You are on page 1of 8

THE 4TH INTERNATIONAL CONFERENCE ON WATER SUPPLY

MANAGEMENT SYSTEM AND SOCIAL CAPITAL


Makassar, Indonesia, July 16-17, 2012

A Trip Mode Choice Model with Endogeneity in


Explanatory Variables
Kakuya MATSUSHIMA Kiyoshi KOBAYASHI Hiroshi FUKUI

1. INTRODUCTION
1.1 Endogeneity and Sorting
The importance and the complexity of the
relationship between land use and travel behavior
has been recognized for long years in the field of
transportation
planning.
In
conventional
transportation mode,transportation system attributes
are often treated as exogenous variables in models
and the emotional
Part
art of utility for travel has been ignored. In
reality, each individual has own preference for travel
mode and obtains the emotional utility if they
choose the mode consistent with their preference.
The preferences also affect residential choice
behavior of households. The tendency of people to
choose locations based on their preferences, referred
to residential sorting would be occurred. Therefore,
travel mode choice behavior and residential choice
behavior are considered to be mutually interrelated.
If residential sorting effects are ignored when
estimating mode choice of individuals, the
estimation results would be biased because of the
endogeneity in the model.
In this paper,, the mechanism of residential sorting
and its effect on the economy are analyzed by
building a theoretical model that explicitly treats
emotional part of utility for choosing specific travel
mode. We also verify the residential sorting effects
by using person
rson trip survey data in Japan. The paper
concludes with a discussion of model findings for
policy planning. In particular, we found
implementing soft transport policy measures that
effects of individuals preference for travel mode
have to be considered ass one of the tools of city
planning policy measures.

1.2 Endogeneity Bias


When the error term and explanatory variables
correlated each other, estimated parameters may be
biased without any consideration of endogeneity
problem. This type of bias is called as endogeneity
bias. Researchers should pay attention to the
endogeneity because implemented policies form the
analysis without any consideration of the
endogeneity issues. Let us start from an example in
linear regression model in order to understand the
endogeneity bias. See the following linear
regression equation.

y = X +

(1)

Here y is a dependent variable with ( n 1)


X=(X1,X2)
2) is a matrix of explanatory variables with
(n k), and is an error term vector with (n 1).
The conditional expected value of estimated
parameters is derived as follows by the least
square method.

E X = ( X X )1 X E y X
= + ( X X ) 1 X X E X

(2)

The estimated parameters are consistent estimator


when E X = 0 holds. However, when some
explanatory variables X2((n k2)) correlated with
the error term, that is, E X 2 0 , this result is
not true. X2 is called as entogenous explanatory
variable in this case.
Observational
error,
Omitted
variable,

4th International
ional Conference on Water Supply Management System and Social Capital
Makassar, Indonesia, 16-17 Juli 2012

112

synchronization are main factors to bring about the


endogeneity issues. Dataset used for the estimation
may contain observational error when those are the
result after taking means or approximation. The
following equation holds with observational error

X = X + u

(3)

Here X is a matrix with observed explanatory


variables, X is the matrix of true explanatory
variables, and u is a vector with error term. Let us
assme that reseachers try to esstimate the following
relation.

y = X +

(4)

As we may have observational error, the following


equation will be used for the estimation.

y = X + u
= X +

(5)

In this case, the error term is correlated with


explanatory variables X , which gives us a biased
estimated parameters because it is eqipped with
the endonegeity. If we apply general the least square
estimation for these cases, attenuation may occur
(Green 2007). Next, let us consider omitted variable
with the following equation.

y = X + U +

(6)

estimate parameters by this methodology. It plays an


important role when we consider the case with
discrete choice model. The control variable shoud
satisfy the following conditions.
(I) It should not have any correlation with the error
term.
(II) It should be correlated with the endogenous
variable.
Let us consider the following linear regression
model.

