probability behind spatial stochastic frontier model · probability behind spatial stochastic...

35
Probability behind Probability behind Spatial Stochastic Frontier model Spatial Stochastic Frontier model Spatial Stochastic Frontier model Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

Upload: others

Post on 16-Aug-2020

26 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Probability behind Probability behind Spatial Stochastic Frontier modelSpatial Stochastic Frontier modelSpatial Stochastic Frontier modelSpatial Stochastic Frontier model

Dmitry Pavlyuk

The Mathematical Seminar,The Mathematical Seminar,

Transport and Telecommunication Institute, Riga, 08.05.2015

Page 2: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Outline

• Classical stochastic frontier model

• A case of truncated normal inefficiency

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

• A case of truncated normal inefficiency

• Spatial stochastic frontier model

• Moment generating and characteristic functions

• Closed/Unified skew-normal distribution

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015 2

Page 3: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Economic background of stochastic

frontier model

We consider a company, which uses K inputs, indexed k

= 1, 2, …, K, to produce M outputs, indexed m = 1, 2, …,

M:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

= 1, 2, …, K, to produce M outputs, indexed m = 1, 2, …,

M:

Than production possibility set is defined as:

( )( ).,...,,

,,...,,

21

21

M

K

yyyy

xxxx

==

{ }yxyxPPS producecan :,=

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

The set of feasible outputs for an input vector:

3

{ }yxyxPPS producecan :,=

( ) ( ){ }PPSyxyxP ∈= ,:

Page 4: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Production possibility frontier

A production possibility frontier is defined as a

function:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

( ) ( ) ( ){ }xPyxPyyxf yy ∉∀∈= > ',: '

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015 4

Page 5: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Technical efficiency

Debreu-Farrell definition of technical efficiency of

output vector y is:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

output vector is:

can be presented in a form

of equation:

( ) ( ){ }1

:sup,−

≤= xfyyxTE θθθ

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

of equation:

5

( ) ( )yxTExfy ,⋅=

Page 6: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Classical stochastic frontier

For estimation purposes the technical efficiency term is

usually transformed as:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

usually transformed as:

Introducing the random disturbances v into the

formula, we consider a classical stochastic frontier model:

( ) ( ) .0,exp, ≥−= uuyxTE

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

where β is a vector of unknown coefficients.

6

( ) ( ) ( )iiii uvxfy −⋅⋅= expexp, β

Page 7: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Classical stochastic frontier

The model is frequently presented in the logarithmic

form:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

form:

So the only probabilistic feature of the stochastic

frontier model is a composed error term:

( ) iiii uvxfy −+= β,lnln

uv −=ε

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015 7

iii uv −=ε

Page 8: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Outline

• Classical stochastic frontier model

• A case of truncated normal inefficiency

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

• A case of truncated normal inefficiency

• Spatial stochastic frontier model

• Moment generating and characteristic functions

• Closed/Unified skew-normal distribution

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015 8

Page 9: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Stochastic frontier: a case of

truncated normal inefficiency

The distribution of the random disturbances v is

usually set to independent identically distributed normal

with zero mean and constant deviation σ :

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

usually set to independent identically distributed normal

with zero mean and constant deviation σv:

The inefficiency term u can be modelled with different

distributions, e.g. truncated normal:

( )2,0~ vi Nv σ

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

distributions, e.g. truncated normal:

9

( )2,0 ,~ ui TNu σµ+∞

Page 10: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Stochastic frontier: a case of

truncated normal inefficiency

The probability density function for truncated normal

distribution is (truncation limits are set to 0 and +∞ to

match the non-negativity requirement of ):

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

distribution is (truncation limits are set to 0 and +∞ to

match the non-negativity requirement of u):

( )( )

<

−−

Φ

=

0,0

0,2

exp2

12

21

i

iu

i

uui

u

uu

uf σµ

σπσµ

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015 10

Page 11: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Stochastic frontier: a case of

truncated normal inefficiency

According to the convolution formula, if u and v are

independent, then:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

( ) ( ) ( )∫∞+

∞−

+=

−=

iiuiivi

iii

duufuff

uv

εε

ε

ε

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015 11

Page 12: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Stochastic frontier: a case of

truncated normal inefficiency

Transformations:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

( ) ( ) ( ) =+=+∞

∫ iiuiivi duufuff εεε

( ) ( )

