N identiestheprodu -
ingunits (inour ase voivodships)intimet=1;:::;T.Therandomtermsare
independent of ea h other, a rosstime and provin es.
Inouranalysisweuse atranslogprodu tionfun tion.This formisadopted
by Koopet al.[Koop, Osiewalski, Steel1999℄ intheir GDP growth analysis of
1
Wearenegle tingallte hnologiesthat areunder patentprote tion.
aset oftheOECD ountries.Translogformallows ustoree tthevariationof
the data ina better way than the Cobb-Douglasfun tion, whi h is one of the
spe ial ases of the translog form. Moreover, we assume that our produ tion
fun tion is time-dependent and we will onsider two ases, with and without
stru ture imposed.
In a general ase the translogprodu tionfrontier an be written
y
it
=x 0
it
t u
it +v
it
; (2.2)
where
u
it
= log(
it )
is anon-negativerandom variable,
v
it
=log(z
it )
is asymmetri allydistributed random variablewith mean zero,
x
it
=
1; k
it
; l
it
; l
it
k
it
; k 2
it
; l 2
it
;
0
;
and a ve tor of parameters
t
=(
t0
;:::;
t5 )
0
:
Lower ase letters y, l, k indi ate logarithms of the e onomi variables (e.g.
y=log(Y)).Weimposenon-negativity onstraintsonlabour and apitalelas-
ti ities
y
it
l
it
=
t2 +
t3 k
it +2
t5 l
it
0;
y
it
k
it
=
t1 +
t3 l
it +2
t4 k
it
0 (2.3)
foralliandt.Asalo almeasureof e onomiesof s ale(whi hisalsoemployed
by Koop et al. [1999℄) we use the elasti ity of returns to s ale (ER TS, see
[Varian 1992℄) whi h,for atranslog produ tionfun tion, isgiven by 2
ER TS
it
t1 +
t2 +(
t3 +2
t4 )l
it +(
t3 +2
t5 )k
it
: (2.4)
Thusweobtainthe onstantreturnstos ale by imposingtherestri tion
t1 +
t2
= 1,
t4
=
t5
and
t3
= 2
t4
. The translog fun tion will redu e to the
Cobb-Douglas formfor
t3
=
t4
=
t5
=0.
In this paper we onsider two models originally proposed by Koop et al.
[1999℄.In both of themwe assumenormallydistributed randomerrorv
it with
avarian ethat remains onstanta ross time and voivodships, v
it
N(0; 2
).
Fortherandome ien yindi ator,u
it
= log(
it
),weassumetheexponential
distributionwithitsexpe tedvalue onstanta rosstimeandvoivodships,u
it
Exp (;0) (for justi ation of the exponential form see [Ritter, Simar 1997℄,
[Koop, Osiewalski,Steel 1999℄; for denitions see Appendix A.1).
Asfarastheprodu tionfun tionparametersare on erned(thereareJ =6
of them),we onsider two versions.Wedene the following (NT 1)ve tors:
y = (y 0
1
;:::;y 0
t
;:::;y 0
T )
0
u = (u 0
1
;:::;u 0
t
;:::;u 0
T )
0
(2.5)
v = (v
1
;:::;v 0
t
;:::;v
T )
0
;
wherey
t
=(y
1t
;:::;y
Nt ),u
t
=(u
1t
;:::;u
Nt
)andv
t
=(v
1t
;:::;v
Nt
)are(1N)
ve torsand a (N J) matrix is given by
X
t
=(x
1t
;:::;x
Nt )
0
: (2.6)
2
Given the produ tion fun tion y = f(x) and a s alar t > 0 onsider the fun tion
y(t)=f(tx).Thentheelasti ityofreturnstos aleisdenedas
e(x)= dy(t)
dt
t
y(t)
t=1 :
In the rst version (indi ated as A version hereafter) we assume that the
parameters are independent of ea h other for every t, hen e our estimation
obje t is a(TJ1) ve tor
A
=( 0
1
;:::; 0
T )
0
: (2.7)
The model an bewritten as
y=X A
A
u+v; (2.8)
wherey,uand v are ve tors dened in2.5, is dened asin2.7and X is the
(NT TJ) matrix
X A
= 2
6
6
6
6
6
6
6
6
6
6
6
6
4 X
1
.