Yi = X i + i , i = 1,..., n

(7)

Assume that an explanatory variable Xi is correlated


with the error term i. The endogeneity bias can be
observed if we apply the general OLS estimation.
Researches get a variable Zi which satisfies the
conditions above and estimate parameters by the
control variable method. The following equation is
satisfied by the condition (I).

1
Zi (Yi X i )
n

(8)

The estimated control variable is diffined as follows.

IV =

(Z Z )(Y Y )
(Z Z )( X X )
i

=+

S
S

Z
ZX

(9)

Here E XU = 0 holds. Let us assme that U are


not observed and taken as omitted variable. The
error term is caluclated as = U + , which
brings about a 0 . The estimation result of is
biased when U and X are correlated. Thirdly,
consider the case with synchronization. It may
happen when at least one explanatory variable and
the dependent variable are simultaneously chosen.
For example, we can observe this phenomenon in
the relation which shows the demand and supply in
markets. If we estimate parameters by ignoring the
structure, estimated parameters are biased.

When we take a limitation of estimated control


variable, the following equation is derived with
conditions (I) and (II).

2 ENDOGENEITY IN DISCRETE CHOICE


MODEL

2.2 Endogeneity in Discrete Choice Models


Generally speaking, the methodology cannot be
applied to the case of discrete choice models due to
the non-linearity. Several alternative methodologies
to estimate parameters in discrete choice models are
proposed to conquer it (Louviere et al. 2005). the
BLP approach and the control function approach are
proposed as major methodologies. Berry(1994).
Berry. Levinsohn and Pakes(1995) proposed a

2.1 A Case of Linear Regression Model


One of the most general methodology which
consider the endogeneity in linear regression model
is the control function method. A control function
(IV: Instrumental Variable) will be found when we

p lim IV = +
=

Cov( Z , )
Cov( Z , X )
(10)

Then the estimated control variable satisfied


consistency. The problem is which data is used for
the control variable. Researches should explain the
appropriateness about the adopted control variable.

4th International Conference on Water Supply Management System and Social Capital
Makassar, Indonesia, 16-17 Juli 2012

113

methodology (hereafter BLP) to estimate parameters


with endogeneity issues. The consumption behavior
to purchase differentiated goods is estimated by
BLP approach based upon the discrete choice
model. Let us explain about BLP model with the
following. Assume that a good in market m = 1,
,M is supplied by firms j = 1, , J. A consumer n
in each market purchases a good from one of the
firms. The utility when households purchase a good
in firm j at market m is shown as follows.
u

njm = n q jm + x jm + jm + njm

(11)

where qjm is the price of good n, xjm is the


observed characteristics of the good, jm is the
unobservedcharacteristics, n, n are unknown
parameters to be estimated, and njm is an error
term. n,n are expressed as follows.

n
= + Dn + vn
n

x.m = ( x1m ,..., xJm ) ,


q.m = (q1m ,..., qJm ) ,
and
.m = (1m ,..., Jm ) are satisfied. Let us also assume
that D, v, and are independent. The market share of
good j in market m can be derived as follows.
s jm ( x.m , q.m , .m ; 2 ) =

dP( D, v, ) dP ( )dP (v)dP


v

A jm

m
D

( D)

(16)

A jm

P shows the distribution of population for each. By


assuming njm follows Gumbel distribution with
i.i.d., the following will be derived.
s jm ( x.m , q.m , .m ; 2 ) =

v D

exp( jm + vnjm )
1 + l =1 exp( lm + vnlm )
J

dPDm ( D ) dPv (v)

(17)

2.3 The Estimation of Parameters


Only the error term in the equation is jm as

(12)

( Dn , vn , n) are already integrated in equation (17).