( ) ( )

=

+−−

+−

Φ=

=

−−+−

Φ=

=

−−

Φ

+−=

∞+−

∞+−

∞+ −

∞−

221

02

2

2

21

02

21

2

2

/1exp

1exp

1

22exp

2

1

2exp

2

1

2exp

2

1

vuii

iu

i

v

ii

uvu

iu

i

uuv

ii

v

iiuiivi

du

duuu

duuu

µσσεµεµ

σµ

σε

σπσσµ

σµ

σπσµ

σε

σπ

ε

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015 12

( )

( )

++

+−Φ

++

Φ

+=

=

++

+−−

+−

Φ=

222222

1

22

0222222

/

/1

/2exp

2exp

2

uvvuuv

vui

uv

i

uuv

i

uvvuuv

vui

uv

i

uvu

du

σσσσµ

σσσσε

σσµεϕ

σµ

σσ

σσσσσσσσσπσσ

Page 13: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Univariate Extended skew normal

distribution

So

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

( )( )

++

+−Φ

++

Φ

+=

222222

1

22

/1 vuiiif

σσσσµ

σσσσε

σσµεϕ

σµ

σσεε

This density function is

known as an

univariate

extended skew normal

( )( )

+

++

−Φ

+

Φ+

=22222222 / uvvuuvuvuuv

ifσσσσσσσσ

ϕσσσ

εε

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

extended skew normal

distribution function,

introduced by Azzalini (1985).

13

Page 14: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Outline

• Classical stochastic frontier model

• A case of truncated normal inefficiency

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

• A case of truncated normal inefficiency

• Spatial stochastic frontier model

• Moment generating and characteristic functions

• Closed/Unified skew-normal distribution

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015 14

Page 15: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Spatial stochastic frontier model

A basic assumption of the classical stochastic frontier:

Observations are independent!

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

Observations are independent!

( )( )nu

nvn

IMVTNu

IMVNv

uvXy

2~,0

2~

,~

,,0~

,

σµσ

β

+∞

−+=

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

In practice, there are a lot of theoretical and empirical

evidences of relationships between observations.

15

Page 16: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Spatial stochastic frontier model

One of the simplest forms of this relationship is a linear

combination:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

combination:

In spatial econometrics this is assumed that

coefficients wij can be explained with a distance between

objects i and j, e.g.

∑≠ ji

jiji ywy on depends

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

objects i and j, e.g.

16

( )jiij objectobject

w,distance

1=

Page 17: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Spatial stochastic frontier model

Spatial contiguity matrix:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

...0 112 nww

Matrix form of dependence (spatial lags):

=

0...

............

...0

...0

21

221

112

nn

n

n

ww

ww

ww

W

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

Matrix form of dependence (spatial lags):

17

Wyyywyji

jijini on depends on depends ..1 ⇔∀ ∑≠

=

Page 18: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Spatial stochastic frontier model

The spatial stochastic frontier model specification:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

,)( uvXWXβYWρY s −+++= β

where

.~,~

,)(

uuWρu

vvWρv

uvXWXβYWρY

uu

vv

sXYY

+=+=

−+++= β

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

• WY and ρY are contiguity matrix and coefficient for endogenous spatial effects (spatial

dependency),

• WX and β(s) are contiguity matrix and coefficients for exogenous spatial effects,

• Wv and ρv are contiguity matrix and coefficients for spatially correlated random

disturbances (spatial heterogeneity),

• Wu and ρu are contiguity matrix and coefficients for spatially related inefficiency.