.
.
X
t
.
.
.
X
T 3
7
7
7
7
7
7
7
7
7
7
7
7
5
; (2.9)
where X
t
is the matrix given in 2.6.
In the se ond version of our model (indi ated as B) we impose a linear
trend restri tion onthe produ tion fun tionparameters
t
=
+t
: (2.10)
This model an be written asin equation2.8
y=X B
B
u+v; (2.11)
where B
=((
) 0
(
) 0
) 0
isa (2J 1)ve tor and
X B
= 2
6
6
6
6
6
6
6
6
6
6
6
6
4 X
1 X
1
.
.
. .
.
.
X
t tX
t
.
.
. .
.
.
X
T TX
T 3
7
7
7
7
7
7
7
7
7
7
7
7
5
is the (NT 2J) matrix. X
t
isgiven in 2.6.
Hen e to des ribe the produ tion frontier in the A version we use (NT +
TJ+2) parameters, whereas the dimension of the version B of the modelis
(NT +2J+2).
3 Growth de omposition
Theprodu tionfrontierallowsustode omposethee onomi growthintothree
omponents: e ien y hange 3
, input hange and te hni al hange. Assume
that the produ tion frontiers for all produ ing units in periods t and t+1
and all inputs in these periods are given. Then the expe ted value of growth
(further abbreviated toexpe ted or predi ted growth) an be writtenas
E(y
i;t+1 y
i;t )=(x
0
i;t+1
t+1 x
0
i;t
t )+(u
ti u
i;t+1
): (3.12)
The rst expression inbra kets indi ates te hni al and input hangewhereas
the se ond one isthe e ien y hange.Let uswrite the rst expression as
x 0
i;t+1
t+1 x
0
i;t
t
= 1
2 (x
i;t+1 x
i;t )
0
(
t+1 +
t )+
1
2 (x
i;t+1 +x
i;t )
0
(
t+1
t
): (3.13)
The rst omponent of equation 3.13 is the hange in inputs for the average
te hnology. The se ondterm indi ateste hni al progress given averageinputs
in two periods onsidered. Te hni al hange for the ith voivodship an be
measuredasexp [x 0
i (
t+1
t
)℄foragiven ve torofinputsx
i
.Duetothefa t
that inputs vary over time, Koop et al. [1999, 2000℄ propose to measure the
impa t of global te hni al hange on the produ tivity of the ith voivodship
3
As far as the e ien y of the produ tion pro ess is on erned we an dierentiate
betweente hni ale ien y that results from theproperuse ofthe produ tionte hnology
giveninputsandallo ativee ien ywhi hisa onsequen eoftheproperallo ationofinputs
(see [Osiewalski 2001℄). The subje t matter of the sto hasti frontier model is te hni al
e ien y.