A factor jm which is not observed in market m are

Here Dn is a vector with d dimension about social


economic variables for households, and is a
matrix with (K + 1) d, and is an adjusted matrix
with (K + 1) (K + 1). Let us assume that = (1,
2) is a vector which contains all parameters to be
estimated. Then 1 = (, )2 = (,, ) is satisfied.
Eq. (11) can be rewritten as follows.

correlated withe the endogenous variable qjm, thats


why an estimation methodology with considering
endogeneity issues is necessary to be adopted. One
of the advantages of BLP methodology for
parameter estimation is to treat this type of issues
explicitly. A control variable matrix Z is introduced
for the estimation, which satisfies the following
condition.

u njm = jm (q jm, x jm, jm ;1 ) + (q jm , x jm , Dn ,vn ; 2 ) + njm ,

(13)

jm is a mean utility which households in market


purchase outside good , which is homogenous for
all households in the market. Each household has an
option to purchase other goods. It means that
households prefer to purchase outside good. The
utility which she/he gets by purchasing it can be
normalized as follows.
un0m = n0m

(14)

Households purchase a good which gives the


highest utility. Define a set of households
characteristics
who purchase good j in market m as follows.
Ajm ( x.m , q.m , .m ; 2 ) = {( Dn , vn , n 0 m ,..., nJm ) unjm u nlm l = 0,...J }

(18)

Here () is a function of parameters to show the


error term. is the true value for parameters. The
GMM estimator is expressed as follows.

where jm .vnjm are expressed as follows.


jm = x jm + q jm + jmvnjm = q jm , x jm (Dn + vn

E Z ( * ) = 0

(15)

= arg min ( )Z 1Z ( ),

(19)

where is a consistent estimator for E [ Z Z ]


which is callsed as a weight matrix, an inverse
matrix of variance-covariance matrix of the
moment. It is necessary to express the error term as
a function of parameters in order to derive the
GMMestimator by equation (19). Berry(1994)
defines jm as an error term, which is different form
the difference between the observed share S and the
estimated share s. The meaning is easily
interpretedeconomically by defining jm as an error
term. The mean utility is a function of defined in
equation (13). It is necessary to derive the mean
utility jm to have jm. In order to achieve it, let us

4th International Conference on Water Supply Management System and Social Capital
Makassar, Indonesia, 16-17 Juli 2012

114

solve thefollowing equations where the observed


share and the estimated share of goods in each zone
are equal.

s.m ( x.m , q.m , .m ; 2 ) = S.m m = 1, ,M

(20)

e nj ynj = E nj W ( xn , zn , ) + nj
(24)
= E nj nj
The error term nj explains some parts of the
endogenous variable ynj which is correlated with

BLP methodology is not applied to cases with small


samples as market share should be observed. It
cannot be applied when a share for each product is
small enough, because market share should be
observed with only small sample error. In the field
of demand estimation models (e.g. Fox(2007)),
there are also other cases which BLP estimation
cannot be applied to models which are not
consistent with the framework of BLP. Control
function method is effective for these cases. It is
widely adopted compared with BLP approach, and
parameters are easily estimated. Hereafter, the
outline of controlfunction method is explained
followed by Petrin et al.(2010). Assume that
household n select one of the choices J. Her utility
by choosing the choice j is expressed as follows.

U nj = V ( ynj , xnj , n ) + nj ,

Divide the error term nj into the condtitional mean


and its difference.

nj = E nj unj + nj

(25)

The conditional mean of nj, E[nj|nj], is called as a


control function which is represented by CF(nj; ).
CF(nj; ) = E[nj|nj]

(26)

is a parameter for estimation. When a household


select a choice j, her utility is derived as follows.
Unj = V (ynj, xn, n) + CF(nj; ) + nj

(27)

(21)

where ynj is the observed endogenous variable (e.g.,


prices) and xnj, n are observed exogenous
parameter vector which shows the preference, and
nj is unobservable error. Assume also that the
endogenous variable ynj are correlated with the error
term nj. Control function approach tries to create
substitutable variables for endogenous variable ynj
correlated with the error term nj. The rest of the
endogenous variable is independent from the
unobserved error term, so the general estimation
methodology can be applied. The endogenous
variable ynj is expressed as the function of
exogenous variables, the control variable and the
error term.

ynj = W ( xn , zn , n ) (22)
in which n , nj are independent from xn , zn and
dependent on n , nj . Let us consider a simple case
where the unovserved error is a function with single
varialbe nj.

ynj = W ( xn , zn ; ) + nj .

nj

(23)

Here is a parameter. As a control variable satisfies


the condition for exogenous variables, the following
equation is satisfied.