18

Page 19: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Spatial stochastic frontier model

Assuming that

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

( )( )nu

nvn

IMVTNu

IMVNv2~,0

2~

,~~,,0~~

σµσ

+∞

and using a simple transformation

we obtain random component distributions:( )MVNv ,,0~ Σ

( )nu IMVTNu ~,0 ,~ σµ+∞

( )( ) ,~

,~

1

1

uWρIu

vWρIv

uu

vv

−=

−=

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015 19

( )( ) ( )[ ]

( )( ) ( )[ ]Tuuuuuu

u

T

vvvvvv

vn

WρIWρI

MVTNu

WρIWρI

MVNv

112~

,0

112~

,,~

,,0~

−−

+∞

−−

−−=Σ

Σ−−=Σ

Σ

σ

µσ

Page 20: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Spatial stochastic frontier model

The main problem is a distribution of the composed

error term:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

error term:

Using of the convolution formula for derivation of the

probability density function for ε is quite complicated

( ) ( ).,~,,0~ ,0 uvn MVTNuMVNv

uv

ΣΣ−=

+∞ µε

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

probability density function for ε is quite complicated

(see in my PhD thesis).

20

Page 21: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Outline

• Classical stochastic frontier model

• A case of truncated normal inefficiency

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

• A case of truncated normal inefficiency

• Spatial stochastic frontier model

• Moment generating and characteristic functions

• Closed/Unified skew-normal distribution

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015 21

Page 22: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Moment generating and

characteristic functions

Let f(x) is a continuous probability density function of a

random variable X.

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

random variable .

Moment generating function for a continuous

probability density function:

( ) ( )∫+∞

∞−

= dxxfetMGF txX

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

Characteristic function for a continuous probability

density function:

22

( ) ( ) ( )12 −== ∫+∞

∞−

idxxfetCF itxX

Page 23: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Moment generating and

characteristic functions

Note that

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

MGFX(-t) is the two-sided Laplace transformation of

the probability density function,

CFX (t) is an inverse Fourier transformation of the

probability density function.

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

probability density function.

23

Page 24: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Moment generating and

characteristic functions

Very useful properties of MGF:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

1. Uniqueness theorem: if two distributions have the

same moment-generating function, then they are

identical at almost all points.

2. Convolution theorem: if random variables X and Y are

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

2. Convolution theorem: if random variables X and Y are

independent and Z=X+Y, then:

MGFZ(t)=MGFX(t)MGFY(t)

24

Page 25: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Moment generating and

characteristic functions

Classical stochastic frontier:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

( )

=2

exp22t

tMGF vv

σ( )2,0~ Nv σ

So the composed error term:

( )

=2

exptMGFvi( ),0~ vi Nv σ

( )2,0 ,~ ui TNu σµ+∞

iii uv −=ε

( )

Φ

+=

tt

ttMGF uuu

uui

σσµ

σµσµ

122

2exp

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015 25

( ) ( ) ( )

( )

−Φ

Φ

++−=

=−−=−

tt

t

tMGFtMGFtMGF

uuu

uv

uv iii

σσµ

σµσσµ

ε

1222

2exp

iii uv −=ε

Page 26: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Moment generating and

characteristic functions

Similarly for the spatial stochastic frontier model

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

( )vnMVNv Σ,0~

( )

Σ= tttMGF vT

v 2

1exp

( )vnMVNv Σ,0~

( ).,~ ,0 uMVTNu Σ+∞ µ

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015 26

( ) ( )[ ] ( )uunuTT

unnu tttttMGF Σ−ΣΦ

Σ+Σ−Φ= − ,,2

1exp,,0 1 µµ

Page 27: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Moment generating and

characteristic functions

MGF of the composed error term:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

uv −=ε

which is a specific case of the Closed Skew Normal

( ) ( ) ( )

( )[ ] ( ) ( )uunuvTT

un

uv

tttt

tMGFtMGFtMGFiii

Σ−Σ−Φ

Σ+Σ+−ΣΦ=

=−=

− ,,2

1exp,,0 1 µµ

ε

iii uv −=ε

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

which is a specific case of the Closed Skew Normal

distribution.

27

Page 28: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Outline

• Classical stochastic frontier model

• A case of truncated normal inefficiency

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

• A case of truncated normal inefficiency

• Spatial stochastic frontier model

• Moment generating and characteristic functions

• Closed/Unified skew-normal distribution

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015 28

Page 29: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Closed Skew Normal distribution

The Closed Skew Normal distribution:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

( ), ,,,,,~ ∆′′Γ′Σ′′nnCSNx νµ( )

( )

( ) 111

1

,

,

,

,

,

,,,,,~

−−−

Σ+Σ=∆′

−=′Σ+ΣΣ−=Γ′

Σ+Σ=Σ′−=′

∆′′Γ′Σ′′

uv

uvu

uv

nnCSNx

µν

µµνµ

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

Introduced by González-Farías, Domínguez-Molina, and Gupta

(2004).