by the geometri mean ofx
it andx
i;t+1
.Hen e weobtain the te hni al hange
TC
i;t+1
given by
TC
i;t+1
=exp
1
2 (x
i;t+1 +x
it )
0
(
t+1
t )
(3.14)
and the hange in inputsIC
i;t+1
IC
i;t+1
=exp
1
2 (x
i;t+1 x
it )
0
(
t+1 +
t )
: (3.15)
The entire produ tivity hange of the ithvoivodship an be al ulated as
PC
i;t+1
=TC
i;t+1
IC
i;t+1
: (3.16)
Thethird growth omponentisthee ien y hange EC
i;t+1
that an bewrit-
ten as
EC
i;t+1
=exp [(u
i;t u
i;t+1 )℄=
i;t+1
i;t
: (3.17)
Cumulated te hni al, input and e ien y hanges for all periods onsidered
are given by
CTC
i
= T 1
Y
t=1 TC
i;t+1
; (3.18)
CIC
i
= T 1
Y
t=1 IC
i;t+1
; (3.19)
and
CEC
i
= T 1
Y
t=1 EC
i;t+1
=exp (u
i;1 u
i;T
): (3.20)
The average hange in all periods are al ulated as a geometri mean of all
annual hanges. Thus, for te hni al hangewe an write
ATC
i
=(CTC
i )
1
T 1
; (3.21)
and for e ien y and input hangewe have
AEC
i
= (CEC
i )
1
T 1
; (3.22)
AIC
i
= (CIC
i )
1
T 1
(3.23)
respe tively. The expe ted (predi ted) average annualgrowth an be derived
fromequation 3.12
AGC
i
=ATC
i
AIC
i
AEC
i
; (3.24)
whereas the average annualprodu tivity be omes
APC
i
=ATC
i
AEC
i
: (3.25)
In order to fa ilitate interpretation, the nal results are given in per entage
points:ATG=100(ATC
i
1)forte hni al hange,AEG=100(AEC
i 1)
for e ien y hange, et .
4 Bayesian model
To investigate the produ tion growth in Polish voivodships in 2000 - 2004
from models 2.8 and 2.11 we use a Bayesian framework. We have N = 16
voivodships and T = 5 periods (annual data from years 2000 - 2004). The
observation ve tor is anoutput ve tor y with the exogenous variables matrix
X given (dened asX =fX A
;X B
g fora A or B version of the model):
y =(y;X):
Ve torof parameters that are the obje t of inferen eis given by
=
0
2
u
;
where is dened (depending onthe modelversion) as 4
A
or B
. Given the
above mentioned notationwe dene the likelihoodfun tion as 5
L(yj;X)=f TN
N
(yjX u; 2
TN I
TN
): (4.26)
The prior density for the ve tor of produ tion fun tion parameters p() is
an(improper) trun ated uniform distribution,whi htakes the value p()=1
whentheregularity onditionsgivenin2.3aresatisedandp()=0otherwise.
For the posterior density to be well-dened our prior densities of parameters
2
and must beproper (informative), otherwise the posteriordensity is not
- omplete 6
(see[Fernandez,Osiewalski,Steel1997℄).Inthemodelweuse the
following prior densitiesfor 2
and
p(
2
)=(
2
) n
0
=2 1
exp a
0
2
2
; (4.27)
and
p(
1
)=f
G (
1
j
01
;
02
) (4.28)
with predened hyperparameters a
0 , n
0 ,
01
and
02
. Koop et al. [1999℄ take
a
0
> 0,
01
>0,
02
> 0 and n
0
0. For n
0
= 0 the prior is improper yet it
impliesaproperposteriordistribution (seeFernandez etal.[1997℄).Following
Koop et al. [1999℄ we assume a
0
= 10 6
. The resulting prior density is lose
to the non-informative prior p() / 1
but it assigns small weights on big
realizations of 2
. The hyperparameter
01
= 1 was hosen for the prior
4
Introdu tionofsu hnotationshouldnot ompli atetheentirereasoningsin einalmost
allfun tionsdis ussedX and appearastheprodu tX,thedimensionofwhi hisequal
(NT 1)forbothmodels(A andB).In aseof ajointdistribution ofallparametersora
distributionoftheparameter weshouldrememberthatformodelAthesedimensionsare
equal(NT+TJ+2)andTJ respe tively,whereasforB theyare(NT+2J+2)and2J.
5
FordenitionsseeAppendixA.1
6
Ameasuredenedonaspa eX is- ompleteithere existsa(innite)sequen eof
subsetsX
1 ,X
2
,:::summinguptoX,that(X
i
)isniteforalli.
density of to be at.