The probability for household n to choose j is


represented by

Pnj = Pr ob(U nj > U nk )


= I (Vnj + CFnj + > Vnk + CFnk + nk k j ) f ( n , n ) d d n

(28)

nj

Where n = n j, n = njj, and the


simultaneous density function is f().
Vnj = V (ynj, xnj, n)

(29)

CFnj = CF(nj, )
is also satisfied. Equation (28) represents the
probability in general choice models except for that
a control function is introduced as an explanatory
variable. It should be pointed out that the integration
in equation (28) is taken for the conditional
distribution , not for the error . is correlated
with the endogenous variable ynj, but is not
correlated with that. We can estimate the model with
the following procedure.
1. A control function CF(nj; ) is specified.
2. the residual nj is estimated.
3.Equation (28) where a control function is
introduced as one of explanatory variables is
estimated.
Take a simple example to specify a control function

4th International Conference on Water Supply Management System and Social Capital
Makassar, Indonesia, 16-17 Juli 2012

115

CF(; ). If we assume that nj and nj follows the


normal distribution with mean 0, the following
equation is satisfied form the characteristic of
normaldistribution.
CF(nj; ) = E[nj|nj]

explanatory variable. Define the estimated


parameter vector as *(b).
(6) Calculate the estimated values { (b)}, b = 1,
,B by repeating B times from (1) to (5).

(30)

2. A CASE STUDY
= nj

(31)

When the household choose j, her utility is discribed


as follows
Unj = V (ynj, xn, n) + nj + nj

(32)

In order to estimate parameters in equation (32), nj


should be ovserved. As it is not ovserved ingeneral,
it is substituted by the estimated ynj. Regard the
endogenous variable ynj as the dependent variable
in equation (35a), and take a regression by taking
exogenous variables and the control variable as
explanatory variables. The residual nj = ynj W(xn,
zn; ) is used for estimation.
2.4 Adjustment of standard error with Bootstrap
Method
As a generated regressor (Pagan 1984) is used for
the second estimation in control function method,
the standard error of estimated parameters are
under-estimated if the general methodology is
applied to derive the standard error. The effect of
additional error should be considered by applying
the generated regressor. The standard error is
adjusted by applying the bootstrap method (Pertin
and Train (2010)) in this paper. After the residual is
calculated with samples thorough bootstrap method,
parameters in a discrete choice model with the
residual are estimated. This process is repeated and
then derive the variance of estimated parameters for
each. Here is the procedure.
(1) The least square method is applied to yn = zn+n,
n = 1, ,N at the first step of the control
function method, and derive the residual { u n, n
= 1, ,N}.
(2) Calculate likelihood distribution F where the
probability 1/N are allocated for each point of the
residual. Then extract the bootstrap sample { u n,
i = 1, ,N} with size N.
(3) Calculate the bootstrap sample yn , i = 1, ,N
of the endogenous variable yn with yn = n+ u n.
(4) Derive the residual {n(b), n = 1, ,N} by
the first step of the control function method with
{ y n , zn}, n = 1, ,N.
(5) Parameters of the discrete choice model are
estimated by adding the residual in (4) as an