29

( )Σ+Σ=∆′ uv

Page 30: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Unified Skew Normal distribution

There are a lot of different variants of multivariate

skewed distributions:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

skewed distributions:

1. Closed Skew Normal

2. Hierarchical Skew Normal (Liseo, Loperfido, 2003)

3. Fundamental Skew Normal (Arellano-Valle, Genton,

2005)

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

Arellano-Valle and Azzalini (2006) introduced a Unified

Skew Normal distribution, which cover many private

cases, including CSN.

30

Page 31: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Estimation of the spatial stochastic

frontier model

Maximum likelihood estimator:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

( ) ( ),,0ln,,,,,,,ln 22)( +Σ−Φ−=s ρρρβL µµσσβ( ) ( )( ) ( ) ( )( ) ( )

( )( ) ( )( )

,

,

,,,ln,,ln

,,0ln,,,,,,,ln

112~

)(

1111

2~

2~

)(

−−

−−−−

−−=Σ

−−−=

Σ+Σ−+Σ+Σ−+Σ+ΣΣ−Φ+

+Σ−Φ−=

vvn

T

vvnvv

sXYY

uvnuvuvun

unuvYuvs

WρIWρI

XβWXβYWρYe

ee

ρρρβL

σ

µϕµµ

µµσσβ

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015 31

( )( ) ( ) ., 112~

−− −−=Σ uun

T

uunuu

vvnvvnvv

WρIWρIσ

Page 32: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Estimation of the spatial stochastic

frontier model

Inefficiency component u can be estimated via

conditional distribution, which is multivariate truncated

normal:

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

conditional distribution, which is multivariate truncated

normal:

( )

( ) ( )( ) 1

1

,0

where

,,~

−−−

+∞

Σ+Σ=Σ

+Σ+ΣΣ−=

Σ

uvuu

uuMVTNu

ε

εε

µεµµ

µε

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

Corresponding theoretical moments of the multivariate

truncated normal distribution are well-known.

32

( ) 111 −−− Σ+Σ=Σ uvu ε

Page 33: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Alternative research paths• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

Sum of normal and

truncated normal

Convolution

Obtained probability

density of CSN

Moment generating

functions

Probability density

functions

Moments of CSN

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015 33

density of CSN

Maximum Likelihood

estimator

Method of Moments

estimator

Page 34: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Related literature

1. González-Farías, G., Domínguez-Molina, J.A., Gupta, A. (2004). The closed

skew-normal distribution, in Skew-elliptical distributions and their applications: a

journey beyond normality, Chapman & Hall/CRC, Boca Raton, FL, pp. 25–42.

• Classical stochastic frontier

• Normal-Truncated Normal Case

• Spatial stochastic frontier

• MGF and CF

• Closed/Unified skew normal

journey beyond normality, Chapman & Hall/CRC, Boca Raton, FL, pp. 25–42.

2. Domínguez-Molina, J.A., González-Farías, G., Ramos-Quiroga, R. (2004). Skew-

normality in stochastic frontier analysis, Skew-elliptical distributions and their

applications: A journey beyond normality, pp. 223–241.

3. Arellano-Valle, R. B. and Azzalini, A. (2006). On the unification of families of

skew-normal distributions. Scand. J. Statist., 33, 561-574.

4. Aziz M. (2011) Study of Unified Multivariate Skew Normal Distribution with

Applications in Finance and Actuarial Science, PhD thesis.

The Mathematical Seminar, Transport and Telecommunication Institute, Riga, 08.05.2015

Applications in Finance and Actuarial Science, PhD thesis.

34

Page 35: Probability behind Spatial Stochastic Frontier model · Probability behind Spatial Stochastic Frontier model Dmitry Pavlyuk The Mathematical Seminar, Transport and Telecommunication

Thank you for your attention!Thank you for your attention!Thank you for your attention!Thank you for your attention!

Questions are very appreciated