02
= log(
) through
ree ts prior beliefs about
the median e ien y. Koop et al. [1999℄ perform a sensitivity analysis of the
model with respe t to dierent values of
. The results show that '(...) the
model displays an impressive degree of robustness a ross these very dierent
priors'.Furtheronin the analysis we assume
=0:75.
The prior density for u is assumed to be the Gamma distribution (see
Se tion2)
T
Y
t=1 N
Y
i=1 f
G (u
it j1;
1
): (4.29)
The jointprior density is aprodu tof priordistributions ofallparameters
p(; 2
; 1
;u)=p()p(
2
)p(
1
)p(u): (4.30)
Having dened all priorswe an writethe posteriordensity as:
p(jy;X)/f TN
N
(yjX u; 2
I
TN
)p()p(
2
)p(
1
) T
Y
t=1 N
Y
i=1 f
G (u
it j1;
1
):
(4.31)
4.1 Gibbs sampler
In order to approximate our posterior density we use Gibbs sampler, the
Markov Chain Monte Carlo algorithm (for details see e.g. [Tierney 1994℄,
[Bernardo, Smith 1994℄, [Casella, Robert 1999℄, [Roberts, Rosenthal 2004℄).
First we spe ify onditional densities of all parameters. The derivations here-
afterare basedon the hintsgiven by Koopet al.[1999℄.
To obtain the onditional density for 7
p(ju; 2
; 1
;y;X) note that
^
=(X 0
X) 1
X 0
(y+u)
and
(y+u) 0
X
^
=
^
0
X 0
X
^
:
7
FormodelAitis30-andformodelB 12-dimensionaldensity.
Then we al ulate
p(ju; 2
; 1
;y;X)/exp
"
(y X+u) 0
(y X+u)
2
2
#
p()
= exp
"
(y+u) 0
(y+u) 2(y+u) 0
X+ 0
X 0
X
2
2
#
p()
= e h
(y+u) 0
(y+u) 2(y+u) 0
X
^
+2
^
0
X 0
X
^
2(y+u) 0
X+ 0
X 0
X
2
2
i
p()
= exp
"
(y+u X
^
) 0
(y+u X
^
)+(
^
) 0
X 0
X(
^
)
2
2
#
p()
/ exp
"
(
^
) 0
X 0
X(
^
)
2
2
#
p() (4.32)
sin e(y+u X
^
) 0
(y+u X
^
)isindependentof (itisapartofanormalising
onstant).Obviously4.32isanormaldistributionwithmean
^
and ovarian e
matrix 2
(X 0
X) 1
.The next onditional density isp(
2
j;u; 1
;y;X).De-
noting as an(NT 1)ve tor of ones we al ulate
p(
2
j;u; 1
;y;X)/(
2
) TN =2
e h
(y X+u) 0
(y X+u)
2
2
i
2
e 10
6
2
2
= (
2
) TN
2 1
exp
1
2 [10
6
+(y X+u) 0
(y X+u)℄
2
:(4.33)
Thedistributiongivenin4.33isaGammadistributionwithashapeparameter
TN
2
and s aleparameter 10 6
+(y X+u) 0
(y X+u) (see the denition
A.3 inAppendix A.1). Forour ine ien y ve tor u the onditional density is
trun ated normal.
p(uj; 2
; 1
;y;X)/
/ exp
"
(y X+u) 0
(y X+u)
2
2
#
T
Y
t=1 N
Y
i=1
exp ( u
it
1
)I(u
it
0)
= exp
"
(y X+u) 0
(y X+u)
2
2
#
exp ( u 0
1
)I(u
it
0)
= exp
"
(y X+u) 0
(y X+u)+2
2
u 0
1
2
2
#
I(u
it
0)
/ e
(y X+u) 0
(y X+u)+2
2
u
0
2
2
(X y)
0
+(
2
)
2
2
2
I(u
it
0):