3.1 Background
A large variety of policies for sustainable urban
development are investigated in developed countries
to attack important issues like global environmental
problems, and sprawl in urban area. The concept of
compact city is proposed in Europe, which plays an
important role for the discussion about the
sustainable regional structure. In Japan, in which
population started to decrease with an ageing
society, the development of sustainable area and the
intensive style of regional structure are key issues
when we consider policies for urban and regional
development. Japanese central government supports
municipalities to develop policies for low-carbon
society by introducing the guideline for them.
Policies, such as the change of regional structure to
more intensive way and the equipment of public
transportation system are taken as examples in the
guideline. The discussion where the regional
structure will be changed from that highly depend
on automobiles by navigating the urban structure
more compact assumes one way casual relation that
living environment affects travel behavior of
households. However, it is not clear whether this
relation holds or not because the relation between
the living environment and travel behavior. Most of
proposed trip mode choice models do not consider
the households preference for trip mode.
Households may decide their residence by
considering their preference toward trip mode,
which is a different assumption from that where
they decide trip behavior given the living
environment. They may decide their residence
where they can enjoy their preferable travel mode
easily. We may define it as residential sorting. In
this case, the service level of transportation which
they enjoy is interpreted as the result of their
selection about the residence. Observed data are
considered as endogenous (not exogenous) variables
which are correlated with the error term as a result.
If we estimate parameters without any consideration
of residential sorting, estimated parameters might be
biased.
3.2 Target Area
We may develop a trip mode choice model with
dataset derived from the 4th person trip survey in
Kansai area, Japan. We had extract trip data with

4th International Conference on Water Supply Management System and Social Capital
Makassar, Indonesia, 16-17 Juli 2012

116

origins as Suma, Tarumi, and Nishi ward in Kobe


city and destination as Chuo ward in Kobe city.
Here we consider two type of trip mode (railway
and automobile) for simplicity. Only individuals
who have the driver license are extracted. The
sample size is 758.

considering the preference for trip mode.


3.4 Control Function
As we have only one endogenous variable, we need
to install a control variable other than
thesevariables. In this paper, declared land price of
thier residence is used as the control variable

3.3 The Model


Consider a discrete choice model for commuting
trip. By setting her utility when individual n choose
railway as Un1, and that for automobile as Un2, the
probability to choose raiwaly is expressed as
follows.
Pn(1) = prob{Un1 Un2 > 0}

(33)

We assume a linear utility function as follows.


Unj = V (ynj, xnj, n) + nj,

(34)

where ynj is an endogenous variable, is an


estimated parameter vector, xn is explanatory
variable vector, and n is the error term. The
following explanatory variables are considered;
Sex(Male:0 and Female:1), Age(More than 50:0 and
less than 50:1), a dummy variable to show the
possession of cars(households without cars:0, those
with cars:1), trip time, access time, and egress time.
Trip time is a common variable for both modes,
while a dummy variable for car possession is only
for automobiles, access time and egress time are
only for railways. We assume that individuals have
their own preference toward trip mode which is not
observed. They may choose their residence given
their preference, which may bring about selfselection mechanism about their residence.
Individuals with the preference toward railway may
select their residence which is close to stations.
Access time is considered as one of endogenous
variables in this case. Assume that the unobservable
error term which affects the endogenous variable is
expressed as nj and that nj and nj follows the
normal distribution with mean 0. The following
equations are satisfied.
ynj = W(xn, zn; ) + nj

(35a)

CF(nj; ) = E[nj|nj] = nj

(35b)

Figure1 Correlation between the endogenous


variable and the control variable
Figure1 shows the relation between the endogenous
variable and the control variable. We can observe
negative correlation where residential area with high
land price is close to the nearest station. See the area
with its land price 20 thousand yen/m2. We can see
that access time to the closest station is distributed
from 10 to 20 minutes from the figure. Some areas
are far from the station, while the others are close to
the station, which shows that households choose
their residence by considering their preference
toward trip mode given the same land price. the
residual in the regression analysis is interpreted as a
variable about the preference to trip mode as
households who live in areas with shorter access
time may prefer to choose railway for commuting.
Households who has positive the residual (it means
longer access time) may have preference toward
automobiles, while those with negative the
Table1 The Comparison of Two Methodologies

Her utility when she choose mode j is


Unj = V (ynj, xn, n) + nj + nj.

(36)

The control function CFnj is considered as a


psychological utility for the trip mode choice.
explains the effect to the choice of access time by

Sex (Female:1
)

Normal Estimation
tStanda valu
e
rd
Erro
Param
Estim
r
eter
ated
5.27
1.559
0.295
7

4th International Conference on Water Supply Management System and Social Capital
Makassar, Indonesia, 16-17 Juli 2012

Control function method


Estim
ated
Param
eter

Stand
ard
Error

tvalu
e

1.696

0.303

5.6

117

Age (More tha


n 50:1)

0.248

0.225

1.10
5

0.086

0.235

0.36
4

Possesion of c
ars

0.89
8

0.36
8

2.4
4

1.02
4

0.37
4

2.7
42

Trip time

0.349

0.141

0.698

0.17
3

-4.0

Access time

1.383

0.167

0.104

0.28
7

Egress time

-0.2

Constant

4.175

The residual

0.223

2.47
6
8.27
4
0.89
7

0.296

0.513

8.14

2.796

0.23
9
0.55
9

2.476

0.45
3

0.36
3
1.23
9
5.00
5
6.15
3

residual (it means shorter access time) may prefer to


use railway. This hypothesis will be examined by
checking the sign in the estimation.
3.5 The Result of Estimation
Table{1 summarizes the result of estimation both by
general methodology and control function method.
The estimated parameter about access time is
different for both cases. The parameter is not
significant by the control function method in
addition. As the residual is significant, this
variable is found to be endogenous variable. The
sign of the parameter is also appropriate as we
assume that positive the residual shows the
preference toward automobile and negative residual
shows the preference for public transportation.
3.6 Policy Implications
We can say form the estimated result that Access
time is an endogenous variable for the trip mode
choice model and correlated with unobservable
variable such as the preference toward public
transportation. It is effective to ally control function
method for the estimation in order to get rid of the
endogeneity bias. Households may choose their
residence by considering the preferable trip mode in
advance. That is, the residential sorting can be
observed thorough the mechanism of residential
selection. It means that intensive regional structure
may not be realized only by introducing policies to
decrease access time, such as the increase in the
frequency of public transportation to the station. It is
also necessary to introducing policies which affect
their preference in order to change the regional
structure thorough residential sorting mechanism.
Some municipalities may attract individuals with
strong preference toward public transportation,
which brings about the increase in population who
are possibly commute by public transportation.
Thanks to the scale economy, such municipalities
can supply rich public transportation service with

cheaper cost, which also attract more individuals


with the preference toward public transportation.
4 Conclusion
The importance and the complexity of the
relationship between land use and travel behavior
has been recognized for long years in the field of
transportation
planning.
In
conventional
transportation mode, transportation system attributes
are often treated as exogenous variables in models
and the emotional part of utility for travel has been
ignored. In reality, each individual has own
preference for travel mode and obtains the
emotional utility if they choose the mode consistent
with their preference. The preferences also affect
residential choice behavior of households. The
tendency of people to choose locations based on
their preferences, referred to residential sorting
would be occurred. Therefore, travel mode choice
behavior and residential choice behavior are
considered to be mutually interrelated. If residential
sorting effects are ignored when estimating mode
choice of individuals, the estimation results would
be biased because of the endogeneity in the model.
This paper proposes a methodology to tackle with
endogeneity issues in discrete choice models. Two
methods, BLP approach and control function
approach are explained and their characteristics are
summarized. The hypothesis that access time in trip
mode choice model is endogenous is proposed and
tested with empirical analysis with person trip
survey data in Kansai area, Japan. The estimation
result shows that access time for the public
transportation is endogenous variable, which
explains
the importance of control function method to prove
the existence of residential sorting. A political
implication is derived from the result, that is,
policies which try to attract individuals who has
similar preference toward trip mode.
REFERENCES
Berry, S. : Estimating Discrete-Choice Models of
Product Differentiation, RAND Journal of Economics , Vol.25, pp.242-262, 1994
Berry, S., J.Levinsohn, and A.Pakes : Automobile
Prices in Market Equilibrium, Econometrica ,
Vol.63, pp.841-889, 1995.
Berry, S., J.Levinsohn, and A.Pakes : Differentiated
Products Demand Systems from a Combination of
Micro and Macro Data: The New Car Market,
Journal of Political Economy, Vol.112, pp.68-105,
2004.
Boarnet, M. G., and R. Crane : The influence of land
use on travel behavior: specification and estimation
strategies, Transportation Research Part A: Policy

4th International Conference on Water Supply Management System and Social Capital
Makassar, Indonesia, 16-17 Juli 2012

118

and Practice, Volume 35, pp823- 845, 2001.


Cao, X., Mokhtarian, P. L., and Handy, S. L. :
Examining the Impacts of Residential Self-Selection
on Travel Behaviour: A Focus on Empirical
Findings, Transport Reviews, Vol.29: 3, pp359-395,
2009.
R. Cervero : Transit Oriented Development Gs
Ridership Bonus: A Product of Self-Selection and
Public Policies, Environment and Planning A,
Vol.39, pp.2068-2085, 2007.
R. Cervero and R. Gorham : Commuting in Transit
Versus Automobile Neighborhoods, Journal of the
American Planning Association, Vol.61, pp.210-225,
1995.
R. Cervero and K. Kockelman : Travel demand and
the 3Ds: Density, diversity, and design, Transportation Research Part D: Transport and
Environment, Vol.2, pp.199-219, 1997.
Crane, R. and Crepeau, R. : Can Land Use Policy
Really Affect Travel Behavior? A Study of the Link
between Non-Work Travel and Land Use
Characteristics, Transportation Research Part D:
Transport and Environment, Vol.3, pp.225-238,
1998.
Durbin, J. : Errors in Variables, Review of the
International Statistical Institute, Vol.22, pp.2332,1954.
Fox, J. : Semi-parametric Estimation of Multinomial
Discrete-Choice Models Using a Subset of Choices
The RAND Journal of Economics ,Vol.38, pp.10021019, 2007
Greene, W. : Econometric Analysis, Sixth Edition,
Prentice Hall, 2007.
Hausman, J. A. : Specification Tests in
Econometrics, Econometrica, Vol.46, pp1251-1271,
1978.
Louviere, J. et al. : Recent Progress on Endogeneity
in Choice Modeling, Marketing Letters,Vol16
pp.255-265, 2005.
Mokhtarian, P. L. and Cao, X. Examining the
impacts of residential self-selection on travel
behavior: A focus on methodologies, Transportation
Research B, Vol.42 (3), pp204-228, 2008.
Pagan, A. R. : Econometric Issues in the Analysis of
Regressions
with
Generated
Regressors,
International Economic Review, Vol.25, pp221-247,
1984.
Petrin, A and Train, K. : Omitted Product Attributes
in Discrete Choice Models, NBER Working Paper,
2003.
Petrin, A and Train, K. : A Control Function
Approach to Endogeneity in Consumer Choice
Models, Journal of Marketing Research, Vol. 47,
pp3-13, 2010.
J.Rajamani, C.R. Bhat, S. Handy, G. Knaap, Yan S. :
Assessing Impact of Urban Form Measures on

Nonwork Trip Mode Choice After Controlling for


Demographic an Train,K. : Discrete Choice Model
with Simulation ,Cambridge University Press ,
2003.
Wu, D. : Alternative Tests of Independence between
Stochastic
Regressors
and
Disturbances,
Econometrica, Vol.41, pp.733-750, 1973. d Levelof-Service Effects, Transportation Research Record:
Journal of the Transportation Research Board,
Vol.1831, pp.158-165, 2003.

4th International Conference on Water Supply Management System and Social Capital
Makassar, Indonesia, 16-17 Juli 2012

119

You might also like